Ijicic Complte Book

International Journal of
Innovative Computing, Information and Control
Editor-in-Chief
Peng Shi, Faculty of Advanced Technology, University of Glamorgan, Pontypridd, CF37 1DL, United
Kingdom (E-mail: pshi@glam.ac.uk)
Co-Editor-in-Chief
Jeng-Shyang Pan, Department of Electronic Engineering, National Kaohsiung University of Applied
Sciences, Kaohsiung 807, Taiwan (E-mail: jspan@cc.kuas.edu.tw)
Executive Editor
Yan Shi, School of Industrial Engineering, Tokai University, 9-1-1, Toroku, Kumamoto 862-8652,
Japan (E-mail: yshi@ktmail.tokai-u.jp)
Advisory Board
Ramesh Agarwal, USA Steve P. Banks, UK Tianyou Chai, China
Chin-Chen Chang, Taiwan Hanfu Chen, China Wai-Chi Fang, USA
Tom Heskes, Netherlands Lakhmi C. Jain, Australia V. Lakshmikantham, USA
Jerry M. Mendel, USA Masaharu Mizumoto, Japan Chengsheng Pan, China
Witold Pedrycz, Canada Bion L. Pierson, USA Jianrong Tan, China
Junzo Watada, Japan Xiaopeng Wei, China Takeshi Yamakawa, Japan
Naoyoshi Yubazaki, Japan Lotfi A. Zadeh, USA
Associate Editors
Michael V. Basin, Mexico Shyi-Ming Chen, Taiwan Ozer Ciftcioglu, Netherlands
Liya Ding, Macau Vasile Dragan, Romania Hugang Han, Japan
Tzung-Pei Hong, Taiwan Chao-Hsing Hsu, Taiwan Xiangpei Hu, China
Hsiang-Cheh Huang, Taiwan Yunfu Huo, China Gerardo Iovane, Italy
Bin Jiang, China Katsuari Kamei, Japan Jien Kato, Japan
Lei Li, Japan Bin-Yih Liao, Taiwan Wenyuan Liu, China
Magdi Mahmoud, Saudi Arabia Anatolii Martynyuk, Ukraine Angelo Miele, USA
Jagadis C. Misra, India Saeid Nahavandi, Australia Masahiro Nakano, Japan
Sing Kiong Nguang, New Zealand Ngoc Thanh Nguyen, Poland Nikos Nikolaidis, Greece
Pavel Pakshin, Russia Charles E. M. Pearce, Australia Fuji Ren, Japan
Xiangshi Ren, Japan Hirofumi Sasaki, Japan M. N. S. Swamy, Canada
Kok-lay Teo, Australia Shaocheng Tong, China Shuoyu Wang, Japan
Yuanqing Xia, China Mingqing Xiao, USA Edwin Engin Yaz, USA
Jianqiang Yi, China Lixian Zhang, China Lindu Zhao, China
2011 ICIC INTERNATIONAL ISSN 1349-4198 PRINTED IN JAPAN

ISSN 1349-4198
IJICIC Volume 7, Number 3, March 2011
International Journal of Innovative

Computing, Information & Control
Sponsored by Science Communication Institute, China

Dalian University, China
Published by ICIC International

http://www.ijicic.org
Volume 7, Number 3, March 2011
CONTENTS
The Improvement of all-Digital Amplitude-Locked Loop Separation Analysis

Combined MIMO System 1001
Gwo-Jia Jong and Gwo-Jiun Horng
Value Based Intelligent Requirement Prioritization (VIRP): Expert Driven Fuzzy
Logic Based Prioritization Technique 1017
Muhammad Ramzan, M. Arfan Jaffar and Arshad Ali Shahid
Generalization of Proxy Signature Based on Factorization 1039
Cheng-Chi Lee, Tzu-Chun Lin, Shiang-Feng Tzeng and Min-Shiang Hwang
A Novel High Capacity 3D Steganographic Algorithm 1055
Meng-Tsan Li, Nien-Ching Huang and Chung-Ming Wang
Real-Time Decentralized Neural Block Control: Application to a Two DOF Robot
Manipulator 1075
R. Garcia-Hernandez, E. N. Sanchez, E. Bayro-Corrochano, V. Santibanez
and Jose A. Ruz-Hernandez
A Supervising Authenticated Encryption Scheme for Multilevel Security 1087
Chien-Lung Hsu, Liang-Peng Chang and Tzong-Chen Wu
Aesthetic Emotions in Human-Robot Interaction. Implications on Interaction Design
of Robotic Artists 1097
Catalin Buiu and Nirvana Popescu
Modeling, Design and Experiment of Improved Super-Mini Underwater Robot 1109
Xin Song, Zao-Jian Zou and Jia-Wei Ye
Design and Approximation Capabilities Analysis of Time-Variant Fuzzy Systems 1121
Degang Wang, Wenyan Song and Hongxing Li
Encoding Prior Knowledge into Data Driven Design of Interval Type-2 Fuzzy
Logic Systems 1133
Chengdong Li, Jianqiang Yi and Tiechao Wang
Generation and Application of Decision Rules within Dominance-Based Rough Set
Approach to Multicriteria Sorting 1145
Liping An, Zengqiang Chen and Lingyun Tong
Automatic Classification of Driving Mental Fatigue with EEG by Wavelet Packet
Energy and KPCA-SVM 1157
Chunlin Zhao, Chongxun Zheng, Min Zhao, Jianping Liu and Yaling Tu
3D Triangle Mesh Compression Based on Vector Quantization with k-Ring Vector
Prediction 1169
Lin-Lin Tang, Zhe-Ming Lu, Fa-Xin Yu, Ping-Hui Wang and Zhen Li
(Continued)

CONTENTS (Continued)
An Efficient Simulated Annealing with a Valid Solution Mechanism for TDMA

Broadcast Scheduling Problem 1181
Gon Kim, Sung Soo Kim, Il-Hwan Kim, Dong Hoi Kim, V. Mani and Jae-Ki Moon
Scalable and Systolic Dual Basis Multiplier over GF(2m) 1193
Liang-Hwa Chen, Po-Lun Chang, Chiou-Yng Lee and Ying-Kuei Yang
A Fuzzy-Based Rough Sets Classifier for Forecasting Quarterly PGR in the Stock
Market (Part II) 1209
Ching-Hsue Cheng and Yoush-Yang Chen
Re-Evaluation of Adaptive X Control Charts: A Cost-Effectiveness Perspective 1229
Yan-Kwang Chen, Hung-Chang Liao and Hsu-Hwa Chang
Inventory Management in a (Q,r) Inventory Model with Two Demand Classes
and Flexible Delivery 1243
Shuo Huang, Bo Liao and David J. Robb
A Novel Text Watermarking Algorithm Using Image Watermark 1255
Zunera Jalil, M. Arfan Jaffar and Anwar M. Mirza
An Efficient and Secure Three-Pass Authenticated Key Agreement Elliptic Curve
Based Protocol 1273
Zeyad Mohammad, Chien-Lung Hsu, Yaw-Chung Chen and Chi-Chun Lo
Indicator of the Quality of the Sensor Set Up: A Study Using Surface EMG
on Sub-Band ICA 1285
Ganesh R Naik and Dinesh K Kumar
H Filtering for Markovian Jump Systems with Time-Varying Delays 1299
Jinliang Liu, Zhou Gu and Songlin Hu
Surface Texture Characterization of Fibers Using Fractional Brownian Motion Model 1311
Jiun-Jian Liaw, Chuan-Pin Lu and Lin-Huang Chang
Power Grid Node and Line Delta Centrality Measures for Selection of Critical Lines
in Terms of Blackouts with Cascading Failures 1321
Hwachang Song, Rodel D. Dosano and Byongjun Lee
Optimizing MIMO-SDMA Smart Antennas by Using Linear Array Phase Perturbations
Based on Particle Swarm Optimization 1331
Chao-Hsing Hsu
An Environmental Visual Features Based Navigation Method for Autonomous
Mobile Robots 1341
Fairul Azni Jafar, Yasunori Suzuki, Yuki Tateno, Kazutaka Yokota
and Takeshi Matsuoka
(Continued)

Observer-Based Iterative Learning Control with Evolutionary Programming Algorithm

for MIMO Nonlinear Systems 1357
Yan-Yi Du, Jason Sheng-Hong Tsai, Shu-Mei Guo, Te-Jen Su and Chia-Wei Chen
Reversible Data Hiding for VQ Indices Using Prediction Errors 1375
Zhi-hui Wang, Chin-Chen Chang, Huynh Ngoc Tu and Ming-Chu Li
A Novel Technique for Designing Decentralized Stabilizers for Robust Control
in Power Systems Using an H Criterion 1387
Alexandre C. Castro, Jose M. Araujo, Eduardo T. F. Santos, Fabricio G. S. Silva
and Clivaldo S. de Araujo
A Study of Three Novel Line-Based Techniques for Multi-Target Selection 1397
Jibin Yin and Xiangshi Ren
A Novel Modal Series Representation Approach to Solve a Class of Nonlinear
Optimal Control Problems 1413
Amin Jajarmi, Naser Pariz, Ali Vahidian Kamyad and Sohrab Effati
Adaptive Frame Length Method for Hardware Context-Switching in Dynamic Partial
Self-Reconfigurable Systems 1427
Trong-Yen Lee, Che-Cheng Hu, Yang-Kun Huang and Chia-Chun Tsai
An ID-Based Access Control in a Hierarchical Key Management for Mobile Agent 1443
Chia-Hui Liu, Yu-Fang Chung, Tzer-Shyong Chen and Sheng-De Wang
Meaningful Shadow Based Multiple Gray Level Visual Cryptography without Size
Expansion 1457
Pei-Yan Pai, Chin-Chen Chang, Yung-Kuan Chan and Chi-Cheng Liao
Service-Oriented Routing and Clustering Strategies for Vehicle Infotainment
Dissemination 1467
Chenn-Jung Huang, Chin-Fa Lin, Ching-Yu Li, Che-Yu Lee, Heng-Ming Chen
Hung-Yen Shen, You-Jia Chen and I-Fan Chen
An Optimal Fuzzy Controller for a High-Performance Drilling Process Implemented
Over an Industrial Network 1481
Rodolfo E. Haber, Agustin Gajate, Steven Y. Liang, Rodolfo Haber-Haber
and Raul M. del Toro
New Secret Key Traitor Tracing Scheme with Dispute Settlement from Bilinear Maps 1499
Kuo-Yu Tsai, Tzong-Chen Wu and Chien-Lung Hsu

AIMS AND SCOPE OF IJICIC

The primary aim of the International Journal of Innovative Computing, Information and
Control (IJICIC) is to publish high-quality papers of new developments and trends, novel
techniques and approaches, innovative methodologies and technologies on the theory and
applications of intelligent systems, information and control. The IJICIC is a peer-reviewed
English language journal and is published monthly.
The scope of the IJICIC is in the broad areas of intelligent systems, information and control,
includes but not limited to:
1. Intelligent Systems and Innovative Computing
Artificial intelligence and expert systems, approximate reasoning, fuzzy systems and rough
sets, fuzzy optimization and decision making, evolutionary computation, intelligent
information systems, image and speech signal processing, human-machine learning, neural
networks, knowledge discovery, data mining, clustering and data analysis, decision support
systems, intelligent fault detection and identification, knowledge-based systems, hybrid
intelligent systems, soft computing, management science
2. Information and Control

Information theory and applications, optimal, adaptive, predictive, robust, fuzzy and reliable
control and filtering on linear and nonlinear systems, stochastic systems, hybrid systems,
supervisory systems, guidance and navigation systems, real-time systems, discrete event
systems, signal and image processing, system identification, optimization techniques,
robustness analysis, learning systems, fault detection, robotics and motion control,
manufacturing systems, power systems, networked control systems, environmental and
bio-engineering, complex and large scale systems, bioinformatics, control education,
transportation systems
SUBSCRIPTION
Library/Institutional/ Individual Rates:
Hard copy (Printing): JPY66,000
CD: JPY33,000
Hard copy with CD: JPY90,000
The charge includes postage, packing, handling and surface delivery of all issues.
IJICIC Editorial Office

Tokai University, Kumamoto Campus
9-1-1, Toroku, Kumamoto 862-8652, Japan
Phone: +81-96-386-2666
Fax: +81-96-386-2666
E-mail: office@ijicic.org

Computing, Information and Control ICIC International 2011
c ISSN 1349-4198
Volume 7, Number 3, March 2011 pp. 10011015
THE IMPROVEMENT OF ALL-DIGITAL AMPLITUDE-LOCKED

LOOP SEPARATION ANALYSIS COMBINED MIMO SYSTEM
Gwo-Jia Jong and Gwo-Jiun Horng

Department of Electronic Engineering
National Kaohsiung University of Applied Sciences
Chien Kung Campus 415, Chien Kung Road, Kaohsiung 807, Taiwan
gjjong@cc.kuas.edu.tw
Received July 2009; revised June 2010
Abstract. The mobile communication is often interfered by any type of noises. The all
digital phase-locked loop (ADPLL) system has been successfully used for decades in order
to track the carrier phase of a frequency modulation (FM) signal. In this paper, we com-
bine the ADPLL and the all-digital amplitude-locked loop (ADALL) system structure for
modulation signals of the separation co-channel transmission system. It is demonstrated
co-channel separation signals from FM system with the results of simulation experiments.
The least mean square (LMS) theory is introduced more eciently to eliminate the noise
interference and to separate the multi-channel signals. This paper is also shown and
demonstrated the multi-channel signal in FM system which is combined the ADPLL and
the ADALL algorithms with an adaptive finite impulse response (FIR) filter. The multi-
channel transmission with the additive white Gaussian noise (AWGN) interference has
been solved by using the proposed algorithm. The demodulated signals are operated and
demodulated by the LMS algorithm. The separation signals of the multi-channel can be
obtained the prefect performance and process for high speed and low cost. The proposed
system of this paper is the virtue and advantage to implement the communication security
field. In general words, it can be substituted and replaced the encryption (or decryption)
system.
Keywords: ADPLL, ADALL, FM, FIR, LMS, Communication security, Encryption
1. Introduction. Over the last decades, digital communication has become one of the
basic technologies for out modern life. Only when using digital transmission, information
can be transported with moderate power consumption, high flexibility and especially over
long distance, with much higher reliability than by using conventional analog modulation.
Thus, the communication world has been going digital [1].
When regarding digital transmission, we have to consider two dominant impairments.
Firstly, the signal is corrupted by additive white Gaussian noise environment, which can
be thermal noise of the receiver front-end or crosstalk caused by other users transmitting
in the same frequency band. Secondly, the transmission media is dispersive. This situation
can be described as a linear system with some specific transfer function, where attenuation
and phase coecients are varied with frequency. This property causes the frequency
components to be aected dierently. The channel transmission signal is distorted for the
transmitted pulses in the time domain [2-6].
The ADPLL system has been successfully used for decades in order to track the car-
rier phase of an FM signal. It is proposed to analysis suppress the in-band interference
problem eciently. The ADPLL demodulated outputs contain large in-band spikes and
get some unintelligible turbulence. In this paper, a novel separation algorithm is pro-
posed for canceling co-channel interference (CCI) communication system by the ADALL
1001
1002 G.-J. JONG AND G.-J. HORNG
method. This technique is adopted to eliminate the interference of the ADPLL in wire-
less transmission channel. It includes to solve many extrinsic factors such as inter-symbol
interference (ISI), multi-path interference, flat-fading channel and AWGN channel. The
ADALL system can separate the dominant and the subdominant signals each other. This
scheme is obtained got the low complex computation and compressed the CCI eect.
2. A Digital Procedure of All-Digital PLL. Phase-locked loops (PLL) are widely
used in modern electronic systems. The applications include synchronization, tracking
demodulation, ranging, etc. The continued progress in increasing performance, speed and
the simultaneous in size and cost of integrated circuits have resulted in strong interest in
the implementation of the ADPLL, which can be made more flexible, versatile, compact,
stable and reliable than its analog counterpart.
The classic PLL consists of three major function units: a phase detector (PD), a loop
filter (LF) and a voltage-controlled oscillator (VCO). For ADPLL, the voltage-controlled
oscillator is named numerically-controlled oscillator (NCO) or digitally-controlled oscilla-
tor (DCO) [7].
The synchronization concept plays a major role in the fields of digital data transmission,
radar and navigation system. The phase detection is an important topic in the synchro-
nization problems of digital communication systems. Based on the maximum likelihood
estimation principle, the conventional analog PLL system has been developed to treat
the nonlinear phase detection problem. The ADPLL has been also studied over year.
However, this conventional ADPLL is a gradient-based searching algorithm to solve the
highly nonlinear problem, it may lead to a local result. Furthermore, the conventional
ADPLL is sensitive to the initial conditions, variation of parameters [8].
A. Basic Model of All-Digital Phase-Locked Loop. A block diagram of ADPLL

and its phase model are shown in Figures 1 and 2, respectively [7]. The loops output
signal is desired and named by the phase error = i o .
Figure 1. Block diagram of ADPLL
Figure 2. Phase model of ADPLL

The discrete input and output signals of Figure 1 can be written as
Vi (k) = Vi sin [0 T k + i (k)] , k = 0, 1, 2, . . . (1)
THE IMPROVEMENT OF ADALL SEPARATION ANALYSIS 1003
Vo (k) = Vo cos [0 T k + o (k)] , k = 0, 1, 2, . . . (2)

where 0 is the central angular frequency of the input signal to be tracked, Vi (k) and
Vo (k) are the maximum amplitudes of the signals sampled by k. The phase detector
considered is a multiplier whose output is given by
Vd (k) = Kd Vi (k) Vo (k)
Kd Vi Vo (3)
= {sin [i (k) o (k)] + sin [2o T k + i (k) + o (k)]}, k = 0, 1, 2, . . .
2
where Kd is represented the phase detector gain.
Equation (3) is defined a second-harmonic term added to a nonlinear function of (k).
It is a common assumption that the second-harmonic term of (3) and Figure 2 will be
filtered by the discrete-time loop filter.
The discrete-time loop filter is considered and characterized by a first-order transfer
function, shown as:
C2
D(z) = C1 + (4)
1 z 1
Since the order of loop is the order of its loop filter plus one. The ADPLL is a second-
order loop whose dynamics can be described by the following dierence equation:
(k + 2) 2 (k + 1) + (k) = (k + 2) 2 (k + 1) + (k) rK1 sin [ (k + 1)]
(5)
+ K1 sin [ (k)] , k = 0, 1, 2, . . .
where K1 = Ko Kd C1 , r = 1 + C2 / C1 and Ko is represented the digitally-controlled
oscillator gain.
When the phase error is small, then the approximation sin is valid and the
sinusoidal nonlinear can be removed. ADPLLs are being implemented in the coherent
communication systems for a long time in carrier regenerating and clock time recovery
purposes. The theory and the design techniques of an ADPLL are well known. ADPLLs
are drawing the attention of research workers till now because they are basically nonlinear
feedback control loops in the discrete time domain and their dynamical behavior is not
completely known [9].
3. All-Digital Amplitude-Locked Loop of CCI.
A. Frequency Modulation (FM). A narrowband digital FM with frequency detection

has been widely used. A conventional [10-12] PLL has been often operated as a frequency
detector. The PLL detector (PLD) has the native frequency tracking ability as a tracking
filter and is possible to track very rapid center frequency variation of the input FM signal
such as Doppler shifts from low earth orbit (LEO) satellites. We adopt the FM scheme
to analyze the FM separation, it can be rewritten as follows:
y (t) = Ac cos [c t + 1 (t)] + mAc [cos c t + 2 (t)] = r (t) cos [1 (t)] (6)
where
d t = 2 (t) 1 (t) (7)
B. Co-channel Interference under AWGN Channel. Typical methods for reducing

co-channel FM interference, it estimates the stronger signals instantaneous frequency
amplitude and phase. CCI is the interference due to the mixture of signals with similar
carrier frequencies. It is necessary to find ecient techniques to reduce the harmful eects
of CCI in FM analogue or digital communication system.
The mobile users often operate in the presence of cumbersome interference along with
multi-path, Rayleigh fading channel and AWGN channel that leads to signal distortion
and signal fading at the receiver. The interference between signals from these cells is
called CCI.
The co-channel signals are transmitted by the same carrier frequency fc in the AWGN
channel. Accordingly, the complex representation of the received signal is defined as
y(t) = y1 (t) + my2 (t)
[ t ] [ t ]
= Ac cos 2fc t + 2kf m1 ( )d + mAc cos 2fc t + 2kf m2 ( )d (8)
0 0
= Ac cos [2fc t + 1 (t)] + mAc cos [2fc t + 2 (t)]
and t
i (t) = 2kf mi ( )d , i = 1, 2 (9)
0
The parameter m is defined as the interference to carrier ratio (ICR). It is denoted in
Equation (8) and the vector diagram is shown in Figure 3.
y(t) = Ac {cos [2fc t + 1 (t)] + mAc cos [2fc t + 2 (t)]} = r(t) cos 1 (t) (10)
where
d t = 2 (t) 1 (t) (11)
Figure 3. Phase diagram of phase modulation with noise
From Figure 3, it can be found a(t) = mAc cos d t, b(t) = mAc sin d t for the following
equations:.

r(t) = [Ac + a(t)]2 + b2 (t)

= A2c + 2Ac a(t) + a2 (t) + b2 (t)
(12)
= A2c + 2mA2c cos d t + A2c m2 cos2 d t + A2c m2 sin2 d t

= Ac 1 + 2m cos d t + m2
b(t) mAc sin d t m sin d t

(t) = tan1 = tan1 = tan1 (13)
Ac + a(t) Ac + mAc cos d t 1 + m cos d t
1 (t) = 2fc t + 1 (t) + (t) (14)
In this paper, the FM transmitted signals are operated at the same carrier frequency,
the transmission channel results in the co-channel interference (CCI). For the convenient
reason, we adopt the ADPLL FM model for the FM scheme to analyze the FM separation
system.
In wireless transmission channel system, it is included many extrinsic noise interference
factors such as flat-fading channel, AWGN channel, multi-path channel, etc. This com-
posite signals go through ideal band-pass filter (BPF) completely and the wideband noise
naturally because the narrownband noise. The band-limited noise is obtained in terms of
in-phase and quadrature components and represented by
n(t) = nI (t) cos(2fc t) nQ (t) sin(2fc t) (15)
where nI (t) and nQ (t) are the in-phase and quadrature noise components of n(t), respec-
tively. Equivalently, we may express n(t) in terms of its envelope and phase as
n(t) = rn (t) cos [2fc t + (t)] (16)
where
rn (t) = n2I (t) + n2Q (t)
and
nQ (t)
n (t) = tan1
nI (t)
Figure 4. ADPLL FM model with noise

The phase diagram of phase modulation with noise signal is shown as Figure 4. The
mathematical representation of this two modulated co-channel FM signals under AWGN
channel can be described as:
yn (t) = s(t) + i(t) + n(t)
= Ac cos [2fc t + 1 (t)] + mAc cos [2fc t + 2 (t)] + rn (t) cos [2fc t + n (t)]
(17)
= r(t) cos (t) + rn (t) cos [2fc t + n (t)]
= R(t) cos 2 (t)
2 (t) = 2fc t + 1 (t) + (t) + n (t) (18)
where
d t = 2 (t) 1 (t)
d = 2 (t) 1 (t) (19)
and
n t = n (t) 1 (t) (t)
n = n (t) 1 (t) (t)
(20)
Figure 5. System-level depiction of ALL application
From Figure 5, it can be found c(t) = rn (t) cos n t, d(t) = rn (t) sin n t

R(t) = [r(t) + c(t)]2 + d2 (t)

= r2 (t) + 2r(t)rn (t) cos n t + rn2 (t) cos2 n t + rn2 (t) sin2 n t (21)

= r2 (t) + 2r(t)rn (t) cos n t + rn2 (t)
d(t) rn (t) sin n t

n (t) = tan1 = tan1 (22)
r(t) + c(t) r(t) + rn (t) cos n t
The envelope and phase of the receiver are represented (21) and (22), respectively.
The phase mathematical representation of two modulated co-channel FM signals under
AWGN channel for ADPLL model is shown as Figure 5.
C. ADALL Analysis in the Discrete-Time. The conventional amplitude-locked loop

(ALL) configuration is a feedback-based control system, the dual of the familiar phase-
locked loop (PLL). The ALL functions have been provided for applications such as tuning
the Q-factor in integrated bandpass filters, noise cancellation, suppression of co-channel
FM interference and demodulation of suppressed carrier signals [1]. More recently, ALLs
have begun to adopt in high data rate wireless communication systems. Many modulation
schemes used to achieve high data rates have non-constant envelopes, creating stringent
linearity requirements for power amplifiers (PAs) [13].
To proceed with the ADALL analysis, we replace each blocks of Figure 5. It is demon-
strated the system-level depiction of ALL application for the frequency-domain equivalent,
individually. It is shown and represented as Figure 6. Here ADES is the desired amplitude
input to the comparator and AOU T is represented by a summing junction and the gain
block Kc. The loop filter is assumed to have a transfer function F (s), the variable gain
amplitude (VGA) is assumed to have a linear gain characteristic with a slope of KV GA ,
and the peak detector is approximated as having a one-pole response with a low-frequency
gain of KP D . In this ADALL analysis, we assume that the amplitude of the input signals
to the KV GA and AIN , remains constant and thus is incorporated into the gain of the
KV GA .
Figure 6. Frequency-domain analysis model of ADALL
The transfer function of Figure 6 for the frequency-domain for of the system is defined
as:
s
KF (s)(1 + )
AOU T (s) P D
= H(s) = s (23)
ADES (s) 1+ + KF (s)
P D
where it is assumed KP D = 1 and K = KC AIN KV GA . It is also expressed the loop gain
of the system as:
KF (s)
G(s) = s (24)
1+
P D
We substitute in the transfer function for various choices of loop filters, and examine
their eects on the stability and transient response of the ALL. Consider the case for the
loop filter is a simple gain, which can be incorporated into the constant K. From (24),
we consider the loop is unconditionally stable and from (23) we express the closed-loop
transfer function as:
s
1+
K P D
H(s) = (25)
1 + K 1 + s (1 + K)
P D
From this, we see that the setting time will be determined by the pole of the peak
detector, P D , in conjunction with the gain factor K. It is also clear that the output
will have a steady-state amplitude error, determined by the size of K. As K increases,
the steady-state amplitude error reduced. To eliminate the steady-state amplitude error,
we replace the simple gain with an integrator. From (24), we see that the loop remains
unconditionally stable; however, it may be necessary to add a zero to the loop filter to
maintain a desired phase margin. From (23), it is replaced the new closed-loop transfer
function as:
s
KP D (1 + )
P D
H(s) = 2 (26)
s + sP D + KP D
where we have again incorporated the loop filter gain into the constant K.
By equating the denominator to the standard form of the second-order characteristic
equation, we can solve the damping factor and the natural frequency of the loop:

1 P D
= , n = KP D (27)
2 K
Using (27), we design the loop to achieve the desired loop dynamics. More complex loop
filter choices are also possible, and can be analyzed for stability and transient behavior in
a manner to that carried out above.
D. The Separation Signal for the ADPLL and ADALL Analysis for Adaptive
FIR Filter. The separation system of transmission communication system block model
is shown in Figure 7. The adaptive FIR filter is proposed for detecting the variation
of the dominant and subdominant signals. Both of dominant and subdominant signals
are through the ADPLL and ADALL separation model. We developed and derived new
algorithm for the discrete and digital mathematical representations of the ADPLL and
ADALL discrete output signals.
Figure 7. The separation system for the ADPLL and ADALL analysis for
adaptive FIR filter
It is demodulated by an ideal ADPLL demodulator for the output signal fADP LL ( ).

The detail derivation mathematical functions are shown as the following equations and
Appendix 1 [35].
( )2
v ( ) 1
u ( )2 1 + 4m cos 2m + m2
u 1 2
fADALL ( ) = t2 1 cos d
2d
(m 1)2
( ( ) )
1 m 1 2m + m2
EllipticE cos , 2 ( ) (28)
2d (m 1)2 1
sin
2 d
( )2
1
1 + 4m cos 2m + m2
2d

The dominant signal is derivative as follows:
1
fADP LL ( ) fADALL2 ( ) = 2
s1 ( )
d di log(m eid + 1)
1 (29)
+ ( id ) fADALL1 ( ) s2 ( )
2
e
d di log
m
The subdominant signal is derivative as follows:

fADP LL ( ) f(ADALL1) ( )
1
fADALL1 ( )
= 2
s1 ( )
d di log(m eid
+ 1) fADALL2 ( )
(30)
1
f(ADALL1) ( ) fADALL1 ( )
+ ( id )
2
s2 ( )
e fADALL2 ( )
d di log
m
4. Multi-Channel FM Signals with Multi-Input Multi-Output (MIMO) Model

for Adaptive FIR Filter ll-Digital Amplitude-Locked Loop of CCI. We present
the theory and application of optimum linear filters and predictors. It is concentrated on
linear filters for the optimum in the sense of minimizing the mean square error (MSE). The
minimum MSE (MMSE) criterion leads to a theory of linear filtering that is elegant and
simple, involves only second-order statistics and is useful in many practical applications.
The optimum filter is designed for a given set of second-order moments can be used for
any realizations of stochastic processes with the same moments [25].
The MIMO communication techniques have been an important area of focus for next-
generation wireless systems because of their potential for high capacity, increased diversity
and interference suppression. For applications such as wireless and mobile communica-
tions, MIMO systems will likely be deployed in environments where a single base be
communicated with many users simultaneously [30].
Wireless systems with transmit and receive diversity have become a very attractive
area of research due to the potential of achieving extraordinarily high data bit rates and
high spectral eciency. In this section, we suggest a systematic design procedure to the
adaptive equalization problem for time-varying channels [26].
For given adaptive finite impulse response (FIR) MIMO channels with maximum al-
lowable average input powers, the transmitter and receiver adaptive FIR MIMO filters
are jointly optimized such that the MSE between the desired and reconstructed signals is
minimized [29]. It is well known that in a MIMO system with nT transmit and nR receive
antennas. The communication channel capacity grows linearly with min (nT , nR ). It is
also desirable to transmit the same information-bearing signal over several channels. The
mode of transmission is used primarily in situations where there is high probability that
one or more of the channels will be unreliable from time to time [32].
The MIMO channel with nT transmitters and nR receivers is typically represented as
a matrix H of dimensions nR nT , where each of the coecients [H]i,j represents the
transfer function from the j th transmitter to the ith receiver. We denote the signal or
symbol transmitted from the j th transmitter yj , and collect all such symbols into an nT
dimensional vector y. With this notation, the matrix model of the channel is
s = Hy + w (31)
where w is a vector of additive noise with a variance of n2 and s is the vector of received
data, with an element in w and s for each receive antenna. In a point-to-point MIMO link
(single-user case), all outputs are available to the user for processing. In the multi-user
case, the nR receivers are distributed among dierent users; for example, if each user has
only one antenna, each user has access to only one element of s.
H is the narrowband transmission matrix of the MIMO channel, the complex correlation
coecient between antenna elements i and j at the transmitter or at the receiver side is
given in accordance with as:

Tx ,Rx
Ri,j
Tx ,Rx
i,j = (32)
T ,R2 T ,R2
i x x j x x
Antenna correlation Ti,jx Rx depends on antenna covariance Ri,j Tx Rx
between the interacting
antenna elements i and j at the transmitter or at the receiver and the respective variances
T R2 T R2
i x x and j x x . Antenna correlations at the Tx and Rx sides are combined in correlation
matrices RT x and RRx as follows:
TR
11x x T1nx Rx
Tx Rx Tx Rx
21 2n
RTx ,Rx = .. .. . .. (33)
. . . . .
Tn1x Rx Tnnx Rx nTx ,Rx nTx ,Rx
For this reason, RTx and RRx explicitly determine the correlations within the regarded
channel model. Subsequently, the eigen-analysis method is used to compute the eigen-
values k with k = 1, . . . , K of the instantaneous matrix HHH that represent the power
gains of the respective subchannels, where []H denotes Hermitian transpose. Each
( chan-)
nel matrix is given by nTx transmit and nRx receive antennas, it oers K min nTx , nRx
parallel subchannels.
It is also provided a uniform power allocation at the Tx antenna array, the information
theoretic MIMO channel capacity may be evaluated as a sum of the singular capacity
values for each of the K subchannels:
K ( )
Pk
C= log2 1 + k 2 (34)
k=1
n
In (34), Pk denotes the amount of power assigned to the k th subchannel and n2 is the
total noise power at the receiver. The signal-to-noise ratio (SNR) amount to:
E [PTx ]
SN R = (35)
n2
where E [] is represented the expectation operator to compute the expectation of the
total signal power PTx at the transmitter.
An important tool for characterizing any communication channel is capacity. In a single-
user channel, capacity is the maximum amount of information that can be transmitted as
a function of available bandwidth given a constraint on transmitted power. In single-user
MIMO channels, it is common to assume that there is a constraint on the total power
broadcast by all transmitted antennas. For the multi-user MIMO channel, the separation
signal model for the ADPLL and ADALL analysis with MIMO system.
5. Simulated Performance. The CCI problem could be become a serious impairment

for any mobile communication system. This simulation results demonstrate the per-
formance of the ADALL combined with the ADPLL separation model using the novel
adaptive FIR filter. This simulations present to verify how co-channel FM signal can be
separated. The simulations of sinusoidal and mobile signals are indicated to improve the
separation performance capability comparing the ADALL and novel separation system
models [1-3,13].
The proposed co-channel FM signal demodulation and separation systems with noise
interference environment consists the following subsystems: the FM signal generator, the
ADPLL combined ADALL model and the adaptive FIR filter with noise canceller. Figure
8 is shown the dominant and subdominant signals are separated using proposed model at
m = 0.1. The simulations demonstrate the performance of the proposed co-channel signal
separation model design with AWGN channel based on the adaptive FIR filter algorithm.
The MMSE criteria using the step size and the optimum weight coecient values are
defined by Wiener-Hopf with steepest descent method. The demodulated output signals
are approximately approached to the desired signals. The impairments of the channel
noise aect the MMSE value. By assuming the input SNR ratio ranged between 20dB
and 20dB step 10dB, it is operated at the random statistic process in this paper.
Figure 8. The separation dominant and subdominant signals for adaptive

FIR filter at m = 0.1
Two important improvement performance cases are discussed in simulation results.
The proposed algorithm of this paper can be suited to implement the perfect separation
performance index for both of the low SNR and the large ICR conditions for the proposed
algorithm. The simulated performances can be presented the ADPLL with ADALL model
for the adaptive FIR filter better then ADPLL with ADALL for the butterworth filter.
The performance of MMSE via SNR ratio is shown as Figure 9 by varying dierent m
values. In Table 1, we summarize the MMSE of proposed and conventional methods with
dominant signal by varying ICR (i.e. m) and the SNR ratios.
Figure 9. The performance of MMSE via SNR ratio

Table 1. The MMSE with dominant signal for SNR via ICR
(a) Proposed method for ADPLL and ADALL

XXX
XXXSNR (dB)
XXX 20dB 10dB 0dB 10dB 20dB
ICR XXX
0.1 0.0017 0.0009 0.00015 0.000098 0.000011
0.5 0.0029 0.0012 0.00023 0.000105 0.000023
0.9 0.0047 0.0014 0.00030 0.000160 0.000037
(b) Traditional method for PLL and ALL

XXX
XXXSNR (dB)
XXX 20dB 10dB 0dB 10dB 20dB
ICR XXX
0.1 0.3705 0.1855 0.0391 0.0108 0.0015
0.5 0.4294 0.2421 0.0502 0.0115 0.0032
0.9 0.5053 0.2999 0.0546 0.0125 0.0061
6. Conclusions. In this paper, the proposed algorithm combined the ADPLL with the
ADALL system for all-digital analysis that is more perfect to separate the CCI or multi-
channel FM signals in the same carrier frequency environment. It is also found to suppress
the distortion that caused by the AWGN channel transmission.
The new combined all-digital method can achieve the expectative goal by modifying
the conventional ALL system (i.e. analog ALL) with the adaptive filter algorithm. For
CCI FM signal with AWGN interference, the dominant and subdominant signals have
been separated completely. The proposed algorithms recover the multiple original or
modulating signals and suppressed the distortion caused by conventional all analog ALL
or PLL systems [1].
It also solve the co-channel or multi-channel transmission problems for suppressing the
AWGN interference in the noisy transmission channel. The major contribution of paper
adopts to separate co-channel and multi-channel transmissions by using MIMO system.
In this paper, the performance of the fully digital separation system can be obtained more
eectively.
The advantage of the adaptive FIR filter is proved to obtain the better tracking ca-
pability and all of the weights are updated to MMSE value. We compare ADPLL and
ADALL with all analog PLL and ALL for channel SNR ratio between 20dB and 20dB,
the performance index dierence is approximately 260 times.
The proposed method can be adopted to the monitor and security systems for the
signal intercepted applications for the future. The proposed system of this paper is the
virtue and advantage to implement the security field. In general words, it is substituted
and replaced the encryption and decryption systems but operated at no multi-access
communication systems to save the cost and complex circuits.
Appendix 1.
n n
1 1
y [n] = ( )d + eL [n] = 2fc + 1 ( ) + ( ) + n ( )d (A.1)
Kd n1 Kd n1

fADP LL ( ) = [2fc + 1 ( ) + ( ) + n ( )] d
[ (A.2)
1 ( iw ) 1
= fc + 1 ( )d + i log e
2 d
+ 2
2 wd di log(m eiwd + 1)
1 1 ]
2( 1
+ )+ 2 ( ) wd 2
wd di log
iw
e md wd log (eiwd + m) log
iw
e md 4
[
1 ( ) 1 1
+ i log eiwn (2 ) + (2 )
2 rn ( )eiwn
wn di log r( ) wn di log r(r)e
iwn
n ( )
1 ]
1
+ 2 ( ) wn 2
wn log (r ( ) eiwn + rn ( )) log r(r)e
iw n
4
n ( )

1 ( iw ) 1
fADP LL ( ) = i log e d
+ fc + 1 ( )d
2 4
2 wn log (eiwn )2
1 1
(
2 ) ( 2 )
wd di log mer( ) wn log (m eiwn + r ( )) log mer( )
iwn iwn
1
+ (2 )
rn ( )eiwn
wn di log r( )
1
+ 2 ( iwn
)
wn log (r ( ) eiwn + rn ( )) log r( )em
1 1
2 2( )
+ +
wd di log (m eiwd + 1) wd di log e md
iw
1 ]
1
+ 2 ( iw ) wd 2
wd log (eiwd + m) log e d 4

m
(A.3)
1
fADP LL1 ( ) = 2
s1 (A.4)
d di log(m eid + 1)
1
2( ) s2
fADP LL2 ( ) = i
(A.5)
d di log e md
REFERENCES
[1] A. M. Pettigrew, The Amplitude-Locked Loop the Theory and Operation of the FM 201/5 High
Performance FM Demodulator, Ampsys Company Document, 1996.
[2] G. J. Jong, T. J. Su and T. J. Moir, The performance of the amplitude-locked loop with co-channel
interference, Electronics Letters, vol.34, no.8, pp.719-720, 1998.
[3] G. J. Jong, T. J. Moir, A. M. Peettigrew and T. J. Su, Improvement of FM demodulator with
co-channel FM interference, Electronics Letters, vol.35, no.20, pp.1758-1759, 1999.
[4] J. E. Ohlson, Phased-locked loop operation in the presence of impulsive and gaussian noise, IEEE
Trans. on Communications, vol.Com21, no.9, pp.991-996, 1973.
[5] T. J. Moir and A. M. Pettigrew, Inband noise canceling in FM systems: The white noise case,
Electronics Letters, vol.28, no.9, pp.814-815, 1992.
[6] B. C. Sarkar, Phase error dynamics of a first-order phase locked loop in the presence of co-channel
tone interference additive noise, IEEE Trans. on Communications, vol.38, no.7, pp.962-965, 1990.
[7] Y. L. Dong and G. Y. He, On spurious phase modulation caused by second-harmonic terms for
digital phase-locked loop, Proc. of the 6th International Conference on ITS Telecommunications,
pp.1180-1183, 2006.
[8] M.-T. Liu, A Fuzzy Rule-based DPLL for Maximum Likelihood Phase Estimation in Signal Trans-
mission, Master Thesis, National Tsing Hua University, 1995.
[9] T. Banerjee and B. C. Sarkar, Phase error dynamics of a class of DPLLs in presence of cochannel
interference, Signal Processing, vol.85, no.6, pp.1139-1147, 2005.
[10] W. Li and J. Meiners, Introduction to phase-locked loop system modeling, Analog Applications
Journal, pp.5-10, 2000.
[11] S. Haykin, Communication System, 4th Edition, Wiley, 2001.
[12] G. J. Jong, T. J. Moir and A. M. Pettigrew, Co-channel FM interference suppressed by using the
amplitude-locked loop for digital communication, Electronics Letters, vol.34, no.8, pp.719-720, 1998.
[13] C. T. Charles and D. J. Allstot, Amplitude-locked loop analysis for calibration applications, Elec-
tronics Letters, vol.42, no.11, pp.614-615, 2006.
[14] Y.-J. Chen, G.-J. Horng and G.-J. Jong, The separated speech signals combined the hybrid adaptive
algorithms by using power spectral density and total harmonic distortion, Multimedia and Ubiquitous
Engineering, pp.825-830, 2007.
[15] Y.-J. Chen, G.-J. Horng, G.-J. Jong and T.-J. Su, The amplitude-locked loop system of co-channel
interference in AWGN channel, Proc. of the 6th Annual Conference on Wireless Telecommunications
Symposium, Pomona, CA, 2007.
[16] G.-J. Jong, P.-J. Liao, C.-Y. Jung and T.-J. Su, Multi-channel interference separation for the AWGN
channel, Intelligent Signal Processing and Communication Systems, pp.581-584, 2005.
[17] J. Gaspar, S. F. Chen, A. Gordillo, M. Hepp, P. Ferreyra and C. Marques, Digital lock in amplifier:
Study, design and development with a digital signal processor, Microprocessors and Microsystems,
vol.28, no.4, pp.157-162, 2004.
[18] Y. Iwanami, A DSP DPLL demodulator with sequence estimator for CP-FSK signals in the presence
of large Doppler shift, Universal Personal Communications, vol.1, pp.351-356, 1998.
[19] Y. Iwanami, A DSP DPLL demodulator with sequence estimation for MSK signal in the presence
of Doppler shift on Rician fading channel, Personal, Indoor and Mobile Radio Communications,
Wireless Networks-Catching the Mobile Future, vol.1, pp.332-337, 1994.
[20] B. C. Sarkar and S. Chattopadhyay, Symmetric lock-range multilevel quantized digital phase locked
FM demodulator, IEEE Trans. on Communications, vol.38, no.12, pp.2114-2116, 1990.
[21] A. Jameel, M. Y. Siyal and N. Ahmed, Transform-domain and DSP based secure speech communi-
cation system, Microprocessors and Microsystems, vol.31, no.5, pp.335-346, 2007.
[22] I. Galton, Analog-input digital phase-locked loops for precise frequency and phase demodulation,
IEEE Trans. on Circuits and Systems II: Analog and Digital Signal Processing, vol.42, no.10, pp.621-
630, 1995.
[23] T. Banerjee and B. C. Sarkar, Phase error dynamics of a class of modified second-order digital
phase-locked loops in the background of cochannel interference, Signal Processing, vol.85, no.8,
pp.1611-1622, 2005.
[24] J. G. Proakis, Digital Signal Processing, John Wiely & Sons, pp.258-272, 2000.
[25] D. G. Manolakis and V. K. Ingle, Statistical and Adaptive Signal Processing: Spectral Estimation,
Signal Modeling, Adaptive Filtering, and Array Processing, Artech House Inc., 2005.
[26] H. M. Karkhanechi, E. A. Fain and B. C. Levy, An investigation of carrier recovery techniques for
space-time coded MIMO wireless systems, Wireless Communications and Networking Conference,
vol.1, pp.454-458, 2002.
[27] T. K. Yeung and S. F. Yau, A multi-input multi-output adaptive FIR filter with guaranteed conver-
gence for feedforward ANC system, Proc. of the 1999 IEEE International Symposium on Circuits
and Systems, vol.3, pp.17-20, 1999.
[28] A. Maleki-Tehrani, B. Sayyarrodsari, B. Hassibi, J. How and J. Cio, Estimation-based synthesis of
H -optimal adaptive equalizers over wireless channels, Proc. of GLOBECOM, vol.1A, pp.457-461,
1999.
[29] A. Hjorungnes, Minimum MSE transmitter and receiver FIR MIMO filters for multi-user uplink com-
munications, Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing,
vol.4, pp.497-500, 2004.
[30] Q. H. Spencer, C. B. Peel, A. L. Swindlehurst and M. Haardt, An introduction to the multi-user
MIMO downlink, IEEE Communications Magazine, vol.42, no.10, pp.60-67, 2004.
[31] O. Klemp and H. Eul, Analytical approach for MIMO performance and electromagnetic coupling in
linear dipole arrays, Wireless Communication Systems, pp.586-590, 2005.
[32] J. G. Proakis, Digital Signal Processing, John Wiely & Sons, pp.709, 2000.
[33] M. Hassani, Approximation of the dilogarithm function, Journal of Inequalities in Pure and Applied
Mathematics, vol.8, no.1, pp.3-10, 2007.
[34] E. S. Ginsberg and D. Zaborowski, The dilogarithm function of a real argument, Communications
of the ACM, vol.18, no.4, pp.200-202, 1975.
[35] G.-J. Horng, All-Digital Amplitude-Locked Loop Separation System for Adaptive FIR Filter Anal-
ysis, Master Thesis, Institute of Electronic Engineering National Kaohsiung University of Applied
Sciences, 1997.
[36] C.-H. Hsu, Downlink MIMO-SDMA optimization of smart antennas by phase-amplitude perturba-
tions based on memetic algorithms for wireless and mobile communication systems, International
Journal of Innovative Computing, Information and Control, vol.5, no.2, pp.443-460, 2009.
[37] H.-F. Huang, C.-W. Chan, C.-H. Lin and H.-W. Wang, A low-computation conference key system for
mobile communications, International Journal of Innovative Computing, Information and Control,
vol.5, no.2, pp.461-466, 2009.
[38] K. Yamada, K. Kimura, H. Yuki and K. Yoshida, The home network system by mutual complement
of wireless and wired communications, ICIC Express Letters, vol.2, no.1, pp.73-79, 2008.
[39] H. Chen, Q. Ding, L. Ding and X. Dong, Experimental study on secure communication of dierent
scroll chaotic systems with identical structure, ICIC Express Letters, vol.2, no.2, pp.201-206, 2008.
c ISSN 1349-4198
VALUE BASED INTELLIGENT REQUIREMENT PRIORITIZATION

(VIRP): EXPERT DRIVEN FUZZY LOGIC BASED
PRIORITIZATION TECHNIQUE
Muhammad Ramzan, M. Arfan Jaffar and Arshad Ali Shahid

Department of Computer Science
National University of Computer and Emerging Sciences (FAST-NU)
A. K. Brohi Road, H11/4 Islamabad, Pakistan
{ muhammad.ramzan; arfan.jaar; arshad.a.shahid }@nu.edu.pk
Received July 2009; revised November 2009
Abstract. Requirement Prioritization is a very critical but often neglected area of re-
quirement engineering. Experience has shown that without proper prioritization of re-
quirements presented by various stakeholders, the end product usually fails to meet its
objectives optimally. In fact in many instances, the product is considered a failure be-
cause it fails to meet its core objectives. Several requirement prioritization techniques
have been presented by various researchers over the past years. Working with these tech-
niques has exposed several limitations when applied in software projects. In this paper,
we have presented a novel multi-level value based intelligent requirement prioritization
technique using fuzzy logic. We have introduced and applied the concept of requirement
value to prioritize requirements. We have performed extensive experimentation using our
proposed technique along with existing techniques. Results have shown that our technique
has achieved superior prioritization results and consistency. The experiments have also
shown that proposed technique is capable of delivering impressive prioritization under
various circumstances.
Keywords: Requirement engineering, Requirements prioritization, Fuzzy systems, In-
telligent requirements prioritization
1. Introduction. Software Engineering is one of the youngest engineering domains which

emerged as recently as somewhere in the middle of 1980 as an accepted engineering disci-
pline. The aim of SE is to create software products, services or their artifacts in order to
meet the requirements posed by stakeholders while meeting quality constraints imposed
on them. In order to meet both these objectives, any software development derives its pur-
pose and meaning from the requirements posed by various stakeholders. In this context,
better elicitation, modeling and analysis of requirements plays a very critical role towards
development of a quality software. Requirement Engineering is an established domain
of knowledge within software engineering which establishes practices and principles for
eective requirement elicitation, modeling, specification, documentation, etc. One very
important but often neglected practice of software requirement engineering is requirement
prioritization. Requirement prioritization is the process of establishing worth and value
of various requirements posed by multiple stakeholders based on certain established cri-
teria of their utility for the ultimate software product. Several requirement prioritization
techniques have been presented by authors. These techniques are both quantitative and
qualitative in their nature. Some well known requirement prioritization techniques include
Analytical Hierarchy Process (AHP), Cumulative Voting, Numerical Assignment, Rank-
ing, Theory W, Requirement Triage and Wiegers Method, etc. And there are several
other techniques which we shall discuss in this paper.
1017
1018 M. RAMZAN, M. A. JAFFAR AND A. A. SHAHID
Requirement prioritization enables us to understand the significance of requirements

vis-`a-vis the system to be developed and among requirements as well. With requirement
prioritization, we can identify the focus areas which need most of our attention in order to
develop a product which optimally meets the requirements of the stakeholders. In most
of the situations, due to budget and time constraints, it becomes impossible to implement
all the requirements posed by stakeholders. Also the nature of many projects is such that
requirements are implemented in a staged environment. In both of these scenarios, we need
requirement prioritization. We can prioritize requirement to realize which requirements
can be delayed or altered so that other urgent requirements can be implemented and to
what degree. We can also use requirement prioritization to determine which requirements
to be implemented in earlier stages or later stages. We have been working with several
funded projects during our research. These projects are faced with both of the above
mentioned situations. We have found it very important to prioritize requirements in their
true sense in order to develop a meaningful and successful product.
Requirement prioritization was a new practice in our specific development environment.
So, our project needs required us to study further into various requirements prioritiza-
tion techniques so that we can select one which can best suit our peculiar development
environment. Our finding was that there is severe deficiency of any experimental results
to determine which technique to prefer. Consequently, during this period of research and
development, we studied various requirement prioritization techniques and tried to imple-
ment them on experimental level at various projects. We soon realized that all of these
techniques worked well within certain situations but had some inherent problems attached
with them which made it impossible to implement any one of these across the organization
for all dierent kinds of projects. The main hindrances faced by us while implementing
these techniques were related to cost, time and handling of evolving and creeping require-
ments. It was very important for us to work with a time and cost ecient prioritization
technique. Since we were testing these techniques for small to medium software projects
which needed a very ecient prioritization mechanism, our ultimate aim was to decide
upon a technique which yielded better accuracy in requirement prioritization while at the
same being ecient in terms of cost and time. We also were in search of a technique which
could reduce the dependence on human factor due to certain obvious reasons which we
shall elaborate in later sections. Unfortunately any of existing techniques could not satis-
factorily answer all these problems. If some techniques like cumulative voting or ranking
were ecient in terms of cost and time, their results were not very impressive. On the
other hand, better techniques like AHP required very skilled human resource and a lot of
capital to be applied in a proper environment.
In order to overcome these problems, one solution before us was to develop an artificially
intelligent expert driven requirement prioritization technique. In one of our previous
works, we had presented the initial sketch of a value based requirements prioritization
technique [52]. This technique was very much similar to Theory W. In this approach, the
end users and experts were asked to prioritize their requirements based upon the value
that accomplishment of this requirement may have for the system. The salient feature of
this technique was a combination of end users and experts in the process of requirement
prioritization process. However, while implementing this technique, we encountered two
major problems.
The technique generated a lot of conflicts at the end of requirement prioritization

process. Conflict resolution was a very long and time consuming process which
needed to be followed at the end of every prioritization session.
VALUE BASED INTELLIGENT REQUIREMENT PRIORITIZATION 1019
The technique was completely manual. The prioritization was done through human
endeavor and element of human bias was noticeable.
While applying the technique proposed in [48], we realized the need for an automated
requirement prioritization technique in order to overcome both the above mentioned lim-
itations. In this paper, we have proposed and elaborated upon a fuzzy logic based intel-
ligent requirement prioritization technique. This technique uses fuzzy logic to prioritize
requirements presented by various stakeholders. This modified scheme is basically a mul-
tilevel prioritization where end users, experts and intelligent system perform requirement
prioritization at various levels. By using this technique, we achieved multiple benefits.
On one hand, we achieved certain degree of automation which ultimately resulted in a
more ecient technique. On the other hand, we were also able to reduce the emphasis
of our approach on skilled requirement engineer solely as now intelligent component as
well as end user was also contributing towards ultimate prioritization. At the same time
developing this technique also oered us with the opportunity of eliminating the role of
requirement engineer altogether since we are currently working on incorporating a neural
network based component to replace engineer during second stage of prioritization.
In order to establish the utility of intelligent requirement prioritization technique, we
applied this as well as a representative group of other techniques on several projects and
determined the degree of success. We have also presented these findings and observations
in this paper. Based on these, we have established the case for an artificially intelligent
hybridized technique on which we are currently working. We have also presented a frame-
work for future evaluations of results acquired by application of our proposed mechanism
with existing requirements prioritization techniques. We have presented preliminary re-
sults of application of this framework on various techniques.
The paper is arranged as follows: After a brief introduction in Section 1, we have
described a comprehensive literature review in Section 2. In Section 3, we have presented
an elaborate overview of existing requirement prioritization techniques. In Section 4, we
have described the concept and implementation of fuzzy logic based intelligent requirement
prioritization technique. These techniques have been applied on several projects to observe
their performance. Findings and evaluations of these experiments have been presented in
Section 5. Conclusion and future work is described in Section 6.
1.1. Major contributions. Following are the major contributions:

In this paper, we have proposed and elaborated upon a fuzzy logic based intelligent
requirement prioritization technique. This technique uses fuzzy logic to prioritize
requirements presented by various stakeholders.
This modified scheme is basically a multilevel prioritization where end users, experts
and intelligent system perform requirement prioritization at various levels.
In order to establish the utility of intelligent requirement prioritization technique, we
applied this as well as a representative group of other techniques on several projects
and determined the degree of success. We have also presented these findings and
observations in this paper. Based on these, we have established the case for an
artificially intelligent hybridized technique on which we are currently working.
We have also presented a framework for future evaluations of results acquired by ap-
plication of our proposed mechanism with existing requirements prioritization tech-
niques. We have presented preliminary results of application of this framework on
various techniques.
1.2. Paper organization. The paper is arranged as follows: After a brief introduction
in Section 1, we have described a comprehensive literature review in Section 2. In Section
3, we have presented an elaborate overview of existing requirement prioritization tech-

niques. In Section 4, we have described the concept and implementation of fuzzy logic
based intelligent requirement prioritization technique. These techniques have been ap-
plied on several projects to observe their performance. Findings and evaluations of these
experiments have been presented in Section 5. Conclusion and future work is described
in Section 6.
2. Literature Review. Requirement Engineering (RE) is one of the earliest and very
critical phases of software engineering. RE as a knowledge stream is basically aimed
at acquisition, modeling and documentation of requirements for the software product.
Requirement Engineering is a unique discipline in the sense that it not only incorporates
the concepts of engineering but also of human and social sciences. Sometimes, referred
to as requirements analysis, RE is treated as a sub discipline of system engineering and
software engineering. Requirement engineering aims to define precisely the requirements
that need to be met. This is not an ordinary task. According to Fred Brooks, deciding
what needs to be built is the most dicult part of software development [1]. We can
visualize one software requirement as one documented need that software product should
accomplish. Usually requirements are classified as either as process based and product
based or functional and non functional requirements. Software requirement can best be
defined as the description of system functionality along with its quality concerns.
Requirement prioritization is the next logical task once requirements have been elic-
itated and properly analyzed. In most cases, it is really dicult to meet all the re-
quirements that have been given by various stakeholders. Most of the times, elicitated
requirements are vague, conflicting or out rightly false. Over period of time, as our un-
derstanding of the system becomes more and more clear, the requirements start attaining
their actual or specific shape. Similarly, in many cases, requirements are implemented in
a staggered fashion. In such circumstances, it becomes important to arrange the require-
ments in a prioritized order to develop the system in more realistic way. This task becomes
even more dicult when performed early in the lifecycle. According to Karlsson et al. [2],
one of the greatest problems of software engineers is development of such a product which
doesnt satisfy the needs and expectations of stakeholders. To overcome this problem,
same authors came up with the idea of prioritizing the requirements according to their
value in the paper [3] titled A Cost-Value Approach for Prioritizing Requirements. Sub-
sequently, many other researchers [4,5] emphasized upon the significance of requirement
prioritization. According to Ed Yourdon, prioritization of requirements is an extremely
important issue [6] where as Lubars et al. stated that prioritization of requirements was
one major topic of discussion during the survey that they undertook[7].
There are various methods for prioritizing requirements [8]. Some major techniques
are Analytical Hierarchy Process (AHP) [9-11], Binary Search Trees [12-15], 100 points
method [16,17], planning game [18,19], numerical assignment technique [20,21] and theory
W 20 [22], etc. The consensus of all of these studies is that the projects success or failure
is strongly determined by how eectively, we can prioritize the requirements.
Computational Intelligence and soft computing are established techniques which have
helped resolve many real world problems. These methodologies include Artificial Neural
Networks, Fuzzy Logic, and Evolutionary Computing etc. Fuzzy logic is a technique
centered on fuzzy set theory. Thus it is considered as an extension of classical set theory
[23,24]. The concept of fuzzy sets as introduced by Lotfi Zadeh [25] can be considered
as generalization of the classical sets which are crisp in their nature. The purpose of
fuzzy logic is to both reduce the complexity of existing solutions as well as increasing the
accessibility of control theory [26]. Computational intelligence based techniques including
fuzzy logic have been widely used recently to tackle many real world problems. Some of
the recent applications of computational intelligence can be found in [27,28]. Fuzzy logic
has also found its way in software engineering where it has most recently been used in
eort estimation [29], software project similarity [30], software development [31], project
evaluation [32], software evolution [33] etc. Fuzzy logic has been used in requirement
engineering as well for various tasks [34,35]. Chengdong Li et al. [36] presented a novel
approach for using prior knowledge and sampled data in fuzzy systems. Some other
recent developments in the domain of fuzzy logic [37-39] have also presented new vistas of
research in software engineering. What is evident after studying extensively is the fact that
researchers in software engineering need to apply artificial intelligence in various domains
of knowledge of SE to propose such techniques which are evolvable and can intelligently
generate ecient results. In this paper, we have proposed another application of fuzzy
logic in software engineering. We have proposed to use fuzzy logic in the domain of
requirement engineering. We suggest introducing fuzzy logic to determine the priority of
requirements. The next section is devoted to elaboration of this proposed technique.
In this paper, we proposed another application of fuzzy logic in the domain of re-
quirement engineering. We suggest introducing fuzzy logic to determine the priority of
requirements. The next section is devoted to elaboration of this proposed technique.
3. Requirement Prioritization Techniques: An Overview. As mentioned in the

literature review, there are various requirement prioritization techniques. However, no
evaluation of these techniques has been made so far so that their utility and relevance can
be determined. In this section, we give a comprehensive overview of various requirement
prioritization techniques.
3.1. Analytical hierarchy process (AHP). AHP is a relative assessment based sta-
tistical technique to prioritize requirements for software products. If we have n number
of requirements, AHP makes nx(n 1)/2 comparisons at each hierarchy level. In real life,
we are usually working with requirements which have multiple objectives. AHP works
as an ecient technique in these kinds of situations by making pair wise comparison to
calculate relative value and cost of each requirement against the other one. This sig-
nificantly large number of comparisons makes the technique less eective as increase in
number of comparisons always takes place at the rate of O(n2 ). AHP is regarded as a five
step method.
1. Establish completeness of requirements.
2. Apply the pair-wise comparison method to assess the relative value.
3. Apply the pair-wise comparison method to assess the relative cost.
4. Calculate each candidate requirements relative value and implementations cost, and
plot each on a cost-value diagram.
5. Use the cost-value diagram as a map for analyzing the candidate requirement.
A sample cost value diagram is shown in Figure 1. A large number of studies have been
made in recent past to determine the eectiveness of AHP for requirements prioritization.
Karlsson [3,40] has made a number of studies which have shown the eectiveness of this
technique in industrial settings. At the same time, some other studies [19,41] have talked
about AHP as being dicult, less ecient and time consuming. AHP can be considered
as a highly sophisticated and complex technique which can establish prioritization at the
level of individual requirements.
Eorts have been made to reduce the number of comparisons. However, this has al-
ways enhanced the margin of error. In our opinion, this tradeo is necessary since some
comparisons may actually never bee needed.
Figure 1. Sample cost-value diagram for AHP
3.2. Cumulative voting (CV). Also referred to as 100 $ test or 100 point method
sometimes, resembles very much to voting mechanism of brainstorming sessions. Each
stakeholder is given 100 points that he or she can distribute among the requirements as
they seem fit. It seems like very straightforward mechanism but it becomes complex as
the number of requirements increases or the stakeholders involved become too many. This
scheme also has several drawbacks associated with it. Firstly this scheme treats all the
requirements as equal opportunity candidates. Secondly, the element of bias can never be
over ruled. It has been observed that in second or subsequent voting, stakeholder assign
more votes to their favorite requirements in order to move them up. Many researchers [42-
44] have pointed out shortcomings in cumulative voting mechanism. Cumulative voting
technique can also be considered as one which is complex in its nature but attempts to
prioritize requirements at their individual level.
3.3. Numerical assignment (NA). It is probably the most common prioritization tech-
nique which is also very easy to use. In the first step, requirements are classified into
dierent groups. These requirements are given to each stakeholder. Each requirement
within these groups is assigned a number on a scale of 1 to 5 by individual stakeholders.
The final ranking is determined by calculating average of all the ranking given to each
requirement by every stakeholder. This technique because of its ease of sue has also been
suggested by IEEE Std. 830-1998. Since the requirements are first classified into groups
and then prioritized so we can say that this technique does not prioritize requirements at
the level of individuality. Instead one level of abstraction is introduced.
Despite its wide applicability, this technique also poses several problems. Clear def-
inition of the groups is one major drawback. Second problem is that even with clear
definitions, stakeholders will tend to put most of their requirements into critical groups
because of their bias (which cant be overruled). Another fact that we have to be mindful
about is that within each group, all the requirements are initially at the same priority level.
Most of these drawbacks in numerical assignment technique have been well documented
in [42,45,46].
3.4. Ranking. This technique is more suitable in the environment where a single stake-
holder is involved. If there are n number of requirements, these requirements are ranked
from 1 (most significant) to n (least significant). This ranking is exclusive in its nature
because requirements are not ranked relative to other requirements as is the case of AHP
or cumulative voting. Various techniques like bubble sort, quick sort or binary search
techniques can be used to achieve this ranking.
There are two major drawbacks associated with this technique. First major problem
is that it can cause more conflicts than agreements when applied in an environment of
multiple stakeholders. The second drawback is that requirements are viewed and ranked
in isolation. The impact of one requirement over the other doesnt play any role in overall
prioritization. Since requirements can have multiple dimensions to them so researchers
have devised a mechanism of combining all of these dimensions and calculating a mean
priority for each requirement [47]. This modification has its own limitations as well as
has been shown in [47].
3.5. Top-ten requirements. This technique prioritizes only the most important re-
quirements into a set of top-ten from a larger set of requirements. Selection of the most
important requirements is subjective to the project environment and so it can be erro-
neous if based on human judgment. Since we create only a set of top-ten requirements, no
prioritization within this set takes place. This can be termed as a shortcoming in many
situations. The technique can be applied in conjunction with other techniques to achieve
better results. According to lausen [48], it is mostly helpful in situations where there are
multiple stakeholders with uniform or similar significance.
3.6. Theory W. The main proponent of this theory is Dr. Barry Boehm who introduced
this concept [49] in 1989. Popularly known as Win-Win model, this technique relies heavily
on negotiation to resolve any dierences of opinion among various stakeholders. The
negotiations are conducted in such a way that each stakeholder is in a Win situation.
The principles of this technique are progress based on predefined plan, risk assessment
and risk handling. In this technique, users are asked to rank their requirements before
actual negotiations start. Users are asked to carefully consider which requirements they
are willing to negotiate and which they are not. Theory W has been an active area of
research among scholars which has been applied in not only requirement engineering but
also in other domains of software engineering. Theory W is a major constituent of Value
Based Software Engineering (VBSE) agenda and principle as well.
3.7. Planning game (PG). This specific requirement prioritization technique is very
suitable to extreme programming. In this specific technique, requirements are prioritized
in consultation with customers. This is a variation of numerical assignment technique as
discussed in Section 3.3. However, it oers more flexibility than numerical assignment
where users are asked to essentially divide the requirements into three groups.
Some other new and innovative techniques to emerge recently include Requirement
Triage (RT) [50] and Wiegers Method (WG) [51]. In requirement Triage, each require-
ment is prioritized relative to the resources that are necessary to meet that specific require-
ment. In this way, a subset of requirements is selected which can optimize the probability
of success of product while using the allocated resources eciently. In Wiegers method,
the priority of each requirement is set by determining the utility of that requirement to
the customer as well as penalty incurred by the customer if that requirement remains
unfulfilled.
In this section, we have presented a brief overview of existing requirements prioritiza-
tion techniques. In this next section, we shall present a theoretical evaluation of these
techniques as well as present the brief idea of our proposed and implemented approach.
4. Value Based Intelligent Requirement Prioritization (VIRP): The Proposed

Technique. Value based intelligent requirement prioritization (VIRP) as shown in Fig-
ure 2 is basically a multilevel prioritization and classification technique. This technique
involves the use of stakeholders, experts and automated fuzzy logic based system at var-
ious stages to iteratively prioritize the requirements. This iterative prioritization ensures
that requirements are evaluated again and again by dierent actors and a more meaningful
and realistic result is achieved.
Figure 2. Flow graph of fuzzy based intelligent requirement prioritization process
4.1. Requirement elicitation and stakeholder level prioritization. Requirement

elicitation process starts at the early stages of requirement engineering. This is a stan-
dard process that is followed in any requirement engineering technique. In order for our
technique to gather its first level prioritized requirements, we introduce two new concepts
in our requirement elicitation.
The requirement elicitation is done electronically. All stakeholders submit their re-
quirements on specially designed web scripts. Stakeholders submit their requirements
according to their own priority.
A new concept of stakeholder profiling is introduced. While submitting their require-
ments, stakeholders also give a brief description regarding themselves, their expectations
about the system, system functionalities of their choice etc. Requirements engineers
assigned to gather requirements also write down their observations regarding those stake-
holders independently. Later both of these are combined to form a stakeholder profile.
Stakeholder profile template is given in Figure 3. These elicitated requirements as well as
stakeholders profile are stored in specially designed requirements database.
4.2. Expert level prioritization. Once all the requirements from stakeholders along
with their profiles have been elicitated, these are submitted to the experts for second level
prioritization. In this stage, experts perform three tasks.
(a) Success in prioritization achieved for dierent process models
(b) Success ratio achieved for two projects with varying degree of requirement clarity
(c) Success ratio achieved for two projects with dierent level of requirement change
(d) Success ratio for two projects with varying number of stakeholders and their commitment
(e) Success ratio achieved for two projects with dierent level of constraints i.e. budget,
time etc.
Figure 3. Experimental evaluation results of various prioritization techniques
In the first task experts quickly review all the requirements given by stakeholders in-
dependent to the stakeholders profile and the requirements given by other stakeholders
and modify the prioritizations given by stakeholders in a quick manner if some glaring
problem emerges.
In the second task, experts review each stakeholder profile and allot it a score in the
range of 1-10 based on his significance and the overall impact that requirements posed by
him may have on the success of the project.
In the third and most significant task, experts assign value to requirements in order
to prioritize them. This valuation is based on certain requirements classification factors
(RCFs) which can enhance the overall value of the requirement with their presence. RCFi
can be defined as one factor whose degree of presence or absence in a requirement can
have a direct bearing on its value. We have identified ten requirement classification factors
whose names and brief descriptions are given in Tables 1 and 2. These RCF are divided
into two groups.
4.2.1. Project specific RCFs. These RCF illustrate the value of requirement in relation
to the project for which it has been elicitated. The importance of requirement for the
project is determined through these RCFs. These include feasibility, modifiability, ur-
gency, traceability and testability. Better value for these RCFs means that this require-
ment can better enhance the quality of end product. These are represented as pRCF .
Their detailed description is given in Table 1.
Table 1. Project specific requirement specification factors (pRCF)
Name Description
The requirement is capable of being implemented within the
1 Feasibility
constraints and resources
Requirement can undergo change to optimize the system with-
2 Modifiability
out aecting the system adversely
Degree of necessity of the requirement for system to be consid-
3 Urgency
ered succesful
Requirement is such that subsequent function of the system can
4 Traceability
be traced to it. Requirements are less compound
Requirement can be tested and validated during testing phase.
5 Testability
Independent test cases for the requirement can be generated
4.2.2. Requirement specific RCFs. These RCFs relate to those attributes of requirements
which if implemented increase the overall value of requirement from the perspective of its
description quality. These include completeness, consistency, understandability, within
scope and non-redundancy. Collectively, these are represented as rRCF . Their detailed
description is given in Table 2.
Table 2. Requirement specific requirement specification factors (rRCF)
Name Description
The requirement statement has enough information to pro-
1 Completeness
ceed to the nest development phase
Requirement specifications use standard terminology and
2 Consistency there are minimum conflicts due to statement and specifi-
cations
Requirements are easy to describe and review. Require-
3 Understandability ments are grammatically correct with single and clear
meaning
Requirement does not envisage something which is not de-
4 Within Scope
scribed in original statement of scope
5 Non-Redundant Requirement is not duplicated in complete or partially
Each factor is given a certain score in the range of 0 5 depending upon the eect of
that factor on the value of given requirement. 0 denotes the total absence of that specific
RCF in any requirement whereas a value of 5 denotes the perfect presence of RCF in
that requirement. All RCFs are given scores between 0 and 5. The scores of these RCFs
are used to determine the final value of the given requirement. Using the equation given
below, the requirement value (RV) is calculated.
{ 5 }
5
RV = 0.35 + 0.02 pRCFi + rRCFi (4.1)
i=1 i=1
Requirements are prioritized using their RV which can be a number between 0.35 and
1.35. This exercise gives the prioritized requirements with respect to each factor. There
are two exceptions where we have introduced additional measures.
First Scenario
In case of situations where any two or more requirements have similar RV, we use pRCF
as deciding measure to assign priority using following rule;
If we have two requirements
I and j with value RVi and RVj such that RVi = RVj , then
if pRCFi > pRCFj
then
high priority RVi
else
high priority RVj
Second Scenario
In case of requirements where RVi = RVj and pRCFi = pRCFj as well, we use
prioritized stakeholders profile to determine which stakeholder is more important among
the two competing ones.
The requirement posed by more significant stakeholder is prioritized higher in that case.
In case of both competing requirements coming from the same stakeholder, we leave it to
the judgment of experts to assign higher priority to any requirement which he deems fit.
At the completion of these three tasks, we have a complete unified list of requirements
in their prioritized order and a prioritized list of stakeholders. Both of these are submitted
to Fuzzy logic based component to perform third level prioritization.
4.3. Fuzzy logic based requirement prioritization. Following steps are executed in
this third and final level of prioritization:
Fuzzy c-means (FCM) is a method of clustering which allows one piece of data to be
in the right position to two or more clusters. This method (developed by Dunn in 1973
and improved by Bezdek in 1981) is frequently used in pattern recognition. However,
the same model can be applied after suitable modifications for requirement prioritization
which we have accomplished. FCM starts with an initial guess for the cluster centers,
which are proposed to mark the mean location of each cluster. The initial guess for these
cluster centers is most likely incorrect. Additionally, FCM assigns every data point a
membership rank for every cluster. By iteratively updating the cluster centers and the
membership grades for each data point, FCM iteratively moves the cluster centers to the
correct place within a dataset. This iteration is based on minimizing an objective function
that symbolizes the distance from any given data point to a cluster center weighted by
that data points membership rank.
The FCM algorithm assigns requirements to each group by using fuzzy memberships.
Let XZ(x1, x2, . . . , xN ) indicates a requirement with N variables to be partitioned into c
clusters, where xi represents multispectral (features) data. The algorithm is an iterative
optimization that minimizes the cost function defined as follows:

N
C
J= Uijm xj Vi 2 (4.2)
j=1 i=1
where uij represents the membership of feature xj in the ith cluster, vi is the ith cluster
center, ||..|| is a norm metric, and m is a constant. The parameter m controls the fuzziness
of the resulting partition, and m = 2 is used in this study. The cost function is minimized
when requirements close to the centroid of their clusters and are assigned high membership
values, and low membership values are assigned to requirements with data far from the
centroid. The membership function represents the probability that a requirement belongs
to a specific cluster. In the FCM algorithm, the probability is dependent solely on the
distance between the requirement and each individual cluster center in the feature domain.
The membership functions and cluster centers are updated by the following:
1
Uij = ( ) m1
2 (4.3)

c xj vi
k=1 xj vk
and

N
Uijm xj
j=1
vi = (4.4)

N
Uijm
j=1
Starting with an initial guess for each cluster center, the FCM converges to a solution
for vi representing the local minimum or a saddle point of the cost function. Convergence
can be detected by comparing the changes in the membership function or the cluster
center at two successive iteration steps.
5. Theoretical and Experimental Evaluation of Requirement Prioritization

Techniques. In Section 3, we have given a brief overview of existing requirement pri-
oritization techniques. In Section 4, we have briefly described the implementation of
intelligent requirements prioritization technique. Several of these techniques are being
applied in software industry for a while now. Some research studies and surveys to deter-
mine the usability of some of these techniques have also been conducted (as mentioned in
previous section). So a significant mass of literature on these requirement prioritization
techniques exists.
However, no significant comparative study has so far appeared where all or most of
the above mentioned techniques might have been applied to the same set of projects.
This can be a very valuable study as it can determine that in what kind of development
environment which specific requirement prioritization technique can yield best results.
As we have already mentioned in the introduction section, we faced a severe problem
of selecting suitable requirement prioritization technique for our projects. This search
ultimately culminated in proposing and implementing our own intelligent requirement
prioritization technique. We feel that it is need of the hour to catalogue the pros and cons
of all existing as well as proposed technique at one place so that it becomes easier for
software engineering community to evaluate and select one technique which better meets
its needs.
This section is dedicated to theoretical evaluation of requirement prioritization tech-
niques. It is further subdivided into three subsections. In the first section, we have
presented a theoretical evaluation of existing requirement prioritization techniques. In
the second part, we have given a brief introduction of our intelligent requirement priori-
tization technique. In the last subsection, we have presented experimental results for our
analysis of this technique with all existing ones.
5.1. Analysis of existing requirement prioritization techniques. Literature oers

us many insights into the intricacies and working of various requirement prioritization
techniques. We have gathered certain valuable findings while working with these tech-
niques as well. In this section, we shall briefly elaborate upon those techniques.
Our experience has shown that for large projects with multi-objective requirements,
AHP is a more preferred approach among the professionals. This technique yields stat-
ically very reliable results. The experience has also shown that the cost of conducting
AHP is also marginal as compared to various other techniques. This technique however is
not suitable in the situation where requirements are fast evolving and new requirements
are being introduced at a much higher pace. The purely statistical nature of AHP (and
it is true for other techniques described above as well) makes it dicult to generate a
prioritization which accommodates these changes taking place. The technique due to
its highly time consuming nature also becomes an unfit solution for development models
where several iterations take place (unless a sucient time box is available for prioritiza-
tion process at each iteration). Cumulative voting is a human intensive exercise. In our
experience, while working with iterative process models, we were able to plan our projects
in such a way that cumulative voting could be conducted at every iteration. Since cumu-
lative voting involves human insight apart from statistical techniques, we also experienced
a more flexible and accommodating prioritization when handling changing and creeping
requirements. The cost factor for cumulative voting becomes an inhibiting factor when
we deal a project of several hundred requirements and a very tight budget. The element
of bias was also visible in some of our experiments as experts inadvertently prioritized
those requirements which they thought were more important from their perspective. We
believe that AHP is more suitable for projects with medium number of requirements and
waterfall or prototyping model. Cumulative counting on the other hand, can manage iter-
ative development quite eciently provided enough budget and expertise is provided. We
have also observed that some kind of automation is required for both of these techniques.
This automation can reduce the time requirement of AHP and make it more suitable for
iterative development while it can also reduce the element of bias for cumulative counting
and lend it more credibility.
Numerical assignment technique was one of the most dicult to work with in our experi-
ence. Despite it being the most commonly used technique, we face almost insurmountable
problems while working with numerical assignment. This technique was rendered useless
when working in iterative environment. It was dicult to identify and gather all the stake-
holders in each iteration, determining the exact status of their requirements (including all
changes, creeps and incomplete) and then performing classification based prioritization.
Our experience has shown us that numerical assignment is very unreliable when software
is to be developed in iterations, has several stakeholders and fluidity of the requirements
is very high. Some degree of success was achieved where the stakeholders are very few and
highly oriented. Second problem while working with numerical assignment was the much
greater degree of bias that was exhibited by stakeholders when prioritizing than the bias
we experienced in the case of cumulative voting. Third problem that we experienced was
that classification posed problems instead of solutions. It was because a large majority
of requirements posed by various stakeholders were actually placed in the highest classifi-
cation by their owners while requirements posed by other stakeholders were put in lower
classifications as those were considered less important.
Ranking was another prioritization mechanism which had very low potential in modern
day development in its true sense. Exact problems as mentioned in Section 3.4 were faced
while working with ranking technique. Top ten techniques is good at establishing a set of
the most critical requirements. Our experience has shown that all these three techniques
are very dicult to work with and cant meet the objectives of requirement prioritization
in an optimal way.
Theory W is a very valuable requirement prioritization technique. It has a two tier
prioritization system which works within predefined limits. Stakeholders are given the
opportunity to prioritize their own requirements which are then further studied and ad-
justed by experts before those are presented to all for negotiations. These negotiations
last until we have asset of requirements in such a prioritized order that every stakeholder
is a winner. We were able to get much better results by applying theory W than any other
technique. The major problem that we faced while working with theory W was when re-
quirements were fluid beyond certain degree. It was impossible to perform negotiations
at each iteration. So the utility of this technique was somewhat diminished in iterative
development with highly evolving and changing requirements. Planning game is also a
better variant of numerical but the same problems persist (with somewhat less intensity).
Wiegers method and requirement triage are relatively new entrants in the field of re-
quirement prioritization. These techniques oer solutions to the problem of requirement
prioritization which are more realistic and are more in sync with ground realities. These
techniques are good in both linear and iterative process models. Our experimentations
and observations have shown that AHP and cumulative voting are best existing techniques
for linear and iterative models respectively. However, AHP being a time consuming and
purely statistical technique needs certain improvements. On the other hand, cumulative
voting is very costly and has the margin of bias and error due to total dependence on hu-
man experience. Apart from these shortcomings, the aspect of automation also demands
certain attention. All these techniques are either statistical or human based. In order to
be more reliability and flexibility in our requirement prioritization mechanism, we need
to develop a computer based approach which can prioritize requirements to a large ex-
tent on its own. Keeping all of these observations in mind as we applied all the above
mentioned techniques, we conceptualized and implemented a fuzzy logic based intelligent
requirement prioritization approach. In the following sub-section, we briefly describe our
requirement prioritization technique. After that, we shall present our experimental results
comparing the performance of all existing as well as new technique.
5.2. Experimental results for requirement prioritization techniques. The theo-

retical analysis that we have presented in Section 4.1 is based upon the application of these
techniques as well as our proposed technique. The experiments were conducted on ten
dierent projects. These projects were ongoing university level projects. Some of these
projects were sub-contracted from industry while some others were government sponsored
projects. These projects were carefully selected to represent the vast spectrum of dierent
natures of requirements. The brief description of all of these selected projects is given in
Table 1. All nine requirement prioritization techniques presented in Section 3 as well as
proposed technique presented in Section 4.2 were applied on all of these techniques. The
utility of each technique was determined by the degree of success in accurate requirement
prioritization determined by a panel of experts at the completion of the project. Selection
criterion description is as follows:
Process Model
As we have described earlier, the process model has significant bearing on the success
or failure of selected process model. The major factor to consider in process model for
requirement prioritization is whether the process model is linear or iterative in its nature.
After careful analysis we selected Mobile Business Intelligent Tool as the project which
was developed in iterative fashion while ALIF was selected for linear development model.
Level of Ambiguity
A couple of projects were selected on the basis of degree of ambiguity in their require-
ments posed by various stakeholders. This was aimed at determining the eciency of
various techniques in the presence of high ambiguity requirements as well as when the
requirements are very clear and formal. Sign Language Recognition and Interpretation
was selected as a project which had a high degree of ambiguous and faulty requirements
whereas Semantic Web Based Clinical Decision Support System had high degree of pre-
cision and clarity in its requirements.
Requirements Fluidity
Most of the times we are faced with situations where requirements are highly fluid.
Requirements change too often. Many creeping requirements emerge during the devel-
opment process. Sometimes existing requirements lose their relevance or simply cease to
exist. This aects the overall quality of requirement prioritization. We selected a repre-
sentative project titled Content Based Image Retrieval System which had highly fluid
requirements while another project beyond RADIX was selected due to low degree of
fluidity in its requirements.
Stakeholders
Number of stakeholders and their degree of involvement in the project also impacts
the overall requirement prioritization eort. We selected site management system as
a project which had few stakeholders but their involvement in the project was high. On
the other hand i Pakistan Sign Language was selected as a representative project where
several stakeholders are involved and their involvement in the project is significantly low.
Project Constraints
Project constraints (time, budget, human resource etc.) have a high impact on deter-
mining the utility of any requirement prioritization technique. We selected two project
Table 3. Selected projects and selection criteria
Project Project
Project Name Selection Criteria
ID Acronym
1 Mobile Business Intelligent tool MBIT Iterative development
Academic Learning and Informa- Linear Development
2 ALIF
tion Framework
Sign Language Recognition and In- Ambiguous Requirements
3 SLR
terpretation
Semantic Web Based Clinical Deci- Clear and Formal require-
4 semclin DSS
sion Support System ments
5 Beyond RADIX BRADIX Low fluidity requirements
Content Based Image Retrieval High fluidity requirements
6 CIRS
System
Intelligent Site Management Sys- Few stakeholders, High
7 ISM
tem Commitments
Several Stakeholders, Low to
8 i Pakistan Sign Language iPSL
Medium Commitment
9 Semantic Bees SeBe High Project Constraints
10 IMS Intrusion Detection System IIDS Relaxed Project Constraints
Figure 4. Performance comparison of leading prioritization techniques
constraints for our experimentation. These were budget and schedule. Based on this
criteria, semantic bees was selected as project which had high constraints (low budget
and tight schedule) whereas IMS Intrusion Detection System was selected as project
with low constraints (ample budget and sucient time for scheduling).
The experiments were conducted on two dierent levels. On first level, the techniques
were applied on dierent projects to determine their degree of success in precise prioriti-
zation under dierent environments. The results have been shown in Figures 3, 4 and 5
respectively. Major findings as highlighted in Figure 3 were as follows:
Projects 1 and 2 were selected to determine the performance of various techniques
on projects that follow iterative and linear process models respectively. We have ob-
served that results of prioritization often get faulty when we adopt iterative process
models since requirements change quite often. This is true for almost all requirement
prioritization techniques presented in literature. However, the results have shown
that intelligent requirement prioritization was able to best prioritize requirements
Figure 5. Empirical evaluation of prioritization techniques

when iterative model was adopted and it was followed by cumulative voting. In case
of linear model, intelligent requirement prioritization, AHP and theory W yielded
best results with intelligent requirement prioritization able to accurately prioritize
more than 90 percent of the requirements. The reason for theory W not performing
better in case of iterative development was that with every iteration, the degree of
involvement of stakeholders reduced so many ambiguities were introduced in later
stages. On the whole, it was observed that quality of all techniques deteriorated in it-
erative development model. Overall intelligent requirement prioritization techniques
performed better and more consistently in both kinds of process models.
Projects 3 and 4 were put to test with special emphasis on degree of clarity of require-
ments. It is dicult to understand the utility and ultimate value of requirements if
requirements are not very well elaborated. In that case it becomes very important
to use such mechanism which can support the application of domain knowledge to
clarify the requirements to a greater extent. It was observed that in case of project
3, where requirements were highly ambiguous, all the techniques delivered relatively
poor prioritization results. Many of the well established techniques showed around
50% of error. The better performing techniques in this case were cumulative voting,
wieger method and intelligent requirement prioritization techniques. However, the-
ory W performed the best in this case with 68% accurate prioritization. The success
of Theory W, cumulative voting as well as intelligent requirement prioritization can
be attributed to the wise and meaningful utilization of end users and customers in an
environment where perhaps customers themselves can better understand their given
requirements. Similar techniques exhibit leading results in case of project 4 where
requirements are much clear and formally written. However, it is interesting to note
that precision and formality in requirement specifications can significantly increase
the accuracy of requirement prioritization (90% for project 4 against 60% for project
three in the case of intelligent requirement prioritization). Overall, intelligent re-
quirement prioritization showed much better results than almost all techniques for
projects 3 and 4.
Projects 5 and 6 dealt with degree of change of requirements during the course of
project execution. It was observed that in case of project 5, where degree of change
in requirements was very low, almost all the leading techniques showed very good
results with theory W achieving 90% accurate prioritization. Many other leading
techniques like AHP, cumulative voting, wieger method and intelligent requirement
prioritization were also able to achieve 85% or more accurate prioritization. Project
6 was a project with higher degree of change in project requirements. The results
degraded for all of these leading techniques except for intelligent requirement pri-
oritization. In fact intelligent requirement prioritization was able to prioritize re-
quirements to a degree of 88 percent accuracy which is quite impressive. This was
due to multilayer nature of intelligent requirement prioritization technique which
involved stakeholders, experts and fuzzy logic at the same time. The resultant pri-
oritization exhibited more farsightedness and degree of anticipation of change than
all other techniques. This factor is important in modern day software development
as requirements change and evolve exceptionally quickly now.
Projects 7 and 8 experimented with degree of involvement of stakeholders. Intelligent
requirement prioritization achieved above 90% accuracy for project 7 which was much
better than all other techniques. High commitment of stakeholders was instrumental
in achieving these results. Theory W also exhibited impressive results by achieving
80% accuracy. In case of project 8 where level of commitment of stakeholders was
low, the quality of prioritization achieved by all techniques degraded. Intelligent
requirement prioritization achieved 73% accuracy followed by AHP which achieved

69%. Better performance on intelligent requirement prioritization was due to the
fact that its prioritization mechanism did not rely too heavily on stakeholders. It
used experts and fuzzy logic as well to achieve prioritization. AHP also showed
better results than other techniques because it was a statistical method with very
low reliance on stakeholders opinion. The experimental results substantiated the
fact that in several cases lack of proper involvement of stakeholders can actually
degrade the quality of end product.
The last two projects were used for experimentation to determine the impact of
project constraints on the accuracy of prioritization. These included constraints
such as time, human resource, capital etc. The results have shown that intelligent
requirement prioritization was able to prioritize requirements accurately to a degree
of more than 80% in both the cases. Prioritization techniques yielded better results
on the whole in the case of relaxed project constraints. In case of high constraints, all
techniques showed relatively poor performance as compared to intelligent require-
ment prioritization. However, when the project teams were allowed to work in a
relaxed environment, the performance and accuracy of all prioritization techniques
improved considerably.
The comparative analysis of leading techniques is shown in Figure 4. As the results
show, in most of the environment, intelligent requirement prioritization was able to per-
form better than any other requirement prioritization. We have created a profiling of
various requirement prioritization approaches based on our experimentation. The brief
profile of these techniques is given in Table 4. This profile shows that when all the fac-
tors are combined together, value based intelligent requirement prioritization technique
performs better than any existing approach.
Another evaluation of all techniques was conducted from the perspective of time con-
sumed, average cost, precision and consistency. The results have been graphically repre-
sented in Figure 5. The results clearly demonstrate that intelligent requirement prioriti-
zation techniques is less time consuming with more precision and consistency than any
other requirement prioritization technique. However it is costlier than some other tech-
niques owing to the fact that stakeholders and experts are involved in first and second
level prioritization. When the increase in cost is viewed in conjunction with other factors,
the overall cost increase is quite negligible.
As was mentioned earlier, this evaluation exercise was conducted on a limited number of
projects and was on academic basis. The need is to apply all these techniques on industrial
level in future and to gather data accordingly. This can further improve visibility into the
working and performance of these techniques. At the same time, a more comprehensive
evaluation and profiling can be generated.
6. Conclusion and Future Work. Requirement prioritization is one important activity

of requirement engineering phase in software development. There are various requirement
prioritization techniques in literature and practice. However, no significant comparative
evaluation of these techniques has been made so far. In this paper, a new intelligent re-
quirement prioritization technique has been proposed and described. This new technique
for requirement prioritization is based on fuzzy logic and is a multilevel approach. In this
technique, stakeholders, experts and fuzzy logic based system perform separate prioriti-
zations at three dierent levels. A comparative analysis based on experimental results
conducted on several projects ahs also been presented. This analysis shows that in almost
all dierent environments, intelligent requirement prioritization is able to exhibit better
and impressive results.
Table 4. Requirement prioritization techniques profile
Technique
Description
name
Better results when used in linear development model, low flui-
AHP dity in requirements and less constrained projects. Not suitable
iterative model and ambiguous requirements.
Works good in all the environments where end users and stake-
Cumulative
holders involvement is vital. Better than other techniques when
Voting
requirements are ambigious.
This technique is not recommended except for the situation wh-
Numerical en there is a low degree of change in the requirements. However,
Assignment many techniques outperform numerical assignment in this situa-
tion as well.
Only marginally eective when using linear development model.
Ranking In all other situations, the performance of this technique is mu-
ch lower than most other approached.
Partially eective when there is low degree of change in require-
Top Ten
ments. Many other techniques perform much better than this a-
Requirements
pproach in most situations.
The best among currently established approaches. Work better
than any other technique currently in practice. Performance of
Theory W
the prioritization slightly deteriorates in iterative development
model and when project constraints are too high.
Planning Not an aective prioritization approach. Yields relatively better
Game results when requirement specifications are very clear.
A new technique with encouraging results in various situations.
Best suited when requirement specifications are very clear and
Requirement
there is low degree of change in these requirements. However,
Triage
some other techniques like Theory W and Intelligent Require-
ment Prioritization perform better.
Another such technique which yields better prioritization when
Wieger
requirements are clearly specified and there is low degree of ch-
Method
ange. Not much eective in other situations.
The best performing technique in almost all situations. Only o-
Intelligent
ther technique to generate comparative results is theory W. Th-
Requirement
e only situation when quality of prioritization is less than satis-
Prioritization
factory is when requirements specifications are not very clear.
There is an urgent need to extend this work towards classification of the prioritized
requirements. Work is already underway towards extending this application in such a
way that it can automatically classify requirements as critical, essential, peripheral etc.
Work can also be done in such a way that prioritized requirements can be classified as
non-negotiable and negotiable requirements. We are also in the process of developing
neural networks to perform expert level prioritization. Using neural networks can make
this system highly evolvable and self optimizing.
Acknowledgments. The authors, Mr. Muhammad Ramzan and Mr. Arfan Jaar would
like to acknowledge the Higher Education Commission (HEC), Govt. of Pakistan and NU-
FAST for providing funding and required resources to complete this work. It would have
been impossible to complete this eort without their continuous support.
REFERENCES
[1] F. Brooks, No silver bullet: Essence and accidents of software engineering, IEEE Computer, vol.20,
no.4, pp.10-19, 1987.
[2] J. Karlsson and K. Ryan, Supporting the selection of software requirements, Proc. of the 8th Inter-
national Workshop on Software Specification and Design, 1996.
[3] J. Karlsson and K. Ryan, A cost-value approach for prioritizing requirements, IEEE Software, vol.14,
no.5, pp.67-75, 1997.
[4] X. Liu, C. C. Veera, Y. Sun, K. Noguchi and Y. Kyoya, Priority assessment of software requirements
from multiple perspectives, Computer Software and Applications Conference, vol.1, pp.410-415, 2004.
[5] L. Fellows and I. Hooks, A case for priority classifying requirements, The 3rd International Confer-
ence on Requirements Engineering, pp.62-65, 1998.
[6] E. Yourdon, Death March Projects, Prentice Hall, 1997.
[7] M. Lubars, C. Potts and C. Richter, A review of the state of the practice in requirements modeling,
Proc. of the IEEE Intemational Symposium of Requirements Engineering, pp.2-14, 1993.
[8] V. Ahl, An Experimental Comparison of Five Prioritization Methods, Master Thesis, School of
Engineering, Blekinge Institute of Technology, Ronneby, Sweden, 2005.
[9] T. Saaty, The Analytic Hierarchy Process: Planning, Priority Setting, Resource, Allocation,
McGraw-Hill, New York, 1980.
[10] T. L. Saaty and G. Hu, Ranking by eigenvector versus other methods in the analytic hierarchy
process, Applied Mathematical Letter, vol.11, no.4, pp.121-125, 1998.
[11] F. Hartwich, Weighting of agricultural research results: Strength and limitations of the analytic
hierarchy process, Research in Development Economic and Policy, Discussion Paper, Grauer Verlag,
Stuttgart, no.9, 1999.
[12] F. Hivert, J. Novelli and J. Thibon, The algebra of binary search tree, Theoretical Computer Science,
vol.339, no.1, pp.3-10, 2005.
[13] L. Xiang, K. Ushijiam, T. Zhao, T. Zhang and C. Tang, O(1) time algorithm on BSR for constructing
a random binary search tree, Proc. of the 4th International Conference on Parallel and Distributed
Computing, Applications and Technologies, pp.599-602, 2003.
[14] I. Al-furaih, S. Aluru, S. Goil and S. Ranka, Parallel construction of multidimension binary search
tree, IEEE Trans. on Parallel and Distributed Systems, vol.11, no.2, pp.136-148, 2000.
[15] C. Lee, L. Hung, M. Chang, C. Shen and C. Tang, An improved algorithm for the maximum
agreement subtree problem, Information Processing Letters, vol.94, no.5, pp.211-216, 2005.
[16] D. Lengwell and D. Widrig, Managing Software Requirements: A Use Case Approach, 2nd Edition,
Addison-Wesley, 2003.
[17] B. Regnell, M. Host, J. N. och Dag, P. Beremark and T. Hjelm, An industrial case study on dis-
tributed prioritization in market-driven requirements engineering for packaged software, Require-
ments Engineering, vol.6, no.1, pp.51-62, 2001.
[18] K. Beck, Extreme Programming: Explained, 7th Edition, Addison-Wesley, 2001.
[19] L. Karlsson, P. Berander, B. Regnell and C. Wohlin, Requirements prioritization: An exhaustive
pair-wise comparison versus PG partitioning, Empirical Assessment in Software Engineering, 2004.
[20] J. W. Brackett, Software Requirements (SEI-CM-19-1.2, ADA235642), Software Engineering Insti-
tute, Carnegie Mellon University, Pittsburgh, 1990.
[21] J. Karlsson, Towards a strategy for software requirements selection, Licentiate, Thesis 513, Linkoping
University, 1995.
[22] B. W. Boehm and R. Ross, Theory-W software project management: Principles and examples, IEEE
Trans. on Software Engineering, vol.15, no.4, pp.902-916, 1989.
[23] J. Park, D. Port and B. Boehm, Supporting distributed collaborative prioritization for win-win
requirements capture and negotiation, Proc. of the 3rd International World Multi-conference on
Systemics, Cybernetics and Informatics, Orlando, FL, vol.2, pp.578-584, 1999.
[24] A. Idri and A. Abran, Towards a fuzzy logic based measures for software projects similarity, MCSEAI
2000, Morocco, 2000.
[25] M. F. N. Ramalho, Application of an Automatically Designed Fuzzy Logic Decision Support Sys-
tem to Connection Admission Control in ATM Networks, Ph.D. Thesis, Department of Electronic
Engineering, Queen Mary and Westfield College, University of London, 1996.
[26] M. Razaz and J. King, Introduction to Fuzzy Logic, Information Systems Signal and Image Pro-
cessing Group, http://www.sys.uea.ac.uk/ king/restricted/boards/.
[27] M. A. Jaar, A. Hussain, A. M. Mirza and A. Chaudhry, Fuzzy entropy and morphology based
fully automated segmentation of lungs from CT scan images, International Journal of Innovative
Computing, Information and Control, vol.5, no.12, pp.4993-5002, 2009.
[28] A. Hussain, M. A. Jaar, A. M. Mirza and A. Chaudhry, Detail preserving fuzzy filter for impulse
noise removal, International Journal of Innovative Computing, Information and Control, vol.5, no.10,
pp.3583-3591, 2009.
[29] M. W. Nisar, Y.-J. Wang and M. Elahi, Software development eort estimation using fuzzy logic A
survey, Proc. of the 5th Internation Conference on Fuzzy Systems and Knowledge Discovery, 2008.
[30] A. Idri and A. Abran, Towards a fuzzy logic based measures for software project similarity, Proc. of
the 6th Maghrebian Conference on Computer Sciences, Morroco, pp.9-18, 2000.
[31] Shradhanand, A. Kaur and S. Jain, Use of fuzzy logic in software development, Issues in Information
Systems, vol.8, no.2, 2007.
[32] N. D. P. Shull, Project Evaluation Using Fuzzy Logic and Risk Analysis Techniques, Master Thesis,
Department of Industrial Engineering, University of Puerto Rico, 2006.
[33] L. Cerulo, R. Esposito, Tortorella and L. M. Troiano, Supporting software evolution by using fuzzy
logic, Proc. of the 7th International Workshop on Principles of Software Evolution, 2004.
[34] L. Mathai, V. Ramchandran and S. Sivanesana, Requirement analysis a fuzzy logic process, Inter-
national Journal of Software Computing, vol.3, no.6, pp.467-474, 2008.
[35] J. Yen, X. T. Liu and S. Hor, A fuzzy logic-based methodology for the acquisition and analysis
of imprecise requirements, Internation Journal of Concurrent Engineering, vol.2, no.4, pp.265-277,
1994.
[36] C. Li, J. Yi and D. Zhao, Design of interval type-2 fuzzy logic system using sampled data and prior
knowledge, ICIC Express Letters, vol.3, no.3(B), pp.695-700, 2009.
[37] X. Chen, C. Shen and C. Xu, Application of fuzzy neural network for FOG zero point drift modeling,
ICIC Express Letters, vol.3, no.3(B), pp.847-852, 2009.
[38] O. Castillo and P. Melin, Intelligent systems with interval type-2 fuzzy logic, International Journal
of Innovative Computing, Information and Control, vol.4, no.4, pp.771-784, 2008.
[39] J. Mockor, Models of fuzzy logic in a category of sets with similarity relations, International Journal
[40] J. Karlsson, Software requirements prioritizing, Proc. of the 2nd International Conference on Re-
quirements Engineering, pp.110-116, 1996.
[41] L. Lehtola and M. Kauppinen, Empirical evaluation of two requirements prioritization methods in
product development methods, Proc. of the 11th European Conference on Software Process Improve-
ment, pp.161-170, 2004.
[42] D. Lengwell and D. Widrig, Managing Software Requirements A Unified Approach, Addison-
Wesley, Upper Saddle River, 2000.
[43] P. Berander and C. Wohlin, Dierence in views between development roles in software process
improvement A quantitative comparison, Empirical Assessment in Software Engineering, 2004.
[44] B. Regnell, M. Host, J. N. och Dag, P. Beremark and T. Hjelm, An industrial case study on dis-
tributed prioritization in market-driven requirements engineering for packaged software, Require-
ments Engineering, vol.6, no.1, pp.51-62, 2001.
[45] G. Kotonya and I. Sommerville, Requirements Engineering Processes and Techniques, John Wiley
& Sons, NY, 1998.
[46] P. Berander, Using students as subjects in requirements prioritization, International Symposium on
Empirical Software Engineering, pp.167-76, 2004.
[47] P. Berander, Prioritization of stakeholder needs in software engineering, understanding and evalua-
tion, Licentiate Series, Blekinge Institute of Technology, no.12, 2004.
[48] S. Lauesen, Software Requirements Styles and Techniques, Addison-Wesley Professional, 2002.
[49] B. Boehm and R. Ross, Theory-W software project management: Principles and examples, IEEE
Trans. on Software Engineering, vol.15, no.4, pp.902-916, 1989.
[50] A. Davis, The art of requirements triage, IEEE Computer, vol.36, no.3, pp.42-49, 2003.
[51] K. E. Wiegers, Software Requirements, 2nd Edition, Microsoft Press, Redmond, 2003.
[52] M. Ramzan, S. Anwar, A. Rauf and A. A. Shahid, Need to redefine Value and case for a new
Software Valuation technique: An analytical study, Proc. of 2009 International Conference on
Computer Engineering and Applications, Manila, Philippine, 2009.
c ISSN 1349-4198
GENERALIZATION OF PROXY SIGNATURE BASED

ON FACTORIZATION
Cheng-Chi Lee1 , Tzu-Chun Lin2 , Shiang-Feng Tzeng3

and Min-Shiang Hwang4,
1
Department of Photonics and Communication Engineering
Asia University
No. 500, Lioufeng Raod, Wufeng, Taichung 41354, Taiwan
cclee@asia.edu.tw
2
Department of Applied Mathematics
Feng Chia University
No. 100, Wenhwa Road, Seatwen, Taichung 40724, Taiwan
3
Department of Library and Information Science
Fu Jen Catholic University
510 Jhongjheng Rd., Sinjhuang City, Taipei County 24205, Taiwan
cclee@mail.fju.edu.tw
4
Department of Management Information Systems
National Chung Hsing University
No. 250, Kuo Kuang Road, Taichung 402, Taiwan

Corresponding authors: mshwang@nchu.edu.tw
Received October 2009; revised March 2010
Abstract. A rich set of proxy signature schemes have been widely researched and dis-
cussed so far. However, they have been mainly focusing on dealing with one or two
separate proxy situations each. In this article, the authors proposed the generalization of
the (t1 /n1 t2 /n2 ) proxy signature scheme based on the factorization of the square root
modulo of a composite number. This scheme can be applied to every proxy situation. The
(t1 /n1 t2 /n2 ) proxy signature scheme allows the original group of original signers to
delegate their signing capability to a designated proxy group of proxy signers. Any verifier
can verify the proxy signatures on the messages with the knowledge of the identities of
the actual original signers and the actual proxy signers. Furthermore, all possible attacks
that have been analyzed so far have failed to break the proposed scheme.
Keywords: Digital signature, Proxy signature, Multi-proxy multi-signature scheme,
Proxy multi-signature scheme, Threshold proxy signature
1. Introduction. Digital signatures [3, 19, 37] are widely used to replace hand-written
signatures in the digital world. However, simple digital signature schemes are not enough
to satisfy todays practical conditions. For example, suppose a chairman in a department
needs to go on a business trip. She/He has to find a proxy person to deal with her/his
work at the oce. Traditional digital signature schemes [1, 21, 33, 48] do not meet this
requirement. To remedy this weakness, the proxy function has been added to digital
signature schemes, and this new type of digital signature is called proxy signature [4, 28,
40, 44].
In 1996, Mambo et al. [30, 31] proposed the first proxy signature schemes. Their
schemes allow the original signer to delegate her/his signing capability to a designated
proxy signer so that the designated proxy signer can generate a signature on behalf of the
original signer.
1039
1040 C.-C. LEE, T.-C. LIN, S.-F. TZENG AND M.-S. HWANG
As discussed by the authors [20, 30, 31], proxy signature schemes have four delegation
types: full delegation, partial delegation, delegation by warrant and partial delegation
with warrant. With full delegation, a proxy signer receives the same private key as the
original signer has. In this case, a valid proxy signature by a proxy signer is indistinguish-
able from a signature generated by the original signer.
With partial delegation, the original signer uses her/his own private key to create a
proxy signature key and gives this proxy signature key to the proxy signer. It is compu-
tationally infeasible for the proxy signer to obtain the original signers private key from
the given proxy signature key. The proxy signer can use the proxy signature key to sign
messages on behalf of the original signer.
With delegation by warrant, a proxy signer receives a proxy warrant. The proxy warrant
is used to certify that the proxy signer is delegated by the original signer. The proxy
warrant records the identities of the original signer and the proxy signer, the valid proxy
period, etc.
Finally, the last delegation kind combines the benefits of both partial delegation and
delegation by warrant. The advantages of partial delegation with warrant are fast process-
ing speed and the message qualification. Furthermore, neither delegation by warrant nor
partial delegation with warrant needs an additional proxy revocation protocol. Among
these four delegation kinds, partial delegation with warrant seems to be the best choice.
In this article, we focus on the proxy signature for partial delegation with warrant.
In general, a proxy signature for partial delegation should satisfy the following nec-
essary conditions [30, 31]: strong unforgeability, verifiability, proxy signers deviation,
distinguishable property, strong identifiable property, secret-keys dependence and strong
undeniable property.
So far, quite a number of proxy signature schemes have been widely discussed [5, 10,
12, 20, 23, 36, 41, 45, 46, 47]. The proxy signature schemes where the original signer
can delegate only one proxy signer cannot be used for group proxy situations. As a
result, threshold proxy signature schemes ((1t/n) proxy signature scheme), proxy multi-
signature schemes ((n 1) proxy signature scheme), and multi-proxy multi-signature
schemes ((n1 n2 ) proxy signature scheme) have also been proposed.
(1 t/n) proxy signature schemes have received much attention [6, 10, 11, 12, 14, 20,
23, 35]. In such a design, the original signer is able to delegate her/his signing capability
to n signers of the proxy group. Any t or more proxy signers in the proxy group can
cooperatively generate the proxy signature.
(n 1) proxy signature schemes can be found in [2, 9, 24, 25, 38, 42, 47]. In these
schemes, a proxy signer is allowed to generate a proxy signature on behalf of two or more
original signers. However, these schemes have a common disadvantage: the size of the
proxy signature on the proxys warrant is proportional to the size of the original signer
group.
Recently, a new type of proxy signature, called the (n1 n2 ) proxy signature, was
proposed [7, 8, 15, 29, 43]. In this scheme, the original group can delegate a proxy group
under the agreement of all the signers in both the original group and proxy group.
In general, the (n 1) proxy signature scheme is a special case of the (t1 /n1 t2 /n2 )
proxy signature scheme where n out of n original signers in the original group and the
proxy group has only one single proxy signer. On the other hand, the (1 t/n) proxy
signature scheme is also a special case of the (t1 /n1 t2 /n2 ) proxy signature scheme where
the original group has only one single original signer. As for the case, where the original
group and the proxy group consist of only a single original signer and a single proxy
singer, it is also a special case of the (t1 /n1 t2 /n2 ) proxy signature scheme. The concept
of the (t1 /n1 t2 /n2 ) proxy signature scheme was first introduced by Li et al. [26]. It is
GENERALIZATION OF PROXY SIGNATURE 1041
an important technique to generalize the proxy signature schemes. Li et al. proposed a

generalization of proxy signature based on discrete logarithms that can be applied to every
proxy situation. However, Li et al.s scheme has a security weakness that an adversary can
forge illegal proxy signatures being likely generated by the proxy group on behalf of an
adversary. To overcome this weakness, Hwang and Chan [17] proposed an improvement
on Li et al.s generation scheme. Based on elliptic curves, Hwang et al. proposed another
generalization of proxy signature scheme [16]. However, [18, 39], respectively, showed
that the scheme has security flaws and then proposed an improvement on Hwang et al.s
generation scheme.
Many asymmetric cryptosystems have been proposed, however, only three types of
systems should be considered both secure and ecient. These systems, classified according
to the mathematical problem on which they are based, are:
Factorization problem (such as the RSA cryptosystem)
Discrete Logarithm problem (such as the ElGamal cryptosystem)
Elliptic Curves Cryptosystem (ECC)
Up to now, two generalization of proxy signature schemes based on discrete logarithms
and elliptic curves are proposed by Li et al. and Hwang et al., respectively. In this article,
we shall propose another generalization of the (t1 /n1 t2 /n2 ) proxy signature scheme with
known signers based on a new type called factorization. The proposed scheme is suitable
for all the proxy situations mentioned above.
This article is organized as follows. In Section 2, we shall first briefly review Lee et al.s
scheme (the Lee-Hwang-Li scheme). Then, in Section 3, our new secure (t1 /n1 t2 /n2 )
proxy signature scheme and security analysis will be proposed and presented. After that,
we propose some special cases of the (t1 /n1 t2 /n2 ) proxy signature in Section 4. Finally,
some concluding remarks will be in the last section.
2. Review of the Lee-Hwang-Li Digital Signature Scheme. Lee et al. [22] proposed
a digital signature scheme based on the Ohto-Okamoto scheme [32]. Initially, each user
randomly generates two large primes, p and q , and computes p, q, N and (N ) as
follows: p = 2p + 1, q = 2q + 1, N = p q and (N ) = 2p q . Next, each user selects
a secret random number si between 1 and (N ). Then she/he computes a private key
xi = si mod N and the corresponding public key yi which is certified by the certificate
authority (CA) as yi = si L = xL i mod N , where L is a random number ( 1050 ) with
GCD(L, (N )) = 1, is an element which is primitive in both GF (p) and GF (q), and
h() is a one-way hash function. The parameters N , L and h() are published, and p, q,
p , q , and (N ) are kept secret.
Assume the message M needs to be signed by the signer. The signer randomly selects
a random number k between 1 and (N 1). Then, she/he computes
r = k L mod N,
e = h(M, r) mod N,
z = kxei mod N.
Then, (e, z) is a signature of M . The signer sends (e, z) with M to the verifier.
On receiving the digital signature (e, z) of M , the verifier can compute r as r =
z L yie mod N . Then the verifier can check the equation, e?= h(M, r ). If the equation holds,
(e, z) is a valid signature of M .
3. The Proposed Scheme. In this section, we propose and analyze a new (t1 /n1 t2 /n2 )
proxy signature scheme with known signers. This scheme is based on the Lee-Hwang-Lis
digital signature scheme as described previously.
According to the proxy warrant, a subset of original signers allows a designated proxy
group to sign on behalf of the original group. A message M has to be signed by a subset
of proxy signers who can represent the proxy group. Then, the proxy signature is sent to
a verifier to verify the validity of it. In other words, some threshold values will be given to
indicate the number of persons to represent the group to delegate the signing capability
or to sign the proxy signature.
Assume that there is a share distribution center (SDC) which is responsible for dis-
tributing shares for the system. Initially, the system parameters (N , L, h(), p, q, p ,
q , , (N )) are the same as those of the Lee-Hwang-Li digital signature scheme back in
Section 2. Now, we will set some system parameters as:
Mw : a warrant which records the identities of the original signers in the original
group and the proxy signers in the proxy group, the parameters (t1 , n1 ), (t2 , n2 ) and
the valid delegation time, etc.
AOSID: (Actual Original Signers ID) the identities of the actual original signers.
AP SID: (Actual Proxy Signers ID) the identities of the actual proxy signers.
Next, SDC computes the original signers private key xOi and public key yOi as follows:
xOi = sOi mod N,
yOi = xLOi mod N, (1)
where sOi is a random integer such that GCD(sOi , (N )) = 1. Similarly, each proxy
signer UPi also owns a private key xPi = sPi mod N , where sPi is a random integer such
that GCD(sPi , (N )) = 1, and a public key:
yPi = xLPi mod N, (2)
which is also certified by that CA. Finally, SDC distributes private keys to all the original
signers and proxy singers. The notations GO and GP are defined as the original signer
group and the proxy signer group, respectively. Let GO = {UO1 , UO2 , , UOn1 } and
GP = {UP1 , UP2 , , UPn2 } be groups of n1 original signers and n2 proxy signers.
Based on the definition of the (t1 /n1 t2 /n2 ) proxy signature scheme, any t1 or more
out of n1 original signers (1 t1 n1 ) can represent GO to delegate signing capability.
Any t2 or more out of n2 proxy signers (1 t2 n2 ) can represent GP to sign a message
on behalf of GO . The procedure of the proposed scheme contains three phases: the proxy
share generation phase, the proxy signature generation phase and the proxy signature
verification phase.
3.1. Proxy share generation phase. Assume that any t1 or more original signers want
to delegate their signing capability to GP so that GP can sign messages on behalf of GO .
Let DO = {UO1 , UO2 , , UOt } be the actual original signers with t1 t1 n1 , and DO
1
as a group executes the following steps to delegate the signing capability to GP .
1. Choose a random number kOi with 0 < kOi < N , and broadcast rOi to other t1 1
original signers in DO and the designated clerk.
L
rOi = kO i
mod N (3)
2. For each received rOj (j = 1, 2, , t1 ; j = i), each UOi DO computes K and Oi
as follows:

t1
K = rOi mod N, (4)

i=1
K yO h(Mw ,K,AOSID)
Oi = kO x i
i Oi
mod N. (5)
3. Send Oi to the designated clerk via a public channel.

4. After receiving Oi , the designated clerk checks whether the following equation holds:
L K yO h(Mw ,K,AOSID)
O i
= rO y i
i Oi
mod N. (6)
If it does, the designated clerk computes O as their proxy share.

t1
O = Oi mod N. (7)
i=1
5. Broadcast (Mw , K, O , AOSID) to GP .

After receiving (Mw , K, O , AOSID), each UPj GP checks whether the following
equation holds:

t1
yO h(Mw ,K,AOSID)
L K
O =K yOi i mod N. (8)
i=1
After each UPj confirms the validity of (Mw , K, O , AOSID), she/he computes her/his
proxy signature key Pj as
yP
Pj = O xPj j mod N. (9)
3.2. Proxy signature generation phase. Without loss of generality, the proposed
scheme allows any t2 or more proxy signers to represent GP to sign a message M coop-
eratively on behalf of GO . Let DP = {UP1 , UP2 , , UPt } be the actual proxy signers
2
with t2 t2 n2 , and DP as a group executes the following steps to generate a proxy
signature.
1. Choose a random number kPj with 0 < kPj < N and broadcast rPj to other t2 1
proxy signers in DP and the designated clerk.
rPj = kPLj mod N.
2. For each received rPk (k = 1, 2, , t2 ; k = j), each UPj DP computes R and sPj
as follows:

t2
R = rPj mod N, (10)

j=1
h(M, R, AP SID)
sPj = kPRj Pj mod N. (11)
Here, sPj is an individual proxy signature which is sent to the designated clerk.
3. After receiving sPj , the designated clerk checks whether the following equation holds:
h(M, R, AP SID)
t1
yO h(Mw , K, AOSID) yP
sLPj = rPRj K K yO i i y Pj j mod N. (12)
i=1
If it does, the designated clerk computes

t2
S= sPj mod N. (13)

j=1
Then, the proxy signature of M is (Mw , K, AOSID, M , R, S, AP SID).

3.3. Proxy signature verification phase. After receiving the proxy signature (Mw ,
K, AOSID, M , R, S, AP SID) for M , any verifier can verify the validity of the proxy
signature by following the steps below:
1. According to Mw , AOSID and AP SID, the verifier obtains the public keys of the
original signers and proxy signers from CA and knows who the actual original signers
and actual proxy signers are.
2. The verifier can check the validity of the proxy signature on the message M through
the following equation:
t2 h(M, R, AP SID)
t1
t2
S L = RR K K yOi i yPj j mod N. (14)

i=1 j=1
If the above equation holds, the message M is authenticated, and the proxy signature
(Mw , K, AOSID, M , R, S, AP SID) is valid.
In the following paragraphs, we shall prove that the proposed scheme can work correctly.
Theorem 3.1. In the proxy share generation phase, the designated clerk can verify the
validity of Oi sent from UOi by Equation (6).
K yO h(Mw , K, AOSID)
Proof: From Equation (5), we have Oi = kO x i
i Oi
mod N . By raising
both sides of the above equation to exponents with L, we have
L LK LyOi h(Mw , K, AOSID)
O i
= kO i
xOi mod N.
According to Equations (1) and (3), the above equation can be rewritten as
L K yO h(Mw , K, AOSID)
O i
= rO y i
i Oi
mod N.

Theorem 3.2. If the proxy share is constructed correctly, it will pass the verification of
Equation (8).
Proof: From Equations (5) and (7), we have

t1
O = Oi ,
i=1

t1
yO h(Mw , K, AOSID)
K
= kO x i
i Oi
mod N.
i=1
By raising both sides of the above equation to exponents with L, we have

t1
LyOi h(Mw , K, AOSID)
L LK
O = kO i
xOi ,
i=1

t1
yO h(Mw , K, AOSID)
K
= rO y i
i Oi
mod N.
i=1
In accordance with Equation (4), the above equation can be rewritten as

t1
yO h(Mw , K, AOSID)
L
O = KK yOi i mod N.
i=1

Theorem 3.3. In the proxy signature generation phase, the designated clerk can verify
any individual proxy signature sPi sent from UPi by Equation (12).
Proof: Taking in Equation (9), we can rewrite Equation (11) as
( yP
)h(M, R, AP SID)
sPj = kPRj O xPj j mod N.
Raising both sides of the above equation to exponents with L, we have
( )
yPj Lh(M, R, AP SID)
sLPj = kPLR
j
x
O Pj ,
( )h(M, R, AP SID)
L yPj
= rPRj O y Pj mod N.
According to Equation (8), the above equation can be rewritten as

h(M, R, AP SID)
t1
yO h(Mw , K, AOSID) yPj
sLPj = rPRj K K yOi i yPj mod N.
i=1

Theorem 3.4. If the proxy signature is constructed correctly, it will pass the verification
of Equation (14).
Proof: In Equations (9), (11) and (13), we have

t2
S = sPj ,
j=1

t2 ( yP
)h(M, R, AP SID)
= kPRj O xPj j mod N.
j=1
By raising both sides of the above equation to exponents with L, we have

t2 ( )h(M, R, AP SID)
L L yPj
S = rPRj O yPj mod N.
j=1
According to Equations (8) and (10), the above equation can be written as
t2 h(M, R, AP SID)
t1
t2
S L = RR K K yOi i yPj j mod N.

i=1 j=1

3.4. Security analysis and discussions. In this subsection, we analyze the security of
the proposed scheme and prove that the proposed scheme can satisfy the conditions in
[30, 31].
The security of the proposed scheme is based on the well-known diculty of computing
one-way hash function [13] and the factorization [34] of the square root modulo of a
composite number N . In the following, we consider some possible attacks and see how
the proposed scheme works against them. We shall prove that the proposed scheme can
successfully withstand these attacks, maintaining the security provided by the one-way
hash function and the factorization problem of N .
In the proposed scheme, all of the actual original signers private keys xOi are used in
the proxy share generation phase to create a proxy share O . Therefore, for the proxy
group, it is necessary to verify the proxy share O by checking Equation (8), and the
verifier also has to check the proxy signature verification equation, namely Equation (14),
by using all of the actual original signers public keys. All of the actual original signers
public keys have been certified by CA. Without knowing the original signers private keys,
an adversary is unable to create the proxy share O in the proposed scheme. Assume that
an adversary wants to obtain the original signers private key xOi in Equation (1). It
means the adversary has to face the diculty of solving the factorization of N . According
to the warrant Mw , the original group cannot deny delegating their signing capability to
a proxy group because of the proof of the proxy share.
Similarly, all of the actual proxy signers private keys xPi are used to generate the proxy
signature in the proxy signature generation phase. Thus, it is necessary for any verifier
to verify the proxy signature verification equation, which is Equation (14), by using all of
the actual proxy signers public keys. These actual proxy signers public keys have also
been certified by CA. If an adversary wants to get some proxy signers private key xPi
from Equation (2), then she/he has to solve the factorization of N . Therefore, the proxy
group cannot deny generating the proxy signature on behalf of the proxy group and the
original group.
To fight against the equation attack, we have to consider the security of the proxy
signature verification equation:
t2 h(M, R, AP SID)
t1
t2
S L = RR K K yOi i yPj j mod N.

i=1 j=1
In this case, an adversary may try to forge a valid proxy signature to pass the proxy
signature verification equation. Suppose
t2
t1
yO h(Mw , K, AOSID)
VO = K K yOi i mod N, (15)
i=1
t2
yP
VP = yPj j mod N.
j=1
We can rewrite the proxy signature verification equation as

S L = RR (VO VP )h(M, R, AP SID) mod N.
The value VO depends on the parameters Mw , K and AOSID, while VP is a fixed value
as the actual proxy signers public keys are certified by CA. Given M , AP SID , VO
and VP , it is dicult to determine R and S based on the one-way hash function and
the factorization of N . Again, under given M , AP SID , VP , R and S , VO is to be
computed so that the equation holds. However, it is dicult to find Mw , K and AOSID
such that Equation (15) holds. It also requires solving the one-way hash function and the
factorization of N . Therefore, the proxy signature verification equation is secure against
the equation attack.
Consider the insider forgery attack in [27, 38]. Without losing of generality, the mali-
cious original signer UOk (UOk GO ) without any private keys of the other original signers
can forge a valid proxy share for an arbitrary warrant. To forge the proxy share, assume
UOk waits until she/he receives any t1 1 original signers public keys yOi . She/He sub-
stitutes her/his public key yOk . UOk selects an integer xOk as her/his private key. Then,

UOk has to make her/his public key yO k
satisfy the following equation:
(t 1 )1

yO 1
y
yOk k = xL
O
Ok yOi i mod N. (16)
i=1
Next, UOk selects a random integer kOk and computes K and O as follows:
L
K = kO k
mod N,
K h(Mw , K, AOSID)
O = kO x
k Ok
mod N.
Thus, O is a valid proxy share. It is because the forged proxy share can pass Equation
(8) as follows:

t1
yO h(Mw , K, AOSID)
L
O = KK yO i i ,
i=1
(t 1 )1 t 1 h(Mw , K, AOSID)
1 1
= K K xL yOi i
yO yO
Ok yO i i ,
i=1 i=1
( )L
K h(Mw , K, AOSID)
= kO x
k Ok
mod N.
However, UOk cannot succeed in generating the proxy share O . From Equation (16),
1 1 yOi
suppose the values of xOk and ti=1 yOi mod N are determined, it is extremely dicult

to find a yOk satisfying the equation. If UOk determines the integer yO k
first, she/he has

to solve the factorization of N to find the value of xOk . Thus, the insider forgery attack
launched by the malicious original signer in the original group cannot succeed in forging
the proxy share for the arbitrary warrant.
Similarly, GP might also try the insider forgery attack. Without loss of generality,
suppose that the malicious proxy signer UPk (UPk GP ) wants to update her/his public
key yPk . UPk also waits until she/he obtains any t2 1 proxy signers public keys yPj .
Then, she/he changes her/his public key yPk . First, UPk chooses a random integer xPk
and makes her/his public key yP k as follows:
(t 1 )1

yP
2
y
yPk k = xL
Pk
P
yPj j mod N. (17)
j=1
Later, UPk selects a random integer rPk and computers R and S as follows:
R = rPLk mod N,
( t2 )h(M, R, AP SID)
S = rPRk O xPk mod N.
Thus, (Mw , K, AOSID, M , R, S, AP SID) is a valid proxy signature. It is because the
forged proxy signature can satisfy Equation (14) as follows:

t2 h(M, R, AP SID)

t1

t2
yP
S L = RR K K y Pj j
yO h(Mw , K, AOSID)
yO i i ,
i=1 j=1

t2 (t 1 )1 t 1 h(M, R, AP SID)

t1
2
yP 2
yP
= RR K K xL y Pj j
yO h(Mw , K, AOSID)
yOi i Pk y Pj j ,
i=1 j=1 j=1

t2 h(M, R, AP SID)

t1
= RR K K xL
yO h(Mw , K, AOSID)
yOi i Pk ,
i=1
( t2
)L
= rPRk (O xPk )h(M, R, AP SID) mod N.
However, UPk cannot create a valid proxy signature. From Equation (17), if the values
2 1 yPj
of xPk and tj=1 yPj mod N are determined, it is an extremely dicult problem to find
a yPk satisfying the equation. If UPk determines the integer yP k first, she/he has to solve

the factorization of N to find the value of xPk . Thus, the insider forgery attack launched
by a malicious proxy signer cannot work.
As for the direct forgery attack in [38], the security analysis is the same as that of the
insider forgery attack for the proposed scheme. A malicious original signer can imperson-
ate any t1 1 or more original signers in the original group. The malicious original signer
can forge the proxy share without relying on the agreement of the other original signers.
Suppose UOk waits until she/he receives any t1 1 original signers public keys. Then,
UOk selects a random number kOk and computes K and O as follows:
(t 1 )1
1
yO h(Mw , K, AOSID)
K K = kO
L
k
yOi i mod N, (18)
i=1
yOk h(Mw , K, AOSID)
O = kOk xOk mod N.
Thus, O is a valid proxy share because the forged proxy share can pass Equation (8) as
follows:

t1
yO h(Mw , K, AOSID)
L K
O = K yOi i ,
i=1
(t 1 )1
1
yO h(Mw , K, AOSID) yO h(Mw , K, AOSID)
t
1 1
yO h(Mw , K, AOSID)
L
= kO k
yOi i yOkk yO i i ,
i=1 i=1
L yO h(Mw , K, AOSID)
= kO y k
k Ok
,
( )
yO h(Mw , K, AOSID) L
= kOk xOkk mod N.
However, UOk cannot succeed in generating the proxy share O . From Equation (18),
it is extremely dicult to find a K satisfying the equation. If UOk determines the integer
K first, she/he has to solve the factorization of N to find the value of kOk . Thus, the
direct forgery attack launched by the malicious original signer in the original group cannot
succeed.
On the other hand, a malicious proxy signer can also impersonate any t2 1 or more
proxy signers in the proxy group. UPk can also forge the proxy signature without the
agreement of the other proxy signers. Suppose that UPk waits until she/he receives any
t2 1 proxy signers public keys. Then, UPk selects a random number rPk and computes
R and S as follows:
(t 1 )1
2
yP h(M, R, AP SID)
RR = rPLk y Pj j mod N, (19)
j=1
t2 yPk h(M, R, AP SID)
S = rPk (O xPk ) mod N.
Thus, (Mw , K, AOSID, M , R, S, AP SID) is a valid proxy signature because the forged
proxy signature can satisfy Equation (14) as follows:
t2 h(M, R, AP SID)
t1

t2
y
S L = RR K K yPj j
yO h(Mw , K, AOSID) P
yOi i ,
i=1 j=1
(t 1 )1
t2
2
yP h(M, R, AP SID)
t1
= rPLk y Pj j K K yO h(Mw , K, AOSID)

yOi i
j=1 i=1
t
)h(M, R, AP SID)
2 1
yP yP
yPkk y Pj j ,
j=1

t2 h(M, R, AP SID)

t1
= rPLk K K yPyPk
yO h(Mw , K, AOSID)
yOi i k
,
i=1
( ( )h(M, R, AP SID) )L
t2 yPk
= rPk O xPk mod N.
However, UPk still cannot create a valid proxy signature. From Equation (19), it is
extremely dicult to find a R satisfying the equation. If UPk determines the integer R
first, she/he has to solve the factorization of N to find the value of rPk . Thus, the direct
forgery attack launched by the malicious proxy signer cannot work.
The proposed scheme satisfies the proxy signature scheme requirements in [30, 31]. The
analysis is given below.
Strong unforgeability: A designated proxy group can generate a valid proxy signature
for the original group. According to the above security analysis, any original signer
in the original group and any third party cannot generate a valid proxy share and
a valid proxy signature. So, the proposed scheme satisfies the strong unforgeable
requirement. Furthermore, the proposed scheme provides proxy-protection property.
Verifiability: By the proxy signature, any verifier can be convinced of the original
groups agreement on the signed message. The verification of the proxy signature
can be used to guarantee that the original group actually authorizes the proxy group.
Hence, the proposed scheme satisfies the verifiability requirement.
Proxy signers deviation: The proxy signers cannot generate a valid proxy signature
without being detected as a proxy signature. The proxy signers cannot generate the
valid signatures of the original signers. In the other words, the proposed scheme
provides the proxy signers deviation.
Distinguishability: The valid proxy signature and the self-signing signature are ap-
plied to dierent congruences for verification. Here, a self-signing signature refers to
an ordinary signature generated by the original signers. Therefore, these verification
congruences can distinguish the proxy signature from the self-signing signature. The
proposed scheme therefore achieves distinguishable requirement.
Strong identifiability: From the proxy signature, anyone can determine the identity
of the corresponding proxy signers. Furthermore, from the AP SID, anyone gets
to know with ease who actually signed the message on behalf of the proxy group.
Besides, from AOSID, we easily know who actually delegated the signing power
on behalf of the original group. Hence, the proposed scheme satisfies the Strong
identifiable requirement.
Secret-keys dependence: The proxy signature is dependent on the proxy signers
private keys and the proxy share O . The proposed scheme provides the secret-keys
dependence through the proxy share.
Strong undeniability: The actual proxy signers generate a valid proxy signature for
the original signer group and the proxy signer group. The actual proxy signers cannot
repudiate the creation of the signature. Therefore, the proposed scheme provides the
strong undeniable property.
4. Special Cases of the Proposed Scheme. In this section, we propose some special
cases in the (t1 /n1 t2 /n2 ) proxy signature scheme. Basically, these special cases are
realized in the framework of the (t1 /n1 t2 /n2 ) proxy signature scheme. We propose three
kinds of proxy signature schemes: the (t/n 1) proxy signature scheme, the (1 t/n)
proxy signature scheme, and the (1 1) proxy signature scheme.
4.1. The (t/n1) proxy signature scheme. In the (t/n1) proxy signature scheme,
a designated proxy signer is allowed to generate a proxy signature on behalf of the original
signer group. Any t or more of the original signers can cooperatively delegate their signing
capabilities to the designated proxy signer.
The system parameters are the same as those in Section 3. Let UP be the designated
proxy signer. The major dierence here is in the proxy signature generation phase.
In the proxy share generation phase, the steps are the same as those in Section 3. DO
broadcasts (Mw , K, O , AOSID) to UP . After receiving (Mw , K, O , AOSID), UP checks
whether Equation (8) holds. If it does, UP confirms the validity of (Mw , K, O , AOSID).
She/He computes her/his proxy signature key P = O xyPP mod N .
In the proxy signature generation phase, UP is allowed to generate a proxy signature for
the message M on behalf of GO . First, UP randomly chooses a number r and computes
R = rL mod N,
h(M,R)
S = r R P mod N.
Then, the proxy signature of M is (Mw , K, AOSID, M , R, S).
Finally, the verifier checks the validity of the proxy signature and identifies the actual
original signers of the original group through the following equation:
( t
)h(M,R)
y h(M , K, AOSID)
yPyP
w
S L = RR K K
O
yOi i mod N
i=1

with t t n. If the above equation holds, the verifier believes in the validity of the
proxy signature and know the identities of the actual original signers from AOSID.
4.2. The (1t/n) proxy signature scheme. In the (1t/n) proxy signature scheme,
any t or more proxy signers from a designated proxy group of n proxy signers can sign
messages on behalf of the original signer. The major dierence here is in the proxy share
generation phase.
The system parameters are the same as those in Section 3. Let UO be the original
signer. In the proxy share generation phase, UO delegates the signing capability to GP .
First, UO chooses a random number k and computes
K = k L mod N,
y h(Mw , K)
O = k K xOO mod N.
Then, the original signer broadcasts (Mw , K, O ) to GP . After receiving the (Mw , K,
O ), each UPj GP checks whether the following equation holds:
L y h(Mw , K)
O = K K yOO mod N. (20)
If it does, each UPj uses O to compute her/his proxy signature key Pj in Equation (9).
In the proxy signature generation phase, any t or more proxy signers, as a group, signs
a message M cooperatively on behalf of GP . The procedure is similar to that in Section 3.
Each UPj computes R and sPj in Equations (10) and (11) and sends sPj to the designated
clerk. The clerk can verify the validity of sPj according to the following equation:
( )h(M, R, AP SID)
y h(M , K) yP
sLPj = rPRj K K yOO w yPj j mod N.
If the equation holds, then the clerk derives S from Equation (13), and the proxy signature
on message M is (Mw , K, M , R, S, AP SID).
In the proxy signature verification phase, the verifier checks the validity of the proxy
signature and identifies the actual proxy signers of the proxy group through the following
equation:
( )h(M, R, AP SID)
( )t
t
y
y h(M , K) P
S L = RR K K yOO w
y Pj j
mod N,
j=1
with t t n. If the above equation holds, the verifier can make sure of the validity of
the proxy signature and identify the actual proxy signers from AP SID.
4.3. The (1 1) proxy signature scheme. The (1 1) proxy signature scheme is the
basic proxy signature scheme. The original signer delegates her/his signing capability to
a designated proxy signer. The proxy signer generates a proxy signature on behalf of the
original signer.
The system parameters are also the same as those in Section 3. Let UO be the original
signer and UP be the proxy signer. The steps of the proxy share generation phase are
the same as those in Section 4.2. The original signer gives (Mw , K, O ) to a designated
proxy signer. After confirming the validity of (Mw , K, O ) by checking Equation (20),
UP performs the same steps as those of the proxy signature generation phase in Section
4.1. Then, the proxy signature of message M is (Mw , K, M , R, S). Finally, the verifier
checks the validity of the proxy signature in the following equation:
( )h(M, R)
y h(M , K)
S L = RR K K yOO w yPyP mod N.
If the equation holds, the proxy signature (Mw , K, M , R, S) of M is valid.

The security analysis and necessary conditions of the (t/n 1), (1 t/n) and (1 1)
proxy signature schemes presented above are the same as those in Section 3. Those attacks
in the above schemes will not work here because it is dicult to obtain the proxy share
or the proxy signature. The proposed schemes also satisfy those necessary conditions.
5. Conclusions. In this paper, we have proposed a new generalization of the (t1 /n1
t2 /n2 ) proxy signature scheme based on factorization to overcome security weaknesses.
Based on the (t1 /n1 t2 /n2 ) proxy signature scheme, three special cases, namely the
(t/n 1) proxy signature scheme, the (1 t/n) proxy signature scheme and the (1 1)
proxy signature scheme, have been explored. For avoiding the abuse of signing capability,
the proposed scheme provides the ability to identify the actual original signers and the
actual proxy signers. These actual original signers cannot deny delegating the warrant.
Furthermore, these actual proxy signers cannot deny signing the proxy signatures. Some
possible attacks such as equation attacks, insider forgery and direct forgery attacks have
been considered. None of them can successfully break the proposed scheme. In practice,
the proposed proxy signature scheme can be applied to a company or an organization.
If a manager wants to go on a business trip, he/she has to find a proxy person to deal
with her/his work at the oce. The manager can delegate her/his signing capability to a
designated proxy signer so that the designated proxy signer can generate a signature on
behalf of the manager. Until to now, generalization of proxy signature schemes based on
discrete logarithms, elliptic curves, and factorization are not ecient. In future, we can
try to improve their eciency.
Acknowledgment. The authors would like to thank the anonymous referee for their
valuable discussions and comments. This work was supported in part by Taiwan In-
formation Security Center (TWISC), National Science Council under the grants NSC
NSC96-2219-E-009-013 and NSC98-2221-E-468-002.
REFERENCES
[1] F. Bao, C. C. Lee and M. S. Hwang, Cryptanalysis and improvement on batch verifying multiple
RSA digital signatures, Applied Mathematics and Computation, vol.172, no.2, pp.1195-1200, 2006.
[2] T. S. Chen, Y. F. Chung and G. S. Huang, Ecient proxy multisignature schemes based on the
elliptic curve cryptosystem, Computer and Security, vol.22, no.6, pp.527-534, 2003.
[3] Y.-F. Chung and K.-H. Huang, Chameleon signature with conditional open verification, International
[4] M. L. Das, A. Saxena and D. B. Phatak, Proxy signature scheme with eective revocation using
bilinear pairings, International Journal of Network Security, vol.4, no.3, pp.312-317, 2007.
[5] M. L. Das, A. Saxena and D. B. Phatak, Algorithms and approaches of proxy signature: A survey,
International Journal of Network Security, vol.9, no.3, pp.264-284, 2009.
[6] C.-S. Feng, Z.-G. Qin and D. Yuan, Improvement of a threshold proxy signature scheme, Interna-
tional Conference on Machine Learning and Cybernetics, pp.3168-3172, 2007.
[7] L. Guo and G. Wang, Insider attacks on multi-proxy multi-signature schemes, Computers and Elec-
trical Engineering, vol.33, no.2, pp.88-93, 2007.
[8] C. L. Hsu, K. Y. Tsai and P. L. Tsai, Cryptanalysis and improvement of nonrepudiable threshold
multi-proxy multi-signature scheme with shared verification, Information Sciences, vol.177, no.2,
pp.543-549, 2007.
[9] C. L. Hsu, T. S. Wu and W. H. He, New proxy multi-signature scheme, Applied Mathematics and
Computation, vol.162, no.3, pp.1201-1206, 2005.
[10] C.-L. Hsu, T.-S. Wu and T.-C. Wu, New nonrepudiable threshold proxy signature scheme with
known signers, The Journal of Systems and Software, vol.58, pp.119-124, 2001.
[11] H. F. Huang and C. C. Chang, A novel ecient (t, n) threshold proxy signature scheme, Information
Sciences, vol.176, no.10, pp.1338-1349, 2006.
[12] M. S. Hwang, I. C. Lin and E. J.-L. Lu, A secure nonrepudiable threshold proxy signature scheme
with known signers, International Journal of Informatica, vol.11, no.2, pp.1-8, 2000.
[13] M.-S. Hwang, C.-C. Chang and K.-F. Hwang, A watermarking technique based on one-way hash
functions, IEEE Transactions on Consumer Electronics, vol.45, no.2, pp.286-294, 1999.
[14] M.-S. Hwang, E. J.-L. Lu and I.-C. Lin, A practical (t, n) threshold proxy signature scheme based
on the RSA cryptosystem, IEEE Transactions on Knowledge and Data Engineering, vol.15, no.6,
pp.1552-1560, 2003.
[15] M.-S. Hwang, S.-F. Tzeng and S.-F. Chiou, A non-repudiable multi-proxy multi-signature scheme,
ICIC Express Letters, vol.3, no.3(A), pp.259-263, 2009.
[16] M. S. Hwang, S. F. Tzeng and C. S. Tsai, Generalization of proxy signature based on elliptic curves,
Computer Standards and Interfaces, vol.26, pp.73-84, 2004.
[17] S. J. Hwang and C. C. Chan, Improvement on Li et al.s generalization of proxy signature schemes,
Computers and Security, vol.23, pp.615-619, 2004.
[18] S. J. Hwang and C. W. Huang, Improvement on Hwang et al.s generalization of proxy signature
schemes based on elliptic curves, Applied Mathematics and Computation, vol.170, pp.941-947, 2005.
[19] R. S. Katti and R. G. Kavasseri, Nonce generation for the digital signature standard, International
Journal of Network Security, vol.11, no.1, pp.23-32, 2010.
[20] S. Kim, S. Park and D. Won, Proxy signatures, revisited, Proc. of ICICS97, LNCS, vol.1334,
pp.223-232, 1997.
[21] C. C. Lee, M. S. Hwang and W. P. Yang, Untraceable blind signature schemes based on the discrete
logarithm problem, Fundamenta Informaticae, vol.55, pp.307-320, 2003.
[22] N.-Y. Lee, T. Hwang and C.-M. Li, (t, n) threshold untraceable signatures, Journal of Information
Science and Engineering, vol.16, no.6, pp.835-845, 2000.
[23] N. Y. Lee, T. Hwang and C. H. Wang, On Zhangs nonrepudiable proxy signature schemes,
ACISP98, LNCS, vol.1438, pp.415-422, 1998.
[24] F. G. Li, S. J. Zhou and R. Sun, Cryptanalysis of an identity based proxy multi-signature scheme,
IEICE Tran. on Fundamentals of Electronics Communications and Computer Sciences, vol.E91A,
pp.1820-1823, 2008.
[25] J. G. Li, Z. F. Cao and Y. C. Zhang, Nonrepudiable proxy multi-signature scheme, Journal of
Computer Science and Technology, vol.18, pp.399-402, 2003.
[26] L. H. Li, S. F. Tzeng and M. S. Hwang, Generalization of proxy signature-based on discrete loga-
rithms, Computers and Security, vol.22, pp.245-255, 2003.
[27] Z. C. Li, L. C. K. Hui, K. P. Chow, C. F. Chong, H. H. Tsang and H. W. Chan, Cryptanalysis
of Harn digital multisignature scheme with distinguished signing authorities, Electronics Letters,
vol.36, no.4, pp.314-315, 2000.
[28] E. J.-L. Lu and C.-J. Huang, A time-stamping proxy signature scheme using time-stamping service,
[29] Y. D. Lyuu and M. L. Wu, Cryptanalysis of and improvement on the Hwang-Chen multi-proxy
multi-signature schemes, Applied Mathematics and Computation, vol.167, pp.729-739, 2005.
[30] M. Mambo, K. Usuda and E. Okamoto, Proxy signatures: Delegation of the power to sign message,
IEICE Trans. Fundamentals, vol.E79-A, pp.1338-1353, 1996.
[31] M. Mambo, K. Usuda and E. Okamoto, Proxy signatures for delegating signing operation, Proc. of
the 3rd ACM Conf. on Computer and Communications Security, pp.48-57, 1996.
[32] K. Ohta and T. Oktamoto, A modification of the Fiat-Shamir scheme, Proc. of CRYPTO88, pp.232-
243, 1988.
[33] R. R. Ramasamy and M. A. Prabakar, Digital signature scheme with message recovery using
knapsack-based ECC, International Journal of Network Security, vol.12, no.1, pp.12-17, 2011.
[34] R. L. Rivest, A. Shamir and L. Adleman, A method for obtaining digital signatures and public key
cryptosystems, Communications of the ACM, vol.21, pp.120-126, 1978.
[35] J. Shao, Z. Cao and R. Lu, Improvement of Yang et al.s threshold proxy signature scheme, The
Journal of Systems & Software, vol.80, no.2, pp.172-177, 2007.
[36] Z. Shao, Proxy signature schemes based on factoring, Information Processing Letters, vol.85, no.3,
pp.137-143, 2003.
[37] W.-G. Shieh and M.-T. Wang, An improvement to Kim-Chungs authentication scheme, ICIC Ex-
press Letters, vol.3, no.4(B), pp.1215-1220, 2009.
[38] H. M. Sun, On proxy (multi-) signature schemes, 2000 International Computer Symposium, Chiayi,
pp.65-72, 2000.
[39] Z. Tan, Improvement on a generalized scheme of proxy signature based on elliptic curves, Interna-
tional Conference on Computational Intelligence and Security, pp.677-681, 2007.
[40] Z.-W. Tan, Improvement on nominative proxy signature schemes, International Journal of Network
Security, vol.7, no.2, pp.175-180, 2008.
[41] G. K. Verma, A proxy blind signature scheme over braid groups, International Journal of Network
Security, vol.9, no.3, pp.214-217, 2009.
[42] Q. Wang and Z. F. Cao, Identity based proxy multi-signature, The Journal of Systems and Software,
vol.80, no.7, pp.1023-1029, 2007.
[43] Z. C. Wang, H. F. Qlan and Z. B. Li, Hybrid proxy multisignature: A new type multi-party signature,
Information Sciences, vol.177, no.24, pp.5638-5650, 2007.
[44] Q. Xie and X.-Y. Yu, Cryptanalysis of two nonrepudiable threshold proxy signature schemes, Inter-
national Journal of Network Security, vol.3, no.1, pp.18-22, 2006.
[45] H. Xiong, Z. Qin and F. Li, A certificateless proxy ring signature scheme with provable security,
[46] Q. S. Xue and Z. F. Cao, Factoring based proxy signature schemes, Journal of Computational and
Applied Mathematics, vol.195, no.1-2, pp.229-241, 2006.
[47] L. Yi, G. Bai and G. Xiao, Proxy multi-signature scheme: A new type of proxy signature scheme,
Electronics Letters, vol.36, no.6, pp.527-528, 2000.
[48] J.-H. Yang and C.-C. Chang, An ecient fair electronic payment system based upon non-signature
authenticated encryption scheme, International Journal of Innovative Computing, Information and
Control, vol.5, no.11, pp.3861-3873, 2009.

!""

" #" $%&

' ( ) )" * ' +" +,

&
"- ( .'' ',
" / ,".'' ',

!""! #
$
%
$ &

#
'

$
( )
* % +
%
' %
," -
))) . /')
0
)
+ '
' +

'

0 " "& 1
&"
0 " #" &

!" # $ %
! & $ ' (! )

%

&
%!

*! +

& % &

& $ & !

+ $ & ,

! +! ! !
-./0 ! ./

! )

!

./ "
1 ./ 2

+ ./ !
!
+ -
0
-
0!

+ ! # $ !
3 ./
$ $ %
./ 4 $

./ $ )+

5
6 $! ! & $ & $

)! $ ./
$ ' ( + # !
7

&
!
)
$ 8 *! $
$

+ ! $ 7 + ./ 7

$
%! $

!

! )
%

!
$ &

+!
.

8 ! 9!:

! $

;
+
! $ ./ &

# + !
3 2 8! $
$
1 *! $ $ $!
+ ./
+ ./ !
$ 3 +
! $ $
+
%
)
$ *! $ $
$
./ ! !
)
$ $
# ! $
3
+ ./ < ./
$
1
# ! $
$ $ $
$! + ./
+ ./ # ! $
3
+

./ %

$ $ # ) ! $
% $
+ $
3

%! $

& + ./ +
! $
8

2 8! $
) $
$! +!

./

$ ) ) *!

)
./ $
+ ! =+

1 $
%!

* !

+ %
+ $
*!

! + %

$ + ./ # $ ) )

) 3
! $
$ & )

./
% 7 $4 $ $ & <

. + > *!
$ $ &
! % ./

/ .?" % $

./ ! $

2 @8 " ) ./ $

% &

' ( '( %

+A
1
% ?>B

+ > C
!

+ -
0
< 2 "

./
%

% 7
;B
!

#
2
@8A $ &! < 2 " @D E
F
-@DEF0! ! + !

%
+ ! 9
!

$
2 < "! $

+ :!?.>
! . ;
2 <
." $ !
< + ! >!>9

!

G 1& >" + !
+
%
$
+ !9!:.
!
;;
+ !
6 :" 3

% +
) !

+ G 1& A 6 A

+ !9.!
;9B $
?? + *! 2 ;"

! ./
%
+ >!>.!>

.

!

< < 9" ) 8

%
-@EF0!
;
@EF D! < < ?"

! $ .
D "

./ !

>.
2 "

$ !
.

$
.
% +

!9!:.
# ! $
= ./
# + ! $ $
= $
; .!9 $ >! >
!;:!.>>

. "
#$% %"# % =
-@20 ./ % $
* $
+ %

$ 4
<

.! $
. ..! % + $

.> $
& .:

% = $

. & #'% ( % ) @2

# ! $
- 0

$ % $ $ $
)
! %

G
8!
E
% ! $ !

'( ! $ )

% # !
$ + -F@0 '( !
$
& F@ <

F@ $ ! $ ) $
)
! $
& ! <
$

. % % !% %
8 *! $
H - 0 $ %!

! $ $

%
8 $ <
! $ 7
*

;; - H

; ;

7

-. 0
F*.
; H ;0 < !
- 0!
8 ) - #/0!
$ #/ H I
- 0! 7

E8 -04

- 0 -
0

- 0 H - 0 - 0 - 0 - 0
-
0 -0

%
-. 0 $ ) 8 ) :
7 ? . ?0 ; ;0!
$ !

-F*0! $

$ H - 0 # ! H

J $
! $ & 1 !

1

' ( ! $
H - 0
K ! ! $

) $
$
&
$
8 ) #/ # ! $ !
+! - 0 ! %! $
- 0 #/ - 0 * $ + $ * !

- .0

' (!
: % $ $ $
#/ : ! $
-. 0
% ) )

- 0 # ! $ )
%

=

! & $! $ & )

% $ *! $ $
7

- 0 E8
-0 ! $ E8 -0 $ ! $

$ ! " %
! 2
E8 -0

- 0 H I I " -0
H - 0 H - 0 -0
% 8 ) $ !
$ )+

8 $ $

-...0
* . ! $
* 7 !

3 $
$

J $
! =
< ! $
8 % 8 !

% 8
% $ $

$
'(
) <
?

$ ; ; -0
-0 -0
! $ $ )

) -E )0 % E ) '( !
$
% 8 E )

! $ = '(

% 7 '(

6 $!
!
!

%
! $
$ $
'(

.. #% !% %

%
$! $ +
-F@0 $ )
<

8
.. ...
$ ! $ $

)
3 ..
..

%
$ $
<
$

*! $ -K0 + -F@0
! $ $
! $ F@

# ! $ ! !
) -E )0 <
$
% !
- 0
E! F@!

! F@
<
%

F@ <
$

* > + # +!

- 0 !
$ + F@ %$

!
- 0 $ #
#! #

%! $ ! ! ) -E* )0
% E* ) ! $
F@ $ !
! $ !
% E* ) '(

* : $ $ '( -, 6 0

$ !

$ $ 7 $

%
) ! $
$ F* #
! $ !
) ! F*
! ! '(
)
.. E $

*! + % +

F

% $ '( $

$ ;!!???
$ 3 +
-F@0 -, 6 0
8 &

+ 2!

+
% $ & +
%

'( J 3

# !
+
)
%!
$

! +

)

% 3
$
+
#

L ! $! +
$
# ! $
)! $ +
$ # ! +

$! $

! +
...
<
$
'( + 2! 3
+ $ 3
!
$ $ = '( <
+
$ + $

* : $ $ '( -,
6 0 9?;!>?;
;!!???

6! $ ; ; - H ; H ;0!
?
< 7

F ./ $

1 #
!
!
F* $

* 7 !

% '( , $ %

,
%

,
$ $
# ! 6 $ $ 8
$ 3 & $
% $ 8 6
!

L
$ 7 ! $

$ % )
' ( + !

2!
* :-0 $ $ ,
%
8

! $

%
$ 8

< $

% 7
$
+!
$ ? ? - H ? H ?0 ; ;
! ;

%
.!? ! .
$
8 ;.:! ; $ >9;!:> 8
; ;
* 8! $ +

K + !
3
8 #

!
8
3
$

+
% $ $ $ '(
6! $

+
-KE0 +

! $ ! $

* F@! KE

-J*0 F@ -0! $

! $ E8 -.0

H H

-.0

J $! -0 +

$ %!
$
# ! $
=+

-
0 =+
!
$ E8 ->0
-0 H - 0
->0
D 3
$ -0 ! $ E8 -:0
2!
$
$

H -0 -:0
# -0 ! !

+

2 8! $

$
+

6 $! $
-0 ! !

+

L ! $ $

#! $ &

$
$
! $
8 # ! $

G <
! $

-0 $ E8 -;0

H - -0 0 -0 -;0
# ) ! $ 8
) '( % )
$ )
$

%
$
=+

-
0 < $ ; ; - H ; H ;0 $
$ %E2 H ;!!???
+
-F@0
J* H 9?;!>?; %
D >.
) : : + $
>
+ /3 =+

% K ! +! J/ H ?!?.
=+

H :!
J/ H >::!

H
* ; 7
6! $ !;:!.>>
$ F/* ) > 7
? ! $ % +
'( % $ ! 9!>? +!

$ =+

6 $! FJ .; G $ =+
:!
) 3 7

/ =+

-
0
F@ +

: 9
K ! "# "!"#! !""$ #%""$% &&%%
,6 ##&% $ " "%!"%# !& %& $ !"%!
6 ""& %$%$$ !"&"& %$ " !%& #" &&"!$

$ #" # %## $ !"!& !#!! #!#! #&!"

7 '( %
) L
=+
-
0 3

-J/0! $ 3

-0 '( ! FJ H . G -
0
-
H :0 J/ H > !:::! FJ H .;; G

-0 -
H 90 J/ H .;! ! FJ H .9.;

G -0 -
H 0 J/ H ;: !:?
.> & ) ( % +

&! .. 6 $!
)
!

8
E 1

$
%! $
) - #/0! +

.: !!! * ! &! < &

+ % ) & -&0 - =
0 % & 3
$
% & -&0 -$
0 % $ & $

! $ $ #/ ! 8
% & -&.0 -
0 % &

# + !
& -&! & &.0 #/!
! 8 + *! &
-&>0 + - = 0 $

% +

$ &
< +!
8
+ # ) ! & 8
+ 6 $! $ & -&. &>0
+
> )(# !! E+ $ < )
7

<
$

- 0
! $
%

- 0
-
0 +!
H :;! H :;!
;

E8 -90 $
-%E20
! $
J*
$ 2!
$ !
3
H - - 00 -90
%

-
0 $ #
$ !
$ $ $
+
$
7 D +

$ $ L
! +
8

! 7

%
$ 8
!

-
0! $ E8 -?0

- 0 -?0

* +! $ $ $ 9 - H 90

7 !

:
%

.

2 ;" - %

$ + ! * @ 0 * ! $
$ $ . - H .0

>

$ !

% $
+ ! $ + *
9 $ + , + !
$ $ J* H 9?;!>?;
%E2 H ;!!???

% +
$ 3 8 +
% 'D( : : >.

< +
, F@ < '(
F ./ 7 :B! ..B! ;;B! :B B
+ +! $ :B +!

3 7 , % 8 -@E0
9.? > FJ $ >: G $ B
+ +! 8
FJ !
8 '(
% FJ : G $ :B +
< $ $
B + !
3

< +

%E2 H !;:!.>>
! $ 8 > 7
? F/* # ) ! $ J*
H ; .!9 $ ;

- H :; H :;0 # ! $ 3 F@ $
>! >! ..!:! >?!>?: ;9! .? ! =+
-
H :0

3 )

,
+
$ 3
$ $

;!!???
9?;!>?; %

%
$ !

! 3
J
$ K $ >! > + !

%E2 H !;:!.>>
! 8
;?9;
+ % )
! $

$
%

+ ./
: +! * " = $

$

) ) !

$ % ) $

)

+ $ +
+ $

;

K $ >! > %

$ ) ! $

#
! $
)
+ ./ + !
$ +
7
!

%

$ 1

# !
$

- 0 $

%

% !

+ ! 1

$

%

% ./ $
G 1& >"! 2 " 6 :" %
$

+ 6 A

:" G +
! 6
A ?
+ $
:>.!:> ! ?9!9; %
;9B

# ! =! $ 7
G $ 9
!
:
!
$ 9>
6 A
% $

+
% &
! +
$ $ &

+
% ) 3 8

+ E+ ./
!
!

% ./

% 3

./ !

$
$
!

$ + %! ! 3
+ 2 8!

# ! $
% $
!

< + ! +
+
./
+

% $
&
+
+ # !
+

2 8! + !
$

*! $ $ %

- 0 $
% 6
A :" -!M0! +

! $ !

< $ $ 9
! ) -:
0
$ < < 9" < < ?" !
;
.

! 6
A :"
!
$ %

+ $ %
:"! ?? +
9>

# !
3
2 A ;"! $
.
+ 2 8! ? .?

$

!" #!! $# $# $#
!" #!!
% % &
'
!("
!(" )!
* )! )#+$,
)#+$,
-, ., ., .,
/+!+" .,
0,,$ .,
1(
%

% + -!"0! $ !

"

+ # ! ;?"
+
1 +
F2 +
%
8 + -! !0! $ !

+ F2 + 6 A :"
! + N-! ! !0! $
;?"
@ ./

+

)

2!

+ 3 $
$ $

% + -0
# !
;?"

$ % + -#0 $ #

+

2 8!

) +

< ! $

! $
6 A *!

+ ! $

= $

! )
) $ + !
+ ! + L $
8 + -F@0
$ ) 2
8! F@ 8
! $

+

&

J! 8 F@

;
!! + ! % ./
$ ./ $

%
1
# !
+ !

$ ./ $

% < $
! $ ) !
$
$
@
+
;. # !

- 0! $ < 1

3
# ! $
;
$ H
:; H :; < $ 9 !
) $ :

+ $

% &
+ ! $

$ ./ %

$
=
& *!

$
8 + % +

$ & < $
& )
./
E !

$ & $
$ &
$ $ ' ( +
$ =

% $

+
$ & $
$

$ &
8 $ *! $

% $

$ ! !
!
8
,%# % $ &
J 2 ! %
$!
J2 :E : :;! J2 ;E : @O
J2 ?E : 9.@O. < & / E F&
% & $
$! $

23 1' 1
*' 4' 5 ' 6' ) " 7 5 %& 1 222
. 1 3 %
'8 '8 ' 9 8 '
23 ' 1 % 1' # & # :/ 5 " "& 222 .
1 %
' ' '(99 8 '
23 '' " +'' ; <'=' " '+' ; *%
" >&

?05@

%
' '( '(9 8'
23 +'' ; '#' + 5 %

" A " "

%
'( ' '9(9(9
'
2(3 B' ; =' ; 1' ; 5 ;C " "& " >

"
&

%
' ' '(8((
8'
293 B' ; 0' D" D' # =' ; C
" " "&/ 5 & , :

%
'(
' ' '
283 D';' + '' " 0 " E " " "

"
%
'( '8
' ( '
23 ' D" <' F 4'' 1 5 %
%
"

%
' ' '(( '
23 =' & C' A 0 " "
222 . 1 %
'(
' ' '
2 3 #' +' D <' ' " 5 %
" 1
2224$ 453 $ '88888 ('
23 ' ' D" <' ' " 5 "
"
&"
)
6 + %
' ' '(9 ('
23 <' ' " ' ' D" 5 % " "
"
&"

7 %
' ' '88 8'
23 <' ' " ' ' D" 5 "& " "
&"

7 %
' ' '((( 9'
23 5' C " G: % ' 6 ' " 0 " "&
&"

6 + %
'8 ' '989 '
2(3 ' ' #" ' +' ; ' ' D" + ,
" &
" "& 222 . 1 8 %
'9 ' ' ( '
293 ' D' ' #' ; ' D' < +' <' ; 5 " & " "&
"
222 7 & 6 %
'( ' '8 '
283 ' ' D" 1' ' D" " "&
" & 9 6
%
' ' '( 9'

23 1' ' D" ' ' D" *%

"
" &
)
. 2 %
' '9 ' 8'
23 #' ; 4' ' 1 F' ' ; #' ' #" *%
"

1
:"" 222 ; 3 .
'8 9'
2 3 <' ' " ' ' D" <' <' + 5 "
"
" &
. 5 .5 :"" 8<. %
' ' 9'
c ISSN 1349-4198
REAL-TIME DECENTRALIZED NEURAL BLOCK CONTROL:

APPLICATION TO A TWO DOF ROBOT MANIPULATOR
R. Garcia-Hernandez1,3, , E. N. Sanchez1 , E. Bayro-Corrochano1

ez2 and Jose A. Ruz-Hernandez3
V. Santiban
1
Departamento de Ingenieria Electrica
CINVESTAV Unidad Guadalajara
Avenida Cientifica 1145, Zapopan, Jalisco, Mexico
{ rhernand; sanchez; edb }@gdl.cinvestav.mx
2
Division de Estudios de Posgrado
Instituto Tecnologico de la Laguna
Apartado Postal 49, Adm. 1, C.P. 27001, Torreon, Coahuila, Mexico
vsantiba@itlalaguna.edu.mx
3
Facultad de Ingenieria
Universidad Autonoma del Carmen
Avenida 56, No. 4, Cd. del Carmen, Campeche, Mexico
jruz@pampano.unacar.mx

Corresponding author: rghernandez@pampano.unacar.mx
Received October 2009; revised April 2010
Abstract. This paper presents a discrete-time decentralized control scheme for trajec-
tory tracking of a two degrees of freedom (DOF) robot manipulator. A modified recur-
rent high order neural network (RHONN) structure is used to identify the plant model
and based on this model, a discrete-time control law is derived, which combines block
control and sliding mode techniques. The neural network learning is performed on-line
by extended Kalman filtering (EKF). The controllers are designed for each joint using
only local angular position and velocity measurements, simplifying computation com-
plexity. The proposed scheme is implemented in real-time to control a two DOF robot
manipulator.
Keywords: Decentralized control, Recurrent neural networks, Extended Kalman filter,
Block control
1. Introduction. Robot manipulators are employed in a wide assortment of applica-

tions. Today most of the applications are in manufacturing to move materials, parts and
tools of various types. Future applications will include nonmanufacturing tasks, such as
construction work, exploration of space and medical care.
In this context, dierent control schemes have been proposed to guarantee ecient tra-
jectory tracking and stability [1,2]. Fast advance in computational technology oers many
ways for implementing control algorithms within the approach of a centralized control de-
sign [3]. However, there is a great challenge to obtain an ecient control for this class
of systems, due to its highly nonlinear complex dynamics, with strong interconnections,
parameters dicult to be measured and dynamics dicult to model. Considering only the
most important terms, the mathematical model obtained requires control algorithms with
great number of mathematical operations, which aect real-time implementation feasibil-
ity. On the other hand, within the area of control systems theory, for more than three
decades, an alternative approach has been developed considering a global system as a set
1075
1076 R. GARCIA-HERNANDEZ ET AL.
of interconnected subsystems, for which it is possible to design independent controllers,

considering only local variables to each subsystem: the decentralized control [4,5].
Decentralized control has been applied in robotics, mainly in cooperative multiple mo-
bile robots and robot manipulators, where it is natural to consider each mobile robot or
each manipulator as a subsystem of the whole system. For robot manipulators, each joint
and activated link is considered as a subsystem in order to develop local controllers, which
consider only local angular position and angular velocity measurements, and compensate
interconnection eects, usually assumed as disturbances. The resulting controllers are
easily implemented for real-time applications [6]. In [7], a decentralized control for robot
manipulators was developed, decoupling the dynamic model of the manipulator in a set of
linear subsystems with uncertainties and simulations for a robot of two joints were shown.
A frequency domain robust decentralized PI (Proportional Integral) controller design for
nominal performance and robust stability is proposed in [8]. In [9], a novel decentralized
control design method for spatially interconnected ane fuzzy systems was presented.
The decentralized control is obtained by solving linear matrix inequalities (LMIs). In [10],
the authors proposed a framework for decentralized design that avoided the knowledge
of local controllers but introduced information about local desired performance. The
resulting analysis equations will provide a framework where currently existing tuning
approaches for single-loop PI/PID controllers can be applied to Two-Input-Two-Output
(TITO) systems within a decentralized approach.
In [11], an approach of decentralized neural identification and control for robot manip-
ulators was presented using models in discrete-time. In [12], a decentralized control for
robot manipulators was reported; it was based on the estimation of independent dynamics
for each joint, using feedforward neural networks. In [13], the authors presented a de-
centralized control strategy using a recurrent neural identifier and block control structure
for both identification and control. This approach is tested only using simulations, with
a two degrees of freedom robot manipulator, and with an Articulated Nimble Adaptable
Trunk (ANAT) manipulator, with seven degrees of freedom.
In this paper, we use an Extended Kalman Filter (EKF)-based training algorithm for
a modified recurrent high order neural network (RHONN), in order to identify the robot
manipulator model. Based on this model, a discrete-time control law is derived, which
combines block control and sliding mode techniques. The block control approach is used
to design a nonlinear sliding surface such that the resulting sliding mode dynamics is
described by a desired linear system [14]. We present the real-time implementation of the
proposed scheme to control a two DOF robot manipulator. It is worth to note that a
preliminary version of this paper was presented in [15].
2. Discrete-Time Decentralized Systems. Let consider a class of discrete-time non-

linear perturbed and interconnected system, which can be presented in the nonlinear
block-controllable (NBC) form [16] consisting of r blocks
( ) ( )
1i (k + 1) = fi1 1i + Bi1 1i 2i + 1i
( ) ( )
2i (k + 1) = fi2 1i , 2i + Bi2 1i , 2i 3i + 2i
..
.
( )
i (k + 1) = fir1 1i , 2i , . . . , r1
r1
i (1)
r1
( 1 2 r1
) r r1
+Bi i , i , . . . , i i + i
r r
( ) r
( ) r
i (k + 1) = fi i + Bi i ui + i
REAL-TIME DECENTRALIZED NEURAL BLOCK CONTROL 1077
[ ]
rT T nij 1
[ j j j T
]
where i ni , i = 1T i 2T
i . . . i and j
i , j
i = i1 i2 . . . il , i =
1, . . . , N ; j = 1, . . . , r; l = 1, . . . , nij ; N is the number of subsystems, ui mi is the

input vector, the rank of Bij = nij , rj=1 nij = ni , ji Dj nij . We assume that
i
fij , Bij and ji are smooth and bounded functions, fij (0) = 0 and Bij (0) = 0. The integers
ni1 ni2 nij mi define the dierent subsystem structures. The interconnection
terms are given by

N
( )
1i = i1 1
=1, =i

N
( )
2i = i2 1 , 2
=1, =i
..
. (2)

N
( )
r1
i = ir1 1 , 2 , . . . , r1

=1, =i

N
( )
ri = ir
=1, =i
where represents the state vector of the -th subsystem with 1 N and = i.
Interconnection terms (2) reflect the interaction between the i-th subsystem and the
other ones.
3. Neural Identifier.
3.1. Decentralized neural identifier. The following decentralized recurrent high order
neural network (RHONN) model is proposed to identify (1):
x1i (k + 1) = wi1 (k)S(1i (k)) + w i 2i (k)
1
( )
x2i (k + 1) = wi2 (k)S 1i (k), 2i (k) + w i (k)3i (k)
2
.. (3)
.
( )
(k) + w i
(r1) r
xr1
i (k + 1) = wir1 (k)S 1i (k), 2i (k), . . . , r1
i i (k)
( )
xri (k + 1) = wir (k)S 1i (k), . . . , ri (k) + w i ui (k)
r
T
where xji = [x1i x2i . . . xri ] is the i-th block neuron state i = 1, . . . , N and j = 1 . . . , r,
wij (k) are the adjustable weights, w ji are the fixed weights with rank(w ji ) = nij , S() is
the activation function and ui (k) represents the input vector.
Comment 1. It is worth to note that model (3) does not consider explicitly the inter-
connection terms, whose dynamics is absorbed into the neural network weights update;
additionally model (3) constitutes a series-parallel identifier [17].
3.2. EKF training algorithm. It is known that Kalman filtering (KF) estimates the
state of a linear system with additive state and output white noises [18]. For KF-based
neural network (NN) training, the network weights become the states to be estimated.
In this case, the error between the NN output and the measured plant output can be
considered as additive white noise. Due to the fact that NN mapping is nonlinear, an
EKF-type is required.
The training goal is to find the optimal weight values wij (k), which minimize the pre-
diction error. We use an EKF-based training algorithm described by:
Kij (k) = Pij (k)Hij (k)Mij (k)
wij (k + 1) = wij (k) + ij Kij (k)eji (k) (4)
Pij (k + 1) = Pij (k) Kij (k)HijT (k)Pij (k) + Qji (k)
with
Mij (k) = [Rij (k) + HijT (k)Pij (k)Hij (k)]1
(5)
eji (k) = [ji (k) xji (k)]
where eji (k) is the identification error, Pij (k) is the prediction error covariance matrix,
wij (k) is the j-th weight (state) of the i-th subsystem, ij is the rate learning parameter
such that 0 ij 1, ji (k) is the j-th plant state, xji (k) is the j-th neural network state,
n is the number of states, Kij (k) is the Kalman gain matrix, Qji (k) is the measurement
noise covariance matrix, Rij (k) is the state noise covariance matrix and Hij (k) is a matrix,
in which each entry of (Hij ) is the derivative of j-th neural network state (xji (k)), with
respect to all adjustable weights (wij ), as follows
[ j ]T
j xi (k)
Hi (k) = (6)
wij (k) wij (k)=wij (k+1)
where i = 1, . . . , N and j = 1, . . . , n. Usually Pij and Qji are initialized as diagonal

matrices, with entries Pij (0) and Qji (0), respectively [19]. It is important to remark that
Hij (k), Kij (k) and Pij (k) for the EKF are bounded [19].
Comment 2. It is worth to remark that the training of the neural network is performed
on-line by extended Kalman filtering (EKF), which means that the respective weights are
update at each sampling time (k).
4. Controller Design. The proposed control strategy is deduced as follows. We begin

defining the tracking error as
zi1 (k) = x1i (k) x1id (k) (7)
where x1id (k) is the desired trajectory signal.
Once the first new variable (7) is defined, one step ahead is taken
zi1 (k + 1) = wi1 (k)S(1i (k)) + w i 2i (k) x1id (k + 1).
1
(8)
Equation (8) is viewed as a block with state zi1 (k) and the state 2i (k) is considered as
a pseudo-control input, where desired dynamics can be imposed. This equation can be
solved with the anticipation of the desired dynamics for this block as
zi1 (k + 1) = wi1 (k)S(1i (k)) + w i 2i (k) x1id (k + 1)
1
(9)
= ki1 zi1 (k)
with |ki1 | < 1, in order to ensure stability of (9).
2i (k) is computed as
1
x2id (k) = 1 1 1 1 1
1 [wi (k)S(i (k)) + id (k + 1) + ki zi (k)]. (10)
wi
Note that the calculated value for state x2id (k) in (10) is not the real value of such
state; instead, it represents the desired behavior for 2i (k). Hence, to avoid confusion this
desired value of 2i (k) is referred as x2id (k) in (10).
Proceeding in the same way as for the first block, a second variable in the new coordi-
nates is defined as
zi2 (k) = x2i (k) x2id (k).
Taking one step ahead for zi2 (k) yields
zi2 (k + 1) = x2i (k + 1) x2id (k + 1).
The desired dynamics for this block are imposed as
zi2 (k + 1) = wi2 (k)S(1i (k), 2i (k)) + w i x3i (k) x2id (k + 1)
2
(11)
= ki2 zi2 (k)
with |ki2 | < 1.
These steps are taken iteratively. At the last step, the known desired variable is xrid (k),
and the last new variable is defined as
zir (k) = xri (k) xrid (k).
As usually, taking one step ahead yields
zir (k + 1) = wir (k)S(1i (k), . . . , ri (k)) + w i ui (k) xrid (k + 1).
r
(12)
[ ]
System (3) can be represented in the new variables zi = zi1T zi2T zirT as
zi1 (k + 1) = ki1 zi1 (k) + w i zi2 (k)
1
zi2 (k + 1) = ki2 zi2 (k) + w i zi3 (k)

2
.. (13)
.
zir1 (k + 1) = kir1 zir1 (k) + w i
(r1) r
zi (k)
w i ui (k) xrid (k + 1).
r
zir (k + 1) = wir (k)S(1i (k), . . . , ri (k)) +
For sliding mode control implementation [20], when the control resources are bounded
by u0i as
|ui (k)| u0i (14)
a sliding manifold and a control law, which drives the states toward such manifold must be
designed. The sliding manifold is chosen as SDi (k) = zir (k) = 0. System (12) is rewritten
as follows:
SDi (k + 1) = wir (k)S(1i (k), . . . , ri (k)) + w i ui (k) xrid (k + 1).
r
(15)
Once the sliding manifold is defined, the next step is to find a control law which takes
into consideration the bound (14); therefore, the control ui (k) is selected as [21]:
{
ueqi (k) for ueqi (k) u0i
ui (k) = ueqi (k)
ui0 u (k) for ueqi (k) > u0i (16)
eqi
where the equivalent control ueqi (k) is calculated from SDi (k + 1) = 0 as
1 r 1 r r
ueqi (k) = r [wi (k)S(i (k), . . . , i (k)) + xid (k + 1)]. (17)
wi
The whole proposed identification and control scheme is displayed in Figure 1.
Neural Block
Controller Robot Manipulator
Controller 1 Link 1
Controller 2
ui (k ) Link 2 c i (k )
Controller N Link N
ei (k )
xid (k ) EKF
wi (k )
References
Neural Network 1
Neural Network 2
xi (k )
Neural Network N
Neural Identifier
Figure 1. Decentralized neural block control scheme
5. Two DOF Robot Manipulator Application. This section presents the real-time
application for a two DOF robot manipulator.
5.1. Robot manipulator description. In order to evaluate, via real-time implemen-

tation, the performance of the proposed control algorithm, we use a two DOF robot
manipulator moving in the vertical plane as presented in Figure 2. The considered robot
manipulator consists of two rigid links; high-torque brushless direct-drive servos are used
to drive the joints without gear reduction. This kind of joints present reduced backlash
and significantly lower joint friction as compared to actuators with gear drives. The
motors used in the experimental arm are DM1200-A and DM1015-B from Parker Com-
pumotor, for the shoulder and elbow joints, respectively. Angular information is obtained
from incremental encoders located on the motors, which have a resolution of 1,024,000
pulses/rev for the first motor and 655,300 for the second one (accuracy 0.0069 for both
motors), and the angular velocity information is computed via numerical dierentiation
of the angular position signal.
5.2. Control objective. To identify the robot model, from (1) and (4), we propose the
following decentralized series-parallel neural network for each joint
(k)S(1i (k)) + w i 2i (k)
1
1
x1i (k + 1) = wi1
(18)
(k)S(2i (k)) + w i ui (k)
2
x2i (k + 1) = wi1
2
(k)S(1i (k)) + wi2
2
j
where x1i (k) and x2i (k) identify 1i (k) and 2i (k); i = 1, 2, respectively; wip are the ad-
justable weights, p is the number of adjustable weights with p = 1; j = 1 for the first NN
state and p = 1, 2; j = 1, 2 for the second one; w 1i and w 2i are fixed parameters.
Due to the fact that RHONN weights are time varying, we need to guarantee the
controllability of the system by assuring that weights w 1i and w 2i are never zero. To
j
update the weights wip , an EKF-based training algorithm (4) is implemented.
Figure 2. Robot manipulator
The goal is to track a desired reference signal, which is achieved by designing a control
law based on the sliding mode technique. The tracking error for each joint is defined as
zi1 (k) = x1i (k) x1id (k) (19)
where x1id is the desired trajectory signal. Using (18) and introducing the desired dynamics
for zi1 (k), we have
(k)S(1i (k)) + w i 2i (k) x1id (k + 1)
1
zi1 (k + 1) = wi1
1
(20)
= ki1 zi1 (k).
The desired value x2id (k) for 2i (k) is calculated from (19) as
1
x2id (k) = 1
[wi1 (k)S(1i (k)) + x1id (k + 1) + ki1 zi1 (k)]. (21)
w 1i
At the next step, let define a new variable as
zi2 (k) = x2i (k) x2id (k). (22)
Taking one step ahead, we have
(k)S(2i (k)) + w i ui (k) x2id (k + 1)
2
2
zi2 (k + 1) = wi1 2
(k)S(1i (k)) + wi2
(23)
= ki2 zi2 (k).
The manifold for the sliding mode is chosen as SDi (k) = zi2 (k) = 0. The control law is
given by
{
ueqi (k) for ueqi (k) imax
ui (k) = ueqi (k) (24)
imax ueq (k)
for ueqi (k) > imax
i
where ueqi (k) is obtained from SDi (k + 1) = 0 as

1 [ ]
ueqi (k) = 2 (wi12 2
(k)S(1i (k)) + wi2 (k)S(2i (k))) + x2id (k + 1) (25)
wi
and the control resources are bounded by imax .
5.3. Real-time results. For the experiments, we select the following discrete-time tra-
jectories
3 3
x11d (k) = b1 (1 ed1 kT ) + c1 (1 ed1 kT ) sin(1 kT ) [rad]
3 3
x12d (k) = b2 (1 ed2 kT ) + c2 (1 ed2 kT ) sin(2 kT ) [rad]
where b1 = /4, c1 = /18, d1 = 2.0 and 1 = 5 [rad/s] are parameters of the desired
position trajectory for the first joint, whereas b2 = /3, c2 = 25/36, d2 = 1.8 and
2 = 1.0 [rad/s] are parameters of the desired position trajectory for the second joint.
The sampling time is selected as T = 2.5 milliseconds.
These trajectories incorporate a sinusoidal term to evaluate the performance for rel-
atively fast periodic signals, where the non-linearities of the robot dynamics are really
important; additionally, they include a term, which smoothly grows for maintaining the
robot in an operation state without saturating actuators. The actuator saturation limits
are 150 [Nm] and 15 [Nm], respectively.
Comment 3. It is worth to remark that the paper reports real-time implementation of the
proposed control scheme. For the real system the initial conditions are the same that those
of the neural identifier, both restricted to be equal to zero in according to the experimental
prototype architecture; therefore transient errors do not appear.
Real-time results for identification and trajectory tracking using the proposed decen-
tralized neural identification and control scheme are shown in Figures 3 and 4.
0.9
0.8
0.7
Joint position (rad)
0.6
0.5
0.4
0.3
0.2
x11d
11
0.1
x11
0
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Figure 3. Identification and tracking for joint 1 x11d (k) (solid line), 11 (k)
(dashdot line) and x11 (k) (dashed line)
The tracking error performance for each joint are presented in Figure 5; The applied
torques for each joint are shown in Figure 6. These torques present oscillations at some
time instants due to gains and fixed parameters selected for each controller. It is easy
to see that both control signals are always inside of the prescribed limits given by the
actuators manufacturer; that is, their absolute values are smaller than the bounds 1max
and 2max , respectively.
Time evolution of the position error e1i reflects that the control system performance is
very good. The performance criterion considered in this paper is the mean square error
3.5
2.5
Joint position (rad)

1.5
0.5
0.5
x12d
1
12
1.5
x12
2
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Figure 4. Identification and tracking for joint 2 x12d (k) (solid line), 12 (k)
(dashdot line) and x12 (k) (dashed line)
0.1
0.05 e11
Error (rad)
0.05
0.1
0 2 4 6 8 10 12 14 16 18 20
Time (s)
0.2
e12
0.1
Error (rad)
0.1
0.2
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Figure 5. Tracking errors for joints 1 and 2
(MSE) value of the position error computed as

v
u n
u1
MSE[ei ] = t
1
e1 2 T (26)
t k=0 i
where T is the sampling time and t = 20 seconds.

The respective mean square errors for the proposed scheme are included in Table 1.
According to the mean square errors presented above, the proposed scheme based on the
block control and sliding mode techniques present a very good performance for trajectory
tracking.
100
1
Torque (Nm)
50
50
100
0 2 4 6 8 10 12 14 16 18 20
Time (s)
10
2
Torque (Nm)
10
0 2 4 6 8 10 12 14 16 18 20
Time (s)
Figure 6. Applied torques to joints 1 and 2
Table 1. Mean square error for the real joint positions
Control algorithm MSE[e11 (k)] MSE[e12 (k)]

DNBC 8.3325e-6 0.0015
6. Conclusions. This paper presents a decentralized control scheme which is designed

with a modified recurrent high order neural network in order to identify the robot model.
The training of each neural network is performed on-line using an extended Kalman
filter in a series-parallel configuration. Based on this neural identifier and applying the
discrete-time block control approach, a nonlinear sliding manifold is formulated.
The obtained real-time results present good performance for trajectory tracking when
applied to a two DOF robot manipulator.
Acknowledgment. Authors thank CONACYT-Mexico for the financial support dur-

ing this research through projects 57801Y, 82084 and 86940. The first author thanks
to Universidad Autonoma del Carmen (UNACAR) and the Programa de Mejoramiento
del Profesorado (PROMEP) for supporting this research. The fourth author thanks to
DGEST-Mexico for supporting this research through project 2166.09-P. They also thank
the anonymous reviewers for helping to improve this paper.
REFERENCES
[1] E. N. Sanchez and L. J. Ricalde, Trajectory tracking via adaptive recurrent neural control with input
saturation, Proc. of International Joint Conference on Neural Networks, Portland, OR, pp.359-364,
2003.
[2] V. Santibanez, R. Kelly and M. A. Llama, A novel global asymptotic stable set-point fuzzy controller
with bounded torques for robot manipulators, IEEE Trans. on Fuzzy Systems, vol.13, no.3, pp.362-
372, 2005.
[3] R. Gourdeau, Object-oriented programming for robotic manipulator simulation, IEEE Robotics and
Automation, vol.4, no.3, pp.21-29, 1997.
[4] Z. P. Jiang, New results in decentralized adaptive nonlinear control with output feedback, Proc. of
the 38th IEEE Conference on Decision and Control, Phoenix, AZ, pp.4772-4777, 1999.
[5] S. Huang, K. K. Tan and T. H. Lee, Decentralized control design for large-scale systems with strong
interconnections using neural networks, IEEE Trans. on Automatic Control, vol.48, no.5, pp.805-810,
2003.
[6] M. Liu, Decentralized control of robot manipulators: Nonlinear and adaptive approaches, IEEE
Trans. on Automatic Control, vol.44, no.2, pp.357-363, 1999.
[7] M. L. Ni and M. J. Er, Decentralized control of robot manipulators with coupling and uncertainties,
Proc. of the American Control Conference, Chicago, IL, pp.3326-3330, 2000.
[8] A. Kozakova, Tuning decentralized PID controllers for performance and robust stability, ICIC Ex-
press Letters, vol.2, no.2, pp.117-122, 2008.
[9] Z. He and Y. Lu, Decentralized control of spatially interconnected ane fuzzy systems with appli-
cation to a morphing airfoil, ICIC Express Letters, vol.3, no.3(B), pp.593-598, 2009.
[10] R. Vilanova, P. Balaguer, A. Ibeas and C. Pedret, Expected interaction analysis for decentralized
control on TITO systems: Application to IMC based PI/PID controller selection, International
Journal of Innovative Computing, Information and Control, vol.5, no.10(B), pp.3439-3456, 2009.
[11] A. Karakasoglu, S. I. Sudharsanan and M. K. Sundareshan, Identification and decentralized adaptive
control using dynamical neural networks with application to robotic manipulators, IEEE Trans. on
Neural Networks, pp.919-930, vol.4, no.6, 1993.
[12] R. Safaric and J. Rodic, Decentralized neural-network sliding-mode robot controller, Proc. of the 26th
Annual Conference on the IEEE Industrial Electronics Society, Nagoya, Aichi, Japan, pp.906-911,
2000.
[13] E. N. Sanchez, A. Gaytan and M. Saad, Decentralized neural identification and control for robotics
manipulators, Proc. of the IEEE International Symposium on Intelligent Control, Munich, Germany,
pp.1614-1619, 2006.
[14] A. Y. Alanis, E. N. Sanchez, A. G. Loukianov and G. Chen, Discrete-time output trajectory tracking
by recurrent high-order neural network, Proc. of the 45th Conference on Decision and Control, San
Diego, CA, pp.6367-6372, 2006.
[15] R. Garcia-Hernandez, E. N. Sanchez, V. Santiba nez, M. A. Llama and E. Bayro-Corrochano, Real-
time decentralized neural block controller for a robot manipulator, Proc. of International Conference
on Intelligent Robots and Systems, Saint Louis, MO, pp.906-911, 2009.
[16] V. Utkin, Block control principle for mechanical systems, Journal of Dynamic Systems, Measure-
ment, and Control, vol.122, no.1, pp.1-10, 2000.
[17] P. A. Ioannou and J. Sun, Robust Adaptive Control, Prentice Hall, New Jersey, 1996.
[18] S. Haykin, Kalman Filtering and Neural Networks, John Wiley & Sons, New York, 2001.
[19] Y. Song and J. W. Grizzle, The extended Kalman filter as local asymptotic observer for discrete-time
nonlinear systems, Journal of Mathematical Systems, Estimation and Control, vol.5, no.1, pp.59-78,
1995.
[20] V. Utkin, Sliding mode control design principles and applications to electric drives, IEEE Trans. on
Industrial Electronics, vol.40, no.1, pp.23-36, 1993.
[21] V. Utkin, J. Guldner and J. Shi, Sliding Mode Control in Electromechanical Systems, Taylor &
Francis, Philadelphia, 1999.
[22] R. Garcia-Hernandez, E. N. Sanchez, M. Saad and E. Bayro-Corrochano, Discrete-time decentralized
neural block controller for a five DOF robot manipulator, Proc. of International Joint Conference
on Neural Networks, Atlanta, GA, pp.925-931, 2009.
c ISSN 1349-4198
A SUPERVISING AUTHENTICATED ENCRYPTION SCHEME FOR

MULTILEVEL SECURITY
Chien-Lung Hsu2 , Liang-Peng Chang1 and Tzong-Chen Wu1

1
Department of Information Management
Chang-Gung University
259 Wen-Hwa 1st Road, Kwei-Shan, Tao-Yuan, Taiwan
clhsu@mail.cgu.edu.tw
2
National Taiwan University of Science and Technology
43 Section 4, Keelung Road, Taipei, Taiwan
kevinchang@ntu.edu.tw; tcwu@cs.ntust.edu.tw
Received October 2009; revised February 2010
Abstract. Access to secret data should be managed to allow authorized people only. An
authenticated encryption scheme can be used to protect valuable information or secret
data for data confidentiality, authentication, non-repudiation and integrity. In this pa-
per, the authors propose a new supervising authenticated encryption scheme for multilevel
security which deals with the monitor and access control problems found in hierarchical
organization, for protecting valuable authenticated encryption messages from being dis-
closed by malicious adversary. The proposed scheme provides two eective access control
mechanisms, one is the partial access control and the other is the complete access control,
which allows management superiors to monitor or access authenticated encryption mes-
sages received by inferiors within a hierarchical organization. Considering user privacy,
the partial access control mechanism allows only a superior to access the intended au-
thenticated encrypted information. In case of some special scenario (e.g., the monitored
inferior is dead or dismissed), the complete access control mechanism allows the superior
to access all authenticated encrypted messages received by the monitored inferior.
Keywords: Multilevel security, Authenticated encryption, Access control, Monitor
1. Introduction. A cryptographic scheme for enforcing multilevel security systems, whe-

re the hierarchy is represented by a partially ordered set, was introduced by Akl and Taylor
[1]. Many cryptography techniques could be used to solve hierarchical access control prob-
lems, which mainly include access control [1-6], key escrow [7-9], and key recovery/key
backup [10-12]. Although established methods could achieve the goal of multilevel secu-
rity [13-17], these methods allow superiors access with the users private key, or it requires
additional operating costs to guarantee superiors only access to all users specific infor-
mation. Considering the balance between the users privacy and superiors access rights,
the system must be able to restrict each superior access to information, and prevent them
from exceeding their level of authority access to information. However, under certain sit-
uations, some superiors must be authorized to access all users information, for example,
a user leaves the job.
An authenticated encryption scheme integrates encryption, decryption and digital sig-
nature techniques, and can be regarded as a combination of a data encryption scheme
and a digital signature scheme for achieving privacy, integrity, and authentication [18-26].
However, the digital signature must be verified by a specified recipient, while the message
remains completely confidential.
1087
1088 C.-L. HSU, L.-P. CHANG AND T.-C. WU
The proposed scheme provides two access control mechanisms for multilevel security,
which eectively address the monitor and access control problems in hierarchical orga-
nizations, for the protection of valued authenticated encrypted information from being
misused, stolen, hidden, etc. by others. The first is partial access control, and the second
is complete access control. The proposed access control mechanisms are based on the
partially ordered set (POSET for short) architecture [1] by Akl and Taylor. The partial
access control mechanism allows only a superior to monitor or access the intended au-
thenticated encrypted information. In an occurrence of a special scenario (e.g., inferior
employee of a hierarchical company is dismissed), the complete access control mecha-
nism allows the superior to monitor or access all authenticated encrypted messages and
information received by the inferior in the hierarchical organization.
The remainder of this paper is organized as follows. Section 2 describes the system
model. Section 3 proposes a new supervising authenticated encryption scheme for mul-
tilevel security. Section 4 analyses the security features. Finally, conclusion is given in
Section 5.
2. System Model. This section will first describe the system model of the proposed
access control mechanisms, as shown in Figure 1. There are five major roles in this system
model: the system authorization center (SA), the sender (signer U), the receiver (specific
verifier), the superior, and the arbitrator (A). The system model provides two eective
access control mechanisms, one is the partial access control and the other is the complete
access control, which allow a superior to apply to the SA for access to all or some of his
inferiors authenticated encryption messages. The partial access control mechanism allows
only a superior to access the intended authenticated encrypted message. The complete
access control mechanism allows the superior to access all authenticated encrypted
messages received by the monitored inferior.
Figure 1. System model of the proposed access control mechanisms
In case of a dispute or the authenticated encryption message is no longer confidential,

this authenticated encryption message can be decrypted by the receiver (specific verifier)
or their superiors. It will be converted to a digital signature with message recovery,
and then be sent to the arbitrator A to verify that the signature of this authenticated
encryption message was signed by the sender (signer U).
The relation between the verifier and the superior follow the POSET architecture, as
proposed by Akl and Taylor. The notation represents the binary relation of partial
A SUPERVISING AUTHENTICATED ENCRYPTION SCHEME 1089
order, where Vi Vj means that the security class of Vj is higher than Vi , or their security
classes are equal in the POSET (VG, ). Specifically, Vj owns the right of access and
verifies the received authenticated encryption message of Vi . For example, following the
POSET architecture of Figure 2, V2 V1 represents security class V1 that higher than
V2 , and V1 can access authenticated encryption messages and information received by
V2 .
An example of the POSET architecture in Figure 2 shows five users V1 , V2 , . . . , V5
and the relation of POSET between them. V1 is the superior of {V2 , V3 , V4 }, V2 and
V3 are the superiors of V5 , and so on. In the access architecture of POSET, the superior
could access his/her subordinates authenticated encryption message. In this model, the
signer U could generate an authenticated encryption message to any users in the group
VG. For example, in Figure 2, only V2 could authenticate and decrypt the authenticated
encryption message. After that, V2 could recover Us signature and signed message, which
he received from U. In case of dispute or the signed message is no more confidential, V2
could send Us signature to arbitrator, to conform the validity of Us signature. For the
purposes of monitor and internal control, this model allows the superior of V2 (i.e. user
V1 who is in Figure 2) to access the authenticated encryption message, which received by
V2 following POSET. The access control mechanisms could be divided to partial access
control and complete access control. In partial access control, the superior V1 need to
request from SA in advance and only could access the special authenticated encryption
message, which had been received by V2 . In complete access control, the superior V1
could access all authenticated encryption messages, which had been received by V2 .
Figure 2. An example with five users {V1 , V2 , . . . , V5 } in the POSET

(VG, ) architecture
3. Proposed Scheme. The proposed scheme is divided into six stages: registration, key
assignment, authenticated encryption, authenticated decryption, signature verification,
and access control. The following describes these stages in detail.
Registration stage: It is assumed that p and q are two large primes. xSA Zq is a
private key of SA, and ySA is a public key corresponding to xSA , where:
ySA = g xSA mod p (1)
Then, SA generates a key pair, a private key xi Zq and a public key yi for each verifier
Vi (i = 1, 2, . . . , l) within the POSET architecture VG, where:
yi = g xi mod p (2)
Then, the SA transfers xi to Vi, via secure channel, and publishes yi .
The signer U chooses the private key xu Zq , and then computes the public key by
yu = g xu mod p (3)
Then, the signer sends the yu to the SA for certification. The SA uses a non-interactive
zero-knowledge proof method [27] to prove the signer has the key pair of yu and xu .
Finally, the SA publishes signer Us public key yu .
Key assignment stage: In accordance with VGs POSET architecture, the SA gen-
erates a supervised secret key and a public key for each verifier Vi VG. Then, the
SA assigns a supervised secret key to all Vi s superiors Vj (Vi Vj ). When one of
the superiors applies to the SA for access and receives approval from the SA, he can
use the specified supervised secret key to decrypt the authenticated encrypted messages,
which were received by inferiors in the hierarchical organization. The SA generates Vi s
supervised public key SPK i :
SP Ki = g H(XSA ||Supervised datai ) mod p (4)
and the supervised secret key SSKi :
SSKi = xi H(XSA ||Supervised datai )SP Ki mod q (5)
where H is a one-way hash function and Supervised datai is the information regarding
access control, such as the identifier of the superior and the expiration date of the access
control. The SA publishes SP Ki and Supervised datai , and then sends the SSKi to all
Vi s superiors, Vj (Vi Vj ), by a secure channel. After Vj receives SSKi , he can use
SP KiSP Ki g SSKi = yi (mod p) to verify the correctness of SSKi .
Authenticated encryption stage: Suppose that, M Zp is the message to be
signed with a specific redundancy attached. When signer U attempts to generate an
authenticated encryption message from M for the specific verifier Vi VG, U would
randomly choose w, d Zq and calculate as:
R = M g w mod p (6)
S = w + xu R mod q (7)
C1 = g d mod p (8)
C2 = R (yid mod p) (9)
and then, transfer the authenticated encryption message aumsg = (C1 , C2 , S) to Vi .
Authenticated decryption stage: When verifier Vi VG receives the authenticated
encryption message aumsg from signer U, Vi can recover R from (C1 , C2 ):
R = C2 (C1xi mod p) (10)
and then, recover the signed message M :
M = Rg S yuR mod p (11)
and then check the attached redundancy of M . If the attached redundancy is valid, it
indicates that aumsg = (C1 , C2 , S) is a valid authenticated encryption message of M ,
which is signed by U, and M is the signed information. Hence, (R, S) could be considered
as the valid signature signed by U for M .
Signature verification stage: In case of any dispute or the signed information is
disclosed, the verifier Vi VG can declare (R, S) to the arbitrator. The arbitrator can
recover the information M from (R, S), via (M = Rg S yuR mod p), and then check the
attached redundancy for verifying the validity of (R, S). Provided the redundancy is suc-
cessfully verified, it indicates that (R, S) is a valid signature, signed by U for information
M.
Access control stage: In this stage, the superior of the verifier can manage received
authenticated encrypted messages, through the POSET architecture, which can be cate-
gorized as partial and complete access control mechanisms. When partial access control
is conducted, it indicates that any superior Vj (Vi Vj ) of the verifier Vi can apply
for access control to the SA, in accordance with the specific authenticated encrypted
message. After receiving approval of the SA, Vj would be authorized with access rights to
verify and decrypt the authenticated encrypted messages. When complete access control
is conducted, it indicates that Vj , is the superior of Vi , and can apply to the SA for ac-
cess control in order to verify and decrypt the entire authenticated encrypted message
received by Vi . After receiving approval of the SA, Vj would be authorized with access
rights to verify and decrypt all authenticated encrypted messages of Vi .
Partial access control: When superior Vj applies to the SA for access to a specific
message aumsg=(C1 , C2 , S) that verifier Vi received from U, the SA would have to view
the authority of Vj, in accordance with VGs POSET architecture for approval of Vj
according to VGs security policies. Once the SA has approved the application of superior
Vj , it would calculate aumsgs supervisory information ES i :
H(XSA ||Supervised datai )
ESi = C1 mod p (12)
and then, transfer ES i to Vj by a secure channel. Then superior Vj uses the supervised
secret key SSK i of Vi to compute:
E = C1SSKi mod p (13)

R = C2 (ESiSP Ki E mod p) (14)
Vj could use M = Rg S yuR mod p to recover message M from (R, S), and then, check the
attached redundancy of M . If the redundancy is valid, it indicates that aumsg is the valid
authenticated encryption message, and that, (R, S) is a valid signature, which signer U
signed for information M .
Complete access control: When superior Vj applies to SA to access all messages
of inferior Vi , the SA must view the authority of Vj , in accordance with VGs POSET
architecture, to determine approval for Vj s application, in accordance with VGs security
policies. After the SA has approved the application of superior Vj , the SA would calculate
complete supervisory information i :
i = H(XSA ||Supervised datai ) (15)
and then transfer the information to Vj by a secure channel. Then, superior Vj can use
the supervised secret key SSKi of Vi to drive Vi s private key:
xi = SSKi + i SP Ki mod q (16)
After that, Vj could use Vi s private key to decrypt all the authenticated encrypted
messages of Vi .
Theorem 3.1. If a verifier Vi VG receives a valid authenticated encryption message

aumsg = (C1 , C2 , S) from signer U, he can correctly recover Us partial signature R from
(C1 , C2 , S) by Equation (10).
Proof: From Equations (10), (9) and (8), a deduction is made:

C2 (C1xi mod p)
= (R (yid mod p)) (C1xi mod p)
= (R (g dxi mod p)) (g dxi mod p)
=R
Theorem 3.2. With the knowledge of a valid signature (R, S), the signed message M can
be correctly recovered from (R, S) by Equation (11).
Proof: From Equations (3), (6), (7) and (11), a deduction is made:
Rg S ysR mod p
= Rg wxs R ysR mod p
= Rg w ysR ysR mod p
= M g w g w mod p
=M
Theorem 3.3. Superior Vj uses the supervised secret key SSKi of Vi to recover (R, S)
from the authenticated encryption message aumsg = (C1 , C2 , S).
Proof: From Equations (14), (12), (13) and (5), a deduction is made:
C2 (ESiSP Ki E mod p)
H(XSA ||Supervised datai )SP Ki
= C2 (C1 C1SSKi mod p)
H(XSA ||Supervised datai )SP Ki +SSKi
= C2 (C1 mod p)
= C2 (C1xi mod p)
= (R (yid mod p)) (g dxi mod p)
= (R (g dxi mod p)) (g dxi mod p)
=R
Theorem 3.4. Superior Vj can use the supervised secret key SSKi of Vi to drive Vi s
private key xi for supervising Vi s received authenticated encrypted messages.
Proof: From Equations (16), (5) and (15), a deduction is made:
SSKi + i SP Ki mod q
= xi H(XSA ||Supervised datai )SP Ki
+ H(XSA ||Supervised datai )SP Ki mod q
= xi
which is the private key of Vi .
4. Security Analysis. The proposed scheme whose security is based on the one-way
hash function (OWHF) and discrete logarithm problem (DLP) assumptions. A hash
function is a function that takes a variable-size input and returns a fixed-size result called
hash value. If a hash function H(x) is one-way hash function, it must satisfy the following
conditions:
(1) For any message x, it is easy to compute xs hash value H(x).
(2) It is computationally infeasible to find a message x from its hash value y = H (x).
(3) For any message x1 , it is computationally infeasible to find another message x2 such
that H (x2 ) = H (x1 ) = y.
(4) It is computationally infeasible to find a pair of dierent messages x1 and x2 such that
H (x1 ) = H (x2 ).
This paper considers the subject of monitor and access control in a hierarchical orga-
nization to propose a secure and ecient authenticated encryption scheme for multilevel
security. This section will examine the security of the proposed scheme. The proposed
scheme can achieve multilevel security requirements of confidentiality, unforgeable, and
non-repudiation for assuring group-oriented activities.
Confidentiality: When an attacker intercepts the authenticated encrypted message
aumsg = (C1 , C2 , S), in order to decrypt the signed and encrypted information M =
Rg S yuR mod p of the authenticated encryption message, then he must recover R first.
However, according to C2 = R(yid mod p), it is impossible to obtain R unless the attacker
knows the secret random number d. Based on a discrete logarithm problem assumption,
the attacker is unable to obtain d from C1 = g d mod p, and thus, the attacker is unable
to retrieve R to decrypt the authenticated encrypted message aumsg = (C1 , C2 , S).
Unforgeable: An attacker attempts to forge message M (which has specific redun-
dancy attached), and generates the forged authenticated encryption message (C1 , C2 , S),

which is for one specific verifier Vi VG. The attacker could select any random numbers
w and d, and use R = M g w mod p, C1 = g d mod p and C2 = R (y d mod p) to calculate
i
C1 and C2 . If the attacker manages to find one S to satisfy M = Rg S yuR mod p, then
R,
he can successfully forge a legal authenticated encrypted message (C1 , C2 , S).
Neverthe-
less, based on a discrete logarithm problem assumption, the attacker would be unable
to determine S. Thus, the attacker would not be able to forge a legal authenticated
encryption message.
Non-repudiation: The authenticated encryption message aumsg = (C1 , C2 , S) can
only be decrypted by a specific verifier, or his superior. However, when a dispute occurs,
this specific verifier or his superior may send the signature to an arbitrator to prove
that the signer signed message M . Unforgeable security analysis proves that an attacker
is unable to successfully forge a legal authenticated encrypted message and signature
without the knowledge of the signers private key. Therefore, a signer cannot deny a
produced authenticated encryption message or signature.
5. Performance Analysis. This section presents a performance analysis considering

both the computational complexity as well as the communication cost of the scheme. To
evaluate the computational complexity, we make the following notations. Let Th be the
time for computing one way hash function, TEXP be the time for computing modular
exponentiation, and TM be the time for computing the modular multiplication.
First we analyze the performance according to computational complexity required for
the proposed scheme. Table 1 presents the computational complexity for each user in
authenticated encryption, authenticated decryption, and signature verification phase.
Table 1. Performance analysis (computational complexity)
Signer 3TEXP + 2TM

Specific verifier 3TEXP + 2TM
Arbitrator 2TEXP + 2TM
In Table 2 we present the supervising computational cost of partial access control and
complete access control.
Table 2. Performance analysis (computational complexity)
Partial access control

SA 1TEXP + 1Th
Superior 4TEXP + 3TM
Complete access control
SA 1Th
Superior 3TEXP + 3TM
Next, we analyze the communicational cost of the proposed scheme. Table 3 presents
the number of messages in terms of communication transmission for each phase.
Table 3. Performance analysis (communication cost)
Phase Communication cost

Authenticated encryption message 2|p| + |q|
Signature |p| + |q|
Partial access control |p|
Complete access control |q|
Note: p and q denote two large prime
6. Conclusions. This paper presented a new supervising authenticated encryption sche-

me for multilevel security. The proposed scheme provides two eective access control
mechanisms, which allow a superior to apply to the system authority for access to all
or some of his inferiors authenticated encryption messages. For the requirements of fair
verification and access control, the proposed scheme allows an authenticated encryption
message to be decrypted and converted into a digital signature with message recovery by
the receiver or his superior, and can be sent to an arbitrator to settle any disputes.
Acknowledgment. This work was supported in part by the Chang Gung University
Grants UARPD390111 and UARPD380241, Chang Gung Memorial Hospital Grant CM-
RPD390031, and in part by National Science Council under the grants NSC 98-2410-H-
182-007-MY2. We would like to thank anonymous referees for their valuable suggestions.
REFERENCES
[1] S. G. Akl and P. D. Taylor, Cryptographic solution to a problem of access control in a hierarchy,
ACM Trans. on Computer System, vol.1, no.3, pp.239-248, 1983.
[2] S. J. MacKinnon, P. D. Taylor, H. Meijer and S. G. Akl, An optimal algorithm for assigning crypto-
graphic keys to access control in a hierarchy, IEEE Trans. on Computers, vol.34, no.9, pp.797-802,
1985.
[3] R. S. Sandhu, Cryptographic implementation of a tree hierarchy for access control, Information
Processing Letters, vol.27, no.2, pp.95-98, 1998.
[4] T. C. Wu and C. C. Chang, Cryptographic key assignment scheme for hierarchical access control,
Computer Systems Science and Engineering, vol.16, no.1, pp.25-28, 2001.
[5] C. Yang and C. Li, Access control in a hierarchy using one-way hash functions, Computers and
Security, vol.23, pp.659-664, 2004.
[6] Y. F. Chung, H. H. Lee, F. Lai and T. S. Chen, Access control in user hierarchy based on elliptic
curve cryptosystem, Information Sciences, vol.178, pp.230-243, 2008.
[7] Y. Desmedt, Securing traceability of ciphertexts Towards a secure key escrow system, Advances in
Cryptology, EUROCRYPT, pp.147-157, 1995.
[8] D. P. Maher, Crypto backup and key escrow, Communications of the ACM, vol.39, no.3, pp.48-53,
1996.
[9] Y. Long, Z. Cao and K. Chen, A dynamic threshold commercial key escrow scheme based on conic,
Applied Mathematics and Computation, vol.171, pp.972-982, 2005.
[10] S. T. Walker, S. B. Lipner, C. M. Ellison and D. M. Balenson, Commercial key recovery, Communi-
cations of the ACM, vol.39, no.3, pp.41-47, 1996.
[11] M. Smith, P. V. Oorschot and M. Willett, Cryptographic information recovery using key recovery,
Computers and Security, vol.19, no.1, pp.21-27, 2000.
[12] K. Rantos and C. J. Mitchell, Matching key recovery mechanisms to business requirements, Com-
puters and Security, vol.24, pp.232-245, 2005.
[13] W. P. Lu and M. K. Sundareshan, A model for multilevel security in computer networks, IEEE
Trans. on Software Engineering, vol.16, pp.647-659, 1990.
[14] H. Kim, B. Park, J. Ha, B. Lee and D. G. Park, New key management systems for multilevel security,
ICCSA 2005, LNCS, vol.3481, pp.245-253, 2005.
[15] X. D. Zuo, F. M. Liu and C. B. Ma, A new approach to multilevel security based on trusted
computing platform, IEEE International Conference on Machine Learning and Cybernetics, vol.4,
pp.2158-2163, 2007.
[16] B. Panja, S. K. Madria and B. Bhargava, A role-based access in a hierarchical sensor network
architecture to provide multilevel security, Computer Communications, vol.31, pp.793-806, 2008.
[17] K. Barkaoui, R. B. Ayed, H. Boucheneb and A. Hicheur, Verification of workflow processes under
multilevel security considerations, Proc. of the 3rd IEEE International Conference on Risks and
Security of Internet and Systems, vol.28-30, pp.77-84, 2008.
[18] W. B. Lee and C. C. Chang, Authenticated encryption scheme without using one-way function,
Electronics Letter, vol.31, no.19, pp.1656-1657, 1995.
[19] C. L. Hsu and T. C. Wu, Authenticated encryption schemes with (t,n) shared verification, Proc. of
the IEE Computers and Digital Techniques, vol.145, no.2, pp.117-120, 1998.
[20] Y. M. Tseng, J. K. Jan and H. U. Chien, Authenticated encryption schemes with message linkages
for message flows, Computers and Electrical Engineering, vol.29, pp.101-109, 2003.
[21] T. S. Chen, K. H. Huang and Y. F. Chung, A practical authenticated encryption scheme based on
the elliptic curve cryptosystem, Computer Standards and Interfaces, vol.26, pp.461-469, 2004.
[22] Z. M. Zhao and F. Y. Liu, Method of constructing elliptic curve authenticated encryption scheme,
Applied Mathematics and Computation, vol.168, pp.146-151, 2005.
[23] S. F. Tzeng, Y. L. Tang and M. S. Hwang, A new convertible authenticated encryption scheme with
message linkages, Computers and Electrical Engineering, vol.33, pp.133-138, 2007.
[24] T. S. Wu, C. L. Hsu, K. Y. Tsai, H. Y. Lin and T. C. Wu, Convertible multi-authenticated encryption
scheme, Information Sciences, vol.178, pp.256-263, 2008.
[25] J. H. Yang and C. C. Chang, An ecient fair electronic payment system based upon non-signature
authenticated encryption scheme, International Journal of Innovative Computing, Information and
Control, vol.5, no.11(A), pp.3861-3874, 2009.
[26] T. Y. Chang and M. L. Chiang, Attack and improve the Zhao-Liu authenticated encryption scheme,
International Journal of Innovative Computing, Information and Control, vol.5, no.10(A), pp.3025-
3030, 2009.
[27] D. Chaum and M. E. Pedersen, Transferred cash grows in size, Advances in Cryptology, EURO-
CRYPT, pp.390-407, 1992.

! " # $
%
&'
( )*$ +$

, " " , +$ - #
.$, , / *, $.$, ,

!
" # $
%& ' ( )
*

! "
+ ,

0 # $$ +$ $ 1$$
"$!

!
! !!
!" # $
! ! !!
!
% &

! %

!!
! '(
!

) '( !! !! ! !

*
+ !!

,
!
- %
! ! !

) ! %

! !

-
! .

) '/( 0

0
! & 1

0

0

2
!

*

!

)
'3(

! !! !

4

05

! !
*
! ! '6 7( !

'8(

!! 9
!

! .
!

! ) '(

!

%

! ! ) !
! : ! !!

! !
!
;2 <! 5 !
!

!

" !
5! 0 =* > 2
%

& 5
!
2
!! !

42 ' ( *??@)A B 4C)B =*??@ )
A!, B
!! ! C )! B>
?! !D
: -+ & ; !
& !D !
!
!
5! !!
! !! <!

!D
5
); 8 ?!+

E

; 9" ) =* >

E +
! ! '( ) !
%
"
% 5! 5!

! " 1 E +

%
+ ! !

!

! !
!

!
!5 % F
G

! 2 !D ! @
; !

=!,FF!
> #<$ : 9 =:<9>
!
!
!

;2 5 !!
E H

!! !D

5! ! 0 !!
!

+ % %
)
" !
2
! ) !!
! E &
!" ! "" C

!

0 E
I !

'(
!!, !
% =
>
C
5 !
0

!!
?
0

!

? =
> *
5 !

!
%

% !

; ! !
?

5!

)
& ! : &
J % ! G)E) -
'/(
! %
=*
< ; 2
J4B >

! !
* 5 ! K
? G L

=!,FF
>

! ! M
5!
!
!
2 !

!

! J

! < C*
I

!

!
&

J
!
!

! ! ! I

CG 3 )G =* />

2 % C : K!
! 0 ! I -

!
% ! &

! 9

B C
B
;
+ 0

4 ;

! J
- B

B !
! !
% %

!!!
2 5 ! " ! &
@
=& @> ? ! 1

! . ! !
! => - 4 ; !
& @ ! ! 4
& @+
%

<
2
!

9" % '3(

!

!
2 +

!!
%
!!
) 0 !

! J
5! & '6( !

I =%
>
!
!" !
! 5! ) 0

! 5! ; +2

% !

= ; G !D > )
%
5! 0
2 'M(
, !! ! F ! <
'7(
/J

! )
!
!
? !
0
!!
/ "#"
5
! !

#

.
! !$
'8( C !
0
! & !

4 !!

!
) '(
, #

% $ ) '( !

! % ! ! -
% % ' ( %
& * !!, #C %
0 2
$ '(
2
!D
-
!
& % ! !

!
! ! !
'/(
) '/( !
5!
"
&

0
!
'3(
& 5

'6M( ?

! =
> !

'7( 2 "

'8(
!
5 ! = D ! >
& !! 5

-
=!! >

& ! .

5 ! 5! '(
*

. ! '/ (,
J = >
< ! = . >
/ J ! = >
3 J = >
6 ; ! = . >
M N% = . >
7 9 = . >
8 2 !! = . >
:
= . >
*

" " $
-!! '/(
* '/(
* '6(
2 5 !! '//(
2 D ! '/3(
2 '/6(

D
2! ! '/M(
D !
2 ! '/7(
D !
2 ! '/8(
D !
2 D '/(
! !

4;; !!

4 ; ; 88 '3 (

1

- %
4;;
'3(
2
0

! 0 ! ! % <%

% "
- ! !
! <

'3( ) ! !
! ) 2 % & , #

5

5!$ '3/( 4 ! )
! 2
) % %
% 5!
'33(
5 '36(
&%
I
! 0 =# $ # $
& '( J 2> %
9 = 9; >

G & ! %

! '36( ! !

! 5 !
+ 9
! = ! >

! 9 0
%
! !
, 5 =
> 1
=
> =

>
=! !! > ,

&

9
! I % 5 .
!! '3M(
! E !

E % & '37( )
D )
E
! 0 E 2+
! J% =# $> 9

2
!

! % #
& E $ = #E & $>
! ) - ! %

& - ! %
=> J%
2

2
! 0 ! 2 9
& 2
! ! !
% .
I 2 !
& % &
# $
5!
) ! 5 I
2 E %

=

O O
>
!

) !
) %

=!> =0
>
! =
>
& ! 5!
=
>
5! 9
2
! 2
) !
9+ &
# 5!$

= > ! = !>
%
'38(,
=
>

=
>
=
>
= ! ! !
! >
&
!
!% " & ' "# ()" & '
< = > ? = >
= > ; = >
& = > < = >
2 = > * = >
? = > - = >
* =
>
=
>
J = > 4 =

>
2 = > =
>
& = > < = >
3 " $ ! "" ) ! 0

! =
>
2 !
! !! 2

5!

=
> ) )
!

! -

P
: 5!
G
! '3( - ! !

!
2 !
!!
.
!! '6 ( !! !

! )!
! 5!
E + 5! !
1

. ! '6( - ! 5!

? ?5! -
C? 3 )) !

5!
% '6( 9
5 !
=
!%

> & !
!
!
5 = ! >
=
% %
%>
! 5! =

!! % D ! > & ! !!

7 !
!

0
0 ! !!
!
!!

!

0
5

!
=&<> '6/(
6 %"" *+
!!
!

2 !

! 9 )

! 5!
!
2
!
!
!

!
! <
.
< =
!>
G 5!

5!
5!

* % ! !

5!
= 5!
!! ! " % > !

+,% % !!
;;&)& :?*)&;&: !D

<)) )J?) MF 8 % !
! !

23 %, * - !. * ! ! 0
$
4,
23 , +5
6 7"$ $
$ ! # *
, , ,-448
,
23 , 1, 1 " 9, , :
$ ; $< : * *

,
, ,888- ,
23 6, => ?

!
" ) *
, ,
,
283 ,@, ! " 6,0, A ;B

$ " * $
!5! 55 $
"
! "
%?:

*
,8 , C+D , ,

2-3 +, E, A! 9, = " ', A F ! *

*
,8 ,
,-4 ,
243 A, ! E, G! " #, 1 55
$$
7!
$
- *
, ,C1D , ,
23 ,, G! @,, " @,0, 0$
' $" =!+ 7
$$
!
*
" $ - *
, ,C1D ,8 ,
23 H, 1, 0
! A, H ! 1, " +, 1, B7
=, 1, '55 " +,

%$! $

$! $" "$ $$ & /!
! +F! , -,
2 3 9, 0 " @, +$ *$ $ $ * &
/!
! ,4444 ,
23 , +5
" #, + >$ # < 1
$ * 0 1 . *
) + ! 9, , :
$ " , 1, 1 C"$,D 6 '$$ ,

23 , + 1
! ! I :
! 2

# # , + C",D +$ )*$ '$$ ,
23 0, " , '!
, ' * " , G, 9 !$ > "< 6 7"$ !"
& 3 # + # + @ ,
23 , +5
1, + >$ 9, H , 0 9, + G, %,
" 9, > *
#+ *
,- ,4 ,4-8 ,
283 +,
" *
< : *$ 4 *
,- , ,
8,
2-3 1, 0, 1"$ 6 $ "
! " $
-
*
, ,44- ,
243 , " " 1, '* 0 $ *
$ !$ , , , < +

*
$ $
* & 5 ( ,-
,
23 , ;, 5" +$ $
>"$ $$ " 7 "! &
& *
, , ,- 4,
23 %, ;*$ 6 $ $$ * )
&
*
, , ,48 ,
2 3 %, 0 * 1 0 '$$ 7 A > ,
23 9, ', , 46 * 4 0 ,
23 1, #, %$ 4 . ! ) 0 '
,
23 , ;, 5" " +, ', 1> *
!5
" !
$ "$
$ " , 7$ " 9, , 0*
" C"$,D H
" '
$! ,
23 9, , :
$ : $ $ # + .

4 7 ;, 0"
> " , C"$,D
'> 1 111 '$$
,
283 G, 9$ 0 8 + *
, , , 8 ,
2-3 ', ;> 1 ! $ $ *
,- ,- ,
243 , +5
; " $
" $

*
,8 ,88 ,
23 G, ' & '$
! '$$ '
"
,
23 ', ;> " G, , :$ $$ $$
$ "
&
& *
,4 , , 4,
2 3 ', ;> 1 $ $J & ! *
, ,88 88 ,
23 +, G " , H 1

"*

) , ;, 5" " #, +, =F C"$,D 7 A > "! )*$ '$$
,
23 9, +, G$ ) )*$ ! '$$ ! ,
23 @, ?
" ', , 9 $ " 6 7"$ !* $
*
, ,8 4,
23 ', ;> G, , :$ " ', ;
$7 G ! $ "$ $ $*$

F"!
* J 7 ', ;> C",D 7 A > "!
)*$ '$$ ,
283 G, % !
# & + $ -,

2-3 #, '
> 1 !
$ *
. * !
9 : * #, '
> " #, @
C"$,D 7 A > 1"

,
243 , , 6 >$ 1K # @, #, " ', ;> C"$,D 0
$"

9 ;
,
23 , ;, 5" '
'$$ 7 A > 44,
23 , 1
" &
)*$ '$$ 7 A > - ,
2 3 1, ? H,
" 1,
$ * "! )*$

'$$ "! ,
23 , +> !! ? "
$ "" $ & 0"
9 . # + !

,
23 , , '
*$> !" $ ! "
!!
11 ,84 848 -,
23 , , +F * ; # %@ ' G
" ,
23 ,
" $ " 7$
$ & 0 * , ,
@$7 C",D > 7 1! '
$! 0 $ ,
283 @, 0!!$ 1
< #$ " $ > !$
#
# *
,-8 , ,8 4,
2-3 ,
" " ?, 'F

! & - * " + >$ +$
,
243 ,
" 6 $

$ !" 7 * #
, 4,
23 , ' * & #" ) + " +*
,
23 , %> , G> , ! A, #, , H " @, > # ! L

$
" *
$ &- <1 *
, , ,- ,
28 3 , " 0, 1
!
B $$ $ *
& 7 9, #$$
" 9, :"5%
$ C"$,D "! )*$ '$$
4,
283 , +5
4 ! 6 '$$ ,
283 , @5> 0, 7 A, > , = 0, 6> , #
, , 55 ', %

" 1, 6>$ $$ " $ " B $$
> $
& 3==5 ! # ?
" :1 -,
283 #, + $ " 1, 0 $
"7!$ * *
! $
$
<
! - *
, , , ,
c ISSN 1349-4198
MODELING, DESIGN AND EXPERIMENT OF IMPROVED

SUPER-MINI UNDERWATER ROBOT
Xin Song1 , Zao-Jian Zou2 and Jia-Wei Ye3

1
School of Naval Architecture, Ocean and Civil Engineering
2
State Key Laboratory of Ocean Engineering
Shanghai Jiao Tong University
No. 800, Dongchuan Rd., Shanghai 200240, P. R. China
songxsx@gmail.com
3
Department of Naval Architecture and Ocean Engineering
South China University of Technology
No. 381, Wushan Rd., Tianhe District, Guangzhou 510640, P. R. China
Received November 2009; revised March 2010
Abstract. To overcome the disadvantages appeared in the application of some super-

mini underwater robots, a simple improved design framework is proposed. A mechanical
improvement scheme is proposed. An easy handle, powerful and computer aided software
system including the closed-loop control algorithm is developed for solving the shortcom-
ings that are existed in the manual operation. The modeling of underwater robot plays
an important role in the closed-loop control design. The hydrodynamic derivatives are
estimated by the experiment based on the decoupled motions in dierent directions. A
new self-tuning PID algorithm is proposed for the controller design. The experimental
results in dierent environment demonstrate the performance.
Keywords: Modeling, Autonomous control, Underwater, ROV, Self-tuning
1. Introduction. Until now, many underwater robots including remotely operated vehi-
cle (ROV) and autonomous underwater vehicles (AUV) have been developed and applied
for detection, salvage, mine hunting, fishery study, etc. More recently, there has been a
trend to use the super-mini underwater robots in lakes and rivers. The implementation
of AUV is complex as the individual intelligent platform. AUV is generally expensive
and has not been popularly applied in the practical works. This paper mainly considers
the improvement of ROV. For the low cost and easy manual operation, ROV draw the
attention of many users. Recent results show that the super-mini underwater robots have
significant potential to carry out the tasks such as the target reconnaissance, inspection,
repair of underwater structures, etc.
However, due to the reply from the pilots, the shortcomings of some underwater robots
come out in practical tasks, such as the steady station-keeping near the dyke or bridge
piers. The drift, waves, etc always cause the troubles. The manual operation is often
impossible to catch up the time-varying disturbance, as the microoperation is uneasy to
handle by the pilots. In this paper, an improved design framework is proposed including
two parts. One improvement is focused in mechanical design. The other improvement in-
cludes the underwater robot deployed with integrated, intelligent, close-loop control plat-
form such as the implementation of the obstacles avoidance, the steady station-keeping,
etc. This close-loop control is consisted of a new self-tuning PID controller. The pilots
only need to give easy commands such as navigation, station-keeping or depth-keeping,
etc. Meanwhile, this computer aided control platform requires proper dynamic models.
1109
1110 X. SONG, Z.-J. ZOU AND J.-W. YE
Considering the previous studies [1-3], the dynamic models are crucial to the realization of
their prediction performance, precision autopilots. Although there are many works con-
cerning underwater robot, the modeling and control design remains very dicult. The
existed dynamics of underwater robots are well clear and understood, whereas they are
dicult to handle for practical design and control purposes. The problem includes many
nonlinearities and modeling uncertainties. In addition, unlike space or terrestrial envi-
ronment, the underwater one typically requires more restrictive conditions and sensor
limitations (e.g., weight, leaks, acoustic variability, corrosion), especially for this super-
mini type. In order to ensure the underwater robots accomplish the missions successfully,
they should be robust and reliable both in electrical and mechanical design.
Modeling is considered to be a general way in describing the corresponding physical
terms. The software and hardware configuration are in relation to the functional imple-
mentation. Simulation is generally processed for the dynamics and control algorithms.
This main purpose is to accurately replicate the actual robots response, subject to a vari-
ety of disturbances, such as waves, drift, etc. This is suitable for the theoretical analysis
of the dynamics and motion controller design. In general, modeling techniques tend to
consist of two ways [4]. One is the predictive methods based on either computational fluid
dynamics [5,6]. The other is the experimental techniques including the common identifi-
cation process [3] or the intelligent ways such as the neural network method [7,8], etc. The
general rule for modeling and simulation is to closely match the corresponding physical
system. In this paper, the simulation is directly connected to the system hardware and
software and the control algorithm is applied and tested in real application.
In this paper, the importance of this work is to solve the shortcomings of some previous
underwater robot. The main objective is to summarize the research on the mechanical
design improvement and the configuration of the software architecture including the mod-
eling, data processing, control algorithm, etc. The detailed modeling steps are presented.
A drift disturbance is considered in the experiment, which is always ignored in the previ-
ous works. A new self-tuning PID controller is proposed and compared with other control
algorithms in the same working environment. The experimental results demonstrate its
performance. The rest of this paper is organized as follows: Section 2 develops the me-
chanical improvement work; Section 3 develops the framework of the software; Section
4 describes the modeling and parameter identification technologies; Section 5 presents a
new self-tuning PID control algorithm and the constraints within this method. In Section
6, the results of the experiment in the pool with and without time-varying drift distur-
bance are presented and discussed. The future works are pointed out. The final section
provides concluding remarks.
2. Improvement of the Mechanical Structure. The improvement design is aimed to

solve the problems appeared in the application. The main disadvantages can come into
two categories: 1) the shortcomings in mechanical design; 2) the manual operation is not
enough and the computer aided software system is not provided.
In order to lower the risk from this whole closure design, the supporting frame and
anti-crash metal frame are added to prevent the super-mini underwater robot from the
accidental collisions. Some special technologies are applied to the inner devices protection
and the preventability of the magnetic interference. Moreover, some underwater robots are
equipped with two bilateral forwarding propellers and one propeller for heave motion (see
Figure 1). Based on the theory study, they can realize the surge, yaw motion. Through
the adjustment of the yaw angle, they can realize the motion in every direction. In the
environment with the negligible disturbance, this mode of propulsion is suitable though
the radius of gyration is large. However, when the underwater robot is working near dykes
MODELING, DESIGN AND EXPERIMENT OF UNDERWATER ROBOT 1111
or bridges, etc, the shortcomings become serious and they are impossible to accomplish
the tasks. A simple improvement is to add two lateral propellers (see Figure 2).
Figure 1. Original propul- Figure 2. Improved propul-

sion mode sion mode
The aim of these two propellers is to solve the problems caused from the terrible radius
of gyration due to the complex waves or current disturbance and enhance the propulsion
capabilities, the station-keeping inspection abilities. For example, when the underwater
robot is expected to survey the corrosion situation of the underwater building, it should
keep the same yawing angle and steady posture, as the camera is located in the front part
of the body. Thus, the bilateral forwarding propellers need to resist the drift disturbance.
Unfortunately, the force is obviously not enough for the station keeping. With the im-
proved forwarding propulsion mode, more practical propulsion choices are provided. The
experimental results demonstrate the eectiveness.
3. Software Architecture. For the second shortcoming mentioned in part 2, the com-
puter aided close-loop control software with easy and visual manipulator interface is pro-
posed and developed. Some previous underwater robots are operated with the manual
operation box which includes the power circuit, joystick circuit, depth control switch,
light, camera, LCD, other feedback lights, etc. Though there are a number of devices, the
open control framework remains not enough to help the pilots to master the underwater
robot well. If the underwater robot is cruising in the long distance, it may work well
with the manual operation box. However, if it moves in a relatively small area with some
obstacles avoidance [9] or station-keeping movement, the manipulator should be very pro-
fessional to control it with the operation box through the LCD images and some sensors
feedback data. Generally, it is not applicable for the common user who only needs an
easy handling device to replace the divers.
In order to solve the weakness shown in the application, a closed-loop control software
is developed with respect to integrate the complex manual operations into simple com-
mands such as the station-keeping command, depth holding command, etc. This software
is aimed to support complex tasks. The high-level computer aided system is built upon
a global shared memory that allows the underwater robot to deal with the periodic tasks
and regular tasks. Through the communication with manual operation box, dierent tasks
can be processed in high-level computer software and the commands can be transmitted
into the underwater robot directly. Meanwhile, the pilot can easily choose the intelligent
self-control methods or the semi-manual operation through the human-computer interac-
tion techniques. In this research, the high-level control software is implemented based on
LabVIEW platform, which is a visual programming language from National Instruments
Company. This whole control software consists of several parts including station-keeping
module, sensor data processing module, depth module, etc. The detailed software archi-
tecture is shown in Figure 3.
Figure 3. Software architecture
4. Modeling and Parameter Identification. Underwater robot model is conveniently

represented using nonlinear series of first order dierential equations of motion. Gener-
ally, six degrees of freedom are defined as surge, sway, heave, roll, pitch and yaw. The
globally inertial referenced frame comprises six positions and attitude components. The
mathematical model of an underwater robot can be expressed with respect to a globally
inertial referenced frame, by the nonlinear equation of motion in matrix form:
M ()
+ C(, ) + g() = J T (),
+ D(, ) (1)
= J()v (2)
where = [x, y, z, , , ]T is the position and attitude vector; M () R66 is the
R6 is the centripetal and coriolis
inertial matrix including adding inertial part; C(, )
R is the damping term including the quadratic and linear drag terms;
term; D(, ) 6
g() R is the hydrostatic restoring force terms; J() R66 is the transform matrix
6
between the body-fixed reference and the earth-fixed reference; R61 is the thruster
input vector. More detailed models of the underwater dynamics are reported in [3,10,11].
In these papers, the moments of inertia Ixx , Iyy , Izz , etc are introduced.
The external forces and moments such as the mentioned hydrodynamic drag force,
thruster force, hydrodynamic added mass forces, etc can be described in terms of the
corresponding hydrodynamic coecients. These coecients can be found in the reference
[3]. As the underwater robot in this research is symmetric about x-z plane, close to
symmetric about y-z plane, the motions in surge, sway, pitch and yaw are assumed to
be decoupled [3]. Due to the low speed of the underwater robot, the surge and heave
motions are decoupled as the coupling eects can be negligible though the x-y plane is
not symmetric. Many experiments verified this assumption.
The inertia matrix includes the mass and moments of inertia. In this research, the
common oscillation experiment with a small swing angle around the principle axis has
been performed to determine the moment of inertia Ixx , Iyy , Izz (See Table 1). Generally,
the other hydrodynamic derivatives are calculated experimentally in towing tank tests
or in flumes with controlled flowing water. The hydrodynamic forces are recorded and
the data is sampled by the data acquisition system under the condition of the controlled
flowing water.
Table 1. Moment of inertia

Value Units
Ixx 0.02512 kg.m2
Iyy 0.02687 kg.m2
Izz 0.02875 kg.m2
The thruster input consists of the forces and moments acting on the underwater robot.
Forward and backward thrusters are generally non-symmetrical [12], unless a special de-
sign is implemented. Considering the thrusters energy balance of the control volume, the
simplified nonlinear equations are given as follows:
n = m an |n| , (3)
T = Ct n |n| , (4)
where n is the rotational motor speed vector, m is the input torque provided by the
thrusters motor, , a and Ct are thruster constants, T is the thrust force. In this research,
there are five thrusters in the underwater robot. Each one has the individual driver that
controls the rotational speed. Since the propeller diameter, mass and the motors are
small, the speed of the thruster in Equation (3) is much faster than the underwater robot.
So the dynamics are neglected. Only the parameter Ct needs to be identified. Here, the
least square method is applied to compute the coecients for the dierent thrusters. The
results are shown in Table 2.
Table 2. Thruster coecients
Ct (N )
Thruster
Forward Backward
4
Horizontal 2.6210 10 1.2310 104
Lateral 3.2774 104 1.7314 104
4
Vertical 1.1532 10 0.8134 104
Following the traditional identification methods, a series of experiments have been

done to get the drag coecients, etc. Some coecients that cannot be estimated by the
experiment are estimated by the theoretical calculation. Note that the linear drag is much
smaller than quadratic drag. Therefore, the linear drag can be neglected [3]. Here, some
coecients are listed in Table 3.
Table 3. Coecients
Added mass
Xu 1.8357 Yv 6.1483
Zw 3.5402 Kp 0.0285
Mq 0.0153 Nr 0.0314
Quadratic drag coecients
Xu|u| 7.2056 Yv|v| 27.671
Zw|w| 25.260 Kp|p| 0.0032
Mq|q| 0.0073 Nr|r| 0.0095
5. Close-loop Control Algorithm. The control algorithms of robot include many

parts. Many papers present the corresponding control algorithm such as the PD, PID
method [10], fuzzy logic [14,16], neural network [17] and sliding mode method [13,15,18,19],
etc. Since the implementation of PID control method is easy and eective, PID control
algorithm is applied in this research. Meanwhile, in this experiment, this control algo-
rithm is applied in the realization of keeping the yaw angle. The original super-mini
underwater robot in this research chooses the open control framework. When the outside
disturbance is complex, the motion can hardly be well controlled. In order to satisfy
the actual work requirement, the yaw angle keeping algorithm is applied to enhance the
super-mini underwater robot to work normally in the path keeping or the observation of
the fixed target, etc. In this research, the yaw angle keeping algorithm is developed by
a new self-tuning PID control algorithm and the dynamic motion models. In order to
testify this algorithm, the underwater robot is closely observing some fixed targets with
and without time-varying flow disturbance.
With respect to the dynamic model, the following motivated assumptions are given:
1) The underwater robot moves at low velocity; 2) The damping force is non-coupling;
3) The linear damping force is negligible; 4) The gravity and the buoyancy are nearly
in balance; 5) Water kinetic coecient is constant; 6) In the parts of the roll and pitch
angle, the system is stable, and therefore the hydrodynamic equations are not considered
in these directions.
Refer to the definition of motion Equation (1), the motion parameters of the equations
can be obtained with the following characteristics:
C.1 Inertia matrix is symmetric and positive, defined as M () = M T () R6 . The
positive constants mm and mM compose mm M () mM ;
C.2 M () 2C(, )
is a skew symmetric matrix;
+ C T (, );
C.3 M () = C(, )
C.4 Damping matrix D(, ) is strictly positive and within limited boundary;
is in limited boundary and is linear, as x R6 . C(, )
C.5 Coriolis matrix C(, ), =
C(, x) always exists. There is positive number kc that keeps C(, ) kc .
In general, the underwater robot simulation neglects the outside wave interference and
considers the impact of water flow. The water density is assumed constant and the sand
stone eects are ignored. Meanwhile, the vertical inclination angle is assumed less than
/2 in this experiment. Since the kinetic knowledge of the underwater robot is limited, the
design for precise control and navigation is very dicult. Therefore, the dynamic models of
the underwater robot for the system simulation and control system kinematics are helpful.
The parameters of underwater robot dynamic equation are non-linear, coupled, time-
varying and unstable. Thus, it is dicult to obtain accurate hydrodynamic equations.
On the other hand, if there is the interference of drift, etc, it is also dicult to get the
accurate model.
Here, equation = Q()v is introduced to express the unknown constant forces of the
sensors, etc, where v = [vB vSx vSy vSz ]T . v is an unknown constant parameter vector, vB
is an unknown buoyancy and the three parameters vSx , vSy , vSz represent the unknown
constant force of the sensors. Q() R64 is a regression matrix, R66 is a strictly
positive matrix related to the adaptation rate. Assuming that the underwater robot is
moving near the target and only yaw angle is changing, the unknown buoyancy and the
forces from the fixed sensors are constant. The equation of motion is presented as:
M ()
+ C(, ) + g() + = J T ().
+ D(, ) (5)
The control model is shown as follows:

[ ]
= J () KP KI (t)dtKD + g() + Q()w ,
T
(6)
where KP , KI , KD R66 is the symmetric positive definite matrix, = d (d

is predefined position) is the position error. Supposing that v is the estimate of v, the
adaptive self-tuning parts are given as follow:
w = LQT ()( + b
+ s), (7)
s = , (8)
where L R44 is the positive matrix, b and are the positive constants.
Theorem 5.1. Considering the control system (5) combines with the control law (6) and
the adaptation law (7) (8), if the condition as follows:
KIM b
KDm + Dm (KDM DM CM 2MM ) (MM + CM + KDM + DM ) > 0, (9)
2 2 2
b b b b
bKP m KDM KIM DM CM MM KP M > 0, (10)
2 2 2 2 2 2
KIM b
KIm KIM CM KP M KDM DM > 0, (11)
2 2 2 2 2 2
are satisfied, then the system is globally asymptotically stable except the singular point
= /2.
Proof: Considering the Lyapunov function

1 1
T M () + sT M () + w T L1 w,
V = T M () + T KP + b (12)
2 2 2
with w
=w w. Dierentiating V and substituting dynamics Equation (5) and control
model (6) into V yields
1
V = T [KP KD KI s D(, ) C(, ) + T M () + T KP
]
2
+ b [KP KD KI s D(, )
T
C(, ) T M () (13)
+ b T M () + b
]
+ s T M () + sT M () + sT [Kp KD KI s D(, ) C(, ) ]

+ s)T Q()w + w T L1 w.
+ ( + b
By using character C.2, C.3 and the Equation (7), one can get
V = T KD T KI s T D(, )
b
T KP b
T KD b
T KI s
b
T D(, ) T C T (, )
+ b sT KP
+ b T M () + s T C T (, ) (14)
sT KD sT KI s sT D(, )
.

Considering the following inequalities holds:

1 1 1
T KI s ( T KI + sT KI s) KIM
2 + KIM s2 ,
2 2 2
..
.

sT D(, ) [DM s2 + DM
[sT DM s + T DM ] 2]
2 2
Then, one obtains:

KIM b
V (KDm + Dm (KDM DM CM 2MM )
2 2

(MM + CM + KDM + DM ) 2
2
b b b b (15)
(bKP m KDM KIM DM CM MM KP M ) 2
2 2 2 2 2 2
KIM b
(KIm KIM CM KP M KDM DM ) s2
2 2 2 2 2 2
Finally, considering conditions (9) (10) (11), V is negative and vanishes if and only if
(,
, s) = (0, 0, 0). The theorem is proved.
6. Experiment of Super-mini Underwater Robot. The super-mini underwater ro-

bot experiment was carried out in the pool of South China University of Technology. The
experiment situation is presented in Figure 4. The parameters of this super-mini robot
in this research are given as follows: The size is 31 cm 22.5 cm 21 cm; The weight in
the air is 60 kg; The max navigation speed is 2.6 knot; The max diving depth is 152 cm.
Figure 4. Experiment of the super-mini underwater robot
Considering the underwater robot is working closely to the target, the dimension of the
controller is reduced to four. The parameters KP , KI , KD are defined as follows:
KP = diag(10, 10, 100, 5), KI = diag(1, 1, 2, 0.4) and KD = diag(2, 2, 5, 0.04).
Let L = 12I, b = 0.62, = 0.3, these parameters satisfy the constraint condi-
tion. The underwater robot is moving from the beginning position [x0 , y0 , z0 , 0 ]T =
[0.1, 0.1, 0.75, 100]T to the destination position [x, y, z, ]T = [0, 0, 0.75, 20]T .
Introducing the equation w(t + T ) = w(t)
+ w(t)T , where w(0)
= 0. The adaptive
self-tuning parameter w is calculated by Euler equation. In the experiment, the yaw angle
command is changed from 100 degrees 56 degrees 20 degrees.
To evaluate performance of the proposed control system, three types of experiments
have been done with respect to the algorithms including traditional PID controller, new
self-tuning PID controller and sliding mode control (SMC) with smooth control. SMC
controller is presented in [19] and the parameters are defined as = 0.75, = 9.2.
Each type of experiment includes two dierent situations. One chooses the experiment
without drift disturbance. The other one considers the drift disturbance. The time-
varying drift speed is predefined as VDrift = 0.15(sin(4t)+1)m/s and the wave disturbance
is ignored. The yaw angle results in the experiment are recorded, shown in Figure 5 and
Figure 6. Figure 5 shows the experimental results between self-tuning PID and PID.
Figure 6 shows the results between self-tuning PID and SMC with smooth control.
Figure 5. Controller results comparison between self-tuning PID and PID
Figure 6. Controller results comparison between self-tuning PID and slid-

ing mode controller with smooth control
From the results without drift disturbance, all controllers can reach relatively steady
status. The stable time grows up when the command angle reduces. At 100 degrees
and 56 degrees, these three algorithms can control the yaw angle to the command angle.
However, at 20 degrees, the underwater robot cannot keep the posture at command angle.
It swings in a short range and becomes relatively steady. In the experiment with drift
disturbance, the tendency of yaw angle is similar to the former results. However, the
vibration in this experiment is larger as the angle errors grow up and the stable time
becomes longer.
The weight of this super-mini underwater robot is small. The disturbance from the
cables and others are relatively large. When the yaw angle is at 180 degrees, the cable
is rightly behind the underwater robot. Thus, the smallest disturbance is attached to
the yaw motion. However, when the underwater robot reduces angle, its front part is
moving toward the cable. The angle errors increase and the vibration becomes bigger as
the resistance and its moment of the cable grow up. When the command angle is at 20
degrees, the disturbances are beyond the controllers abilities of regulation. Thus, the
underwater robot keeps a small range of swing in a relatively stable state, rather than
the other two situations that the underwater robot can finally stabilize and keep the yaw
angle after a period of adjustment.
From the results, comparing with other two algorithms, the new self-tuning PID control
algorithm shows obvious advantages as the errors are smallest and the stable times are
shortest. The results of SMC with smooth control are better than PID control. Though
the self-tuning algorithm cannot keep the angle at 20 degrees, the vibrations and the angle
errors are the smallest. Thus, the robustness is the best among these algorithms. These
results can demonstrate the performance and eectiveness of this new algorithm.
Comparing with the original underwater robot, the whole improvement scheme can
largely enhance its working abilities. The operation becomes easy and the steady obser-
vation can be realized quickly. This new design can help the underwater robot meet higher
working requirement, such as the observation task in these experiments. Unfortunately,
the original design can hardly satisfies them.
For the shortcomings in current work, the mathematic models of the cable are impor-
tant. A prediction and compensation algorithm of the cable and other disturbance are
necessary to be considered in future. A fast station-keeping algorithm and the param-
eters optimization algorithm should also be further studied considering more complex
nonlinear disturbance.
7. Conclusion. In this paper, the modeling, design, control algorithms and data process-
ing algorithm of super-mini underwater robot are discussed in detail. The main advantages
of these new improvements are summarized in two categories. One is the mechanical part.
The other is the software part.
An improvement scheme of mechanical structure is proposed. The main features are to
reduce the radius of gyration and increase the capabilities of anti-risk. This design can
make the underwater robot become more applicable in the complex environment. A close-
loop control software is proposed. The easy commands that are instead of the complex
manual operations reduce the risk of low response of pilots and help the underwater robot
easy to handle. A new self-tuning PID controller is proposed for the dynamic yaw angle
control. This new algorithm is aimed to estimate the unknown constant forces, etc. This
control algorithm also realizes a simple disturbance estimation and compensation. The
way of measurement and calculation of hydrodynamic coecients, etc are presented. The
results of coecients in this research are given.
To evaluate the eectiveness and feasibility of the improvement scheme, several ex-
periments with and without the time-varying drift disturbance have been carried out.
The experimental results are discussed and also demonstrate the performance of the new
control algorithm. Finally, shortcomings in current work and further research works are
pointed out.
Acknowledgment. This work is partially supported by the National Natural Science
Foundation of China (Grant No. 50779033) and National 863 High-Tech R & D Program
of China (Grant No. 2007AA11Z250). The authors also gratefully acknowledge the helpful
comments and suggestions of the reviewers, which have improved the presentation.
REFERENCES
[1] J. Yuh, Learning control for underwater robotic vehicles, IEEE Control Systems Magazine, vol.14,
no.2, pp.39-46, 1994.
[2] J. Yuh, Underwater Robotic Vehicles Design and Control, TSI Press, Albuquerque, 1995.
[3] T. I. Fossen, Guidance and Control of Ocean Vehicles, John Wiley & Sons, 1994.
[4] K. Goheen, Techniques for URV modeling, in Underwater Robotic Vehicles Design and Control, J.
Yuh (ed.), Albuquerque, TSI Press, 1995.
[5] H. Suzuki, N. Kato and T. Katayama, Motion simulation of an underwater vehicle with mechan-
ical pectoral fins using a CFD-based motion simulator, Proc. of the International Symposium on
Underwater Technology, Tokyo, Japan, pp.384-390, 2007.
[6] T. Sunlin, N. Takeshi, T. Blair et al., Estimation of the hydrodynamic coecients of the complex-
shaped autonomous underwater vehicle TUNA-SAND, Journal of Marine Science and Technology,
vol.14, no.3, pp.373-386, 2009.
[7] P. V. D. Ven, T. Johansen, A. Sorensen et al., Neural network augmented identification of underwater
vehicle models, Control Engineering Practice, vol.15, no.6, pp.715-725, 2007.
[8] P. V. D. Ven, C. Flanagan and D. Toal, Identification of underwater vehicle dynamics with neural
networks, Proc. of Conference on OCEANS 04 MTS/IEEE, vol.3, pp.1198-1204, 2004.
[9] L. H. Jiang, M. C. Deng and A. Inoue, Obstacle avoidance and motion control of a two wheeled
mobile robot using SVR technique, International Journal of Innovative Computing, Information and
Control, vol.5, no.2, pp.253-262, 2009.
[10] D. A. Smallwood and L. W. Louis, Model-based dynamic positioning of underwater robotic vehicles:
Theory and experiment, IEEE Journal of Oceanic Engineering, vol.29, no.1, pp.169-186, 2004.
[11] T. Fossen, Marine Control Systems Guidance Navigation and Control of Ships Rigs and Underwater
Vehicles, Marine Cybernetics, Norway, 2002.
[12] J. N. Newman, Marine Hydrodynamics, MIT Press, 1977.
[13] X. Song, J. Ye and W. L. Min, Application of the integral sliding mode controller with fuzzy logic
to submersible vehicle, International Journal of Innovative Computing, Information and Control,
vol.3, no.4, pp.897-906, 2007.
[14] Y. H. Cheng, X. S. Wang and R. H. Lei, A fuzzy control system for path following of mobile robots,
[15] J. Guo, F. C. Chiu and C. C. Huang, Design of a sliding mode fuzzy controller for the guidance and
control of an autonomous underwater vehicle, Ocean Engineering, vol.30, no.16, pp.2137-2155, 2003.
[16] M. M. Fateh and A. Azarfar, Improving fuzzy control of robot manipulators by increasing scaling
factors, ICIC Express Letters, vol.3, no.3(A), pp.513-518, 2009.
[17] G. Capi, Application of recurrent neural controllers for robot complex task performance, Interna-
tional Journal of Innovative Computing, Information and Control, vol.5, no.5, pp.1171-1178, 2009.
[18] W. M. Bessa, M. S. Dutra and E. Kreuzer, Thruster dynamics compensation for the positioning
of underwater robotic vehicles through a fuzzy sliding mode based approach, Proc. of the 18th
International Congress of Mechanical Engineering, Ouro Preto, Brazil, 2005.
[19] T. Chatchanayuenyong and M. Parnickkun, Neural network based-time optimal sliding mode control
for an autonomous underwater robot, Mechatronics, vol.16, no.8, pp.471-478, 2005.
c ISSN 1349-4198
DESIGN AND APPROXIMATION CAPABILITIES ANALYSIS OF

TIME-VARIANT FUZZY SYSTEMS
Degang Wang1 , Wenyan Song2 and Hongxing Li1

1
Faculty of Electronic Information and Electrical Engineering
Dalian University of Technology
No. 2, Linggong Road, Ganjingzi Dist., Dalian 116023, P. R. China
wdg0621@sina.com
2
Department of Quantitative Economics
Dongbei University of Finance and Economics
No. 217, Jianshan St., Shahekou Dist., Dalian 116025, P. R. China
agathasong@163.com
Received November 2009; revised May 2010
Abstract. In this paper, the design and analysis for a class of time-variant fuzzy sys-
tems are investigated. Firstly, a novel modeling method for time-variant fuzzy system,
called variable weighted interpolating modeling (VWIM) method is proposed. It is pointed
out that the time-variant fuzzy systems constructed by VWIM method can be represented
by some interpolation functions. Then, VWIM method is applied to the nonlinear dy-
namic systems modeling. It is proved that time-variant fuzzy systems based on VWIM
method are universal approximators to a class of nonlinear systems. And, the approx-
imation error bounds for various classes of time-variant fuzzy systems are established.
Finally, a simulation example is provided to demonstrate how to utilize a time-variant
fuzzy system to approximate a given nonlinear system with arbitrary precision.
Keywords: Time-variant fuzzy system, Variable weighted interpolating modeling method,
Universal approximators, Nonlinear systems
1. Introduction. In face of some complex systems with fuzziness and linguistic variable,
such as dynamic systems, control systems and economic systems, mastering the mathe-
matical model is the prerequisite for the theoretical analysis of the system [1-8]. Fuzzy
system theory, which is introduced by L. A. Zadeh in [9], is a useful tool for system mod-
eling when the exact model is unknown. It is well known that fuzzy system consists of
four principle components: fuzzifier, fuzzy rules, fuzzy inference engine and defuzzifier.
From the viewpoint of mathematics, the main object to construct a fuzzy system is to
approximate a desired model or function within a given level of accuracy.
The existing research on the approximation of fuzzy systems can be classified into two
aspects: the qualitative aspect and the quantitative aspect. On the qualitative aspect,
the authors mainly investigate the universal approximation properties of various classes
of fuzzy systems [10-14] and the theoretical foundation of fuzzy system modeling [15-17].
On the quantitative aspect, the approximation error bounds of TS fuzzy systems and
Mamdani fuzzy systems are established in [13,14,18]. Besides, the perturbation error
bounds of various classes of fuzzy systems are deduced in [19].
So far, the majority of existing results about approximation theory of fuzzy systems are
only suitable to the time-invariant fuzzy systems. They can guarantee their abilities for
approximating a wide class of functions. It should be noted that many practical systems
operate in environment with time-varying characteristics and fuzziness. Most of them are
represented as dierential equations. Naturally, an important question followed is how to
1121
1122 D. WANG, W. SONG AND H. LI
design a fuzzy system to approximate a given nonlinear system with arbitrary accuracy.
The research on this question is essential to the stability analysis and controller design
for fuzzy systems. Until now, only a few results related to this subject can be found
in [2,20,21]. The question given a nonlinear system, how many rules are needed to
guarantee the desired approximation accuracy still remains unanswered. Besides, the
fuzzy rule bases obtained by data or experts knowledge are usually represented as the
discrete-time information, that is, the corresponding rule bases are sparse with respect to
time universe. If an input occurs in an empty space between two times, then no rule will
be fired, thus we can not compute the corresponding output by fuzzy modeling method.
To tackle the question of fuzzy reasoning with sparse rule bases, interpolated reasoning
methods are introduced in [22-25]. However, how to utilize discrete-time fuzzy rule bases
to depict the dynamic behavior of the system is still a question which needs to be studied.
Above facts bring about the motivation of integrating interpolation method and fuzzy
system theory to deal with the design and analysis of time-variant fuzzy systems. In this
paper, we concentrate on the modeling method for time-variant fuzzy systems and the
approximation of a class of nonlinear systems by them. The paper is organized as follows.
In Section 2, we introduce some preliminaries knowledge. In Section 3, we propose a novel
modeling method for time-variant fuzzy systems, called variable weighted interpolating
modeling (VWIM) method. In Section 4, we analyze the approximation accuracy of time-
variant fuzzy systems constructed by VWIM method to a class of nonlinear systems. In
Section 5, we provide a simulation example to demonstrate the validity of the VWIM
method. In Section 6, we give a few concluding remarks.
2. Problem Statement and Preliminaries. In this section, we will introduce several

concepts which are used in this paper and two famous fuzzy systems: Mamdani fuzzy
systems and TS fuzzy systems.

Definition 2.1. Given a universe X, let A = {Ai }1in be a family of normal fuzzy sets
on X and let xi be the peak point of Ai . A is called a fuzzy
n partition of X, if it satisfies
the conditions ((i, j)) (i = j xi = xj ) and (x X)( i=1 Ai (x) = 1).
In [26], if {Ai }1in is the fuzzy partition of X, then the set of functions = {Ai (x), i =
1, , n} is said to be a normal basis set.
Definition 2.2. [26] For a given function f : X R and a normal basis set =

{Ai (x), i = 0, , n}, the function f: X R defined as y = f(x) = ni=0 f (xi )Ai (x),
is called a fuzzy interpolation of f over , and xi is the peak point for Ai (x).
Definition 2.3. The mappings r (r = 1, 2) from 2[a, b] to [0, 1] are variable weights if the
following conditions hold: (a) for any t [a, b], r=1 r (t) = 1; (b) 1 (a) = 1, 2 (b) = 1.
p1
Example 2.1. Let U = [tk , tk+1 ], then the following mappings are respectively variable
k=1
weights on [tk , tk+1 ] (k = 1, , p 1):
tk+1 t t tk
(i) 1k (t) = , 2k (t) = ;
tk+1 tk tk+1 tk
(tk+1 t) (tk+1 t)
(ii) 1k (t) = sin2 , 2k (t) = cos2 .
2(tk+1 tk ) 2(tk+1 tk )
In the following, we take multi-input-single-output (MISO) fuzzy system as an example
to introduce the general expressions of Mamdani fuzzy system and TS fuzzy system. From
the viewpoint of mathematics, fuzzy system can be regarded as a mapping from input
universe X1 Xn to output universe Y , i.e., f : X1 Xn Y .
TIME-VARIANT FUZZY SYSTEMS 1123
The fuzzy rules of Mamdani fuzzy systems are formed as follows:

If x1 is A1j1 and x2 is A2j2 and and xn is Anjn then y is Bj1 jn , (1)
where xi (i = 1, , n) are the input variables of fuzzy systems, y is the output variable
and Aiji F (Xi ), Bj1 jn F (Y ) (ji = 1, , pi ; i = 1, , n).
For a given input x = (x1 , , xn ) X1 Xn , the Mamdani fuzzy system
which is determined by rules (1), singleton fuzzifier, CRI method and centroid defuzzifier
technology can be expressed as
( ( ))
p1
pn p1 pn n
Aili (xi ), Bl1 ln (yj1 jn ) yj1 jn
j1 =1 jn =1 j1 =1 jn =1 i=1
f (x) = ( ( )) .
p1
pn p1 pn n
Aili (xi ), Bl1 ln (yj1 jn )
j1 =1 jn =1 j1 =1 jn =1 i=1
In [18] when both t-norm and implication operator are the product operator, CRI
method is called product inference engine. When they are the min operator, CRI method
is called min inference engine.
Dierent from Mamdani fuzzy systems, the fuzzy rule bases of TS fuzzy systems with
linear consequent are given by:

n
If x1 is A1j1 and and xn is Anjn then yj1 jn = cij1 jn xi , (2)
i=0
where j1 = 1, , p1 ; ; jn = 1, , pn , x0 1.
The TS fuzzy systems with singleton fuzzifier, rules (2), product inference engine and
centroid defuzzifiier can be expressed as following,
(n (n ))

p1
pn
Aiji (xi ) cij1 jn xi
j1 =1 jn =1 i=1 i=0
f (x) = (n ) .

p1 pn
Aiji (xi )
j1 =1 jn =1 i=1
3. VWIM Method for Time-Variant Fuzzy Systems. In this section, we also take
MISO time-variant system as an example to introduce a novel modeling method for time-
variant fuzzy system. Figure 1 shows an n-input-single-output time-variant fuzzy system.
(x1 , , xn , t) - (y, t) -
S
Figure 1. n-input-single-output time-variant fuzzy system
The time-variant system S can be regarded as a mapping from input universe X1

Xn U to output universe Y , which is denoted by F , i.e., F : X1 Xn U Y ,
(x, t) = (x1 , , xn , t) 7 F (x, t), where (x, t) and F (x, t) are the input value and output
value of system at time t respectively. For convenience, we take Xi , Y and U as real
number intervals respectively, where Xi = [ai , bi ] (i = 1, , n), Y = [c, d] and U = [0, T ].
Now, we introduce some basic steps in time-variant fuzzy system modeling.
Step 1. Construct fuzzy rule bases. Firstly, we divide time universe U = [0, T ] as:
0 = t1 < < tp = T . With respect to each index k (k = 1, , p), we make a
partition of interval [ai , bi ]: ai = xki1 < < xkipik = bi , i = 1, , n. Then, we
construct normal fuzzy set Akij (k) F([ai , bi ]) such that xkij (k) is the peak point of
i i
Akij (k) and {Akij (k) }1j (k) pi is a fuzzy partition of [ai , bi ]. Furthermore, through experi-
i i i k
ment or observation method, we can determine the corresponding element ykj (k) jn(k) Y
( ) ( ) 1

with respect to the data xkj (k) jn(k) , tk = xk1j (k) , , xknjn(k) , tk . Similarly, we take
1 1
ykj (k) jn(k) as the peak point to determine the normal fuzzy set Bkj (k) jn(k) F (Y ) such
1 { } 1
that Bkj (k) jn(k) (k) (k)

is a fuzzy partition of Y .
1 1j1 p1k ; ;1jn pnk
Accordingly, a group of discrete-time fuzzy rule bases can be formed as follows:
If x1 is Ak1j (k) and and xn is Aknjn(k) then y is Bkj (k) jn(k) , (3)
1 1
(k)
ji = 1, , pik ; i = 1, , n; k = 1, , p.
Step 2. For any given input (x, t) X1 Xn U . Assume that t [tk , tk+1 ), then
by fuzzy system modeling method, for instance Mamdani fuzzy systems and TS fuzzy
systems, we can compute the output values at time tk and time tk+1 , which are denoted
as Fk (x, tk ) and Fk+1 (x, tk+1 ) respectively.
Step 3. Take weighed sum of Fk (x, tk ) and Fk+1 (x, tk+1 ) to determine the output value
of system at time t, i.e., F (x, t) = 1k (t) Fk (x, tk ) + 2k (t) Fk+1 (x, tk+1 ).
Step 4. Using the characteristic function, the time-variant fuzzy system can be ex-
pressed as

p
F (x, t) = (1k (t) Fk (x, tk ) + 2k (t) Fk+1 (x, tk+1 )) [tk ,tk+1 ) (t), (4)
k=1

where Fk denotes the fuzzy system at time tk (k = 1, , p) and [tp , tp+1 ) = {T }.
In this way, we use the variable weight interpolation technology and fuzzy system
modeling method to obtain the mathematical representation of time-variant fuzzy system.
In this paper, such a modeling method is called variable weight interpolation modeling
(VWIM) method.
Remark 3.1. By Steps 1 4, we find that VWIM method can transfer a group of discrete-
time rule bases into a continuous dynamic model. And, the properties of the corresponding
time-variant fuzzy systems depend on two factors: variable weights and the fuzzy systems
at every sample time.
Naturally, we may propose such a question: how to choose variable weight and time-
invariant fuzzy system in practical application? Next, we will answer this question.
Proposition 3.1. Suppose that the time-invariant fuzzy systems are Mamdani fuzzy sys-
tems with min inference engine, then the time-variant fuzzy system determined by rules
(3) and VWIM method can be expressed as
pnk ( )

p1k
n
A (k) (xi ) ykj (k) jn(k)
i=1 kiji
p
(k) (k) 1
j1 =1 jn =1
F (x, t) = 1k (t)
p1k
n
pnk
k=1 ( Akij (k) (xi ))
(k) (k) i=1 i
j1 =1 jn =1 (5)
p1 pn ( )
k+1 k+1 n
A (k+1) (xi ) y

(k+1) (k+1)
(k+1) (k+1) i=1 (k+1)ij (k+1)j1 jn
i
+2k (t)
j1 =1 jn =1
[t ,t ) (t).
p1

k+1
pn

k+1
(
n
) k k+1
A (k+1) (xi )
(k+1) (k+1) i=1 (k+1)ij
i
j1 =1 jn =1
Proposition 3.2. Suppose that the time-invariant fuzzy systems are Mamdani fuzzy sys-
tems with product inference engine, then the time-variant fuzzy system determined by rules
( pnk ( n )
p

p1k

F (x, t) = 1k (t) Akij (k) (xi ) ykj (k) jn(k) + 2k (t)
i 1
k=1 (k)
j1 =1
(k)
jn =1 i=1
p1k+1 pnk+1 ( ) ) (6)

n
A(k+1)ij (k+1) (xi ) y(k+1)j (k+1) jn(k+1) [tk ,tk+1 ) (t).
i 1
(k+1)
j1 =1
(k+1)
jn =1 i=1
Proposition 3.3. Suppose that the time-invariant fuzzy systems are TS fuzzy systems
determined by product inference engine, then the time-variant fuzzy system based on rules

pnk ( n ) ( n )

p

p1k

F (x, t) = 1k (t) Akij (k) (xi ) ckij (k) jn(k) xi
i 1
k=1 (k)
j1 =1
(k)
jn =1 i=1 i=0
p1k+1 pnk+1 ( )

n
+ 2k (t) A(k+1)ij (k+1) (xi ) (7)
i
(k+1)
j1 =1
(k+1)
jn =1 i=1
( n )]

c(k+1)ij (k+1) jn(k+1) xi [tk ,tk+1 ) (t).
1
i=0
Theorem 3.1. For any k {1, , p}, if Fk (x, tk ) is a fuzzy interpolation function then
the time-variant fuzzy system determined by (4) is a fuzzy interpolation function.
Proof: It can be verified by Definitions 2.2 and 2.3 directly.
Corollary 3.1. If the time-variant fuzzy systems are respectively determined by (5) (7),
then they are fuzzy interpolation functions.
Remark 3.2. When a fuzzy system is represented as an interpolation function, by nu-
merical approximation theory, it possesses the universal approximation property. Theorem
3.1 and Definition 2.3 provide a theoretical foundation for the election of fuzzy system and
variant weight in applying VWIM method.
Further, we will investigate the approximation accuracy of time-variant fuzzy systems to
continuously dierentiable functions. Let C n (X) be the space of all n-times continuously

dierentiable functions on X. Particularly, C(X) = C 0 (X) denotes the space of contin-

uous functions on X. If g C(X), then norm of g is defined as: g = sup |g(x)| .
xX
n
Furthermore, if g = (g1 , , gn ) C(X), then g = gi .
i=1
In order to give a unified and compact expression of the universal approximation prop-
erty of time-variant fuzzy systems,
{ we need to introduce
} some notations.
n
Let i = max xkij (k) xki(j (k) 1) , = i and = max {|tk tk1 |}.
(k)
1kp,2ji pik , i i i=1 2kp
( assume that g C ([a1 , b1 ] [an , bn ] )

1
Theorem 3.2. ( Under the above ) conditions,
(k)
[0, T ]) and g xkj (k) jn(k) , tk = ykj (k) jn(k) ji = 1, , pik ; k = 1, , p; i = 1, , n .
1 1
If the time -variant fuzzy system F is determined by (5), then

n
F g x
g
i
i + g .
t
i=1
Proof: For any (x , t ) = (x1 , , xn , t ) X1 X

n U , without loss of generality,
we suppose that t [tk , tk+1 ), xi [xkil(k) , xki(l(k) +1) ] [x(k+1)il(k+1) , x(k+1)i(l(k+1) +1) ], where
i i i i
i = 1, , n. By dierential mean valued theorem, we can prove that
n
|F (x , t ) g(x , t )| x
g
i + g
i t
.
i=1

n
This implies that F g x
g
i
i + g holds.
t
i=1
Remark 3.3. By the numerical approximation theory, we know that any time-variant
fuzzy system satisfying Theorem 3.2 is a first-order accurate approximator for the desired
continuously dierentiable function.
Theorem 3.3. Under the above conditions, suppose that 1) g C 2 ([a1 , b1 ] [an , bn ]
(k)
[0, T ]) and g(xkj (k) jn(k) , tk ) = ykj (k) jn(k) (ji = 1, , pik ; k = 1, , p; i = 1, , n);
1 1
2) the time-variant fuzzy system F is determined by (6); 3) the membership functions
(k)
Akij (k) (xi ) (ji = 1, , pik ; k = 1, , p; i = 1, , n) are triangle-shaped membership
i
functions; 4) the variable weight functions are chosen to be (i) of Example 2.1, then
1 2g
n
1 2
F g x2 i2 + t2g 2 .
8 i=1 i 8
Proof: For any (x , t ) X1 Xn U , we also suppose that t [tk , tk+1 ),

xi [xkil(k) , xki(l(k) +1) ] and xi [x(k+1)il(k+1) , x(k+1)i(l(k+1) +1) ], where i = 1, , n.
i i ([ i i] [ ] )
We define linear operators on C 1 xk1l1 , xk1(l1 +1) xknln , xkn(ln +1) [tk , tk+1 )
as
tk+1 t t tk
L(k,t ) g(x , t ) = g(x , tk ) + g(x , tk+1 );
tk+1 tk tk+1 tk
(k)
lj +1
( )

Lkj g(x , tk ) = Akjr(k) (xj ) g x1 , , xkjr(k) , , xn , tk , j = 1, , n;
j j
(k) (k)
rj =lj
(k) ( n )
( )
(k)

l1 +1 ln+1
Lk g(x , tk ) = Akir(k) (xi ) g xkr(k) rn(k) , tk .
i 1
(k)
r1 =l1
(k) (k)
rn =ln
(k) i=1
Similarly,
([ we can also ]define linear [ operators L(k+1)j g] (j = 1, ), n) and L(k+1) g on
C 1
x(k+1)1l1 , x(k+1)1(l1 +1) x(k+1)nln , x(k+1)n(ln +1) [tk , tk+1 ) .
Similar to the proof of Theorem 3.2 in [18], we can prove that
1
n
2g 1 2g
|F (x , t ) g(x , t )| x2i
i2 + t2
2 .
8 i=1 8
Hence, the assertion holds.
Theorem 3.4. Under the above conditions, suppose that a) the conditions 1), 3) and
4) of Theorem 3.3 hold; b) the time-variant fuzzy system F is determined by (7), where
( )
g
n
ckij (k) jn(k) = xi ( ) , ck0j (k) jn(k) = g xkj (k) jn(k) , tk ckij (k) jn(k) xkij (k) ,
1 (x,tk )= x (k) (k) ,tk 1 1
i=1 1 i
kj1 jn
(k)
ji = 1, , pik ; k = 1, , p; i = 1, , n, then
1
n n
2g 1 2g
F g xi x 2
+ t2
2 .
2 i=1 j=1 j 8
Proof: Using Theorem 3.4 in [15] and similar to the proof of Theorem 3.3, we can
prove that the assertion holds.
Remark 3.4. Theorems 3.3 and 3.4 mean that choosing fuzzy systems to be Mamdani
fuzzy systems with product engine or TS fuzzy systems with product engine, the member-
ship functions to be triangle-shaped functions, and the variant weights to be (i) of Example
2.1, the time-variant fuzzy systems with second-order accurate approximators can be ob-
tained.
Remark 3.5. Theorems 3.2 3.4 are dierent from the results of [10-16,18]. In the
existing results on universal approximation properties of Mamdani fuzzy systems and TS
fuzzy systems, fuzzy relation generalized by fuzzy rules are continuous functions. However,
in time variant fuzzy system modeling, the fuzzy relation obtained by data are discrete
valued functions. Thus, Mamdani fuzzy systems and TS fuzzy systems are not suitable
to approximate such a time-variant model. From the mathematical point of view, the
approximation mechanism of VWIM method is to combine local fuzzy system at each time
interval to approximate the the desired time-variant model.
4. Approximation of Nonlinear Systems by VWIM Method. In this section, we

will answer the question: given a nonlinear system, how many fuzzy sets are needed for
input and output variable in order to guarantee the desired approximation accuracy?
Consider a nonlinear system as follows:
x = g(x, t), (8)
where x = (x1 (t), , xn (t))T is the state vector, xi (t) (i = 1, , n) are state variables.
X1 Xn and X 1 X n are the universes of x and x respectively, and g = (g1 gn )
is a vector-valued function on X1 Xn U . Without loss of generality, we also assume
that Xi , X i (i = 1, , n) and U are real number intervals respectively, i.e., Xi = [ai , bi ],
X i = [ci , di ] and U = [0, T ].
Similar to the Step 1, we can obtain a group of discrete-time fuzzy rules based on
experiment data. The fuzzy rules at time tk (k = 1, , p) are given by:

If x1 (t) is Ak1j (k) and and xn (t) is Aknjn(k) then x 1 (t) is Bk1j (k) jn(k)

1 1

If x1 (t) is Ak1j1(k) and and xn (t) is Aknjn(k) then x 2 (t) is Bk2j1(k) jn(k)
(9)

If x 1 (t) is A (k) and and x n (t) is A (k) then x
n (t) is B (k) (k)

(k)
k1j1 knjn knj1 jn
ji = 1, , pik ; i = 1, , n.
where {Akij (k) }1j (k) pi and {Bkij (k) jn(k) }1j (k) p1 , ,1jn(k) pn are fuzzy partitions of Xi
i i k 1 1 k k
and X i (i = 1, , n) respectively; and the peak points of fuzzy sets A (k) and B (k) (k) kiji kij1 jn
are denoted by xkij (k) and x kij (k) jn(k) respectively.
i 1
In the following, the time-variant fuzzy system determined by fuzzy rules (9) and VWIM
method is denoted as
x = f (x, t), (10)
where f = (f1 , , fn ).
Further, in order to give compact proof of the conclusions, we need to give a lemma
and an assumption.
Lemma 4.1. [21] Consider two initial-value problems of first-order dierential equations,
x = f (x, t), x(0) = x0 ; (11)
x = g(x, t), x(0) = x0 . (12)
Suppose that f and g are continuous on the region D = {(x, t) ||t t0 | a, x x0
b} and satisfy the Lipschitz condition about variable x respectively. The common

Lipschitz constants of f and g are denoted by L. Let M = max{ f , g },

h = min{a, Mb }, and be the solutions of initial-value problems (11) and (12) on

I = [t0 h, t0 + h] respectively. For any > 0, as long as f g < on D , then
L eLh holds on I.
Assumption (): 1) the partition on [0, T ] is equidistant as tk , k = 1,( , p; 2) for any )
k {1, , p}, i {1, , n} the partition on [ai , bi ] is equidistant; 3) g xkj (k) , ,jn(k) , tk =
( ) 1
(k)
x k1j (k) jn(k) , , x knj (k) jn(k) , ji = 1, , pik ; k = 1, , p; i = 1, , n.
1 1
Theorem 4.1. Suppose that a) the assumption () is satisfied and g C 1 ([a1 , b1 ]

[an , bn ] [0, T ]); b) the time-invariant fuzzy systems of VWIM method are chosen to be
Mamdani fuzzy systems with min inference engine; c) nonlinear systems (8) and (10) have
( / g )
the same initial value. For any > 0, if p g T (n+1) e x T + 1
g
t x
( / )
and pi (bi ai ) g (n+1) e x T + 1, then .
g
k x
Proof: It is easy to verify that f and g satisfies the Lipschitz condition. By the
uniqueness theorem for the initial-value problem, we know that nonlinear systems (8) and
(10) have the sole solution respectively, denoted as and .
Since g C 1 ([a1 , b1 ] [an , bn ][0, T ]), we choose xg
as the Lipschitz constant
of g. By Theorem 3.2 and Lemma 4.1, we have
n
g
xi pii 1i + g T
(b a )
t p1 n

i=1 k
g e g
x
T
< + = .
n+1 n+1
x i=1
Remark 4.1. Theorem 4.1 means that when fuzzy systems of VWIM method are chosen to
be Mamdani fuzzy systems with min inference engine, then the corresponding time-variant
fuzzy systems can approximate a class of nonlinear systems with first-order accuracy.
Theorem 4.2. Suppose that a) the condition a) and condition b) of Theorem 4.1 are
satisfied and g C 2 ([a1 , b1 ] [an , bn ] [0, T ]); b) the time-invariant fuzzy systems
are chosen to be Mamdani fuzzy systems with product inference engine. For any > 0,
2g
2
g

g (bi ai ) x2 g
if p T

2 2
n+1

t2
g
x
e x T + 1 and pik
2 2
n+1

g
i
x
e x T + 1
(k = 1, , p; i = 1, , n), then .
Proof: By Theorem 3.3 and Lemma 4.1, we can prove that the assertion holds.
Theorem 4.3. Suppose that the condition a) of Theorem 4.2 holds and the time-invariant
fuzzy systems are chosen to be TS fuzzy systems with product inference engine. For any

n
n 2g
1
2
x x
k j g
> 0, if pik (bi ai ) n+1

k=1 j=1
g
x
e x T + 1, k = 1, , p; i = 1, , n
2
g
g
and p T

2 2
n+1

t2
g
x
e x T + 1, then .
Proof: By the Theorem 3.4 and Lemma 4.1, we can prove that Theorem 4.3 holds.
Remark 4.2. The results of Theorems 4.2 and 4.3 show that when fuzzy systems of
VWIM are chosen to be Mamdani fuzzy systems or TS fuzzy systems with product inference
engine, then the time-variant fuzzy systems constructed by VWIM method are universal
approximations to a class of nonlinear systems with second order accuracy.
Remark 4.3. The majority of existing results on sucient conditions for fuzzy systems
as universal approximators concentrated on answering the question: given a continuous
function, how many rules and fuzzy sets are needed to achieve predefined approximation ac-
curacy. Theorems 4.1 4.3 give some sucient conditions for time-variant fuzzy systems
as universal approxiamtors to a class of nonlinear systems. These results are meaningful
to the optimum design and stability analysis of fuzzy systems.
5. Simulation Experiment.
Example 5.1. Consider the following nonlinear system:
{
x 1 = 0.1 x22 0.1 x1 + cos t,
(13)
x 2 = 0.3 x2 sin t + 0.1 x21 sin(2t/5),
where the initial state x1 (0) = 0.8, x2 (0) = 0.5 and the time universe is [0, 5].
The simulation steps are shown as follows.
Step 1. Make equidistant partition on [0, 5] as tk = 5(k1)
p1
, k = 1, , p.
Step 2. Determine the input universe. For any k {1, , p}, we substitute tk into
Equation (13). By Matlab, we can determine the maximum values and minimum values
(1) (1) (1) (1) (2) (2) (2) (2)
with respect to x1 and x2 respectively, i.e., xmax = bk , xmin = ak , xmax = bk , xmin = ak .
Step 3. Determine the fuzzy sets on the input universe. For each index k, we also make
(1) (1)
(1) (1) (2) (2) (1) (1) (i1)(bk ak )
equidistant partition on [ak , bk ] and [ak , bk ] as xki = ak + n1
(i = 1, , n)
(2) (2)
(2) (2) (j1)(bk ak )
and xkj = ak + m1
(j = 1, , m). The fuzzy sets Aki and Bkj are defined as
(1) (1) (1) (2) (2) (2)
Aki (x1 ) = trimf (x1 , [xk(i1) , xki , xk(i+1) ]) and Bkj (x2 ) = trimf (x2 , [xk(j1) , xkj , xk(j+1) ]).
(1) (2)
Step 4. Substituting (xki , xkj , tk ) into nonlinear Equation (13), the corresponding
(1) (2)
elements x kij and x kij can be computed respectively.
Step 5. Time-invariant fuzzy system is chosen to be Mamdani fuzzy system with prod-
uct inference engine. The variable weights are chosen to be (i) of Example 2.1.
Based on VWIM method, the time-variant fuzzy systems can be expressed as:
m ( )

p1 n

x 1 = tk+1 t (1) (1)
Ckij x kij + tk+1 tk C(k+1)ij x (k+1)ij k (t)
ttk

tk+1 tk
k=1 i=1 j=1

p1 n m ( )
tk+1 t (2) (2) (14)

x 2 = tk+1 tk
Ckij x + ttk
tk+1 tk
C(k+1)ij x (k+1)ij k (t)

k=1 i=1 j=1
kij

Ckij = Aki (x1 ) Bkj (x2 ), k (t) = [tk ,tk+1 ) (t)
The predefined error accuracy is chosen to be 0.1. Let p = 22, n = 5 and m = 3. The
error curves of variables x1 and x2 are shown in Figures 2 and 3. The solution curves and
the phase plane curve of system (14) and the comparison with corresponding curves on
real system (13) are shown in Figures 4 6, where denotes the curve of system (13)
and denotes the curve of (14).
3
x 10
0.06 7
6
0.04
5
0.02 4
error curve of x1
error curve of x2
3
0
2
0.02 1
0
0.04
1
0.06 2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t t
Figure 2. Error curve of Figure 3. Error curve of

variable x1 variable x2
2 0
0.05
1.5
0.1
0.15
1
0.2
x1(t)
x2(t)
0.5 0.25
0.3
0
0.35
0.4
0.5
0.45
1 0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t t
Figure 4. Comparison curve Figure 5. Comparison curve

of variable x1 of variable x2
0.05
0.1
0.15
0.2
x2(t)
0.25
0.3
0.35
0.4
0.45
0.5
1 0.5 0 0.5 1 1.5 2
x1(t)
Figure 6. Comparison curve of phase plant (x1 , x2 )
From the simulation results, we can see that time-variant fuzzy system (14) based on
VWIM method can approximate system (13) with the desired accuracy.
6. Conclusions. In this paper, we investigate the design and approximation problem of
time-variant fuzzy systems. The main contributions include:
(1) A novel modeling method for time-variant fuzzy systems, called VWIM method
has been proposed. From the mathematical point of view, the time-variant fuzzy systems
based on VWIM method can be regarded as some interpolation functions. And, the
approximation accuracy of various classes of time-variant fuzzy systems has been analyzed.
(2) The time-variant fuzzy systems based on VWIM method have been introduced to
approximate a class of nonlinear systems. Some sucient conditions for time-variant
fuzzy systems as universal approximators to a class of nonlinear systems have been given.
By them, we can determine the numbers of fuzzy sets and fuzzy rules which are needed
for approximating any nonlinear systems with given precision.
From the theoretical analysis and simulation result, we can learn that only if experts
master adequate input-output data information of the systems, the time-variant fuzzy
systems determined by VWIM method could reflect the dynamic behavior of systems
with high precision. This conclusion shows that VWIM method can be used as a new
tool to deal with the design and analysis of fuzzy controller.
Acknowledgment. This work is partially supported by the Fundamental Research Funds

for the Central Universities, the National Natural Science Foundation of China (60774049,
60834004), the National 863 High-Tech Program of China (2006AA042163), Specialized
Research Fund for the Doctoral Program of Higher Education (20090041110003), China
Postdoctoral Science Foundation (20090461182) and the National 973 Basic Research
Program of China (2009CB320602, 2002CB312200). The authors also gratefully acknowl-
edge the helpful comments and suggestions of the reviewers, which have improved the
presentation.
REFERENCES
[1] H. X. Li, Z. H. Miao and J. Y. Wang, Variable universe stable adaptive fuzzy control of nonlinear
system, Science in China (Series E), vol.45, no.3, pp.225-240, 2002.
[2] H. X. Li, J. Y. Wang and Z. H. Miao, Modeling on fuzzy control systems, Science in China (Series
E), vol.45, no.12, pp.1506-1517, 2002.
[3] L. Wang and X. Liu, New relaxed stabilization conditions for fuzzy control systems, International
[4] Y. Shi, P. Messenger and M. Mizumoto, Fuzzy inference modeling based on fuzzy singleton-type
reasoning, International Journal of Innovative Computing, Information and Control, vol.3, no.1,
pp.13-20, 2007.
[5] T. Wang and S. Tong, Adaptive fuzzy output feedback control for SISO nonlinear systems, Interna-
tional Journal of Innovative Computing, Information and Control, vol.2, no.1, pp.51-60, 2006.
[6] T. Wang, S. Tong and Y. Li, Robust adaptive fuzzy control for nonlinear system with dynamic
uncertainties based on backstepping, International Journal of Innovative Computing, Information
and Control, vol.5, no.9, pp.2675-2688, 2009.
[7] X. Ban, X. Z. Gao, X. Huang and H. Yin, Stability analysis of the simplest Takagi-Sugeno fuzzy
control system using Popov criterion, International Journal of Innovative Computing, Information
[8] Z. He and Y. Lu, Decentralized control of spatially interconnected ane fuzzy systems with appli-
cation to a morphing airfoil, ICIC Express Letters, vol.3, no.3(B), pp.593-598, 2009.
[9] L. A. Zadeh, Outline of a new approach to the analysis of complex systems and decision processes,
IEEE Trans. on Systems, Man, and Cybernetics, vol.3, pp.28-44, 1973.
[10] X. J. Zeng and M. G. Singh, Approximation theory of fuzzy systems-MIMO case, IEEE Trans. on
Fuzzy Systems, vol.3, no.2, pp.219-235, 1995.
[11] X. J. Zeng and M. G. Singh, Approximation theory of fuzzy systems-SISO case, IEEE Trans. on
Fuzzy Systems, vol.2, no.2, pp.162-176, 1994.
[12] H. Ying, General SISO Takagi-Sugeno fuzzy systems with linear rule consequent are universal ap-
proximators, IEEE Trans. on Fuzzy Systems, vol.6, no.4, pp.582-587, 1998.
[13] H. Ying, Sucient conditions on uniform approximation of multivariate functions by general Takagi-
Sugeno fuzzy systems with linear rule consequent, IEEE Trans. on Systems, Man and Cybernetics,
Part A, vol.28, no.4, pp.515-520, 1998.
[14] K. Zeng, N. Y. Zhang and W. L. Xu, A comparative study on sucient conditions for Takagi-Sugeno
fuzzy systems as universal approximators, IEEE Trans. on Fuzzy Systems, vol.8, no.6, pp.773-780,
2000.
[15] Y. M. Li, Analysis of Fuzzy Systems, Science in China Press, China, 2005.
[16] D. Wang, W. Song and H. Li, Unified forms of fuzzy similarity inference method for fuzzy reasoning
and fuzzy systems, International Journal of Innovative Computing, Information and Control, vol.4,
no.9, pp.2285-2294, 2008.
[17] H. X. Li, F. You and J. Y. Peng, Fuzzy control and their response functions based on some fuzzy
implication operators, Progress in Nature Science, vol.14, no.1, pp.15-20, 2003.
[18] X. J. Zeng and M. G. Singh, Approximation accuracy analysis of fuzzy systems as function approx-
imators, IEEE Trans. on Fuzzy Systems, vol.4, no.1, pp.44-63, 1996.
[19] W.-Y. Song, D.-G. Wang and W.-G. Wang, Analysis for perturbation of fuzzy systems modeling,
[20] W. Song, D. Wang, H. Li and W. Wang, Stability analysis of a class of nonlinear dynamic systems,
ICIC Express Letters, vol.4, no.1, pp.281-287, 2010.
[21] H. X. Li, W. Y. Song, X. H. Yuan and Y. C. Li, Time-varying system modeling method based on
fuzzy inference, Journal of Systems Science and Mathematical Sciences, vol.29, no.8, pp.1109-1129,
2009.
[22] L. T. Koczy and K. Hirota, Interpolative reasoning with insucient evidence in sparse fuzzy rules
bases, Information Science, vol.71, no.1-2, pp.169-201, 1993.
[23] L. T. Koczy and K. Hirota, Approximate reasoning by linear rule interpolation and general approx-
imation, International Journal of Approximate Reasoning, vol.9, no.3, pp.197-225, 1993.
[24] Y. Shi, M. Mizumoto and W. Z. Qiao, Reasoning conditions on Koczys interpolative reasoning
method in sparse fuzzy rule bases, Fuzzy Sets and Systems, vol.75, no.1, pp.63-71, 1995.
[25] Y. Shi and M. Mizumoto, Reasoning conditions on Koczys interpolative reasoning method in sparse
fuzzy rule bases part II, Fuzzy Sets and Systems, vol.87, no.1, pp.47-56, 1997.
[26] P. Z. Wang, S. H. Tan and F. M. Song, Constructive theory for fuzzy systems, Fuzzy Sets and
Systems, vol.88, no.2, pp.195-203, 1997.
c ISSN 1349-4198
ENCODING PRIOR KNOWLEDGE INTO DATA DRIVEN DESIGN OF

INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS
Chengdong Li1,2 , Jianqiang Yi2 and Tiechao Wang2

1
School of Information and Electrical Engineering
Shandong Jianzhu University
Fengming Road, Lingang Development Zone, Jinan 250101, P. R. China
chengdong.li@ia.ac.cn
2
Institute of Automation
Chinese Academy of Sciences
Beijing 100190, P. R. China
Abstract. In system identification or modeling problems, interval type-2 fuzzy logic

systems (IT2FLSs), which have obvious advantages for handling dierent sources of un-
certainties, are usually constructed only using the information from sample data. This
paper tries to utilize the information from both sample data and prior knowledge to design
IT2FLSs to compensate the insuciency of the information from single knowledge source.
First, sucient conditions on the antecedent and consequent parameters of IT2FLSs are
given to ensure that the prior knowledge can be incorporated into IT2FLSs and three
kinds of prior knowledge bounded range, symmetry (odd and even) and monotonicity
(increasing and decreasing) are explored. Then, design of IT2FLSs using the informa-
tion from both sample data and prior knowledge is transformed to the constrained least
squares optimization problem. At last, to show the superiority of the proposed method,
simulations and comparisons are made.
Keywords: Information fusion, Type-2 fuzzy, Least squares algorithm, Prior knowledge
1. Introduction. Recently, a number of extensions to classical fuzzy logic systems (type-

1 fuzzy logic systems: T1FLSs) have been attracting interest. One of the most widely
used extensions is the interval type-2 fuzzy logic system (IT2FLS) [1-10]. IT2FLSs have
obvious advantages for handling dierent sources of uncertainties, reducing the number of
fuzzy rules and weakening noisy disturbance, etc., as IT2FLSs utilize interval type-2 fuzzy
sets (IT2FSs) which can provide additional degrees of freedom and have more parameters
than type-1 fuzzy sets (T1FSs) in T1FLSs [1,2,6]. Due to these merits, IT2FLSs have
found lots of applications, for example, time-series forecasting [2], control of autonomous
mobile robots [3] and direct model reference control [8].
Until now, IT2FLSs are always constructed only using the information from the single
knowledge source sample data (training data). Sometimes, satisfactory performance
of IT2FLSs can be achieved using this data-driven design method, but, in general, only
sample data is not enough to provide sucient information for system identification,
especially when the sample data is not informative enough or is noisy.
One way to compensate this weakness is to incorporate prior knowledge into IT2FLSs.
Although, in most cases, it is hard to obtain exact physical structure knowledge of some
complex systems, part of their physical properties can be observed easily, such as mono-
tonicity, bounded range, symmetry, etc. Such prior knowledge can partly reflect the
characteristics of the unknown systems and compensate the insuciency of the informa-
tion from sample data. Recently, this topic has gained considerable concern from dierent
1133
1134 C. LI, J. YI AND T. WANG
research areas. Some work has been done to incorporate prior knowledge into support
vector machines [11,12], neural networks [13-15], T1FLSs [16-19], etc.
But, to the authors knowledge, there is no work concerning on how to make full use of
the information from multiple knowledge sources (sample data and prior knowledge) to
compensate the insuciency of the information from single knowledge source in the design
of IT2FLSs. In this paper, we will study this issue and three kinds of prior knowledge
bounded range, symmetry (odd and even) and monotonicity (increasing and decreasing)
will be considered. First, we will present sucient conditions on the parameters of
IT2FLSs to ensure that the prior knowledge can be encoded. Then, we will show how
to utilize the constrained least squares algorithms to incorporate the information from
both sample data and prior knowledge into the design of IT2FLSs. At last, we will give
simulations and comparisons to show the superiority of the proposed method.
2. Interval Type-2 Fuzzy Logic Systems. In this section, we will introduce the in-
ference process of IT2FLSs [1-10] briefly. At first, let us give the definition of trapezoid
IT2FSs. Without detailed specification, in this study, all IT2FSs are the trapezoid IT2FSs.
2.1. Trapezoid IT2FS. Figure 1 shows a trapezoid IT2FS and a triangular IT2FS. The
triangular IT2FS is a special case of the trapezoid IT2FSs. A trapezoid IT2FS A e can be
described by its lower and upper membership functions Ae(x, ) and Ae(x, ) as

h2 bxa 2
2 a2
, a 2 < x b2 ,
h1 xa1 , a1 < x b1 ,
b1 a1
h2 , b2 < x c2 , h1 , b1 < x c1 ,
Ae(x, ) = d2 x and Ae(x, ) = d1 x

h 2 , c 2 < x d 2 ,
h 1 , c1 < x d1 ,
d2 c2 d1 c1
0, else, 0, else,
where a1 b1 c1 d1 , a2 b2 c2 d2 , a1 a2 , d1 d2 , h2 h1 1 and
e
= (a1 , b1 , c1 , d1 , a2 , b2 , c2 , d2 , h1 , h2 ) is the parameter vector of all the parameters in A.
h1 h1
h2 h2
a1 a2 b1 b2 c2 c1 d 2 d1 a b c d e
Figure 1. Examples of trapezoid IT2FSs
2.2. IT2FLS. Consider the following complete rule base of an IT2FLS

[ ]
ej11 , x2 = A
Rj1 j2 jp : x1 = A ej22 , . . . , xp = A
ejp , y = wj1 j2 jp , wj1 j2 jp
p
where p is the number of the input variables, A eji (i = 1, 2, . . . , p, ji = 1, 2, . . . , Ni ) are

i
IT2FSs of the IF-part and [w j1 j2 jp
, wj1 j2 jp ]s are consequent interval weights of the
THEN-part. There are totally pi=1 Ni fuzzy rules in this complete rule base.
Once a crisp input x = (x1 , x2 , . . . , xp )T is applied to the IT2FLS, through the singleton
fuzzifier, the interval firing strength of the rule Rj1 j2 jp can be obtained as
[ j1 j2 jp
] [ ]
x, a ) = f j1 j2 jp (x
F j1 j2 jp (x x, a ), f x, a ) = Ti=1
(x p
Aeji (xi , ji i ), Ti=1
p
Aeji (xi , ji i ) ,
i i
where a is a vector of all the parameters in the antecedent part of the IT2FLS, and T
denotes minimum or product t-norm.
ENCODING PRIOR KNOWLEDGE INTO DATA DRIVEN DESIGN OF IT2FLS 1135
Then, using the center-of-sets type-reducer [1,2,4,6] and the center average defuzzifier,
the crisp output of the IT2FLS can be computed as
1
y(x x, ) = (yl (x x, ) + yr (x x, )),
2
where = (a , c )T in which c is the parameter vector of all the parameters (wj1 j2 jp
and wj1 j2 jp ) in the consequent part of the IT2FLS; yl (x x, ) and yr (x x, ) are the left and
right end points of the type-reduced interval set and can be expressed as
{ N }
1 Np f j1 jp wj1 jp
j1 jp j1 jp
x, ) = min
yl (x j1 =1
N1
jp =1
Np f F j j
x,a ),w 1 p [w 1 p ,w 1 p ] ,
(x j j j j
f j1 jp
{ N }
j1 =1 jp =1

1 Np f j1 jp wj1 jp
j1 jp j1 jp
yr (xx, ) = max j1 =1
N1
jp =1
Np j1 jp
f F x,a ),wj1 jp [wj1 jp ,wj1 jp ] .
(x
f
j1 =1 jp =1
Mendel et al. have given the iterative Karnik-Mendel algorithm based formulas for the
x, ) and yr (x
computation of yl (x x, ). For more detail about this topic, see [1,2,4,6].
3. Parameter Conditions. In this section, we will present sucient conditions on the

parameters of IT2FLSs to ensure that the prior knowledge of bounded range, symmetry
(odd and even) and monotonicity (increasing and decreasing) can be encoded. First, let
us consider the prior knowledge of bounded range.
3.1. Prior knowledge of bounded range. Notice that the bounded range of an IT2FLS
automatically implies the bounded-input-bounded-output (BIBO) stability of the IT2FLS,
which is usually required in many real-world applications. For the prior knowledge of
bounded range, we have the following results for IT2FLSs:
Theorem 3.1. The output y(x x, ) of an IT2FLS falls in the bounded range [b, b], i.e.,
b y(x
x, ) b, if its consequent parameters satisfy that:
min {wj1 jp } b and max {wj1 jp } b.
ji =1, ,Ni ji =1, ,Ni
i=1, ,p i=1, ,p
Proof: f j1 jp F j1 jp (xx, a ) and wj1 jp [wj1 jp , wj1 jp ], we have

N1 Np j1 jp j1 jp
j1 =1 jp =1 f w
N1 Np j j min {wj1 jp } min {wj1 jp }.
j1 =1
ji =1, ,Ni ji =1, ,Ni
jp =1 f
1 p
i=1, ,p i=1, ,p
Thus,
{ N }
1 Np f j1 jp wj1 jp
j1 jp j1 jp
x, ) = min
yl (x j1 =1
N1
jp =1
Np f F j
x,a ),w 1
(x j p [w 1
j j p j
,w 1 j p ]
f j1 jp
j1 =1 jp =1
min {wj1 jp }.
ji =1, ,Ni
i=1, ,p
x, )
Similarly, we can prove that yr (x max {wj1 jp }.
ji =1, ,Ni
i=1, ,p
Therefore, if min {wj1 jp } b and max {wj1 jp } b, then,
ji =1, ,Ni ji =1, ,Ni
i=1, ,p i=1, ,p
1
b y(xx, ) = (yl (x x, )) b.
x, ) + yr (x
2
Remark 3.1. For the prior knowledge of bounded range, we only need to constrain the
consequent parameters c and there is no need to constrain the antecedent parameters.
As a result, in this case, there is no requirement on how to partition each input domain.
3.2. Prior knowledge of symmetry (odd and even). For many applications, espe-
cially control problems, the fuzzy systems designed for them should be odd symmetric or
even symmetric. The following two theorems will show how to constrain the parameters
of IT2FLSs to incorporate the prior knowledge of odd symmetry and even symmetry.
Theorem 3.2. An IT2FLS is odd symmetric, i.e., y(x x, ) = y(x x, ), if the following
conditions are satisfied:
1) i {1, 2, . . . , p}, the input domain of xi is partitioned symmetrically around 0 by
trapezoid IT2FSs A e1 , Ae2 , . . . , A
eNi as shown in Figure 2, where Ni is an odd number.
i i i
2) The
[ j j consequent
] [ (N +1j )(Nof+1j
parameters the rules Rj1 jp and R(N1 +1j ]
1 )(Np +1jp )
satisfy that
j
w 1 p , w 1 p = w 1
j 1 p p )
, w 1
(N +1j 1 )(Np +1jp )
, where ji = 1, 2, . . . , Ni2+1 .
~ ~ ~ ~ ~ ~ Ni
Ai1 Ai2 Ai3 AiN i - 2 AiN i - 1 Ai
0 xi
Figure 2. Interval type-2 fuzzy partition with Ni trapezoid IT2FSs
Proof: Suppose that t fuzzy rules, which are denoted as Rj1 jp , . . . , Rj1 jp , can be
1 1 t t
fired when x is given to the IT2FLS.

As the rule base is complete, and each input domain is partitioned symmetrically
1 1
around 0, hence, when x x is given to the IT2FLS, the fuzzy rules R(N1 +1j1 )(Np +1jp ) ,
t t
. . . , R(N1 +1j1 )(Np +1jp ) can be fired.
Also, from the conditions above, it is obvious that, for s = 1, 2, . . . , t
F j1 jp (x
s s s s
x, a ) = F (N1 +1j1 )(Np +1jp ) (xx, a ),
[ j s j s j s j s ] [ s s s s ]
w 1 p , w 1 p = w(N1 +1j1 )(Np +1jp ) , w(N1 +1j1 )(Np +1jp ) .
Hence,
{t s s
}
s=1 f w s j1s jps
[ j s j s j s j s ]
x, ) = min
yl (x t s
f F x, a ), w w
(x s 1 p ,w 1 p
s=1 f
{t s
}
s
s=1 f (w ) s j1s jps
[ j1s jps j1s jps
]
= max t s f F (xx, a ), w w
s
, w
s=1 f
{t s s
s=1 f w s s s
= max f F (N1 +1j1 )(Np +1jp ) (x x, a ),
t s
s=1 f
}
[ (N1 +1j s )(Np +1j s ) (N1 +1j s )(Np +1j s ) ]
w w
s 1 p ,w 1 p
= yr (x
x, )
x, ) = yl (x
In the similar way, we can prove that yr (x x, ).
Therefore,
x, ) + yr (x
yl (x x, ) yr (x
x, ) yl (x
x, )
x, ) =
y(x = = y(x
x, ).
2 2
Similarly, for the prior knowledge of even symmetry, we have following results:
Theorem 3.3. An IT2FLS is even symmetric, i.e., y(x x, ) = y(x x, ), if the following
conditions are satisfied:
1) i {1, 2, . . . , p}, the input domain of xi is partitioned symmetrically around 0 by
trapezoid IT2FSs A e1 , Ae2 , . . . , A
eNi as shown in Figure 2, where Ni is an odd number.
i i i
2) The
[ j j consequent parameters
] [ (N +1j )(N +1j of the rules Rj1 jp and R(N ]
1 +1j1 )(Np +1jp )
satisfy that
j1 jp p) (N1 +1j1 )(Np +1jp )
w 1 p
,w = w 1 1 p
,w , where ji = 1, 2, . . . , (Ni +1)/
N1 +1 N +1 N1 +1 N +1
p2 p2
2; especially, the consequent parameters of the rule R 2 satisfies that w 2 =
N1 +1 Np +1
w 2 2 .
Remark 3.2. The first condition in Theorems 3.2 and 3.3 means that: i {1, 2, p},
1) If ji {1, 2, . . . , (Ni 1)/2}, then {a1 , b1 , c1 , d1 , a2 , b2 , c2 , d2 }, iji = iNi +1ji
and {h1 , h2 }, iji = iNi +1ji .
Ni +1 Ni +1 Ni +1 Ni +1 Ni +1 Ni +1
2) If ji = (Ni + 1)/2, then a1,i2 = d1,i2 , a2,i2 = d2,i2 , b1,i2 = c1,i2 and
Ni +1 Ni +1
b2,i = c2,i .
2 2
This is the constraint on the antecedent parameters of IT2FLSs to ensure that the prior
knowledge of odd symmetry and even symmetry can be encoded.
3.3. Prior knowledge of monotonicity (increasing and decreasing). The mono-
tonicity is a common kind of prior knowledge [20]. For example, consider the water heating
system [18] and the coupled-tank liquid-level system [21]. The outputs of both systems
will change monotonically with respect to their inputs.
For simplicity, here, we only give useful results about the single-input monotonically
increasing and monotonically decreasing IT2FLSs.
Theorem 3.4. Assume that the input domain U = [u, u] is partitioned ( by N trape- )
zoid IT2FSs A e ,A
1 e ,...,A
2 e , where j {1, 2, . . . , N }, ej (x) = ej x, a1 , bj1 , cj1 , dj1 , hj1 ,
N j
( ) A A
Aej (x) = Aej x, aj2 , bj2 , cj2 , dj2 , hj2 . Then, the IT2FLS is monotonically increasing, if the
following conditions are satisfied:
1) IT2FSs A e1 , A
e2 , . . . , AeN form fuzzy partition as shown in Figure 3 where a1 = b1 =
1 1
1 N
a2 = b2 = u, c1 = d1 = c2 = dN
1 N N
2 = u, bj
1 aj+1
1 , d j
1 c j+1
1 , dj
2 > a j+1
2 and dj1
1 a j+1
1 .
2) The consequent parameters satisfy that w w ... w and w w ... wN .
1 2 N 1 2
~ ~ ~ ~ ~ ~ ~
A1 A2 A3 A4 AN - 2 AN - 1 AN
1
x
t1 S1 t2 S2 t3 S N-1
tN-1 tN
Figure 3. Fuzzy partition with N trapezoid IT2FSs
Proof: Denote j (j = 1, . . . , N ) as

{u},[( ) ( j+1 )] j = 1
j = U \ u, dj1 a1 , u , 1 < j < N
1
{u}, j=N
As dj1
1 aj+1
1 , hence, j = .
For any point tj j (j = 1, . . . , N ), the input domain U = [u, u] can be partitioned

1 j
into N 1 areas as shown in Figure 3, i.e., U = N j
j=1 S , where S = [tj , tj+1 ].
The first condition assures that the trapezoid IT2FSs A e1 , A
e2 , . . . , A
eN are arranged in
the ascending order, and, at each point, no more than two IT2FSs can be fired. Hence,
from (55) and (56) in [1], we can obtain that:
If x Sj (j = 1, 2, . . . , N 1), then
j
f (x, a )wj + f j+1 (x, a )wj+1
yl (x, ) = j , (1)
f (x, a ) + f j+1 (x, a )
j+1
f j (x, a )wj + f (x, a )wj+1
yr (x, ) = j+1 , (2)
f j (x, a ) + f (x, a )
j j
where f j (x, a ) = Aej (x), f (x, a ) = Aej (x). For short, we use f j (x), f (x) to replace
j
f j (x, a ) and f (x, a ) below.
From (1) and (2), we can see that, if x S j , then
wj yl (x, ) wj+1 and wj yr (x, ) wj+1 .
Hence, if x S j , then
wj + wj wj+1 + wj+1
y(x, ) . (3)
2 2

x, x U , suppose that x x and x S j , x S j .
If j < j , from (3) and condition 2) in this theorem, we can see that y(x, ) y(x , );
If j = j , then we have following results:
a) From (1), x x
[ j j
]
j+1
f (x )f (x) f (x )f (x) (wj+1 wj )
j+1

yl (x , ) yl (x, ) = ( j )( j )
f (x ) + f j+1 (x ) f (x) + f j+1 (x)
j
As f j+1 (.) is monotonically increasing and f (.) is monotonically decreasing in S j , hence
j j
f j+1 (x ) f j+1 (x) and f (x) f (x ).
This implies that
j j
f j+1 (x )f (x) f (x )f j+1 (x) 0.
Therefore, yl (x , ) yl (x, ) in S j .
b) In the similar way, x x S j , yr (x , ) yr (x, ).
From a) and b), if both x and x are in S j , then y(x , ) y(x, ).
Thus, from the discussion above, we can conclude that this theorem holds.
Similar results can be obtained for the prior knowledge of decreasing monotonicity.
Theorem 3.5. Assume that the input domain U = [u, u] is partitioned ( by N trape-
)
e1 e2 e
zoid IT2FSs A , A , . . . , A , where j {1, 2, . . . , N }, Aej (x) = Aej x, a1 , b1 , c1 , dj1 , hj1 ,
N j j j
( )
Aej (x) = Aej x, aj2 , bj2 , cj2 , dj2 , hj2 . Then, the IT2FLS is monotonically decreasing, if the
following conditions are satisfied:
1) IT2FSs A e1 , A
e2 , . . . , A eN form fuzzy partition as shown in Figure 3 where a1 = b1 =
1 1
a12 = b12 = u, cN
1 = d N
1 = c N
2 = d N
2 = u, bj
1 aj+1
1 , d j
1 c j+1
1 , dj
2 > a j+1
2 and dj1
1 aj+1
1 .
2) The consequent parameters satisfy that w1 w2 ... wN and w1 w2 ... wN .
In this section, sucient parameter constraints of IT2FLSs have been studied to ensure
the aforementioned prior knowledge. In the following section, we will show how to design
IT2FLSs using the information from both sample data and prior knowledge.
4. Design of IT2FLSs Using Both Sample Data and Prior Knowledge. In the
design of IT2FLSs using the information from both sample data and prior knowledge,
dierent kinds of prior knowledge are used to constrain the antecedent and consequent
parameters of IT2FLSs, and sample data is utilized to train and optimize these parameters
further. For simplicity, in this study, we only consider the design of the single-input
IT2FLSs using both sample data and prior knowledge.
4.1. Problem formulation. Suppose that there are M input-output sample data (x1 ,
y 1 ), (x2 , y 2 ), . . . , (xM , y M ). And, the training criteria is chosen to minimize the following
squared error function:
1 M
E= |y(xi , ) y i |2 , (4)
2 i=1
where y(xi , ) is the output of the IT2FLS, and is the parameter vector of all the
antecedent and consequent parameters of the IT2FLS.
In Section 3, we have studied how to transform dierent kinds of prior knowledge to
the constraints of the antecedent and consequent parameters of IT2FLSs. These con-
straints can be rewritten abstractly as , where represents the constrained feasible
parameter space.
Therefore, design of the IT2FLSs using the information from both sample data and
prior knowledge can be seen as the following optimization problem
{
min 21 M i=1 |y(x , ) y |
i i 2
(5)
subject to .
As the outputs of the IT2FLSs are nonlinear with respect to the parameters of their
antecedent IT2FSs, this optimization problem needs to be solved using constrained non-
linear optimization algorithms, e.g., genetic algorithms, particle swarm algorithms.
4.2. Problem transformation. Although dierent optimization algorithms can be uti-
lized to solve the optimization problem in (5), in this work, we adopt another strategy,
which determines and optimizes the antecedent parameters and consequent parameters
separately. In this strategy, two steps are needed. The first step is to set up the member-
ship functions of the IT2FSs in the antecedent part of the IT2FLSs, and the second step
is to optimize the consequent interval weights under the constraints on these parameters.
The first step can be accomplished by partitioning the input domains intuitively or by the
clustering algorithms. How to use dierent clustering algorithms to obtain optimal or sub-
optimal interval type-2 fuzzy partitions, which satisfy the constraints on the antecedent
membership functions in Theorems 3.1 3.5, will be one of our research directions in the
near future. In this study, we mainly focus on the second step.
Below, we will first show that the output of the IT2FLS is linear with its consequent
parameters, and then, we will demonstrate that the optimization problem (5) can be
changed to a constrained least squares optimization problem.
We suppose that the fuzzy partition of the input domain satisfies that no more than
two IT2FSs can be fired at each point as shown in Figure 3, and the rule base of the
single-input IT2FLS is written as
ej y = [wj , wj ]}N .
{Rj : x = A (6)
j=1
From Theorem 3.3 in [22], we know that the crisp output of the single-input IT2FLS in
(6) is equal to the average of the outputs of two T1FLSs (called Lower T1FLS and Upper
T1FLS). The rules of the two T1FLSs are [22]:
{ }N
Lower T1FLS : RLj : x = AjL y = wj j=1 ,
{ }N
Upper T1FLS : RUj : x = AjU y = wj j=1 ,
where A1L , . . . , AN 1 N
L , AU , . . . , AU are type-1 fuzzy sets. Their membership functions can be
determined by

e (x), x S j1
, w j1
w j
,
Aej (x), x S j1 , wj1 wj ,
Aj
(x), x S j1 , wj1 > wj ,
Aej (x), x S j1 , wj1 > wj , Aej
Aj (x) = (x) =
(x), x S j , wj wj+1 ,
j
L
e (x), x S j
, w j
w j+1
, AU

A (x), x S j , wj > wj+1 . Aej
j
ej
A
Aej (x), x S j , wj > wj+1 .
Therefore, the crisp output of the single-input IT2FLS can be computed as
[ N N ]
i i
1 i=1 AiL (x)w i=1 AiU (x)w
y(x, c ) = N + N = [1 (x), . . . , 2N (x)]c , (7)
2 i=1 Ai (x) i=1 Ai (x)
L U
where
[ ]T
c = w 1 , . . . , w N , w 1 , . . . , w N ,

1
Ai (x)
2 Ni=1 L i (x) , i = 1, . . . , N,
A
i (x) = L

1
iN (x)
A
2 N U (x)
, i = N + 1, . . . , 2N.
i=1 AiN
U
From (7), we can see that the output of the IT2FLS is linear with its consequent
parameters. And, the training criteria (4) can be rewritten as
1 M 1
E= |y(xi , c ) y i |2 = (c y )T (c y ),
2 i=1 2
where
y = [y 1 , y 2 , . . . , y M ]T ,

1 (x1 ) 2 (x1 ) 2N (x1 )
1 (x2 ) 2 (x2 ) 2N (x2 )
= .. .. .. .. RM 2N .

. . . .
1 (xM ) 2 (xM ) 2N (xM )
As we only consider to optimize the consequent parameters, therefore, the abstract
constraint in (5) can be rewritten as c c . Below, we will write the constraints
on the consequent parameters in Theorems 3.1 3.5 into the forms of linear-inequality
constraints Cc b and/or linear-equality constraints Ceq c = b eq .
1) Linear constraint for bounded range
[ ] [ ]
IN 0 bb
c ,
0 IN b
where IN is the N N identity matrix, b = [b, b, . . . , b]T and b = [b, b, . . . , b]T .
| {z } | {z }
N N
2) Linear constraint for odd symmetry
[ ]
IN , R(IN ) c = 0 ,
where R(IN ) denotes the N N matrix obtained by rotating IN 90 degrees clockwisely.

3) Linear constraint for even symmetry

I N 1 0 N 1 1 R(I N 1 ) 0 N 1 0 N 1 1 0 N 1
0 N2 1 2 2 2 2
1 0 1 N 1
2
1 2 1 0 1 N 1
2
0 1 N 1
2 2
c = 0 .
0 N 1 0 N 1 1 0 N 1 I N 1 0 N 1 1 R(I N 1 )
2 2 2 2 2 2
4) Linear constraint for increasing monotonicity

[ ]
V 0
c 0 ,
0 V
where

1 1 0 0 0
0 1 1 0 0
V=
... ... .. .. .. .. R(N 1)N .
. . . .
0 0 0 1 1
5) Linear constraint for decreasing monotonicity
[ ]
V 0
c 0 .
0 V
From the discussions above, we can observe that design of the IT2FLSs using both
sample data and prior knowledge can be transformed to the following constrained least
squares optimization problem
{
min(c y )T (c y )
c
subject to Cc b and/or Ceq c = b eq .
5. Simulations and Comparisons.
5.1. Problem description. To show the usefulness of the prior knowledge, lets consider
the following nonlinear function:
g(x) = (|x|/3)2 tanh(|x|),
where x U = [3, 3]. Obviously, this function is bounded in [0, 1], even symmetric,
monotonically decreasing in [3, 0] and monotonically increasing in [0, 3].
The noisy training data is generated by
e,
y(x) = g(x) + n
where n is the uniformly distributed additive noise in [b, b]. In this simulation, three
dierent levels of noisy disturbance are tested, where b = 20%, 30% and 40%, respec-
tively. In each noisy circumstance, we consider 5 cases. In these cases, the sizes of
both the training data and the evaluation data { 1are1 30, 260, 2 100, 200,
( K 300. )}In case
Ki
i, we denote the training data set as Ti = (x , y ) , (x , y ) , . . . , x , yi
. Fur-
thermore,
{ 1 in each case, the noise-free
( K data set
)} is chosen as the evaluation data set
1 2 2 Ki
Di = (x , g(x )) , (x , g(x )) , . . . , x , g(x ) which can be used to check whether
i
the trained model could follow the characteristics of the original data rather than the
noisy data.
In our simulation, for comparison, we design four fuzzy logic systems (FLSs) to identify
the nonlinear function: sample-Data-and-prior-Knowledge based IT2FLS (DK-IT2FLS),
only sample-Data based IT2FLS (D-IT2FLS), sample-Data-and-prior-Knowledge based
T1FLS (DK-T1FLS), only sample-Data based T1FLS (D-T1FLS). To evaluate the ap-
proximation performance and the generalization performance of dierent FLSs, we con-
sider the following two performance indexes Root of the Mean Squared Errors (RMSE s)
of the training data and evaluation data:
( ( )1
1 Ki ( k ) )2 2
i = y y x
k
,
Ki k=1
( ( )1
1 Ki ( k ) )2 2
i = g(x ) y x
k
,
Ki k=1
where y(xk ) is the output of FLSs for the input xk . The first index i can reflect the
approximation ability of dierent FLSs for the training data in case i in each level of the
noisy disturbance, while the second index i can reflect the generalization characteristics of
dierent FLSs for the original noise-free data in case i in each level of the noisy disturbance.
For both indexes, the less the value is, the better the result is.
For each FLS, we use 7 fuzzy rules, the membership functions of whose antecedent parts
are shown in Figure 4. Note that the fuzzy partitions in Figure 4 satisfy the conditions
on fuzzy sets in Theorems 3.1 3.5. The left task is to tune the consequent weights with
or without constraints.
~1
A ~2
A ~3
A 1 ~4
A ~5
A A~6 ~
A7
-3 -2 -1 0 1 2 3
x
Figure 4. Type-1 (dotted line) and type-2 (gray area) fuzzy partitions
5.2. Simulation results and comparisons. As stated above, three dierent levels of
noise are considered, and, under each noisy circumstance, five cases are tested. What is
more, in each case, the data generation processes and the identification processes of the
four FLSs are run 50 times. After that, the performance indexes in case i are calculated
as the average of the corresponding values obtained in the 50 run times, i.e.,
1 T j
ei = i ,
T j=1
1 T
ei =
ij ,
T j=1
where T = 50, and ij and ij are the first and second performance indexes obtained in
the jth run in case i.
For the aforementioned three noise levels, the comparisons of the four FLSs with respect
to the RMSE s for the training and evaluation data are shown in Figure 5. And, Figure
6 shows us an example of the identification result where 100 sample data are utilized to
train the four FLSs and the noise distributes uniformly in [30%, 30%].
As stated above, the RMSE s i for training data can reflect the approximation abilities
of dierent FLSs. From Figures 5(a) 5(c), D-IT2FLS has the best approximation ability,
and then, DK-IT2FLS, at last, D-T1FLS and DK-T1FLS. In general, according to the
approximation performance in this simulation, the type-2 FLSs perform better than the
type-1 FLSs; the reason for this is that the type-2 FLSs have more parameters and
more freedoms than the type-1 FLSs. On the other hand, D-IT2FLS and D-T1FLS have
better approximation performance compared with DK-IT2FLS and DK-T1FLS, as the
Figure 5. Performance comparisons of the four FLSs
Figure 6. The identification results of the four FLSs
consequent weights of DK-IT2FLS and DK-T1FLS are constrained, and this makes their
feasible parameter spaces smaller than those of D-IT2FLS and D-T1FLS.
Also, as stated above, the RMSE s i for evaluation data can reflect the generalization
abilities of dierent FLSs. From Figures 5(d) 5(f), we can see that DK-IT2FLS and DK-
T1FLS have better generalization ability than D-IT2FLS and D-T1FLS. This means that
DK-IT2FLS and DK-T1FLS can follow the characteristics of the real system dynamics
better than D-IT2FLS and D-T1FLS under noisy training circumstances.
In conclusion, the type-2 FLSs have better approximation ability than the type-1 FLSs,
but the FLSs designed using both sample data and prior knowledge have better general-
ization characteristics. The reason for this is that prior knowledge can prevent impossible
or implausible values, so that the designed FLSs will not be over fitted. Therefore, we
can say that the prior knowledge can improve the performance of the IT2FLSs.
6. Conclusions. This study has presented how to design IT2FLSs using the information
from both sampled data and prior knowledge. From the simulation results and compar-
isons, we can conclude that: 1) DK-IT2FLS is able to prevent impossible or implausible
values which conflict with the prior knowledge; 2) the prior knowledge may improve the
generalization ability of IT2FLSs. Here, we have just explored the prior knowledge of
bounded range, symmetry (odd and even) and monotonicity (increasing and decreasing).
How to utilize the other prior knowledge, such as fixed points, stability, etc., will be one
of our future research directions.
Acknowledgment. This work was partly supported by the NSFC Project under Grant
Nos. 60975060 and 61074149, National 863 Program (No. 2007AA04Z239), China. The
authors also gratefully acknowledge the helpful comments and suggestions of the reviewers.
REFERENCES
[1] Q. Liang and J. M. Mendel, Interval type-2 fuzzy logic systems: Theory and design, IEEE Transac-
tions on Fuzzy Systems, vol.8, pp.535-550, 2000.
[2] J. M. Mendel, Uncertain Rule-Based Fuzzy Logic Systems: Introduction and New Directions,
Prentice-Hall, Upper Saddle River, 2001.
[3] H. Hagras, A hierarchical type-2 fuzzy logic control architecture for autonomous mobile robots, IEEE
Transactions on Fuzzy Systems, vol.12, pp.524-539, 2004.
[4] J. M. Mendel, R. I. John and F. Liu, Interval type-2 fuzzy sets made simple, IEEE Transactions on
Fuzzy Systems, vol.14, pp.808-821, 2006.
[5] H. B. Mitchell, Ranking type-2 fuzzy numbers, IEEE Transactions on Fuzzy Systems, vol.14, pp.287-
294, 2006.
[6] J. M. Mendel, Advances in type-2 fuzzy sets and systems, Information Sciences, vol.177, pp.84-110,
2007.
[7] C. Li, J. Yi and D. Zhao, Design of interval type-2 fuzzy logic system using sampled data and prior
knowledge, ICIC Express Letters, vol.3, no.3(B), pp.695-700, 2009.
[8] M. A. Khanesar and M. Teshnehlab, Design of direct interval type-2 fuzzy model reference, ICIC
Express Letters, vol.3, no.4(A), pp.975-981, 2009.
[9] O. Castillo and P. Melin, Intelligent systems with interval type-2 fuzzy logic, International Journal
[10] J. M. Mendel and D. Wu, Perceptual reasoning for perceptual computing, IEEE Transactions on
Fuzzy Systems, vol.16, pp.1550-1564, 2008.
[11] X. Y. Wu and R. Srihari, Incorporating prior knowledge with weighted margin support vector ma-
chines, Proc. of the 10th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, Seattle, pp.326-333, 2004.
[12] F. Lauer and G. Bloch, Incorporating prior knowledge in support vector machines for classification:
A review, Neurocomputing, vol.71, pp.1578-1594, 2008.
[13] C. W. Chen and D. Z. Chen, Prior-knowledge-based feedforward network simulation of true boiling
point curve of crude oil, Computers and Chemistry, vol.25, pp.541-550, 2001.
et al., Incorporating prior knowledge into artificial neural networks
[14] S. Milanic, S. Strmcnik, D. Sel
An industrial case study, Neuralcomputing, vol.62, pp.131-151, 2004.
[15] W. H. Joerding and J. L. Meador, Encoding a priori information in feedforward networks, Neural
Networks, vol.4, pp.847-856, 1991.
[16] J. Abonyi, R. Babuska, H. B. Verbruggen and F. Szeifert, Incorporating prior knowledge in fuzzy
model identification, International Journal of Systems Science, vol.31, pp.657-667, 2000.
[17] P. F. van Lith, B. H. L. Betlem and B. Roel, Combining prior knowledge with data driven modeling
of a batch distillation column including start-up, Computers and Chemical Engineering, vol.27,
pp.1021-1030, 2003.
[18] P. Lindskog and L. Ljung, Ensuring monotonic gain characteristics in estimated models by fuzzy
model structures, Automatica, vol.36, pp.311-317, 2000.
[19] J. M. Won, S. Y. Park and J. S. Lee, Parameter conditions for monotonic Takagi-Sugeno-Kang fuzzy
system, Fuzzy Sets and System, vol.132, pp.135-146, 2002.
[20] C. Li, J. Yi and D. Zhao, Analysis and design of monotonic type-2 fuzzy inference systems, 2009
IEEE International Conference on Fuzzy Systems, Jeju Island, Korea, pp.1193-1198, 2009.
[21] C. Li, J. Yi and D. Zhao, Interval type-2 fuzzy neural network controller (IT2FNNC) and its ap-
plication to a coupled-tank liquid-level control system, Proc. of the 3rd International Conference on
Innovative Computing, Information and Control, Dalian, China, pp.508-511, 2008.
[22] C. Li and J. Yi, SIRMs based interval type-2 fuzzy inference systems: Properties and application,
International Journal of Innovative Computing, Information and Control, vol.6, no.9, pp.4019-4028,
2010.
c ISSN 1349-4198
GENERATION AND APPLICATION OF DECISION RULES WITHIN

DOMINANCE-BASED ROUGH SET APPROACH TO
MULTICRITERIA SORTING
Liping An1 , Zengqiang Chen2 and Lingyun Tong3

1
Business School
2
College of Information Technical Science
Nankai University
No. 92, Weijin Road, Tianjin 300071, P. R. China
{ anliping; chenzq }@nankai.edu.cn
3
School of Management
Hebei University of Technology
No. 8, Guangrong Road, Hongqiao District, Tianjin 300130, P. R. China
tonglingyun2008@sina.com
Abstract. We are considering the problem of multicriteria sorting. In this problem,

a set of if . . . then . . . decision rules is used as a preference model to sort objects
evaluated by a set of criteria. The rules can be induced from dominance-based rough
approximations of preference-ordered decision classes according to the Dominance-based
Rough Set Approach (DRSA). In this paper, we present an approach to generating the
minimal decision rules based on Boolean reasoning techniques. The concepts of decision
matrices and decision functions of the lower approximations and the boundary regions
are proposed to generate the certain rules and the approximate rules, respectively. Then,
the decision rules can be applied to predict sorting of new objects. An example is used
for illustration of the proposed method.
Keywords: Multi-criteria sorting, Rough sets, Dominance relation, Decision rules
1. Introduction. Both multiattribute classification and multicriteria sorting refer to the

assignment of a set of alternatives into predefined classes, they dier with respect to the
way that the classes are defined. Classification refers to the case where the classes are
defined in a nominal way. On the contrary, sorting refers to the case where the classes
are defined in an ordinal way [1]. The set of alternatives is evaluated by a set of criteria
and regular attributes. The dierence between criteria and regular attributes is such that
domains of criteria are preference ordered, while domains of regular attributes are not [2].
The wide range of real-world applications of the sorting problem has constituted a major
motivation for researchers in developing methodologies for constructing sorting models.
Recently, a complete and well-axiomatized methodology for constructing decision rule
preference models from decision examples based on the rough sets theory [3] has been
proposed. The decision rules are logic statements of the type if . . . , then . . . . The
set of decision rules represents a preference model of the decision maker who made the
classification decisions described by a data set. It is more general than the classical
functional or relational model [4,5] and it is more understandable for the users because of
its natural syntax.
The rough set approach and its extensions [6-8] provide useful tools for reasoning from
data. They are applied to various fields such as medicine, engineering, management and
economy, etc [9,10]. The original rough sets idea has proved to be particularly useful in
1145
1146 L. AN, Z. CHEN AND L. TONG
the analysis of multiattribute classification problems [11]. However, the classical rough
set theory is based on the indiscernibility relation, which is failing when attributes whose
domains are preference ordered (criteria) have to be taken into account. In order to
deal with problems of multicriteria sorting, a number of methodological changes to the
original rough sets theory were necessary [12]. Greco, Matarazzo and Slowinski [2,13]
have proposed an extension of the rough sets approach, called Dominance-based Rough
Set Approach (DRSA). DRSA is able to deal with sorting problems involving criteria. This
innovation is mainly based on substitution of the indiscernibility relation by a dominance
relation in the rough approximation of decision classes. Decision rules with a special
syntax are further induced from these approximations of decision classes.
As we know, the discernibility matrix of a decision table depicts completely the identi-
fication capability of the table. And, all rules of the table are hidden in the discernibility
function induced by the discernibility matrix [14]. For extracting decision rules from a
decision table, the technique of the discernibility function derived from the discernibility
matrix is sucient. Therefore, many theoretical methods for extracting decision rules
from a decision table have been developed around the construction of a suitable and sim-
ple discernibility function equivalent to the original one obtained directly from the table
[15].
In this paper, motivated by the above considerations, we define the outranking and
outranked relations of the single criterion in a decision table. The relations defined by
single criterion can be aggregated into global dominance and dominated relations at the
level of the set of criteria. Then, the lower and upper approximations of the upward
and downward unions of decision classes are defined based on the aggregated relations.
In order to induce the minimal decision rules used to support the decision task, the
decision matrices of a decision table with respect to the lower approximation and boundary
are defined to construct the decision functions which are Boolean functions. The set of
decision rules is decoded from prime implicants of the Boolean functions. The induced
rules can be used to explain relationships in the analyzed data set. The proposed approach
is an extension of the original approach of induction of decision rules from a decision table
in DRSA.
Moreover, the set of rules combined with a particular sorting method may be used to
sort new, unseen objects. Blaszczynski, Greco and Slowinski [16] proposed a new sorting
method based on variable-consistency decision rules. In this paper, both consistency and
inconsistency decision rules are applied to a new object.
The rest of the paper is organized as follows. Section 2 contains multicriteria sorting
problems statement and description of the problem via decision tables. And, we discuss
the outranking and outranked relations for the criteria. Then these binary relations
are aggregated into a global dominance or dominated relation. Section 3 considers the
definitions of rough approximations and some of their properties. A sorting decision
rule model induced from rough approximations is presented in Sections 4. In Section
5, according to the sorting method proposed by Blaszczynski, Greco and Slowinski, we
consider both consistency and inconsistency decision rules to sort an object. Section 6
gives an example illustrating the presented method. The last section concludes the paper.
2. Multicriteria Sorting Problems and Dominance Relations. A decision prob-
lem involves a set of objects described or evaluated by a set of attributes [11]. In par-
ticular, the sorting problem under consideration is to assign a finite set of m objects
U = {x1 , x2 , . . . , xm } into n predefined groups Cl1 , Cl2 , . . . , Cln . The objects are de-
scribed using a vector of p attributes A = (a1 , a2 , . . . , ap ). The performance of object xi
on attribute aj will be denoted by f (xi , aj ). Therefore, each object can be considered as
GENERATION AND APPLICATION OF DECISION RULES 1147
a vector ai = (f (xi , a1 ), f (xi , a2 ), . . . , f (xi , ap )). Furthermore, assuming that the decision
attribute d makes a partition of U into a finite number of decision classes, let CL =
{Clt , t T }, T = {1, 2, . . . , n}, be a set of decision classes of U , such that each x U
belongs to one and only one class Clt CL. In the case where the groups are defined in
an ordinal way, Cln will denote the group consisting of the most preferred objects, while
Cl1 will denote the group of the least preferred objects. We assume that for all r, s T ,
such that r > s, each element of Clr is preferred to each element of Cls [2]. The above
assumptions are typical for consideration of a multicriteria sorting problem.
The rough set theory assumes a representation of the information in a table form
called information table. The table is just an appropriate form for description of decision
situations.
An information table is a 4-tuple S = {U, A, V, f }, where U is a non-empty finite set
of objects, A is a non-empty finite set of attributes, Va is the domain of the attribute a,
V = Va and f : U A V is a total function such that f (x, a) Va for each a A
aA
and x U , called information function.
An attribute a A can be qualitative or quantitative. A qualitative attribute can
be nominal or ordinal. A nominal attribute is a discrete attribute whose values are not
necessarily in any order. The values of an ordinal attribute are preference ordered. A
quantitative attribute can be continuous or integer.
A decision table is any information table of the form S = (U, C {d}, V, f ), where C
is a set of condition attributes, d is a unique decision attribute and C {d} = .
In general, except nominal attributes, the preference order in the domain of a condition
attribute (criterion) has to be provided. The statement f (y, b) f (x, b) means that y is
at least as good as x with respect to criterion b.
Definition 2.1. Let S = (U, C {d}, V, f ) be a decision table. P C, b P , the
outranking relation of b on U , denoted by Db , is defined as: Db = {(x, y) U U :
f (x, b) f (y, b)}.
When (x, y) Db , we say that x is at least as good as y with respect to criterion b,
denoted by xDb y. The outranking class of an object x U , denoted by Db (x), is the set
of objects which are outranking x on b: Db (x) = {y U : yDb x}.
Definition 2.2. Let S = (U, C {d}, V, f ) be a decision table. P C, b P , the
outranked relation of b on U , denoted by Db , is defined as: Db = {(x, y) U U :
f (y, b) f (x, b)}.
When (x, y) Db , we say that x is at most as good as y with respect to criterion b,
denoted by xDb y. The outranked class of an object x U , denoted by Db (x), is the set
of objects which are outranked by x on b: Db (x) = {y U : yDb x}.
Proposition 2.1. Let Db be an outranking relation of b on U , Db be an outranked
relation of b on U , then we have the following properties:
xDb y yDb x x Db (y) y Db (x) (x, y) Db (y, x) Db .
The outranking relations defined by single criterion can be aggregated into a global
dominance relation at the level of the set of criteria.
Definition 2.3. Let S = (U, C {d}, V, f ) be a decision table. The dominance relation
between objects with respect to the set of criteria P C, denoted by DP , is defined as:
DP = {(x, y) U U : xDb y, b P }. The dominated relation of P on U , denoted by
DP , is defined as: DP = {(x, y) U U : yDb x, b P }.
If (x, y) DP , it is said that x dominates y with respect to P , denoted by xDP y. If

(x, y) DP , it is said that x is dominated by y with respect to P , denoted by xDP y.
Given P C and x U , the P -dominating set and P -dominated set with respect to
x, are defined respectively, as: DP (x) = {y U : yDP x}, DP (x) = {y U : xDP y}.
Proposition 2.2. The following statements hold:
DP = Db , DP = Db , DP (x) = Db (x), DP (x) = Db (x).
bP bP bP bP
Proposition 2.3. Let DP be a dominance relation of P on U , DP be a dominated relation

of P on U , then we the following properties:
xDP y yDP x x DP (y) y DP (x) (x, y) DP (y, x) DP .
3. Rough Approximations and Properties. This section deals with some basic con-
cepts of rough approximation and boundary by means of dominance relations and some of
their properties from [13,17]. And, the concept of inconsistent pair of objects is presented.
In DRSA, the sets to be approximated are upward and downward unions of classes and
the granules of knowledge are sets of objects defined using dominance relations.
Let S = (U, C {d}, V, f ) be a decision table. The sets to be approximated are called
the upward union and downward union of decision classes, respectively [2]:
Cl
t = Cls , Cl
t = Cls , t = 1, 2, . . . , n.
st st
The statement x Cl
t means x belongs at least to class Clt , while x Clt means

x belongs at most to class Clt . Observe that Cl1 = Cln = U ; Cln = Cln ; Cl1 = Cl1 .
Definition 3.1. The P -lower and the P -upper approximation of Cl t with respect to P
C are defined respectively, as: P (Clt ) = {x U : DP (x) Clt }, P (Cl

t ) = DP (x).
xCl
t
Definition 3.2. The P -lower and the P -upper approximation of Cl t with respect to P
C are defined respectively, as: P (Cl
t ) = {x U : D
P (x) Cl
t }, P (Cl
t ) = DP (x).
xCl
t
Definition 3.3. The P -boundary of Cl

t and Clt are defined respectively, as: BnP (Clt ) =
P (Cl
t ) P (Clt ), BnP (Clt ) = P (Clt ) P (Clt ).
The following basic properties of the rough approximations and boundary [13,17] are
satisfied:
Properties. For any t T and for any P Q C,
(1) Equivalent definitions of upper approximations: P (Cl
t ) = {x U : DP (x) Clt =
}, P (Cl
t ) = {x U : DP (x) Clt = }.
(2) Rough inclusion: P (Clt ) Clt P (Cl

t ), P (Clt ) Clt P (Clt ).
(3) Complementarity: P (Cl
t ) = U P (Clt1 ), P (Clt ) = U P (Clt1 ).
(4) Identity of boundaries: BnP (Cl
t+1 ) = BnP (Clt ).
(5) Monotonicity: P (Cl
t ) Q(Clt ), P (Clt ) Q(Clt ), P (Clt ) Q(Clt ), P (Clt )

Q(Cl t ).
(6) Monotonicity of boundaries: BnP (Cl
t ) BnQ (Clt ), BnP (Clt ) BnQ (Clt ).
The dierence between the upper and lower approximation constitutes the boundary
region of the rough set. The information about objects from the boundary region is in-
consistent because of limited discriminatory power of criteria or because of hesitation of
the decision maker (DM). The rough set approach is intended to deal with the inconsis-
tency which may convey important information that should be taken into account in the
construction of the preference model [2].
Definition 3.4. Let xm Cl
t and yn Clt1 , the pair (xm , yn ) is said to be inconsistent
if yn DP xm .
xm Cl
t and yn Clt1 mean that the DM comprehensively prefers xm over yn . From
Definition 3.4, when the pair (xm , yn ) is inconsistent, we can not explain why the DM
prefers xm over yn using the values of condition criteria, because yn is at least as good as
xm with respect to each criterion.
Theorem 3.1. Let xm Cl
t and yn Clt1 . If the pair (xm , yn ) is inconsistent, then
xm , yn BnP (Cl
t ).
Proof: From Definition 3.4, we have yn DP xm . From Proposition 2.3, yn DP (xm ).

Since yn Cl t1 , i.e., yn / Cl
t , we have DP (xm ) Clt . From Definition 3.1, xm /

P (Clt ). On the other hand, xm Clt , from Definition 3.1, xm P (Clt ). Then xm
P (Cl
t ) P (Clt ) = BnP (Clt ).

From Definition 3.4, we have xm DP yn . From Proposition 2.3, xm DP (yn ). Since

xm Cl t , i.e., xm / Cl
t1 , we have DP (yn ) Clt1 . From Definition 3.1, yn / P (Clt1 ).
On the other hand, yn Cl t1 , from Definition 3.1, y n P (Cl
t1 ). Then y n P (Cl
t1 )
P (Clt1 ) = BnP (Clt1 ). So, from the Property of Identity of boundaries, yn BnP (Cl

t ).
4. Rule Induction. We distinguish two types of decision rules: certain and approximate,
depending whether they are induced from lower approximations of decision classes or
from the boundaries of these classes composed of inconsistent examples. Certain rules
are generated from the P -lower approximation of unions of classes; approximate rules are
generated from the P -boundary of unions of classes.
Procedures for generating decision rules from a decision table use an inductive learning
principle. In order to induce decision rules with the consequent D, objects concordant
with D are called positive examples while all the others are called negative examples.
The following three types of decision rules can be considered:
(1) D -decision rules, which are generated from P (Clt ), having the following form:
IF (f (x, bj ) rbj ) THEN x Cl

t ,
j
where bj P , rbj Vbj .

(2) D -decision rules, which are generated from P (Cl
t ), having the following form:
IF (f (x, bj ) rbj ) THEN x Cl

t ,
j
where bj P , rbj Vbj .

(3) D -decision rules, which are generated from BnP (Cl
t ) and BnP (Clt ), having the
following form, respectively:
IF (f (x, bj ) rbj ) ((f (x, bk ) rbk )) THEN x BnP (Cl
t ),
j k
where bj , bk P , rbj Vbj , rbk Vbk . j and k are not necessarily dierent.
IF (f (x, bj ) rbj ) ((f (x, bk ) rbk )) THEN x BnP (Cl
t ),
j k
where bj , bk P , rbj Vbj , rbk Vbk . j and k are not necessarily dierent.
Generally, in supervised learning process, rules are generated for each class so as to
cover learning examples belonging to the class of positive examples while covering no, or
very few, learning examples belonging to other classes of negative examples. For example,
when the rules are generated for P (Cl
t ), the learning examples belonging to U P (Clt )
are negative examples.
Theorem 4.1. Suppose P (Cl t ) is the class of positive examples and U P (Clt ) is the

class of negative examples. If xm P (Cl

t ) and yn U P (Clt ), then, there exists
bj P such that: not xm Dbj yn , i.e., f (xm , bj ) f (yn , bj ) and f (xm , bj ) = f (yn , bj ).
Proof: Since xm P (Cl

t ) = {x U : DP (x) Clt }, we have DP (xm ) Clt . On

the other hand, yn U P (Cl / P (Cl

t ) yn / DP (xm ), which implies that:
t ), then yn
not xm DP yn , according to Proposition 2.3. From Definition 2.3, there exists bj P such
that: not xm Dbj yn .
From Theorem 4.1, bj has the ability to discern xm from yn . Analogously, we can prove
the following Theorem 4.2.
Theorem 4.2. Suppose P (Cl t ) is the class of positive examples and U P (Clt ) is the

class of negative examples. If xm P (Cl

t ) and yn U P (Clt ), then, there exists
bj P such that: not yn Dbj xm , i.e., f (yn , bj ) f (xm , bj ) and f (yn , bj ) = f (xm , bj ).
Theorem 4.3. Suppose BnP (Cl t ) is the class of positive examples and U BnP (Clt )

is the class of negative examples. If xm BnP (Cl

t ), yn1 (U BnP (Clt )) Clt and

yn2 (U BnP (Clt )) Clt1 , then there exists bj P such that: (1) not yn1 Dbj xm , i.e.,
f (yn1 , bj ) f (xm , bj ) and f (yn1 , bj ) = f (xm , bj ), and (2) not xm Dbj yn2 , i.e., f (xm , bj )
f (yn2 , bj ) and f (xm , bj ) = f (yn2 , bj ).
Proof: (1) Since yn1 U BnP (Cl / BnP (Cl
t ) y n1

t ) and yn1 Clt , we have
yn1 Cl
t BnP (Clt ) P (Clt )BnP (Clt ) = P (Clt ), which implies that DP (yn1 ) Clt .
Assume that yn1 Dbj xm for bj P , i.e., xm DP (yn ), then xm P (Cl t ), which is a

contradiction to xm BnP (Clt ).
(2) Since yn2 U BnP (Cl t ) yn2 / BnP (Cl
t ) and yn2 Clt1 , we have yn2
Cl
t1 BnP (Clt ) P (Clt1 ) BnP (Clt ) = P (Clt1 ) BnP (Clt1 ) = P (Clt1 ), which
implies that DP (yn2 ) Cl
t1 . Assume that xm Dbj yn2 for bj P , i.e., xm DP (yn2 ),
then xm P (Cl
t1 ), which is a contradiction to xm BnP (Clt1 ) = BnP (Clt ).
Analogously, we can prove the following Theorem 4.4.
Theorem 4.4. Suppose BnP (Cl t ) is the class of positive examples and U BnP (Clt )

is the class of negative examples. If xm BnP (Cl

t ), yn1 (U BnP (Clt )) Clt and

yn2 (U BnP (Clt )) Clt+1 , then there exists bj P such that: (1) not xm Dbj yn1 , i.e.,
f (xm , bj ) f (yn1 , bj ) and f (xm , bj ) = f (yn1 , bj ), and (2) not yn2 Dbj xm , i.e., f (yn2 , bj )
f (xm , bj ) and f (yn2 , bj ) = f (xm , bj ).
The above Theorems show that outranking and outranked relations of criteria can be
used to discern objects belonging to the class of positive examples from those belonging to
classes of negative examples. The ability to discern between perceived objects is important
for decision rule generation. And, the outranking or outranked relation of a criterion can
be obtained from the values of the criterion.
In classical rough set theory, all minimal decision rules can be generated using indis-
cernibility matrix and its Boolean functions [14]. The idea of Boolean reasoning is based
on construction for a given problem P of a corresponding Boolean function fP with the

following property: The solutions for the problem P can be decoded from prime impli-
cants of the Boolean function fP [18]. In this paper, the decision matrix [19] of P (Cl t )
and its decision functions are used to induce the minimal D -decision rules; the decision
matrix of P (Clt ) and its decision functions are used to induce the minimal D -decision
rules; and the decision matrix of BnP (Cl
t ) and BnP (Clt ) and their decision functions are
used to induce the minimal D -decision rules.
Definition 4.1. Let S = (U, C {d}, V, f ) be a decision table, P C, bj P . The
decision matrix of P (Cl
t ) in S, denoted by M(P (Clt )), is defined as: M(P (Clt )) =

(m
t (xm , yn ))|P (Cl )||U P (Cl )| , where mt (xm , yn ) = {bj : xm Dbj yn and f (xm , bj ) = f (yn ,
t t
bj ), where xm P (Cl
t ), yn U P (Clt )}.
From Theorem 4.1, the element m

t (xm , yn ) in M(P (Clt )) consists of all the criteria

that can discern xm from yn . Suppose bj mt (xm , yn ), then f (xm , bj ) can discern xm
from yn . This can be expressed as IF xDbj xm THEN x P (Cl t ). Our purpose is to
obtain the minimal rules that can discern xm P (Clt ) from yn U P (Cl

t ). This
can be expressed as IF (xDbj xm ) THEN x P (Cl t ).
yn
In order to obtain the minimal rules, we can construct the decision function f
t
(xm )

relative to an object xm from the decision matrix M(P (Clt )), as shown below.
Definition 4.2. The decision function of object xm with respect to M(P (Cl t )), denoted

by f t (xm ), is defined as: f t (xm ) = {bj : bj mt (xm , yn ) and mt (xm , yn ) = }, where
yn
bj is corresponding to the criterion bj .
The decision function f
t
(xm ) is a Boolean function that expresses how an object xm
P (Clt ) can be discerned from yn U P (Cl

t ). Turning f t (xm ) into disjunctive normal
form, the prime implicants of f t
(xm ) reveal the minimal subsets of P that are needed to
discern object xm from the objects that are not members of P (Cl t ) and correspond to the
minimal D -decision rules of which form is IF (f (x, bj ) f (xm , bj )) THEN x P (Cl t ),
j
where bj P .
Definition 4.3. Let S = (U, C {d}, V, f ) be a decision table, P C, bj P . The deci-
sion matrix of BnP (Cl
t ) in S, denoted by M(BnP (Clt )), is defined as: M(BnP (Clt )) =

(m
t (xm , yn1 ), mt (xm , yn2 ))|BnP (Cl )||U BnP (Cl )| , where mt (xm , yn1 ) = {bj : xm Dbj yn1
t t
and f (xm , bj ) = f (yn1 , bj ), where xm BnP (Cl

t ), yn1 (U BnP (Clt )) Clt },
and m
t (xm , yn2 ) = {bj : xm Dbj yn2 and f (xm , bj ) = f (yn2 , bj ), where xm BnP (Clt ),
yn2 (U BnP (Cl
t )) Clt1 }.
The element m
t (xm , yn1 ) in M(BnP (Clt )) consists of all the criteria that can discern
xm from yn1 . Suppose bj m t (xm , yn1 ), then f (xm , bj ) can discern xm from yn1 . The
element mt (xm , yn2 ) in M(BnP (Cl

t )) consists of all the criteria that can discern xm from

yn2 . Suppose bj mt (xm , yn1 ), then f (xm , bj ) can discern xm from yn2 . This can be
expressed as IF (xDbj xm ) ( (xDbj xm )) THEN x BnP (Cl t ).
yn1 yn2
Definition 4.4. The decision function of object xm with respect to M(BnP (Cl t )), denoted
by ft (xm ), is defined as: ft (xm ) = {bj : bj m t (x ,
m n1y ) m
t (xm , yn2 ), and
yn1 ,yn2
m
t (xm , yn1 ) = or m
t (xm , yn2 ) = }, where bj is corresponding to the criterion bj .
Turning ft (xm ) into disjunctive normal form, the prime implicants of the simplified
decision function correspond to the minimal D -decision rules.
Analogously, using P (Cl
t ) and BnP (Clt ), we can define the decision matrix M(P (Clt ))
and its decision functions f
t
(xm ), and M(BnP (Cl
t )) and its decision functions ft (xm ),
respectively.
Definition 4.5. Let S = (U, C {d}, V, f ) be a decision table, P C, bj P . The
decision matrix of P (Cl
t ) in S, denoted by M(P (Clt )), is defined as: M(P (Clt )) =

(mt (xm , yn ))|P (Cl )||U P (Cl )| , where mt (xm , yn ) = {bj : xm Dbj yn and f (xm , bj ) = f (yn ,
t t
bj ), where xm P (Cl
t ), yn U P (Cl
t )}.
Definition 4.6. The decision function of object xm with respect to M(P (Cl t )), denoted

by f t (xm ), is defined as: f t (xm ) = {bj : bj mt (xm , yn ) and mt (xm , yn ) = }, where
yn
bj is corresponding to the criterion bj .
Definition 4.7. Let S = (U, C {d}, V, f ) be a decision table, P C, bj P . The deci-
sion matrix of BnP (Cl
t ) in S, denoted by M(BnP (Clt )), is defined as: M(BnP (Clt )) =

(mt (xm , yn1 ), mt (xm , yn2 ))|BnP (Cl )||U BnP (Cl )| , where mt (xm , yn1 ) = {bj : xm Dbj yn2
t t
and f (xm , bj ) = f (yn2 , bj ), where xm BnP (Cl

t ), yn2 (U BnP (Clt )) Clt+1 },
and m
t (xm , yn2 ) = {bj : xm Dbj yn1 and f (xm , bj ) = f (yn1 , bj ), where xm BnP (Clt ),
yn1 (U BnP (Cl
t )) Clt }.
Property 4.1. M(BnP (Cl

t ))=M(BnP (Clt1 )).
Proof: m
t (xm , yn1 ) = {bj : xm Dbj yn1 and f (xm , bj ) = f (yn1 , bj ), where xmBnP (Clt1 ),
yn1 (U BnP (Cl
t1 )) Clt } = mt1 (xm , yn1 ).
m
t (xm , yn2 ) = {bj : xm Dbj yn2 and f (xm , bj ) = f (yn2 , bj ), where xm BnP (Clt1 ), yn2
(U BnP (Cl
t1 )) Clt1 } = mt1 (xm , yn2 ).

Then, M(BnP (Clt )) = (mt (xm , yn1 ), m
t (xm , yn2 ))|BnP (Cl )||U BnP (Cl )| = (mt1 (xm ,
t t
yn1 ), m
t1 (xm , yn2 ))|BnP (Cl = M(BnP (Cl
t1 )).
t1 )||U BnP (Clt1 )|

Definition 4.8. The decision function of object xm with respect to M(BnP (Cl t )), denoted
by ft (xm ), is defined as: ft (xm ) = {bj : bj m t (x ,
m n1y ) m
t (x m , yn2 ), and
yn1 ,yn2
m
t (xm , yn1 ) = or m
t (xm , yn2 ) = }, where bj is corresponding to the criterion bj .
5. Sorting Method. Blaszczynski, Greco and Slowinski [16] proposed a sorting method
which can be applied to sort an object using decision rules. Let R(z) be the set of decision
rules covering a given object z. Condi denotes the set of objects verifying the condition
part of rule i . For object z and set of decision rules R(z) covering it, a value of ScoreR (Clt ,
z) is calculated for each decision class Clt from CL [16]:

R (Clt , z) ScoreR (Clt , z).
ScoreR (Clt , z) = Score+
Object z is then assigned to the class Clt for which the value of ScoreR (Clt , z) is the
greatest. Score+ R (Clt , z) takes into account decision rules i R(z) (i = 1, 2, . . . , k)
that suggest assignment of object z to the union of classes Cl
s and Clq such that Clt
Cl
s and Clt Clq , respectively. ScoreR (Clt , z) is defined as follows: ScoreR (Clt , z) =
+ +
|(Cond1 Clt )(Condk Clt )|

2
, where Cond1 , , Condk are the sets of objects verifying

|Cond1 Condk ||Clt |
condition parts of rules i .
ScoreR (Clt , z) takes into account decision rules i R(z) (i = k +1, . . . , h) that suggest
assignment of object z to the union of classes Cl
s and Clq such that Clt Cls = and
Clt Cl
q = , respectively. ScoreR (Clt , z) is defined as follows:
Score R (Clt , z)
2

(Condk+1 Cl k+1 ) (Cond l Cl
l ) (Cond l+1 Cl
l+1 ) (Cond h Cl h
)
=

,

Condk+1 Condl Condl+1 Condh Cl k+1 Cl
l Cl
l+1 Cl h
where Cl
k+1 , . . . , Cll are all the upward unions of decision classes suggested for assign-
ment of object z by rules i (i = k + 1, . . . , l) and Cl
l+1 , . . . , Clh are all the downward
unions of decision classes suggested for assignment of object z by rules i (i = l+1, . . . , h).
6. An Example. The example [2] shown in Table 1 concerns six students described by
means of four criteria: a, level in Mathematics; b, level in Physics; c, level in Literature;
d, global evaluation (decision class). We denote P = {a, b, c}. According to the decision
attribute, the students are divided into two preference-ordered classes: Cl1 = Clbad =
{x2 , x4 , x5 }, Cl2 = Clgood = {x1 , x3 , x6 }. Thus, the following unions of classes were
approximated: Cl
2 = {x1 , x3 , x6 }, Cl1 = {x2 , x4 , x5 }. The P -dominating set and the
P -dominated set of xi , i = 1, 2, . . . , 6, are computed as shown in Table 2.
Table 1. Decision table

Student a b c d
x1 good good bad good
x2 medium bad bad bad
x3 medium bad bad good
x4 bad bad bad bad
x5 medium good good bad
x6 good bad good good
Table 2. The P -dominating set and the P -dominated set
Student DP (xi ) DP (xi )

x1 {x1 } {x1 , x2 , x3 , x4 }
x2 {x1 , x2 , x3 , x5 , x6 } {x2 , x3 , x4 }
x3 {x1 , x2 , x3 , x5 , x6 } {x2 , x3 , x4 }
x4 {x1 , x2 , x3 , x4 , x5 , x6 } {x4 }
x5 {x5 } {x2 , x3 , x4 , x5 }
x6 {x6 } {x2 , x3 , x4 , x6 }
Then, the P -lower approximations, the P -upper approximations and the P -boundaries
of classes Cl
2 and Cl1 are equal, respectively, to: P (Cl2 ) = {x1 , x6 }, P (Cl2 ) = {x1 , x2 , x3 ,
x5 , x6 }. P (Cl
1 ) = {x4 }, P (Cl1 ) = {x2 , x3 , x4 , x5 }, BnP (Cl2 )=BnP (Cl1 ) = {x2 , x3 , x5 }.

The decision matrix of P (Cl2 ) is the following:
x2 x3 x4 x5
M(P (Cl x1 ab ab ab a .
2 )) =
x6 ac ac ac a
By M(P (Cl
2 )), the decision functions f 2 (xm ), m = 1, 6, are constructed to obtain the
minimal rules. For example, f 2
(x1 ) = (a b) (a b) (a b) a. Turn f 2
(x1 ) into

disjunctive normal form, we have f 2 (x1 ) = a, i.e., the prime implicant of f 2 (x1 ) is a.
Then, the corresponding minimal decision rule is the following: f (x, a) good x
Cl2.
x1 x2 x3 x5 x6
Similarly, M(P (Cl 1 )) = x . The corresponding minimal de-
4 |ab a a abc ac|
cision rule is f (x, a) bad x Cl
1 . And, M(BnP (Cl2 )) = M(BnP (Cl1 )) =

x1 x6 x4
x2 ab ac a
. The decision function, for example, f2 (x2 ) = (a b) (a
x3 ab ac a
x5 a a abc
c)a = a. The corresponding minimal decision rule is the following: f (x, a) medium
f (x, a) medium x BnP (Cl 2 ), i.e., f (x, a) = medium x BnP (Cl2 ).

The final minimal set of certain and approximate decision rules, shown as follows, can
be obtained from the considered decision table.
1 : f (x, a) good Cl 2 , (x1 , x6 )

2 : f (x, a) bad Cl1 , (x4 )
3 : f (x, a) = medium BnP (Cl 2 ), (x2 , x3 , x5 )
We consider two new objects represented by y1 and y2 .
y1 : f (y1 , a) = bad, f (y1 , b) = good, f (y1 , c) = good, and
y2 : f (y2 , a) = medium, f (y2 , b) = bad, f (y2 , c) = good.
The decision rule covering y1 and y2 are, respectively, 2 and 3 .
The sorting method calculates the certainty of assignment of each object to each par-
ticular decision class separately.
|{x4 }{x2 ,x4 ,x5 }|2
The score coecients of y1 for class Cl1 are: Score+ R (Cl1 , y1 ) = |{x4 }||{x2 ,x4 ,x5 }| = 0.33,
and Score R (Cl1 , y1 ) = 0, ScoreR (Cl1 , y1 ) = 0.33. The score coecients of y1 for class Cl2
|{x4 }{x2 ,x4 ,x5 }|2
are: Score+ R (Cl2 , y1 ) = 0 and ScoreR (Cl2 , y1 ) = |{x4 }||{x2 ,x4 ,x5 }| = 0.33, ScoreR (Cl2 , y1 ) =
0.33.
As the value of score for class Cl1 is the greatest, the object is assigned to class Cl1 .

The score coecients of y2 for class Cl1 are: Score+ R (Cl1 , y2 ) = 0.44, ScoreR (Cl1 , y2 ) =
0.11 and ScoreR (Cl1 , y2 ) = 0.33. The score coecients of y2 for class Cl2 are: Score+ R (Cl2 ,

y2 ) = 0.11, ScoreR (Cl2 , y2 ) = 0.44 and ScoreR (Cl2 , y2 ) = 0.33.
The assignment of object y2 to class Cl1 gains an approval.
7. Conclusions. We presented a method for inducing the set of decision rules from
rough approximations of preference-ordered decision classes, according to DRSA. The set
of rules induced in this way constitutes a preference model of the decision maker who gave
sorting examples. The main idea of the presented method resolves itself into incorporating
Boolean reasoning into DRSA. We proposed the concepts of decision matrix and decision
function which are notions analogous to those of discernibility matrix and discernibility
function well-known in the context of classical rough set approach. The decision rules
can be applied to predict sorting of new objects. It is shown that the decision rule model
resulting from the extended rough set approach has an advantage over the usual functional
and relational models because it permits handling inconsistent sorting examples.
The above definitions show basic features from theoretical points of view. Most of the
problems related to generation of the mentioned above entities are NP-complete or NP-
hard. However, it was possible to develop ecient heuristics solutions of the problems.
The ecient algorithms for generating the entities needs further research.
Acknowledgment. This work was supported by National Natural Science Foundation

of China (Nos. 71071079, 70601013).
REFERENCES
[1] C. Zopounidis and M. Doumpos, Multicriteria classification and sorting methods: A literature review,
European Journal of Operational Research, vol.138, no.2, pp.229-246, 2002.
[2] S. Greco, B. Matarazzo and R. Slowinski, Rough set theory for multicriteria decision analysis, Eu-
ropean Journal of Operational Research, vol.129, no.1, pp.1-47, 2001.
[3] Z. Pawlak, Rough sets, International Journal of Computer and Information Sciences, vol.11, no.5,
pp.341-356, 1982.
[4] R. L. Keeney and H. Raia, Decisions with Multiple Objectives: Preferences and Value Trade-Os,
2nd Edition, Cambridge University Press, Cambridge, 1993.
[5] B. Roy and D. Bouyssou, Aide Multicrit`ere ` a La Decision: Methodes Et Cas, Economica, Paris,
1993.
[6] Z. Pawlak and A. Skowron, Rudiments of rough sets, Information Sciences, vol.177, no.1, pp.3-27,
2007.
[7] Y. Matsumoto and J. Watada, Knowledge acquisition from time series data through rough sets
analysis, International Journal of Innovative Computing, Information and Control, vol.5, no.12(B),
pp.4885-4898, 2009.
[8] B. Crossingham, T. Marwala and M. Lagazio, Evolutionarily optimized rough sets partitions, ICIC
[9] J. Komorowski, Z. Pawlak, L. Polkowski and A. Skowron, Rough sets: A tutorial, in Rough Fuzzy
Hybridization: A New Trend in Decision-Making, S. Pal and A. Skowron (eds.), Singapore, Springer-
Verlag, 1999.
[10] C.-Y. Huang, G.-H. Tzeng, C.-C. Chan and H.-C. Wu, Semiconductor market fluctuation indicators
and rule derivations using the rough set theory, International Journal of Innovative Computing,
Information and Control, vol.5, no.6, pp.1485-1504, 2009.
[11] Z. Pawlak and R. Slowinski, Rough set approach to multi-attribute decision analysis, European
Journal of Operational Research, vol.72, no.3, pp.443-459, 1994.
[12] S. Greco, B. Matarazzo and R. Slowinski, The use of rough sets and fuzzy sets in MCDM, in Advances
in Multiple Criteria Decision Making, T. Gal, T. Stewart and T. Hanne (eds.), Dordrecht, Kluwer,
1999.
[13] S. Greco, B. Matarazzo and R. Slowinski, Rough sets methodology for sorting problems in presence
of multiple attributes and criteria, European Journal of Operational Research, vol.138, no.2, pp.247-
259, 2002.
[14] A. Skowron and C. Rauszer, The discernibility matrices and functions in information table, in
Intelligent Decision Support: Handbook of Applications and Advances of the Rough Set Theory,
R. Slowinski (ed.), Dordrecht, Kluwer, 1991.
[15] M. W. Shao and W. X. Zhang, Dominance relation and rules in an incomplete ordered information
system, International Journal of Intelligent Systems, vol.20, no.1, pp.13-27, 2005.
[16] J. Blaszczynski, S. Greco and R. Slowinski, Multi-criteria classification A new scheme for appli-
cation of dominance-based decision rules, European Journal of Operational Research, vol.181, no.3,
pp.1030-1044, 2007.
[17] S. Greco, B. Matarazzo and R. Slowinski, Rough approximation by dominance relations, Interna-
tional Journal of Intelligent Systems, vol.17, no.2, pp.153-171, 2002.
[18] Z. Pawlak and A. Skowron, Rough sets and boolean reasoning, Information Sciences, vol.177, no.1,
pp.41-73, 2007.
[19] L. An and Z. Chen, Generation of dominance-based decision rules, ICIC Express Letters, vol.3,
no.3(B), pp.745-750, 2009.
c ISSN 1349-4198
AUTOMATIC CLASSIFICATION OF DRIVING MENTAL FATIGUE

WITH EEG BY WAVELET PACKET ENERGY AND KPCA-SVM
Chunlin Zhao1,2 , Chongxun Zheng1 , Min Zhao1

Jianping Liu2 and Yaling Tu1
1
Key laboratory of Biomedical Information Engineering of Education Ministry
University of Xian Jiaotong
No. 28, Xianning West Road, Xian 710049, P. R. China
{ zclwj1974; zhaoclzm }@163.com; cxzheng@mail.xjtu.edu.cn; tuyaling 0119@yahoo.com.cn
2
Engineering College of Armed Police Force
Wujing Road, Weiyang District, Xian 710086, P. R. China
liujp96@sohu.com
Abstract. Driving mental fatigue is a main cause of some serious transportation acci-
dent and it has drawn increasing attention in recent years. In this study, an automatic
measurement of driving mental fatigue based on the Electroencephalographic (EEG) is
presented. Fifteen healthy subjects who performed continuous simulated driving task for
90 minutes with EEG monitoring are included in this study. The feature vectors of
ten-channel EEG signal on prefrontal, frontal, central, parietal and occipital regions are
extracted by wavelet packet transform. Kernel principal component analysis (KPCA) and
support vector machines (SVM) are jointly applied to identify two driving mental fatigue
states. The results show that wavelet packet energy (WPE) of EEG is strongly correlated
with mental fatigue level on prefrontal frontal central and occipital regions. Moreover,
the KPCA method is able to eectively reduce the dimensionality of the feature vectors,
speed up the convergence in the training of SVM and achieve higher recognition accuracy
(98.7%). The KPCA-SVM could be a promising candidate for developing robust auto-
matic mental fatigue detection systems for driving safety.
Keywords: EEG, Driving mental fatigue, WPE, KPCA, SVM
1. Introduction. Mental fatigue is one of the major causes of serious accidents, espe-
cially in transportation and aviation area and it is believed to account for 20 30% of all
trac accidents [1]. Experts agree that the actual contributions of driver mental fatigue
to road accidents may be much higher [2]. Developing and establishing an accurate and
non-invasive real-time system for monitoring drivers mental fatigue is quiet important to
reduce the number of fatigue-related crash and to lower social cost in trac safety. Driv-
ing simulator studies have dominated these types of studies, mainly because of safety, low
cost, well-controlled conditions and easiness of data collection [3].
To data, many methods have been developed for detecting the drivers mental fatigue
in the recent few years, including the measurements of physiological features like EEG,
event related potential (ERP), heart rate and pulse rate, heart rate variability (HRV),
electrooculographic (EOG), eyelid movement, gaze, head movement, facial tone and be-
haviors of the vehicle, such as lane deviations and steering movements, etc. Among those
dierent physiological parameters, the EEG is quite promising for monitoring driving
mental fatigue [4].
Recently, the theory of machine learning has developed a wide variety of novel learning
and classification algorithms, which provides an opportunity to develop a non-intrusive
1157
1158 C. ZHAO, C. ZHENG, M. ZHAO, J. LIU AND Y. TU
driving mental fatigue automatic recognition system based on EEG with higher accuracy.
In particular, Kernel-based nonlinear feature extraction and classification algorithms are
a popular new research direction in machine learning [5]. Successful applications of kernel-
based algorithm have been reported from various fields, for instance in the de-noising in
feature space, face recognition, time-series prediction and so on [6,7]. The advantage of
using kernel method is that it is neither necessary to know the form of the function, nor
to calculate the dot product in the higher-dimensional space due to fact that an inner
product in the feature space has an equivalent kernel in the input space [8].
As well known, driving mental fatigue is a complex phenomenon which is relative to
nerve central activity. It must look at activity distributed over the entire scalp in or-
der to detect brain state in the round while mental fatigue is occurring [9]. A large
body of studies reported that wavelet packet is a powerful tool for EEG signal analyzing.
Wavelet packet transform has been widely used in various areas of nonstationary EEG
signal processing [10]. In this study, wavelet packet energy (WPE) is employed to extract
multi-channel EEG signal features. Two kernel-based learning machines are extended
to develop an EEG-based driving mental fatigue monitoring system. Considering the
high-dimensional and nonlinear characters of EEG data, the high-dimensional features
extracted by WPE are reduced by Kernel principal component analysis (KPCA). Then
the dimensionality-reduced feature vectors are considered as input data for support vector
machine (SVM) which is suitable to analyze biomedical data for its good generalization
capacity independent of the input space dimension [11].
This paper is organized as follows. Section 2 describes the experimental materials and
methods from three aspects: the simulated driving tasks and EEG data collection, the
feature extraction with EEG signals based on WPE and the background about KPCA
and SVM. Section 3 reports the results. Finally, Discussion and Conclusion are given in
Sections 4 and 5 respectively.
2. Materials and Methods.
2.1. Simulated driving design. To reduce inter-subject and age dierences, fifteen
male students (between 21 25 years with mean age 23.2 years) were recruited to partic-
ipate in simulated driving study. The driving task is performed on a driving simulator.
It is made based on a real car dimensions and consists of a soft car seat, a steering wheel,
gas pedal, brake pedal, clutch, manual shift, a horn and turn signal, etc. The driving
operation is the same as a real car. A larger display is in front of car seat to display
the virtual reality-base driving environment. Driving training is performed in the previ-
ous day before the experiment. No further training is given if subject could drive three
consecutive laps without any accident.
The experiments began at 9:00 11:30 AM or 3:00 5:30 PM in a sound-attenuated
and temperature-controlled laboratory. This was corresponding with the normal work-
time. Before the experiment, the subjects learned the whole procedure to well understand
the procedure and the instructions, and the psychological self-report measures of mental
fatigue were conducted. The driving task was conducted subsequently. During the driving
task, subjects performed in the simulator for 90 minutes without any break. Subjects
were asked to restrict all unnecessary movements as much as possible and try their best
to maintain constant speed and avoid car accident. There were no any questionnaires
during the experiment. At the end of all experiment sessions, the psychological self-
report measure of fatigue was also carried out. The epochs of the experimental beginning
(0 5min) and end (85 90min) were selected to investigate dierent mental fatigue
states.
AUTOMATIC CLASSIFICATION OF DRIVING MENTAL FATIGUE WITH EEG 1159
2.2. Data acquisition. The Neuroscan Synamps System was used to record the phys-
iological data. Thirty channels of EEG, one channel of ECG and vertical EOG were
measured simultaneously while participants were performing experimental tasks. The 10
20 international standard of electrode placement was applied. The unipolar reference
region was linked at right and left earlobe and the ground electrode was located at AFz.
The impedance was kept below 5k for EEG region and 10k for EOG and ECG region.
All physiological signals were sampled at 500Hz with 0.05Hz 70Hz band-pass filter and
50Hz notched.
The raw EEG data were inspected visually and then the possible residual artifacts were
eliminated. An EEG larger than 100uV was treated as artifact. EOG artifact was removed
by using VEOG signals as predictors of the artifact voltages at each EEG electrode in
a multivariate linear regression. Finally, the EEG data were filtered at 0.5 30Hz to
remove the line noise (50Hz and its harmonic) and other high-frequency noise for further
analysis.
2.3. Feature extraction based on wavelet packet analysis. Wavelet packet analysis
can provide more subtle information about signals not only in approximation space but
also the detail space [12]. Define (t) and (t) as the scaling function and the correspond-
ing mother wavelet function respectively. Supposing 0 (t) = (t), 1 (t) = (t). Then
the wavelet basis can be constructed as follows according to two-scale equations.
( )
1 2i 2j k t
j,k (t) =
2i
j
= i
h(n)j1,2kn (t) (1)
2 2 n
( j )
1 2 k t
2i+1
j,k (t) = 2i+1 j
= i
g(n)j1,2kn (t) (2)
2 2 n
where i is the node number, j is the level of decomposition. h(n) and g(n) = (1)1n h(1
n) are a pair of quadrature mirror filters group. Base on the Equations (1) and (2), the
decomposed wavelet packet coecients with dierent jth level and kth point of signal
f (t) can be calculated by recursion formula as follows:

2i
dj (k) = 2i
f (t)j,k (t)dt = h(n)dij1 (2k n) (3)
n

d2i+1
j (k) = 2i+1
f (t)j,k (t)dt = g(n)dij1 (2k n) (4)
n
The sub-signal of the sub-band space on the jth level can be reconstructed by:

s(t)2i
j = d2i 2i
j (k)j,k (t) (5)
k
2i
Since the wavelet j,k (t) is an orthogonal basis at L2 (R), the jth level scale-energy of
the sub signal s(t)2i
j is calculated by:
2
Ej = d2i
j (k)
(6)
k
In consequence, the total energy can be obtained by:

Etotal = Ej (7)
j
Then the relative WPE for the resolution level j, denoted by pj , is defined as
Ej
pj = (8)
Etotal
which quantifies the probability distribution of the spectral energy [13].
After artifacts elimination, 30-seconds continuous EEG data in the session of beginning
and end are selected to be analyzed. The first 10s EEG is chosen as basic data segment
and steps by 0.2 second size. By shifting the basic data segment with step length, for each
subject 200 data segments are obtained. On the other hand, multi-dimension features are
extracted from EEG after eight-octave wavelet packet decomposition, which consist of rel-
ative WPE in delta (0.5 3.5Hz), theta (4 7Hz), alpha (8 12Hz), beta (13 30Hz) fre-
quency bands and the ratio indices beta/alpha, theta/alpha, (alpha+theta)/beta. These
EEG indices are reported to driver mental fatigue [14]. EEG data of ten electrodes (fp1,
fp2, f3, f4, c3, c4, p3, p4, o1, o2) produced 70 dimensions feature totally.
2.4. SVM classifier. Support vector machines (SVMs) developed by Vapnik are known
to be a powerful tool for generating pattern recognition systems with high generalization
ability [15]. In an SVM, the input space is mapped into a high dimensional feature space
firstly and the optimal hyperplane that maximizes a margin between two classes is de-
termined. SVM is a supervised learning method used for classification and regression.
Unlike other statistical learning methods (such as neural networks and decision trees)
which usually aim only to minimize the empirical classification error, SVM simultane-
ously minimizes the empirical classification error and maximizes the geometric margin in
classification. This is done by constructing a hyperplane which optimally separates sam-
ples from two classes (for minimum empirical classification error) with maximum margin
(for maximum classification margin); hence it is also known as maximum margin classifier
[16].
Firstly, a brief review of SVM is given as follows: Given the training set {(xk , yk )}1kl ,
xk Rq which belongs to a class labeled by yk {1, +1}. Let the weight vector and
the bias of the separating hyperplane be w and b respectively, the objective of SVM is to
maximize the margin of separation and minimize the errors, formulated as the constrained
optimization problem:
C 2
l
1 T
min w w +
w,b, 2 2 i=1 i (9)
subject to yk (wT (xk ) + b) + i 1, i, i 0,
where : x z, is a nonlinear mapping which maps the data into higher dimensional
feature space form the original input space, using a kernel function which satisfies Mercers
condition [15]. i is a slack variable representing the error measures of data. The error
weight C (penalization factor) is a free parameter. The trade-o between the maximal
l
margin and the minimal training error i2 is controlled by the regulation parameter C.
i=1
Usually they can be solved through dual problem

l
1
l
1 2
l
W () = i i j yi yj K(xi , xj )
i=1
2 i,j=1 2C i=1 i
(10)

l
1
l
= i i j yi yj KC (xi , xj )
i=1
2 i,j=1

l
with the constraints: i yi = 0 and i 0.
i=1
According to Equation (10), the modified kernel function is obtained as follows:
[ ]
1
KC (xi , xj ) = [KC ]ij = K + I (11)
C ij
Then the error weight C (penalization factor) acts as the variable of kernel function.
2.5. KPCA-SVM classification system. When SVM maps input data into a high
dimensional feature space, the dimension of input vector directly aects the performance
of the SVM classifier. Therefore, dimension reduction is necessary for an ecient classi-
fier, especially for large data sets. Thus, the first important step for SVM algorithm is
feature extraction (new features are transformed from the original inputs) [17]. Principal
component analysis (PCA) is a well-known method for feature extraction. By calculating
the eigenvectors of the covariance matrix of the original inputs, PCA linearly transforms
a high-dimensional input vector into a low-dimensional one whose components are un-
correlated. KPCA, proposed by Scholkopf et al., is a type of nonlinear PCA developed
by generalizing the kernel method into PCA. The basic idea is to map the original input
vectors into a high-dimensional feature space and then to calculate the linear PCA in
the feature space [18]. In this study, KPCA algorithm is used to reduce the dimension-
ality of features which are extracted by WPE firstly. Then SVM algorithm is applied
to distinguish the driving mental fatigue states. The schematic diagram for KPCA-SVM
framework is shown in Figure 1.
Figure 1. The schematic diagram for KPCA-SVM
3. Results.
3.1. Self-report analysis. Subjects all felt tired, bored and drowsy when the driving
task was over. They were yawning occasionally during the driving task and reported
diculties to concentrate and focus their attention on the driving task. Thirteen of them
reported that above-mentioned feeling appeared just only after 30 min of driving and
became stronger with driving time increasing. These evidences indicated that driving
mental fatigue was induced after 90 min driving task.
3.2. Wavelet packet parameters of EEG. To eliminate the influences of parameters
fluctuation, the mean value of 10 electrodes EEG was calculated to be statistically ana-
lyzed. The comparison results of wavelet packet parameters of EEG between two sessions
are shown in Table 1. (For convenience only statistically significant results are given).
Compared with the beginning epoch, the mean value of relative WPE in theta and alpha
bands significantly increases (P < 0.05), in beta bands significantly decreases (P < 0.05)
on prefrontal, frontal, central and occipital electrodes. The mean value of theta/alpha
and (alpha+theta)/beta increases significantly (P < 0.05) on prefrontal, frontal and cen-
tral electrodes. The mean value of relative WPE in delta band and the mean value of
beta/alpha on all electrodes do not change significantly at the end of the task. The results
indicate that the wavelet packet features of EEG are closely related with driving mental
fatigue.
Table 1. The wavelet packet parameters of EEG (mean, STD)
WPE the beginning the end

Channel name value of t, p
parameters (mean, STD) (mean, STD)
Fp1 0.128 (0.029) 0.148 (0.035)* t = 2.51, p = 0.021
Fp2 0.129 (0.013) 0.153 (0.059)* t = 2.41, p = 0.033
F3 0.148 (0.035) 0.161 (0.040)* t = 2.31, p = 0.038
F4 0.144 (0.032) 0.164 (0.055)* t = 2.22, p = 0.045
Theta
C3 0.138 (0.027) 0.274 (0.026)* t = 2.18, p = 0.049
C4 0.138 (0.026) 0.150 (0.030)* t = 2.53, p = 0.019
O1 0.113 (0.025) 0.117 (0.029)* t= 2.24, p=0.045
O2 0.113 (0.027) 0.121 (0.020)* t = 2.34, p = 0.041
Fp1 0.422 (0.046) 0.400 (0.033)* t = 2.61, p = 0.019
Fp2 0.418 (0.032) 0.398 (0.037)* t = 2.35, p = 0.035
F3 0.380 (0.037) 0.367 (0.046)* t = 2.20, p = 0.047
beta
C3 0.374 (0.057) 0.360 (0.051)* t = 2.93, p = 0.013
C4 0.370 (0.047) 0.363 (0.055)* t = 2.85, p = 0.015
O1 0.443 (0.054) 0.463 (0.041)* t = 3.50, p = 0.004
Fp1 0.104 (0.014) 0.112 (0.013)* t = 2.94, p = 0.012
Fp2 0.106 (0.010) 0.114 (0.018)* t = 2.24, p = 0.045
F3 0.103 (0.027) 0.109 (0.026)* t = 2.32, p = 0.039
alpha
C4 0.108 (0.020) 0.112 (0.018)* t = 2.83, p = 0.015
O1 0.135 (0.013) 0.123 (0.012)* t = 2.36, p = 0.024
O2 0.129 (0.021) 0.130 (0.020)* t = 2.18, p = 0.048
Fp1 1.209 (0.156) 1.487 (0.184)* t = 2.69, p = 0.012
Fp2 1.170 (0.199) 1.516 (0.146)* t = 2.33, p = 0.035
F3 1.429 (0.212) 1.622 (0.151)* t = 2.20, p = 0.043
Ratio1
F4 1.378 (0.236) 1.639 (0.160)* t = 2.53, p = 0.019
C3 1.283 (0.223) 1.414 (0.132)* t = 2.24, p = 0.040
C4 1.307 (0.358) 1.437 (0.159)* t = 2.23, p = 0.039
Fp1 0.602 (0.053) 0.659 (0.057)* t = 2.35, p = 0.035
Ratio2 F4 0.694 (0.073) 0.752 (0.043)* t = 2.93, p = 0.013
C3 0.711 (0.061) 0.738 (0.086)* t = 2.85, p = 0.015
(Ratio1: theta/alpha; Ratio2: (alpha+theta)/beta; *: p < 0.05)
3.3. The classification results. In order to distinguish dierent mental fatigue states,
the EEG sample sets includes 200 data segments for each subject are selected to analyze. A
5-fold cross-validation test is applied. For each subject, the data sample set was randomly
averaged into five portions, each trail uses one portion as testing data and the other four as
training data, and dierent trials used dierent testing sets. The classification accuracy
is calculated over 5 trials. For each subject we constructed a KPCA-SVM to identify
the norm and fatigue states. WPE features are extracted from EEG after eight-octave
wavelet packet decomposition. As a comparison, the EEG features extracted by fast
Fourier transform (FFT) are chosen as the input for the original SVM. Figure 2 shows
the classification result based on SVM without using KPCA (Original SVM) between
WPE and FFT.
The original SVM can dierentiate the norm and mental fatigue states eectively and
the average max-accuracy across all subjects achieves 98.77% (STD 1.35%) with WPE
and 96.16% (STD 2.57%) with FFT respectively (t = 4.04, p = 0.001). It clearly can be
seen that the results of WPE is statistically better than that of FFT.
Figure 2. Classification accuracies for 15 subjects and the average result

across subjects (Dark-grey bars represent the accuracy based on WPE; dark
bars represent FFT)
In order to improve classification performance and speed up the calculation, KPCA is

applied to the EEG data segments. Previous study showed that compared with other
kernel function (such as polynomial and sigmoid kernel functions), the Gaussian kernel
function was the preferred without a prior knowledge of the data set [19]. In this study,
Gaussian function is selected as the kernel function for KPCA, dierent kernel parameter
can aect accuracy of classification. Kernel parameter is chosen from 1 to 20. When
is adding to 18, the average accuracy reaches its maximum to 99.2%.
Figure 3. The various accuracy with features dimensionalities

For SVM algorithm, the kernel function is very important for SVM since it defines the
nature of the decision surface, dierent kernel function and parameters can have widely
varying performance [20]. We compared the results under dierent kernel function (linear,
polynomial and sigmoid kernel functions), for our experimental data, the best result was
obtained under the Gaussian kernel function. The eects of dierent kernel parameters
were small since section of input variables was avoided, the fixed kernel parameter was
chosen in this study. Figure 3 illustrates that the average accuracy across all subjects
varies with the dierent number of the feature dimensionalities reduced by KPCA. When
the dimensionality is more than 30, KPCA-SVM shows a better performance than that of
original SVM. The maximum accuracy (98.8%) is obtained while the number of feature
dimensions equals to 35. On the other hand, LPCA-SVM gains the maximum accuracy
(98.3%) while the number of feature dimensions equals to 30. KPCA-SVM also shows
better performance than that of LPCA-SVM under same dimensionality of features.
The following three measures, accuracy (Ac), specificity (Sp) and sensitivity (Se) are
used to assess the performance of classifiers [21]:
TP + TN
Accuracy = 100%
TP + FP + TN + FN
TN
Specif icity = 100%
TN + FP
TP
Sensitivity = 100%
TP + FN
where T P is the number of true positives, T N is the number of true negatives, F P is false
norm identifications and F N is false fatigue identification. The specificity is defined as
the ability of the classifier to correctly recognize a fatigue state. The sensitivity indicates
the classifiers ability not to generate a false detection (normal state). Table 2 compared
these performances of dierent classifying algorithm.
Table 2. The best performance of dierent classifying algorithm
Algorithm Ac(%) Sp(%) Se(%) n*

KPCA-SVM 98.8 99.3 98.4 35
LPCA-SVM 98.3 99.0 97.7 30
Original SVM 98.7 99.1 98.3
BP 93.2 92.6 93.8
*: n is the number of principle components.
The kernel functions are Gaussian function.
According to the access criteria, SVM shows better performance than that of the back
propagation (BP) neural network. KPCA can extract more ecient features that are
useful for the classification purpose. Therefore, the every access index of KPCA-SVM is
better than that of LPCA-SVM. The KPCA combined SVM can obtain the best perfor-
mance with 98.8% accuracy and decrease computing time greatly. Thus, it is an eective
approach to classify two states of driving mental fatigue.
4. Discussion. The self-report is an important physiological measurement for driving

mental fatigue. The results from this study clearly revealed the decrease in alertness and
increase in mental fatigue, evidenced by significant increase in fatigue feeling at the end
of task. It can be judged that the 90 min simulative driving task induces mental fatigue
to the subjects indeed.
It has been known for many years that the change in brain arousal involves specific
changes in oscillatory brain activity and the EEG can reflect the fluctuation of alertness
level. EEG is a non-stationary signal due to many causes, such as epileptogenic tran-
sients, physiological artifacts and changes of the physiological state of the subject [22].
Wavelet analysis is one of the useful method for the analyzing the non-stationary pro-
cess and we expect it can generate spectral resolution fine enough to meet the problem
requirement. In this study, the most of WPE and relative ratio parameters on the pre-
frontal, frontal, central and occipital regions showed significant dierences between the
beginning and the end of driving task. Smith and Givens have demonstrated regional
dierences in the manner in which particular task demands modulate the EEG spectrum
[23]. Some brain regions are of particular importance in maintaining the primary driving
performance of event detection during a secondary conversation task. Several researchers
have confirmed this conclusion by using both fMRI and MEG methods [24]. Maybe it
is a good explanation for the EEG rhythm changes on dierent scalp regions. At the
same time, relative WPE of alpha and theta rhythms significantly increased and relative
WPE of beta rhythm significantly decreased (P < 0.05). These results were consistent
with the previous research finding that is the increase of alpha energy indicated that the
number of activating cortical neurons was decreasing with driving duration, and the cor-
tical activity was inhibiting; the increase of theta energy was the sign of onset of slight
sleepy; the decrease of beta energy implied that the brain arousal level declined [25]. The
changes of alpha, beta, theta rhythm and their ratio were consistent with driving mental
fatigue status. It is suggested that the alertness ability of subject decreased after long
time driving task, which was proved by increases in slow waves of EEG and decreases in
fast waves of EEG. Compared with EEG features extracted by FFT, WPE overcomes the
drawback of FFT, i.e. the spectral decomposition with Fourier transform cannot detect
the time-varying EEGs rhythms. That is the reason why the accuracy becomes higher
when FFT is replaced by WPE. The higher accuracy of classification based on WPE indi-
cates two things: the WPE parameters are more sensitive to the characteristics of normal
and mental fatigue states than FFT parameters and these parameters are of statistical
significant dierence; the original SVM shows good performance in distinguishing two
mental states, it is suitable tool for EEG classification.
For the high-dimensionality of WPE feature, additional vector space transformation
may be applied on the initial features. The transformation is not only capable of ex-
tracting good WPE features (i.e., reduce the dimensionality of the WPE data), but also
provide additional discriminative information for improving classification performance.
In this study, we compared KPCA and LPCA as pre-processing tools of WPE features.
LPCA is based on second-order correlations between data and it can only be eectively
performed on a set of observations that vary linearly. When the variations are nonlinear,
the data can always be mapped into a high-dimensional space in which they vary linearly.
That is, according to Covers theorem, the nonlinear data structure in the input space is
more likely to be linear after high-dimensional nonlinear mapping [26]. It is well known
that EEG does reflect weak but significant nonlinear structure [27], and that, therefore,
all the interesting EEG characteristics cannot be described by second-order correlations.
KPCA was capable of extracting more eigenvectors, which could contain more relevant
information about the intrinsic nonlinear dynamics of mental fatigue states, than LPCA.
The reason is that for EEG data with random noise, linear PCA only discards the def-
inite variance with noise by projecting EEG data on the main principal components.
For KPCA, the same is obtained in higher-dimensional feature space by using nonlinear
principal components, but KPCA does extract a markedly larger number of nonlinear
principal components and allows spreading the information about the EEG data struc-
ture into more nonlinear principal components. Giving the potential of discarding some
of the eigendirections where the noisy part of data resides. So the classification accuracy
became better when KPCA was employed. When applying KPCA, the kernel parameter
can aect classification accuracy. We should select the appropriate to get better
result. More investigation is needed to demonstrate how to choose kernel parameter to
get the better performance. At the same time, when KPCA and LPCA were applied to
reduce the dimensionality of features, the iteration steps were less than original SVM to
reach the convergence.
The support vector machines (SVM) has been a well-known machine learning method,
and it is based on the structural risk minimization principle which seeks to minimize an
upper bound of the generalization error consisting of the sum of the training error and
a confidence interval. It overcomes the problem for ERM (empirical risk minimization)
posing by optimizing the object consisting of the learning error on the training samples
and the capacity of hypothesis space. A remarkable property of SVM is its good gen-
eralization capacity independent of the input space dimension. The high generalization
ability of SVM makes it particularly suited for high dimensionality data analysis, such
as multi-channel EEG recording, where many features can be redundant due to inherent
redundancy of the data. Meanwhile, another key characteristic of SVM is that training
SVM is equivalent to solving a linearly constrained quadratic programming problem so
that the solution of SVM is always unique and globally optimal, at the same time, it avoids
local minima and over-fitting in the training process. BP network training based on em-
pirical risk minimization principle which requires nonlinear optimization with the danger
of getting stuck into local minima [28]. So the original SVM shows superior performance
to that of BP network. Although the dimensionality of feature is few, the KPCA-SVM
capability is beyond the original SVM. The low-dimensional feature representation pre-
processed by KPCA would accelerate the speed and improve the accuracy of the classifier.
Combined KPCA with SVM, it obtains the best performance with the highest average ac-
curacy (98.8%) over all subjects. The high accuracy of classification indicates two things:
the WPE feature of physiological parameters are sensitive to the characteristics of normal
and mental fatigue states and these parameters are of statistical significant dierence; the
classification algorithm shows good performance in distinguishing two mental states, it is
suitable tool for EEG classification.
5. Conclusions. In this study, we designed a simulated driving task with monotonous

and repetitious highway scene in VR-based laboratory. The relative WPE of dierent
EEG rhythm has shown significant dierence between the beginning and the end of driv-
ing task. There is a close relationship between driving mental fatigue and the physiological
parameters. These indices can reflect driving mental fatigue status objectively and may
provide a feasible approach to identify the state of mental fatigue under simulated en-
vironment. When subjects perform driving task, the findings in simulated driving can
be generalized and yield important suggestions regarding countermeasures against driv-
ing mental fatigue. KPCA is a good feature extractor, and KPCA-SVM shows the best
performance with higher accuracy and lower computation cost. In conclusion, combining
machine learning algorithms with physiological measurements identified during driving
mental fatigue provide useful insights for future development of a countermeasure device
and mental fatigue management.
Acknowledgment. The project is supported by National Science Foundation of China

under grant No. 30870654. The authors also gratefully acknowledge the helpful comments
and suggestions of the reviewers, which have improved the presentation.
REFERENCES
[1] NHTSA, Drowsly Driver Detection and Warning System for Commercial Vehicle Drives: Field Pro-
portional Test Design, Analysis, and Progress, http://www.nhtsa.dot.gov.
[2] S. K. L. Lal, A. Craig, P. Boord, L. Kirkup and H. Nguyen, Development of an algorithm for an
EEG-based driver fatigue countermeasure, J. Saf. Res., vol.34, no.3, pp.321-328, 2003.
[3] M. P. Reed and P. A. Green, Comparison of driving performance on-road and in a low-cost simulator
using a concurrent telephone dialling task, Ergonomics, vol.42, no.8, pp.1015-1037, 1999.
[4] S. K. L. Lal and A. Craig, A critical review of the psychophysiology of driver fatigue, J. Biol.
Psychol., vol.55, no.3, pp.173-194, 2001.
[5] L. J. Cao, K. S. Chua, W. K. Chong, H. P. Lee and Q. M. Gu, A comparison of PCA, KPCA
and ICA for dimensionality reduction in support vector machine, Neurocomputing, vol.55, no.1-2,
pp.321-336, 2003.
[6] A. R. Teixeira, A. M. Tome, K. Stadlthanner and E. W. Lang, KPCA denoising and the pre-image
problem revisited, Digit Signal Process, vol.18, no.4, pp.568-580, 2008.
[7] A. Kocsor and L. Toth, Kernel-based feature extraction with a speech technology application, IEEE
Transactions on Signal Processing, vol.52, no.8, pp.2250-2263, 2004.
[8] K. R. Muller, S. Mika, G. Ratsch, K. Tsuda and B. Scholkopf, An introduction to kernel-based
learning algorithms, IEEE Trans. Neural Network, vol.12, no.2, pp.181-201, 2001.
[9] C. W. Anderson, E. A. Stolz and S. Shamsunder, Multivariate autoregressive models for classification
of spontaneous electroencephalographic signals during mental tasks, IEEE Trans. Biomed. Eng.,
vol.45, no.3, pp.277-286, 1998.
[10] L. Sun, G. Chang and P. J. Beadle, Multiresolution of clinical EEG recordings based on wavelet
packet analysis, Proc. of ISNN, LNCS, vol.4492, pp.1168-1176, 2007.
[11] Z. Chen, W. Hong and C. Wang, RNA secondary structure prediction with plane pseudoknots based
on support vector machine, ICIC Express Letters, vol.3, no.4(B), pp.1411-1416, 2009.
[12] S. Mallat, A Wavelet Tour of Signal Processing, 2nd Edition, Academic Press, San Diego, 1999.
[13] O. A. Rosso, M. T. Martin and A. Plastino, Brain electrical activity analysis using wavelet-based
informational tools, Phys. Stat. Mech. Appl., vol.313, no.3-4, pp.587-608, 2002.
[14] S. K. L. Lal and A. Craig, Driver fatigue: Electroencephalography and psychological assessment,
Psychophysiology, vol.39, no.3, pp.313-321, 2002.
[15] V. N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, New York, 1995.
[16] H. Tamura and K. Tanno, Midpoint validation method for support vector machines with margin
adjustment technique, International Journal of Innovative Computing, Information and Control,
vol.5, no.11(A), pp.4025-4032, 2009.
[17] N. Acir, A support vector machine classifier algorithm based on a perturbation method and its
application to ECG beat recognition systems, Expet. Syst. Appl., vol.31, no.1, pp.150-158, 2006.
[18] B. Scholkopf, A. Smola and K. R. Muller, Nonlinear component analysis as a kernel eigenvalue
problem, Neural Comput., vol.10, no.5, pp.1299-1319, 1998.
[19] B. Scholkopf and A. J. Smola, Learning with Kernels Support Vector Machines, Regularization,
Optimization, and Beyond, MIT Press, 2001.
[20] Y. Xing, X. Wu and Z. Xu, Multiclass least squares auto-correlation wavelet support vector machines,
[21] C. C. C. Pang, A. R. M. Upton, G. Shine and M. V. Kamath, A comparison of algorithms for
detection of spikes in the electroencephalogram, IEEE Trans. Biomed. Eng., vol.50, no.4, pp.521-
526, 2003.
[22] I. Clark, R. Biscay, M. Echeverria and T. Virues, Multiresolution decomposition of nonstationary
eeg signals A preliminary-study, Comput. Biol. Med., vol.25, no.4, pp.373-382, 1995.
[23] M. E. Smith and A. Gevins, Neurophysiologic monitoring of mental workload and fatigue during
operation of a flight simulator, Proc. of the SPIE, vol.5797, pp.116-126, 2005.
[24] S. M. Bowyer, L. Hsieh, J. E. Moran, R. A. Young et al., Conversation eects on neural mechanisms
underlying reaction time to visual events while viewing a driving scene using MEG, Brain Research,
vol.1251, pp.151-161, 2009.
[25] A. S. Gevins and R. E. Schaer, A critical-review of electroencephalographic (Eeg) correlates of

higher cortical functions, CRC Crit. Rev. Bioeng., vol.4, no.2, pp.113-164, 1980.
[26] J.-M. Lee, C. Yoo, S. W. Choi, P. A. Vanrolleghem and I.-B. Lee, Nonlinear process monitoring
using kernel principal component analysis, Chem. Eng. Scie., vol.59, no.1, pp.223-234, 2004.
[27] C. J. Stam, Nonlinear dynamical analysis of EEG and MEG: Review of an emerging field, Clin.
Neurophysiol., vol.116, no.10, pp.2266-2301, 2005.
[28] Y. E. Shao and B.-S. Hsu, Determining the contributors for a multivariate SPC chart signal using ar-
tificial neural networks and support vector machine, International Journal of Innovative Computing,
Information and Control, vol.5, no.12(B), pp.4899-4906, 2009.
c ISSN 1349-4198
3D TRIANGLE MESH COMPRESSION BASED ON VECTOR

QUANTIZATION WITH k-RING VECTOR PREDICTION
Lin-Lin Tang1 , Zhe-Ming Lu2, , Fa-Xin Yu2

Ping-Hui Wang2 and Zhen Li3
1
Harbin Institute of Technology Shenzhen Graduate School
HIT Campus of Shenzhen University Town, Xili, Shenzhen 518055, P. R. China
linlintang2009@gmail.com
2
School of Aeronautics and Astronautics
Zhejiang University
Hangzhou 310027, P. R. China

Corresponding author: zheminglu@zju.edu.cn
3
School of Electrical and Electronic Engineering
Nanyang Technological University
Singapore
LIZH0019@ntu.edu.sg
Abstract. The transmission and storage of large amounts of triangle and vertex ge-
ometry data are required for rendering geometrically detailed 3D meshes. To alleviate
bandwidth requirements, this paper uses vector quantization (VQ) as an eective lossy
vertex data compression technique for triangle meshes with high rate-distortion perfor-
mance. The proposed novel VQ-based vertex encoding algorithm adopts a region growing
based k-ring prediction scheme. During the encoding process, the value k related to the
current vertex to be encoded is estimated by the prediction errors of the preceding encoded
vertices, and then the adaptive prediction rule for the current vertex is obtained according
to the value k. Therefore, the dynamic range of the residual vectors generated by the pro-
posed prediction scheme is reduced. Thus it is expected to achieve a higher quantization
quality with the same codebook size. Experimental results show that, compared with the
vertex encoding algorithm based on the conventional parallelogram prediction scheme, the
proposed algorithm can achieve a higher encoding quality at the same bit rate.
Keywords: Computer graphics, 3D meshes, Vertex data compression, Vector quantiza-
tion, k-ring prediction
1. Introduction. Interactive 3D graphics plays an important role in various fields such

as entertainment, manufacturing and virtual reality. When combining the graphics tech-
nology with the Internet, the transmission delay for 3D graphics data is becoming a major
performance bottleneck, especially for meshes consisting of millions of triangles. Under
the limited network bandwidth, as well as the storage problem within host systems, re-
ducing the amount of data is, go without saying, an eective solution. Consequently, the
interests in compression techniques for the 3D geometry data have surged in recent years.
Although many representation methods for 3D meshes exist, a triangle is the basic
geometric primitive for standard 3D graphics applications, graphics rendering hardware
and many simulation algorithms, and any other surface based representations can be
easily converted into triangle meshes. Triangle meshes are composed of two components:
vertex data and connectivity data. Vertex data mainly include the positional coordinates
1169
1170 L.-L. TANG, Z.-M. LU, F.-X. YU, P.-H. WANG AND Z. LI
of the vertices and, optionally, some other attributes. The connectivity data supply the
topological information that specifies the connectivity of vertices.
Connectivity data compression has been well studied. In 1999, Rossignac proposed the
popular Edgebreaker coder [1], which compressed the connectivity data to be at least 4
bits per vertex (in short, bpv). Since then, its derivatives [2-6], and some other techniques
[7-10] have been also proposed. A complete survey of these approaches is shown in [11].
At present, connectivity compression reaches its limit of less than two bits per triangle
for the connectivity portion of a mesh. On the other hand, relatively few attentions
have been paid to the geometry portion of a triangle mesh. In addition, with hardware-
oriented mesh representations such as triangle strips [12,13] or generalized triangle meshes
[14,15], each vertex is often specified multiple times due to the storage constraints of
hardware decoding. Therefore, it is important to develop eective compression techniques
for the vertex data in order to further reduce overall mesh representation and geometry
bandwidth requirements.
In this paper, we are engaged in the use of vector quantization (VQ) [16,17] as an
eective compression technique for vertex geometry data. VQ is a popular technique for
data compression, and has already been extensively applied in audio, still image and video
coding. It has many advantages over entropy coding, for example, high rate-distortion
performance, real-time decompression for its simple decoding and hardware tractability.
Furthermore, we assume that the new triangle can be reconstructed by the neighboring
previously reconstructed triangles, so the compatibility is maintained for most existing
connectivity compression schemes.
The rest part of the paper is organized as follows. Section 2 briefly reviews the prior
work on vertex data compression. Section 3 describes the proposed VQ based compression
scheme, followed by the simulation results in Section 4. Finally, Section 5 concludes the
paper.
2. Prior Work. Although the uncompressed geometry data is often given in the precise
IEEE floating point representation for vertex positions at 96.0 bpv, some applications
may tolerate the reduction of this precision to achieve higher compression ratios, e.g.
the current Java3D geometry compression specification [18]. In this section, we briefly
summarize several existing schemes for compressing geometry data.
The early works usually quantize, uniformly and separately, the vertex positions for
each coordinate in the Cartesian space. In 1995, Deering et al. first proposed a vertex
data compression scheme [14], where positions are first normalized within an axis-aligned
bounding box. The coordinates are then uniformly quantized to k bits precision with
entropy coding based on a modified Human code. Deering reported results ranging
from 36 bpv for a model with coordinates quantized at 16-bits precision, to 17 bpv for
a model with coordinates quantized at 10-bits precision. Since then, many variations of
Deerings scheme have also been proposed [7,14,15,19,20], and more sophisticated vector
quantization based schemes have also been proposed in [21,22]. Karni and Gotsman [23]
demonstrated the relevance of applying the quantization technique in the space denoted
by spectral coecients. Sorkine et al. [24] addressed the issue on how to reduce the visual
eect due to quantization errors.
As for the prediction scheme, the early work [14] employed simple delta coding or linear
prediction along the vertex order dictated by the coding of the connectivity data [7,19,20].
The approach proposed by Lee et al. [25] achieves better visual appearance by applying
quantization in the angle space after prediction. Inspired by the TG parallelogram pre-
diction scheme [7], Isenburg and Alliez [26] completed the techniques described in [27,28]
by generalizing it to polygon mesh geometry compression. In [29], prediction trees were
3D TRIANGLE MESH COMPRESSION BASED ON VECTOR QUANTIZATION 1171
employed to improve the prediction where the geometry drives the traversal order instead
of the connectivity as before. Cohen-Or et al. [11] suggested a multi-way prediction
technique, where each vertex position is predicted from all its neighboring vertices, as
opposed to the one-way parallelogram prediction. In the approach proposed by Shikhare
et al. [30], the redundancy is removed by detecting similar geometric patterns, involving
much computation complexity.
According to the application practice of VQ in signal compression, we know that pre-
dictive VQ outperforms basic VQ in encoding the vector sequence with high correlation
between consecutive vectors. That is, we should consider the neighboring vertices for
encoding the current vertex. Existing vector quantization based methods are either with
poor rate-distortion performance or with high complexity. The poor rate-distortion perfor-
mance is ascribed to poor prediction schemes typically with only 3 neighboring vertices.
To predict current vertex with more neighboring vertices with a relatively simple VQ
structure, this paper proposes a novel k-ring vector prediction scheme, which uses the
adaptive number of neighboring vertices to predict the current vertex, as given in next
section.
3. Proposed Method. Conventional VQ methods compress vertex positions without

considering minimizing the prediction errors. We propose a vector quantization (VQ)
scheme which optimally employs the correlation between the vertex to be encoded and
its adjacent encoded vertices and thus oers superior performance over conventional par-
allelogram prediction based methods, as verified in the simulation results.
Conceptually, VQ is a generalization of non-uniform scalar quantization to operate
on vectors rather than scalars, oering superior performance over scalar quantization
in terms of rate-distortion [31]. Vector quantization can be defined as a map from a
q-dimensional Euclidian space to a finite subset, i.e. Q: Rq C, where the subset
C = {ci |i = 1, 2, , N } is called a codebook,
( and ci is a) codevector and N is the
codebook size. For each input ( vector x = x0 ,)x1 , , x(q1) , the quantizer selects an
appropriate codevector cp = cp0 , cp1 , , cp(q1) as its decoded vector, guaranteeing that
the codevector cp is the closest vector to x among all the codevectors in C. The distance
metric is usually the Squared Euclidean measure as follows:

q1
d(x, ci ) = x ci 22 = (xj cij )2 (1)
j=0
The above process is lossy since dierent input vectors may be mapped to the same
approximating codevector. The quantized input vector can then be specified by p, i.e.
the index of the codevector cp , which is an integer that can be encoded as a fixed-length
binary index and thus all indices are transmitted over the network in place of the original
geometry data. Compared with traditional entropy geometry compression schemes rely-
ing on variable-length encoding, fixed-length coding can result in better error-resiliency
performance and lower decoder hardware complexity. The VQ decoding process is highly
ecient, requiring only table lookup operations. Recently, other than the application
field of data compression, VQ has also been successfully used in other fields such as image
hashing [32], video watermarking [33], image retrieval [34] and face detection [35].
3.1. Vector quantizer design. The first issue in designing a VQ scheme for compressing
any kind of source is how to map the source data into a vector sequence as the input of
the vector quantizer. For two-dimensional signals such as images, the vector sequence is
commonly formed from blocks of neighboring pixels. The blocks can be directly used as the
input vector for the VQ quantizer. In the case of triangle meshes, neighboring vertices are
also likely to be correlated. However, blocking multiple vertices is not as straightforward

as the case for images. The coordinate vector of a vertex cannot be directly regarded as
an input vector to the quantizer in that if multiple vertices are mapped into the same
vertex, the distortion of the mesh will be unacceptable and the connectivity of the mesh
will also disappear.
Based on above considerations, we employ the approach known as predictive vector
quantization (PVQ) [36]. The block diagram of the PVQ encoder and decoder is shown
in Figure 1. To exploit the correlation between vertices, it is necessary to use a vector
quantizer with memory. Let {vi }N i=0 represent the sequence of vertices and vn be the next
vertex to be encoded. The encoder forms a prediction vn for vn based on the observations
of previously encoded vertices. The residual vector en , representing the prediction error,
is then computed as en = vn vn . This residual vector is then quantized by the vector
quantizer, which generates a codevector en that approximates en . The index identifying
this codevector is then stored or transmitted to the decoder. Each vertex is thus encoded
with log2 N bits, where N is the codebook size.
(a) Encoder
(b) Decoder
Figure 1. PVQ encoder and decoder
However, the prediction error between vn and vn may be accumulated to the subsequent
encoded vertices. When the number of vertices in a mesh is large enough, the accumulated
error may be unacceptable. To permit reconstruction of the vertices by the decoder, the
prediction must only be based on previous reconstructed vertices. Thus, the encoder also
needs to replace each processed original vertex with its reconstructed vertex in order to
compute prediction vectors for subsequent vertices.
The decoder receives the sequence of indices. Given an index, the decoder first performs
a table lookup operation to obtain the residual vector en . The decoder then adds en and
vn to reconstruct the quantized vertex vn . As in the encoder, vn is fed back to the

predictor for computing subsequent vertex predictions.
For the quantizer design, the residual vectors are used as training vectors to generate
the codebook based on the directed-search binary splitting (DSBS) technique [37]. A brief
summary of the DSBS is as follows: the process begins with the centroid of the training
vectors. Then the principal component analysis (PCA) [38] technique is applied to find
the approximate distribution shape for each cluster and the direction along which the
present codevectors are doubled is estimated so their positions are closer to the optimum
positions. When the codebook size reaches the desired size, the codebook generation
procedure is terminated.
3.2. Predictor design. There are two components in a PVQ system: prediction and
residual vector quantization. We first discuss the design of the predictor. The goal of
the predictor is to minimize the variance of the residual vectors, as well as maintaining
low computation complexity, allowing them to be coded more eciently by the vector
quantizer.
The proposed k-ring prediction scheme is based on the principle of the famous parallelo-
gram prediction rule [26], which is intuitively illustrated in Figure 2. In the parallelogram
prediction scheme, a vertex is predicted based on its neighboring triangle, enabling ex-
ploitation of the tendency for neighboring triangles to be roughly coplanar and similar
in size. This is particularly true for high-resolution, scanned meshes, which have little
variation in the triangle size. Suppose that vn1 , vn2 and vn3 in Figure 2 are the
three vertices of a neighboring triangle buered by the encoder. Then, an eective and
computationally inexpensive way to compute the prediction vn for vn is:
vn = vn1 + vn2 vn3 (2)
Figure 2. Parallelogram prediction scheme
However, for more jugged 3D meshes, Equation (2) is not sucient. Inspired by Touma
and Gotsmans idea [7], which describes an estimate of the local face curvature rather
than assuming triangles are coplanar, we propose a k-ring prediction scheme here, in
which en is minimized by observing the k-ring neighboring encoded vertices to achieve a
better prediction.
For vi M , i.e., a vertex from the 3D mesh M whose 3D coordinates are given by the
vector vi , we define its r-ring neighborhood as all the vertices with an r-edge distance
from it:
Nr (vi ) = { vj | |vj vi | = r} (3)
where |vj vi | denotes the least number of edges required to travel from vi to vj , and
r = 0, 1, . The point vi is considered as its own 0-ring neighborhood. Thus we can
represent vn and all vertices in the neighborhoods within k-rings as follows:
(m) (m) (m)
vni = vni + eni m = 0, 1, . . . , Li 1 and i = 0, 1, . . . , k (4)
(m) (m) (m)
where vni is the m-th vertex in the i-ring neighborhood of vn , while vni and eni are
its prediction vector and prediction error based on Equation (2), respectively. Here, Li
denotes the number of vertices in the i-ring neighborhood, and obviously L0 = 1 and
(0)
vn = vn .
Before the deduction, we assume that all the triangles in the mesh are equilateral, which
is an ordinary assumption in 3D mesh analysis. We define ni as the average prediction
(m)
error of eni , m = 0, 1, . . . , Li 1 for all encoded vertices in the i-ring neighborhood of
(0)
vn . It is certainly that if vn is the vertex to be encoded, then n = en .
If the mesh surface is rather planar, we can estimate that ni = 0, i = 0, 1, . . . , k;
While the shape of the surface becomes a bit complex, ni may becomes a constant, i.e.
n = 0, and thus
n = n1 (5)
When the shape of the mesh surface is more complex, namely, ni varies more dra-
matically, 2 n = 0 may occurs, and thus n = n1 i.e.,
n = 2n1 n2 (6)
With the mesh surface being more and more complex, if the average prediction error
satisfies the following condition:
k n = 0 (7)
Based on the mathematical induction, we can easily draw the following conclusion:

k
n = Cki ni (1)i+1 (8)
i=1
where Cki is the combination of i out of k. It is obvious that Equations (5) and (6) are
specific instances of Equation (8). Now we can represent vn as follows when Equation (7)
is satisfied:

k
vn = vn + n = vn + Cki ni (1)i+1 (9)
i=1
where vn is the newly predicted vector of vn , while ni is the average prediction error of
(m)
eni that are computed by Equation (4).
Now Equation (4) is updated to the following equation:
vn = vn + n (10)
where n is the error vector between vn and vn , and is certainly of less magnitude than
en that is generated by the conventional parallelogram prediction scheme. Parallelogram
prediction is evidently equivalent to the case k = 0 in the proposed algorithm. n is
then quantized and represented by its corresponding codevector index. The smaller error
vectors are, the higher quantization quality is achieved.
Up to now, the prediction rule is deduced and an appropriate k is needed to be es-

timated. In this work, we limit k to be 0, 1 or 2 in order to reduce the computational
complexity while maintaining the prediction precision. Thus n1 and n2 are sucient
for estimating k. We first define the average magnitude and the variance of n1 and
n2 as follows:
= (n1 2 + n2 2 )/2 (11)
= n1 n2 2 /2 (12)
If is very small, it is obvious that n1 n2 0; if is very small, it can be
approximated that n = 0; If and are both with larger values, n is nearly a
constant and thus 2 n 0. In other cases, n should be computed with a k larger than
2, for simplicity, we set k to be 0 here. So we regulate the k estimation rule as follows:

0 if < 0

1 else if < 0
k= (13)

2 else if < 1
0 else
where 0 < 1 are two variance thresholds.

3.3. Residual vector quantization. There are many VQ techniques that could be
employed for quantizing the residual vectors. In this paper, we focus on the conventional
unconstrained VQ in which a unique codebook is trained for each mesh as pre-processed
data. This enables the codebook to be tailored to the unique residual distribution of the
mesh.
Codebook training can be performed in either an open-loop or closed-loop manner [32].
In the open-loop approach, the unquantized mesh is traversed and the predictor is applied,
and then the resulting residual vectors are saved as training vectors. The main drawback
of the open-loop approach is that the prediction residual vectors to be quantized during
the actual encoding process may be statistically dierent from those used for training
the codebook. The closed-loop method attempts to mitigate this problem by training
the codebook in multiple iterations. A closed-loop method, known as the asymptotic
closed-loop (ACL) method [39], can be used. In the ACL approach, the predictions
are computed based on the previously quantized vectors. The residual vectors of this
prediction sequence are then used to train a new codebook with DSBS. The same residual
vectors are quantized with the new codebook and then added with the prediction sequence
to obtain the quantized vectors used to generate the predictions for the next iteration.
The process is repeated for a number of iterations until no significant improvement in
the training distortion is observed. The ACL method combines the stability benefits of
the open-loop method with asymptotically equivalent closed-loop performance [39]. The
detailed procedure of the ACL method is illustrated in [21].
In our scheme, 42507 residual vectors were randomly selected from the famous Princeton
3D mesh library [40] for training the approximate universal codebook o-line, and the size
of it ranges from 64 to 8192. In this way, we expect the codebook to be suitable for nearly
all triangle meshes for VQ compression and can be pre-stored in each terminal in the
network. Thus the compressed bit stream can be transmitted alone with convenience.
4. Simulation Results. To evaluate the eectiveness of the proposed k-ring predictive

VQ (KRPVQ) method, the conventional parallelogram predictive VQ (CPPVQ) method
is also implemented for comparison. We use eight 3D triangle meshes to test the PSNR
performance at the fixed bit rate. Here, PSNR is computed between the quantized mesh
M and the original mesh M as follows:
N
PSNR = 10 log10 N (14)
i=1 vi v i 22
where N is the number of vertices of M , vi and v i are the positions of the i-th vertex of
M and M , respectively, and all the vertices in M are previously normalized into a zero
mean sphere with a radius of 1.0. Here we set 0 , 0 and 1 in Equation (13) to be 0.001,
0.0035 and 0.005, respectively. These parameters may be slightly adjusted by the user
to tailor specific meshes. The codebook size is 1024 for all the experimental meshes. In
other words, the bit rate is 10.0 bpv.
Figure 3 depicts the uncompressed Cup and Dolphin meshes, the compressed ver-
sions with the CPPVQ and the proposed KRPVQ methods respectively. Due to its
smoothness and uniformity of triangles, the Cup mesh has a high correlation among
its vertices along the traversal order and thus the proposed method has not many ad-
vantages over the conventional one. Meanwhile, the Dolphin mesh is more jagged and
the proposed method exhibits much less distortion compared to the conventional method
at the same bit rate. Uncompressed meshes other than Cup and Dolphin we used
are depicted in Figure 4. As expected, the proposed k-ring prediction scheme produces
high PSNR values, outperforming the conventional method by 0.1dB 1.2dB, as shown
in Table 1.
(a) (b) (c)
(d) (e) (f)
Figure 3. (a) and (d) are the uncompressed cup and dolphin meshes;
(b) and (e) are the compressed versions with the CPPVQ method; (c) and
(f) are the compressed versions with the proposed KRPVQ method
From above experimental results, we can clearly see that, the proposed technique can
obtain much better rate-distortion performance. The main reason is that the dynamic
range of the residual vectors generated by the proposed prediction scheme is reduced.
In addition, our scheme needs low computation complexity and it is compatible with
existing connectivity compression schemes. Another advantage of our scheme lies in the
possibility for users to adaptively adjust parameters, k and the codebook size, in the
encoding procedure to acquire a higher quantization quality.
Table 1. PSNR comparions of the two methods at the same bit rate with
dierent meshes
Number of Number of PSNR obtained PSNR obtained
Mesh
Vertices Faces by CPPVQ (dB) by KRPVQ (dB)
Cup 1157 1550 39.97 40.23
Dolphin 420 836 36.96 38.29
Chess 802 1600 41.01 40.97
Knife 262 516 26.29 26.44
Face 1996 3564 39.81 39.92
Fist 1198 2392 35.72 35.83
Palm 386 768 33.34 33.45
Hat 958 1500 25.55 27.18
(a) (b) (c)
(d) (e) (f)
Figure 4. Other meshes are used in the experiment
5. Conclusions. Compression is essential for dealing with the growing complexity and
bandwidth requirements resulting from the increasing geometric complexity of 3D meshes
and the rapidly increasing performance of graphics processors. This paper has demon-
strated that the k-ring prediction based vector quantization scheme is a promising VQ
technique for compressing the vertices for triangle meshes, which are used extensively in
computer graphics applications. The key benefits of the proposed technique include the
superior rate-distortion performance, the ecient decompression process that is amenable
to hardware implementation, the compatibility with existing connectivity compression
schemes and the low computation complexity. In addition, it is possible for users to ad-
just parameters in the encoding procedure to acquire a higher quantization quality. Future
work will concentrate on further improvement of the k estimation method to achieve a
better rate-distortion performance.
Acknowledgment. This work was supported by the Program for New Century Excellent
Talents in University of China under grant NCET-04-0329 and Foundation for the Author
of National Excellent Doctoral Dissertation of P. R. China (No. 2003027).
REFERENCES
[1] J. Rossignac, Edgebreaker: Connectivity compression for triangle meshes, IEEE Trans. Visualization
and Computer Graphics, vol.5, no.1, pp.47-61, 1999.
[2] D. King and J. Rossignac, Guaranteed 3.67V bit encoding of planar triangle graphs, Proc. of the
11th Canadian Conference on Computational Geometry, pp.146-149, 1999.
[3] S. Gumhold, New bounds on the encoding of planar triangulations, Technical Report WSI-2000-1,
Univ. of Tubingen, 2000.
[4] D. Cohen-Or, D. Levin and O. Remez, Progressive compression of arbitrary triangular meshes, Proc.
of IEEE Visualization, pp.67-72, 1999.
[5] A. Szymczak, D. King and J. Rossignac, An edgebreaker-based ecient compression scheme for
regular meshes, Computational Geometry, vol.20, no.1-2, pp.53-68, 2001.
[6] M. Isenburg and J. Snoeyink, Spirale reversi: Reverse decoding of the edgebreaker encoding, Proc.
of the 12th Canadian Conference on Computational Geometry, pp.247-256, 2000.
[7] C. Touma and C. Gotsman, Triangle mesh compression, Proc. of Graphics Interface, pp.26-34, 1998.
[8] C. Bajaj, V. Pascucci and G. Zhuang, Single-resolution compression of arbitrary triangular meshes
with properties, Proc. of Data Compression Conf., pp.247-256, 1999.
[9] S. Gumhold and W. Straer, Real time compression of triangle mesh connectivity, Proc. of ACM
SIGGRAPH, pp.133-140, 1998.
[10] M. Isenburg and J. Snoeyink, Mesh collapse compression, Proc. of SIBGRAPI, Campinas, Brazil,
pp.27-28, 1999.
[11] R. Cohen, D. Cohen-Or and T. Ironi, Multi-way geometry encoding, Technical Report, 2002.
[12] F. Evans, S. Skiena and A. Varshney, Optimizing triangle strips for fast rendering, Proc. of IEEE
Visualization, pp.319-326, 1996.
[13] H. Hoppe, Optimization of mesh locality for transparent vertex caching, Proc. of ACM SIGGRAPH,
pp.269-276, 1999.
[14] M. F. Deering, Geometry compression, Proc. of ACM SIGGRAPH, pp.13-20, 1995.
[15] M. M. Chow, Optimized geometry compression for real-time rendering, Proc. of IEEE Visualization,
pp.346-354, 1997.
[16] R. M. Gray, Vector quantization, IEEE ASSP Magazine, 1984.
[17] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression, Kluwer Academic, Norwell,
1992.
[18] Sun Microsystems, The Java3D API Specification, http://java.sun.com/products/java-media/3D/
index.html, 1999.
[19] G. Taubin and J. Rossignac, Geometric compression through topological surgery, ACM Trans.
Graphics, vol.17, no.2, pp.84-115, 1998.
[20] G. Taubin, W. Horn, J. Rossignac and F. Lazarus, Geometry coding and VRML, Proc. of the IEEE,
vol.86, no.6, pp.1228-1243, 1998.
[21] P. H. Chou and T. H. Meng, Vertex data compression through vector quantization, IEEE Trans.
Visualization and Computer Graphics, vol.8, no.4, 2002.
[22] E. Lee and H. Ko, Vertex data compression for triangular meshes, Proc. of Pacific Graphics, pp.225-
234, 2000.
[23] Z. Karni and C. Gotsman, Spectral compression of mesh geometry, Proc. of ACM SIGGRAPH,
pp.279-286, 2000.
[24] O. Sorkine, D. Cohen-Or and S. Toldeo, High-pass quantization for mesh encoding, Proc. of Euro-
graphics Symposium on Geometry Processing, 2003.
[25] H. Lee, P. Alliez and M. Desbrun, Angle-analyzer: A triangle-quad meshes codec, Proc. of Euro-
graphics, pp.383-392, 2002.
[26] M. Isenburg and P. Alliez, Compressing polygon mesh geometry with parallelogram prediction, Proc.
of IEEE Visualization, pp.141-146, 2002.
[27] M. Isenburg, Compressing polygon mesh connectivity with degree duality prediction, Proc. of Graph-
ics Interface, pp.161-170, 2002.
[28] A. Khodakovsky, P. Alliez, M. Desbrun and P. Schroder, Near-optimal connectivity encoding of
2-manifold polygon meshes, Graphical Models, vol.64, no.3-4, 2002.
[29] B. Kronrod and C. Gotsman, Optimized compression of triangle mesh geometry using prediction
trees, Proc. of the 1st International Symposium on 3D Data Processing, Visualization and Trans-
mission, pp.602-608, 2002.
[30] D. Shikhare, S. Bhakar and S. P. Mudur, Compression of large 3D engineering models using automatic
discovery of repeating geometric features, Proc. of the 6th International Fall Workshop on Vision,
Modeling and Visualization, 2001.
[31] R. Gray and D. Neuho, Quantization, IEEE Trans. Information Theory, vol.44, pp.2325-2384, 1998.
[32] V. Cuperman and A. Gersho, Vector predictive coding of speech at 16 kbit/s, IEEE Trans. Comm.,
vol.33, no.7, pp.685-695, 1985.
[33] F.-X. Yu and Z.-M. Lu , A DCT-VQ based multipurpose image hashing scheme for copyright protec-
tion and content authentication, International Journal of Innovative Computing, Information and
Control, vol.5, no.9, pp.2703-2710, 2009.
[34] H.-X. Wang, Z.-M. Lu, Y.-N. Li and S.-H. Sun, A compressed domain multipurpose video wa-
termarking algorithm using vector quantization, International Journal of Innovative Computing,
[35] W.-M. Zheng, Z.-M. Lu and H. Burkhardt, Color image retrieval schemes using index histograms
based on various spatial-domain vector quantizers, International Journal of Innovative Computing,
[36] Z.-M. Lu, X.-N. Xu and J.-S. Pan, Face detection based on vector quantization in color images,
2006.
[37] C. M. Huang, Large Vector Quantization Codebook Generation Analysis and Design, Ph.D. Thesis,
Utah State Univ. Logan, 1991.
[38] C. Chatfield and A. J. Collins, Introduction to Multivariate Analysis, Chapman and Hall, London,
1980.
[39] H. Khalil, K. Rose and S. L. Regunathan, The asymptotic closed-loop approach to predictive vector
quantizer design with application in video coding, IEEE Trans. Image Processing, vol.10, no.1,
pp.15-23, 2001.
[40] Princeton University, 3D Model Search Engine, http://shape.cs.princeton.edu.
c ISSN 1349-4198
AN EFFICIENT SIMULATED ANNEALING WITH A VALID

SOLUTION MECHANISM FOR TDMA BROADCAST
SCHEDULING PROBLEM
Gon Kim1 , Sung Soo Kim2, , Il-Hwan Kim3 , Dong Hoi Kim3
V. Mani4 and Jae-Ki Moon5
1
Division of Architecture
2
Department of Industrial Engineering
3
Department of Electronic and Tele Communication Engineering
Kangwon National University
Chunchon 200-701, Korea
gonkim@kangwon.ac.kr; ihkim@kangwon.ac.kr; donghk@kangwon.ac.kr

Corresponding author: kimss@kangwon.ac.kr
4
Department of Aerospace Engineering
Indian Institute of Science
Bangalore 560-012, India
mani@aero.iisc.ernet.in
5
Department of Information Management Engineering
Korea University, Seoul, Korea
moon8385@nate.com
Received April 2009; revised September 2009
Abstract. This paper presents an ecient Simulated Annealing with valid solution
mechanism for nding an optimum conict-free transmission schedule for a broadcast
radio network. This is known as a Broadcast Scheduling Problem (BSP) and shown as
an NP-complete problem in earlier studies. Because of this NP-complete nature, earlier
studies used genetic algorithms, mean eld annealing, neural networks, factor graph and
sum product algorithm, and sequential vertex coloring algorithm to obtain the solution.
In our study, a valid solution mechanism is included in simulated annealing. Because
of this inclusion, we are able to achieve better results even for networks with 100 nodes
and 300 links. The results obtained using our methodology is compared with all the other
earlier solution methods.
Keywords: Broadcast scheduling problem, Packet radio network, Ecient simulated
annealing, Time-division multiple access (TDMA), Optimum transmission schedule
1. Introduction. Packet Radio Network (PRN) provides high speed wireless packet data
services to a group of stations over a broad geographic region. The stations in the network
share a single radio channel. Each station in the network has a transceiver unit and a
control unit. Stations can either transmit or receive packets using the shared radio channel
at each time instant. The control unit schedules the trac according to a channel access
protocol. Because of this, Packet radio network (PRN) design is an important issue in
communication services and distributed communication applications [1].
Packet radio network adopts time-division multiple-access (TDMA) protocol for nodes
to communicate each other in a single shared radio channel. TDMA is a digital transmis-
sion technology which allows a number of users to access a single radio frequency channel
without interference by allocating unique time slots to each user (station) within each
channel. The time is divided into distinct TDMA frames and the frames consists of a
1181
1182 G. KIM, S. S. KIM, I.-H. KIM, D. H. KIM, V. MANI AND J.-K. MOON
xed number of time slots. A time slot is the transmission time for a single packet to
be sent and received by a pair of neighboring nodes (stations). In TDMA protocol, a
single clock controls the transmission of packets. At each time slot a station may either
transmit or receive at most one packet from one of its neighbors. Several stations can
transmit packets simultaneously at the same time slot if they are far apart to avoid mu-
tual interference; otherwise these stations have to transmit their packets in dierent time
slots.
The design problem that occurs in packet radio network is known as Broadcast Schedul-
ing Problem. The objective of Broadcast Scheduling Problem (BSP) is to nd an optimal
TDMA frame. The optimal TDMA frame should be able to schedule transmission of all
nodes in a minimum TDMA frame length without conict, and should also maximize the
channel utilization or total conict-free transmission. A TDMA frame is a time frame,
which is divided into M equal length time slots. In a TDMA cycle, each node should be
scheduled at least once and is known as no-transmission constraint. Another constraint
in this BSP scheduling problem is no-conflict constraint. No-conict constraint has pri-
mary conict and secondary conict. Primary conict constraint is that a node can not
transmit and receive simultaneously. Secondary conict constraint is that a node is not
allowed to receive more than one transmission simultaneously. With these constraints,
the BSP problem is minimize M (time slots) and maximize the channel utilization or total
conict-free transmission. Another objective is to minimize the average packet delay. The
average packet delay is reduced if the channel utilization is maximized.
It is shown in an earlier studies that this BSP is an NP-complete combinatorial opti-
mization problem. Hence, earlier studies used genetic algorithms, mean eld annealing,
neural networks, factor graph and sum product algorithm, and sequential vertex coloring
algorithm to obtain the solution. We will discuss the methodologies used in earlier studies
[2, 3, 4, 5, 6, 7].
Genetic Algorithms (GA) is the most well-known evolution based search algorithm de-
veloped in an attempt to explain the adaptive process of natural systems and to design
articial systems based upon these natural systems [8]. The GA is a search algorithm
based on the mechanism of natural selection that transforms a population (a set of in-
dividuals) into a new population (i.e., next generation) using genetic operators such as
crossover, mutation and reproduction [8, 9, 10]. In the early works on GA [9], the solution
(string) representation was restricted to only binary digits (0 and 1), and it is shown in
[11] that a natural representation of strings are more ecient and produce better results.
The applicability of GAs to broadcast scheduling problem is shown in [2]. In their study
[2], a randomized algorithm is presented to create elite initial population. This random-
ized algorithm satises all the constraints of the BSP problem and so they obtain valid
initial solutions. A modied crossover operator is also presented in [2], and a proof that all
solution in every generation are valid solution; i.e., all solutions satisfy all the constraints.
Mean Field Annealing (MFA) is a method used in statistical mechanics. The well-
known Simulated Annealing (SA) is based on annealing process in statistical mechanics
[12]. It is known that SA can be used for solving dicult combinatorial optimization
problems. The broadcast scheduling problem is solved using the mean eld analysis in
[3]. Simulated annealing uses a stochastic process whereas MFA uses a set of deterministic
equations.
A two-phase algorithm is presented to obtain the solution for broad scheduling problem
in [4]. In this two-phase algorithm, in the rst phase, the frame length is minimized using a
sequential vertex coloring algorithm. The second phase maximize the channel utilization.
The main draw backs of using MFA to broadcast scheduling problems are given in [4] and
TDMA BROADCAST SCHEDULING PROBLEM 1183
better results are achieved in [4] than the results obtained by using mean eld algorithm
given in [3].
The broadcast scheduling problem in the context of ad hoc mobile network is considered
and a competent permutation genetic algorithm is presented in [5]. A permutation-based
encoding strategy is proposed and it is shown that this strategy is more ecient in ob-
taining the valid solutions [5].
The factor graph and the sum product algorithm is used to solve the broadcast sched-
uling problem is presented in [6]. The methodology presented in [6] is a self-organizing
distributed algorithm using only local collaborative interactions among the neighboring
nodes in the network. The approach used in [6] is divide the constraints of BSP into
many simple rules, and these rules are enforced by a local processing unit in the factor
graph. The optimal broadcast schedule is obtained by using the sum product algorithm
in an iterative manner from the information exchanged among local processing units. A
chaotic neural network for solving the BSP problem is given in [7].
Motivation of our study is to present an an ecient simulated annealing with valid
solution mechanism for TDMA broadcast scheduling problem. First, we present the
scheduling problem and our ecient simulated annealing methodology. In Section 4, we
present the results obtained using our methodology and show that our method performs
better than all other earlier solution methods [2, 3, 4, 5, 6, 7], in terms of utilization and
thus the average delay time.
2. The Broadcast Scheduling Problem. In this section, we briey describe the broad-
cast scheduling problem that arise in packet radio network considered in earlier studies
[2, 3, 4, 5, 6, 7]. A given packet radio network can be represented by graph G = (V, E),
where vertices V = 1, 2, ..., N are the radio network nodes and N is the total number
of nodes in the network. E is the set of transmission links. In Broadcast Scheduling
Problem (BSP), we need to obtain a schedule for transmission for all stations, so that no
collision among packets take place. In the schedule, the time is divided into slots of equal
length. The number of time slots in each frame is xed. We are interested in nding an
optimal collision free transmission schedule for stations in a frame. Once this is obtained
the frame is repeated in time. In the graph G, two nodes i and j (i, j E) are connected
by an edge if and only if they can receive each others transmission. In this situation,
the nodes i and j are one hop away or adjacent neighbors. When two adjacent neighbors
transmit in the same time slot, a primary conflict occurs. When a node receives two
dierent packets from two of its adjacent neighbors in the same time slot, a secondary
conflict occurs [2, 3, 4, 5, 6, 7]. The network topology can be represented by an N N
connectivity matrix C dened as
{
1 if nodes i, j are one-hop apart
Cij = (1)
0 otherwise.
Another N N matrix called compatibility matrix D is dened as
{
1 if nodes i, j are one-hop or two hop apart
Dij = (2)
0 otherwise.
Our objective is to obtain an optimal conict free TDMA frame for packet transmission
that also satises the constraints. This TDMA frame is repeated over time. We use an
M N binary matrix X = xij to denote a TDMA frame, where M is the number of time
slots in the TDMA frame. The matrix X is dened as
{
1 if time slot i in a frame is assigned to node j
xij = (3)
0 otherwise.
For node j the channel utilization j is dened as [2, 3, 4, 5, 6, 7]

M
the number of slots assigned to node j xij
j = = i=1 (4)
TDMA cycle length M
The channel utilization for the entire network is given as
1
N N M
1
= j = xij (5)
N j=1 N M j=1 i=1
Another performance measure for the evaluation of TDMA frame (transmission sched-
ule) is the average time delay. Queueing theory has been in evaluation the performance
of packet radio network [13], availability analysis [14], transmission control in networks
[15, 16]. For obtaining the average time delay, the following assumptions were made in
[3, 4]. These are:
1. The length of a time slot equals the time required to transmit a packet and the
packets have a xed length.
2. The inter-arrival time for each station i is statistically independent from other sta-
tions, and packets arrive according to a Poisson process with a rate of i (pack-
ets/slot). The total trac in station i consists of its own trac and the trac
incoming from other stations. At each station the packets are stored in buers that
have innite capacity.
3. The probability distribution of the service time of station i is deterministic and
statistically independent from other stations. The average service rate of station i
is i (packets/slot).
4. Transmission of packets are done only at the beginning of the time slot.
With the above assumptions, a network can be modeled as N , M/D/1 queues, where
N is the number of stations. Let Di be the average delay for station i, and this can be
obtained as [3, 4]:
2
1 1 ( ii )
Di = + (6)
i i 2(1 i )
i
and

M
i = xji (7)
j=1
The total time delay is given by

N
i Di
D=
i=1
N
(8)
i=1 i
The objective of the BSP is to nd a transmission (TDMA) frame with minimum

number of time slots, which satises all the constraints and also maximize the channel
utilization . The BSP is described as:
BSP: Minimize M and maximize , subject to the following constraints

M
xij > 0, for all j = 1, 2, ..., N (9)
i=1

M
N
N
djk xij xik (10)
i=1 j=1 k=1,k=j
Equation (9) represents the no-transmission constraint and Equation (10) represents the
no-conict constraint.
3. Ecient Simulated Annealing. Dicult combinatorial optimization problems have

been solved in a successful manner by using Simulated Annealing (SA). For the BSP,
the SA algorithm starts with a non-optimal (but a valid) initial TDMA schedule (solu-
tion). Initial valid and non-optimal TDMA schedule can be generated randomly. SA
algorithm improves the initial TDMA schedule (solution) by selecting a neighborhood
solution (TDMA schedule) which increases the objective function value. The acceptance
or rejection of an uphill move is determined by a sequence of random numbers with a
controlled probability. The probability of accepting a move that causes an increase Z
in the objective function value Z is called acceptance function. This is normally set to
exp( Z
T
). Here T is a control parameter that corresponds to temperature in physical
annealing. We can see from the acceptance function that small increases in the objective
function value (Z) are more likely to be accepted than large increases. When the value of
T is high most moves will be accepted but as T 0 most uphill moves will be rejected.
SA algorithm will starts with a large value of T , and the algorithm continues by attempt-
ing a certain number of neighborhood moves at each temperature, while the T is reduced
in a gradual manner.
In Ecient Simulated Annealing (ESA), we rst select the initial valid solution (valid
TDMA schedule) by using the randomized algorithm given in [2]. This initial valid solu-
tion (TDMA schedule) is further modied to improve the utilization () and is used as a
starting solution in SA. We call the modication as valid solution mechanism. The mod-
ications are done by using the information given by compatibility matrix (D-matrix).
The inclusion of valid solution mechanism for improving the solution (TDMA schedule) to
SA is called Ecient Simulated Annealing (ESA) in our study. In the ESA methodology,
the generation of initial feasible solutions has the following steps.
Step 1: Decide the sequence to allot time slots for dierent nodes in the permutation
sequence.
Step 2: Assign time slot 1 to the rst node for the sequence.
Step 3: Check that time slot 1 can be assigned to the second node using matrix D.
Step 4: Assign time slot 2 to the second node if time slot 1 can not be assigned to the
second node. We assume that time slot 2 is assigned to the second node.
Step 5: Check that time slot 1 and 2 can be assigned to the third node using matrix D.
Step 6: Assign time slot 3 to the third node if time slot 1 and 2 can not be assigned to
the third node.
Step 7: Follow the previous steps until all nodes are assigned to minimize the time slot
M if possible.
Step 8: Maximize time slot utilization if possible using matrix D after all nodes are as-
signed.
In the ESA methodology, the generation of neighbor feasible solutions has the following
steps.
Step 1: Change 0 to 1 or 1 to 0 in the current feasible solution probabilistically to generate
neighbor feasible solutions using matrix D. Here 1 refers to if time slot i in a frame
is assigned to node j; 0 refers to otherwise as deed by xij .
Step 2: Use high change probability rate to generate neighbors for diversity and use low
change probability rate to generate neighbors for convergence. Reduce the change
probability rate when generation is increasing.
We would like to mention here that this ecient simulated annealing does not guarantee
that we obtain the optimal solution for a given network. However, in our simulation results
we are able to obtain better TDMA schedules (using ESA) than the TDMA schedules
obtained in earlier studies [2, 3, 4, 5, 6, 7]. The owchart for ESA is given in Figure 1.
Start
Input data C matrix
Generation of D matrix
Selection of T, t for SA
initial solution Vc using VSM
Neighbor solution Vn using VSM
Vn with minimum time slots
No
Probabilistic
f(Vn
f(Vn)
Vn) is better than f(Vc
f(Vc)
Vc) ?
selection
Yes
Yes
Vc Vn No
No t is satisfied ?
Yes
Tnew= (Told-)
No
Stop condition ?
Yes
Stop
Figure 1. Flow-chart for ecient simulated annealing (ESA)
4. Simulation Results and Comparison with Other Methods. In this section,

we present the simulation results obtained by using our Ecient Simulated Annealing
(ESA), and also present a comparison of performance of our ESA algorithm with the
results obtained in earlier methods [2, 3, 4, 5, 6, 7]. As mentioned in [2], there is no
standard benchmark problem set and so we compare our results with the available results
given in earlier studies [2, 3, 4, 5, 6, 7], and show the performance improvement in PR
Figure 2. TDMA schedule for 15 node network

Table 1. Comparison of our results for N=15 node network: Based on
number of time slots M and utilization
Number Our Wang & Yeo Wu chen

of nodes Method Ansari [3] [4] [5] [6]
15 8(0.167) 8(0.15) 8(0.15) 8(0.167) 8(0.167)
networks using our Ecient Simulated Annealing methodology. The performance criteria
used for comparison are M (time slots), (channel utilization for the whole network), and
the average time delay.
We rst consider a small-scale network with N=15. The TDMA schedule obtained
using our ecient simulated annealing methodology is shown in Figure 2. In this Figure
2, a black box represents an assigned time slot. The schedule obtained in our methodology
has 8 (M) time slots and the utilization of the whole network () is 0.167 (20/120). The
same result is obtained in earlier studies [5, 6]. In other earlier studies [3, 4], the value of
M obtained is the same as our result i.e., M=8, but the utilization obtained is 0.15 which
is less than the result obtained in our methodology. These results are shown in Table 1.
We next consider a medium-scale network with N=30 and 40. The TDMA schedule
obtained using our ecient simulated annealing methodology are shown in Figures 3 and
4.
For the network with N=30, the schedule obtained in our methodology has 10 (M)
time slots and the utilization of the whole network () is 0.123 (37/300). The same result
is obtained in earlier studies [5, 6]. For this network N=30, the results obtained in an
earlier study [4], are M=11 and =0.112. In another study [3], the corresponding values
are M=12 and =0.108.
For the network with N=40, the schedule obtained in our methodology has 8 (M) time
slots and the utilization of the whole network () is 0.216 (69/320). For this case, our
methodology produces a schedule that has better (utilization of the whole network)
compared to other studies [2, 3, 4, 5, 6]. The number of time slots M obtained is same
(M=8) in earlier studies [2, 4, 5, 6], but the value of obtained is less than 0.216. This
shows our methodology performs better than earlier studies. The results for N=30 and
40, are shown in Table 2.
Now, we consider a large-scale network with N=100. We consider a 100 node network
with 200 and 300 links. The TDMA schedule obtained using our ecient simulated
annealing methodology are shown in Figures 5 and 6.
For the network with N=100 and 200 links, the schedule obtained in our methodology
has 9 (M) time slots and the utilization of the whole network () is 0.152 (136/900). For
Table 2. Comparison of our results for N=30 and 40 middle-scale network:

Based on number of time slots M and utilization
Number Our Chakra- Wang & Yeo Wu chen

of nodes Method borty [2] Ansari [3] [4] [5] [6]
30 10(0.123) NA 12(0.108) 11(0.112) 10(0.123) 10(0.123)
40 8(0.216) 8(0.203) 9(0.197) 8(0.188) 8(0.209) 8(0.203)
this case, our methodology produces a schedule that has better (utilization of the whole
network) compared to other studies [2, 5]. The number of time slots M obtained is same
(M=9) in earlier studies [2, 5], but the value of obtained is less than 0.152. This shows
our methodology performs better than the earlier studies.
For the case of a network with N=100 and 300 links, the schedule obtained in our
methodology has 10 (M) time slots and the utilization of the whole network () is 0.11
Table 3. Comparison of our results for N=100 large-scale network: Based

on number of time slots M and utilization
Number Our Chakra- Wu

of nodes Method borty [2] [5]
100
(200 links) 9(0.152) 9(0.148) 9(0.151)
100
(300 links) 10(0.11) 11(0.104) 11(0.109)
Figure 5. TDMA schedule for 100 node network with 200 links
(110/1000). For this case, our methodology produces a schedule that has better (uti-
lization of the whole network) compared to other studies [2, 5]. The number of time slots
M obtained is same (M=9) in earlier studies [2, 5], but the value of obtained is less than
0.110. This also shows that our methodology performs better than the earlier studies.
The results obtained for N=100, with 200 and 300 links are shown in Table.3.
It can be seen that increasing the utilization,
will reduce the average time-delay. This
is because, in the queueing models i = M j=1 ji , depends on the utilization. i is the
x
average service rate of station i is increased because of increased utilization and thus
reducing the average time-delay. In our results, we have seen that when some of the
nodes are transmitting, all other nodes are not transmitting. This can be thought of as a
bottleneck station. It may be possible to identify the bottleneck stations in the network,
Figure 6. TDMA schedule for 100 node network with 300 links
using the methodology given for the determination of bottleneck machine identication
problems presented in the context of job shop scheduling presented in [17].
5. Conclusions. The broadcast scheduling problem is considered with the objective of
obtaining the best/optimal TDMA schedule. This schedule is to minimize the number of
time slots (M) and maximize the utilization (). This is known as Broadcast Scheduling
Problem (BSP) and shown as an NP-complete problem in earlier studies. An ecient
Simulated Annealing algorithm with valid solution mechanism for nding an optimum
conict-free transmission schedule for a broadcast radio network is presented to this NP-
complete problem. Simulation results are presented for networks of size 15, 30, 40 and
100. Our results are compared with earlier methods such as genetic algorithms, mean
eld annealing, neural networks, factor graph and sum product algorithm, and sequential
vertex coloring algorithm. It is shown that our ecient simulated annealing algorithm
is able to perform better than the earlier methods both in minimizing the number of
time slots (M) and maximizing utilization (). In our study, a valid solution mechanism
is included in simulated annealing. Because of this inclusion, we are able to achieve
better results even for networks with 100 nodes and 300 links. The results obtained using
our methodology is compared with all the other earlier solution methods, in terms of
utilization and thus the average delay time.
Acknowledgements. This research work was supported by Basic Science Research Pro-
gram through the National Research Foundation of Korea (NRF) funded by the Ministry
of Education, Science and Technology (No. 2009-0087069).
REFERENCES
[1] B. M. Leiner, D. L. Nielson and F. A. Tobagi, Issues in packet radio network design, Proc. of IEEE,
vol.75, no.1, pp.6-20, 1987.
[2] G. Chakraborty, Genetic algorithm to solve optimum TDMA transmission schedule in broadcast
packet radio networks, IEEE Transactions on Communications, vol.52, no.5, pp.765-777, 2004.
[3] G. Wang and N. Ansari, Optimal broadcast scheduling in packet radio networks using mean eld
annealing, IEEE Journal on Selected Areas of Communication, vol.15, no.2, pp.250-260, 1997.
[4] J. Yeo, H. Lee and S. Kim, An ecient broadcast scheduling algorithm for TDMA ad-hoc networks,
Computers and Operations Research, vol.29, pp.1793-1806, 2002.
[5] X. Wu, B. S. Sharif, O. R. Hinton and C. C. Tsimenidis, Solving optimum TDMA broadcast sched-
uling in mobile ad hoc networks: A competent permutation genetic algorithm approach, IEE Pro-
ceedings - Communication, vol.152, no.6, pp.780-788, 2005.
[6] J. C. Chen, Y. C. Wang and J. T. Chen, A novel broadcast scheduling strategy using factor
graphs and the sum-product algorithm, IEEE Transactions on Wireless Communications, vol.5,
no.6, pp.1241-1249, 2006.
[7] L. Wang and H. Shi, A gradual noisy chaotic neural network for solving the broadcast scheduling
problem in packet radio networks, IEEE Transactions on Neural Networks, vol.17, no.4, pp.989-1000,
2006.
[8] H. J. Holland, Adaptation in Natural and Articial Systems, University of Michigan Press, New
York, 1975.
[9] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley,
New York, 1989.
[10] L. David, Handbook of Genetic Algorithms, Van Nostrand Reingold, New York, 1991.
[11] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, AI Series Springer-
Verlag, New York, 1994.
[12] S. Kirkpatrick, C. Gelatt and M. Vecchi, Optimization by simulated annealing, Science, vol.220,
no.4598, pp.671-680, 1983.
[13] L. Kleinrock and J. Silvester, Spatial reuse in multihop packet radio networks, Proc. of IEEE, vol.75,
no.1, pp.156-167, 1987.
[14] V. Mani and V. V. S. Sarma, Queueing network models for aircraft availability and spares manage-
ment, IEEE Transactions on Reliability, vol.R-33, no.3, pp.427-433, 1984.
[15] H.-Y. Chu, K.-H. Tsai and W.-J. Chang, Fuzzy control of active queue management routers for trans-
mission control protocol networks via time-delay ane Takagi-Sugeno fuzzy models, International
[16] X. Chen, T. Liu and J. Zhao, A logic-based switching control approach to active queue management
for transmission control protocol, International Journal of Innovative Computing, Information and
Control, vol.4, no.7, pp.1811-1820, 2008.
[17] R. Zhang and C. Wu, Bottleneck machine identication based on optimization for job shop scheduling
problem, ICIC Express Letters, vol.2, no.2, pp.175-180, 2008.
c ISSN 1349-4198
SCALABLE AND SYSTOLIC DUAL BASIS MULTIPLIER

OVER GF(2m )
Liang-Hwa Chen1 , Po-Lun Chang2 , Chiou-Yng Lee1

and Ying-Kuei Yang3
1
Department of Computer Information and Network Engineering
2
Department of Electrical Engineering
Lunghwa University of Science and Technology
No. 300, Sec. 1, Wanshou Rd., Guishan Shiang, Taoyuan County 33306, Taiwan
{ whallis2000; PP010 }@mail.lhu.edu.tw; whc1223@ms7.hinet.net
3
No. 43, Sec. 4, Keelung Rd., Taipei 106, Taiwan
yingkyang@yahoo.com
Received August 2009; revised June 2010
Abstract. This work presents a novel low-complexity scalable and systolic architec-
ture for dual basis multiplications over GF(2m ). The proposed architecture is derived
by utilizing the block Hankel matrix-vector representation and is feasible for finite fields
generated by irreducible trinomials. By selecting an appropriate digit size d, the proposed
scalable architecture can achieve a satisfactory trade-o between throughput performance
and hardware complexity for implementing cryptographic schemes such as ECDSA in
resource-constrained environments such as smart cards and embedded systems. Analyti-
cal results indicate that both area and time-area complexities of the proposed architecture
are significantly lower than those of the non-scalable architecture schemes. Furthermore,
due to its features of regularity, modularity and concurrency, the proposed architecture
is highly feasible for VLSI implementations.
Keywords: Finite field, Cryptography, Dual basis, Hankel matrix-vector, Scalable mul-
tiplier, Elliptic curve cryptography (ECC)
1. Introduction. Mobile communications and Internet transactions have become in-

creasingly popular in recent years, explaining concern over the integrity of transmitted
data, against eavesdropping or unauthorized data altering. Thus, research advances in
cryptography have received increasing interest as well [1,14,19-21]. Cryptography and cod-
ing theory frequently involve finite field arithmetic operations, especially for the binary
extension field GF(2m ). In particular, the elliptic curve cryptography (ECC) [16], which
has become increasingly popular owing to its ability to realize a robust cryptosystem in
smart cards, mobile phones and other resource-constrained environments, requires finite
field arithmetic operations. The National Institute for Standards and Technology (NIST)
has recommended the elliptic curve digital signature algorithm (ECDSA) standard [17],
in which five GF(2m ), i.e. for m = 163, 233, 283, 409 and 571, are used to achieve ade-
quate security. Among the basic arithmetic operations over GF(2m ), multiplication is the
most important and time-consuming computation. Other complex arithmetic operations,
including exponentiation, division and multiplicative inversion, can be performed by re-
peating multiplications, subsequently explaining the high demand for ecient design and
implementation of the finite field multiplier with a low complexity.
1193
1194 L.-H. CHEN, P.-L. CHANG, C.-Y. LEE AND Y.-K. YANG
For finite field GF(2m ), three conventionally adopted bases are polynomial basis (PB),
normal basis (NB) and dual basis (DB) to represent its elements. Each representation has
its own distinct advantages. For instance, the polynomial basis multiplier does not require
basis conversion and is characterized by its regularity and simplicity. The normal basis
multiplier is quite eective in performing squaring an element in the finite field. The dual
basis multiplier needs the least number of gates that leads to the smallest chip area demand
for VLSI implementation [4]. Pertinent literature in recent years has generally classified
most finite field multipliers on these bases as bit-serial [2,25] and bit-parallel architectures
[10-12,18]. Cryptography applications such as the above-mentioned ECDSA heavily rely
on finite fields of large values of m (160 or higher) to achieve a sucient level of security.
To implement such cryptosystems, although the bit-serial architecture requires less chip
area, its throughput is too slow. Conversely, the bit-parallel architecture is typically faster
yet more complex and requires a significantly larger chip area and power consumption.
Digit-serial architectures [3,5,7] have been proposed to facilitate the trade-o between
throughput performance and hardware complexity. These architectures are based on a
cut-set systolization scheme to accelerate the computational process. However, such an
architecture has a space complexity similar to that of the original bit-level multiplier
design, making it infeasible for implementing cryptosystems that require a large number
of operand bits in resource-constrained environments. Scalable architectures [8,9,23] have
subsequently been developed to satisfy the demands for implementing ECC cryptosystems
such as ECDSA in resource-constrained environments. As a combination of serial and
parallel schemes, the scalable architecture requires smaller chip area than that of the digit-
serial architecture. In the scalable architecture, each m-bit operand word in the GF(2m )
field is separated into k = m/d d-bit sub-words which are also termed as digits. Two
operand words are multiplied in digits with a serial scheme while two digits are multiplied
with a parallel scheme. Digit size d is normally selected to be substantially smaller than
the operand bit length m. Thus, a substantially smaller hardware area multiplier (for
d-bit operands) that can be easily implemented in resource-constrained environments
is sucient to perform all of the m-bit multiplications. Additionally, considering the
trade-o between throughput performance and hardware complexity allows the scalable
architecture to generate an optimal realization in hardware implementation by selecting an
appropriate scalable factor, i.e. digit size d, especially in cases involving implementation
of ECC cryptosystems in resource-constrained environments. Furthermore, the scalable
architecture is flexible in terms of re-usage. Assume that a multiplier has been designed
for 768-bit operand words. This multiplier cannot be applied directly to a system with
an operand length of 1024 bits. The non-scalable (bit-parallel) multiplier must be re-
designed to comply with system requirements. Conversely, the scalable architecture does
not require a change in the core multiplier. By simply adjusting the size of the register
array to satisfy the required longer operand length and reusing the core multiplier, the
scalable multiplier can be applied to the new system. Doing so also makes it possible to
achieve multiplications over the five GF(2m ) used in ECDSA with only one size (such as
d = 20) of the core multiplier in order to obtain dierent levels of security. This feature
is highly attractive for implementing ECDSA in resource-constrained environments such
as smart cards and handset devices. A similar multi-scale concept can also be found in
other fields (e.g., [24]).
Scalable multiplication architectures have been developed for PB [9] and NB [8]. Cor-
respondingly, this work presents a novel scalable architecture for dual basis (DB) mul-
tiplication over GF(2m ). The proposed algorithm is derived based on the block Hankel
matrix-vector representation, from which, a low-complexity scalable and systolic multi-
plier is then derived. The proposed scalable multiplier is feasible for finite fields GF(2m )
SCALABLE AND SYSTOLIC DUAL BASIS MULTIPLIER OVER GF(2m ) 1195
generated by irreducible trinomials. Analytical results indicate that both area and time-
area complexities of the proposed architecture are significantly lower than those of con-
ventional digit-serial and bit-parallel multipliers. Additionally, for a given field size m of
GF(2m ), the optimal digit size d can also be easily obtained from the analytical results to
achieve the optimal trade-o between throughput performance and hardware complexity.
The rest of this paper is organized as follows: Section 2 briefly reviews the dual basis
multiplication over GF(2m ) and a bit-parallel dual basis multiplication algorithm. Section
3 then introduces the proposed novel scalable and systolic algorithm and architecture for
dual basis multiplication over GF(2m ). Next, Section 4 discusses its time and space
complexities. Conclusions are finally drawn in Section 5, along with recommendations for
future research.
2. Preliminaries. As is well known, the finite (Galois) field GF(2m ) can be viewed as
a vector space of dimension m over GF(2), where the field is generated by irreducible
polynomial F (x) = f0 + f1 x + + fm1 xm1 + xm of degree m over GF(2). Assume
that denotes a root of the irreducible polynomial F (x). Then, any element A in the
finite field GF(2m ) can be represented as A = a0 + a1 + a2 2 + + am1 m1 , where
the coordinates ai GF(2) for 0 i m 1. The set {1, , 2 , . . . , m1 } is called the
polynomial basis (PB) of GF(2m ).
Definition 2.1. The trace function T r(x) over GF(2m ) is defined as [2]

m1
i
T r(x) = x2 .
i=0
Definition 2.2. A basis {0 , 1 , . . . , m1 } in GF(2m ) is said to be the dual basis (DB)

of {1, , 2 , , m1 } if the following condition is satisfied:
{
i
1, if i = j,
T r( j ) = (1)
0, if i = j,
where is selected to simplify the conversion between polynomial and dual bases.
For any element A = a0 + a1 + a2 2 + + am1 m1 in GF(2m ), its dual basis
representation can be expressed as A = T r(A)0 + T r(A)1 + + T r(m1 A)m1 .
Besides, for any two elements A and B in GF(2m ) represented in polynomial and dual

m1
m1
basis respectively, i.e., A = ai i and B = bi i , their product C = AB represented
i=0 i=0

m1
in dual basis, i.e., C = ci i , can be computed with the following discrete-time Wiener-
i=0
Hopf equation (DTWHE) [15]:

c0 b0 b1 bm1 a0
c1 b1 b2 bm a1
. = . .. . . .. .. (2)
.. .. . . . .
cm1 bm1 bm b2m2 am1
where

m1
bm+i = fj bi+j = f0 bi + f1 bi+1 + + fm1 bi+m1 , for i = 0, 1, , m 1. (3)
j=0
The above equation is derived as follows: First, from Definition 2.2, the coordinates bi
of B can be obtained as bi = T r(i B) for 0 i m 1. Next, since F () = 0, thus,

m1
m
= fi i = f0 + f1 + f2 2 + + fm1 m1 , (4)
i=0

m1
m+i
= fj i+j = f0 i + f1 i+1 + f2 i+2 + + fm1 i+m1 . (5)
j=0
Let us define that bm+i = T r(m+i B). Then, according to (5),

( (m1 ) ) m1

( i+j ) m1
m+i i+j
bm+i = T r( B) = T r fj B = fj T r B = fj bi+j
j=0 j=0 j=0
(6)
= f0 bi + f1 bi+1 + f2 bi+2 + + fm1 bi+m1 , for i = 0, 1, , m 1.
Based on Definition 2.2, the coordinates ci of C can also be obtained as ci = T r(i C).

m1
Given that C = AB, A = ai i and according to (6), we obtain
j=0
( ( ) )

m1
m1
ci = T r(i C) = T r(i AB) = T r i aj j B = aj T r (i+j B)
j=0 j=0
(7)

m1
= aj bi+j = a0 bi + a1 bi+1 + a2 bi+2 + am1 bi+m1 , for i = 0, 1, , m 1.
j=0
Equation (7) is expressed in matrix form and, then, the DTWHE in (2) is obtained.
Besides, if we define the following vectors: A = [a0 , a1 , , am1 ], F = [f0 , f1 , , fm1 ]
and
B (i) = [bi , bi+1 , , bi+m1 ], for i = 0, 1, , m 1, (8)
then, (6) and (7) can also be expressed as
bm+i = B (i) F (9)
ci = B (i)
A (10)
where denotes the inner product operation of two vectors. Notably, B (0) = B =
[b0 , b1 , , bm1 ]. By applying (9) and (10), the DB multiplication can be carried out by
the following algorithm:
Algorithm 1: [12]
Input: A = [a0 , a1 , , am1 ], B = [b0 , b1 , , bm1 ] and F = [f0 , f1 , , fm1 ]
Output: C = [c0 , c1 , , cm1 ] = AB
1. Initial step
1.1 C = [0, 0, , 0]
1.2 B (0) = B
2. Multiplication step
2.1 For i = 0 to m 1 do
2.2 bm+i = B (i) F
2.3 ci = B (i) A
2.4 B (i+1) = (B (i) 1) + [0, , 0, bm+i ]
2.5 Endfor
3. Return C .
Based on the above algorithm, Lee et al. [12] developed a bit-parallel systolic DB
multiplier architecture consisting of m2 cell, which comprises one AND gate, one XOR
gate and two 1-bit latches. Owing to the regularity of its architecture, this DB multiplier is
feasible for VLSI implementation. However, for a large field size of binary finite fields, e.g.
the five GF(2m ), m = 163, 233, 283, 409 and 571, used in ECDSA, the corresponding large
space complexity O(m2 ) makes such a multiplier infeasible for implementing in constrained
hardware environments such as smart cards and mobile handsets. To overcome this
problem, the next section introduces a novel scalable scheme for DB multiplication that
divides the m-bit operand word into several d-bit digits and, then, iteratively applies a
smaller scale multiplier to achieve the complete m-bit multiplication.
3. Proposed Scalable Systolic Dual Basis Multiplier Over GF(2m ). Deriving the
scalable architecture of DB multiplier initially requires introducing the Hankel matrix-
vector representation.
3.1. Hankel matrix-vector representation.
Definition 3.1. An m m matrix H is referred to as a Hankel matrix if it satisfies the
relation H(p, q) = H(p + 1, q 1), for 0 p m 2, 1 q m 1, where H(p, q)
represents the element in the intersection of row i and column j.
A Hankel matrix can be determined entirely by the 2m 1 entries that are located on
its first row and last column. Namely, the matrix can be defined by the corresponding
Hankel vector H = [h0 , h1 , , h2m2 ]. The Hankel matrix has been extensively adopted
in system identification (e.g., [13]). With the Hankel matrix-vector representation, the
product of a Hankel matrix H and a vector V = [v0 , v1 , , vm1 ] , i.e., HV , is denoted
as H V . With such notation, the DB multiplication in (2) can be expressed as
C = B A, (11)
where B = [b0 , b1 , , bm1 , bm , , b2m2 ] represents the corresponding Hankel vector of
the Hankel matrix in (2).
3.2. Proposed algorithm. For the digit size selected as d-bits, and k = m/ d, Equa-
tion (2) can also be expressed as the following block Hankel matrix-vector form:

C0 B0 B1 Bk1 A0
C1 B B2 Bk A1
C= .
..
= BA = . 1
.. .
.. . . . .. .
.. (12)
.
Ck1 Bk1 Bk B2k2 Ak1
where
Ci = [cid , cid+1 , , cid+(d1) ], for 0 i k 1, (13)
Ai = [aid , aid+1 , , aid+(d1) ], for 0 i k 1, (14)
which are all d 1 vectors, and

bid bid+1 bid+(d1)
bid+1 bid+2 bid+d
Bi = .
.. .
.. . .. .. , for 0 i 2k 2,
(15)
.
bid+(d1) bid+d bid+(2d2)
which are all d d Hankel matrices whose corresponding Hankel vectors are
Bi = [bid , bid+1 , , bid+(d1) , bid+d , , bid+(2d2) ], for 0 i 2k 2. (16)
If m is divisible by d, then m = kd. Otherwise, (k 1)d < m < kd. In this case, the ap-
pending bits am , , akd1 in Ak1 should be set to 0. The redundant bits cm , , ckd1 in
the resulting Ck1 should be disregarded. The additionally required bits b2m1 , , b2kd2
in B2k2 are also calculated according to (3). With the above Hankel matrix-vector
representation, the following equation can be obtained from (12) :
Ci = B i A0 + B i+1 A1 + + B i+k1 Ak1

k1
(17)
= B i+j Aj , for 0 i k 1.
j=0
The above equation implies that each digit of the product word C , i.e., Ci , can be
obtained with the summation of the k Hankel matrix-vector multiplications, i.e., B i+j
Aj , for 0 j k 1.
To compute C0 , according to (17), the Hankel vectors B 0 , B 1 , , B k1 are required.
(0)
These vectors can be generated from vector B = [b0 , b1 , , bm1 , bm , , bkd+d1 ] whose
former part of coordinates, b0 , b1 , , bm1 , are exactly those of the original input vector
(0)
B. Namely, bi = bi , for 0 i m 1. However, its latter part of coordinates,
(0)
bm , bm+1 , , bkd+d1 , i.e., bi for m i kd + d 1, must be derived from its
former part of coordinates. When the generating function is an irreducible trinomial, i.e.,
F (x) = xm + xn + 1, the values of bm , bm+1 , , bkd+d2 , bkd+d1 can be pre-computed
simultaneously as follows: Let be the root of F (x), then m = n + 1. The fact that
bm+i is defined as T r(m+i B) leads to
( )
bm+i = T r(m+i B) = T r(m i B) = T r (n + 1)i B
(18)
= T r(n+i B) + T r(i B) = bn+i + bi , for 0 i (k + 1)d m 1.
When d is selected as d m/ 2 and n satisfies the condition n m 2d, then n + (k +

1)d m 1 m 1. Using (18) leads to a situation in which [bm , bm+1 , , bkd+d1 ] =
[(bn + b0 ), (bn+1 + b1 ), , (bn+(k+1)dm1 + b(k+1)dm1 )]. Namely, the values can be pre-
computed simultaneously from b0 , b1 , , bm1 with (k + 1)d m XOR gates since the
addition of two bits in GF(2) can be performed by a XOR gate.
Computing Ci , 1 i k 1, requires Hankel vectors B i , B i+1 , , B i+k1 which can
be generated from the vector [bid , bid+1 , , bid+m1 , bid+m , , bid+kd+d1 ] which is defined
(i)
as B . Restated,
(i)
[ ]
(i) (i) (i) (i) (i)
B = b0 , b1 , , bkd1 , bkd , , bkd+d1
(19)

= [bid , bid+1 , , bid+kd1 , bid+kd , , bid+kd+d1 ] .
(i) (i1) (i1)

B can be obtained by the operation of d B because the j-th coordinate of d B
is
T r(j d B (i1) ) = T r(j+d B (i1) )

(20)
= b(i1)d+(j+d) = bid+j , for 0 j kd + d 1,
(i) (i) (i1)

which is exactly the j-th coordinate, bj , of B . The d B operation can be divided
(i)
[ ]
(i) (i) (i)
into two parts. For the former part of B , i.e., b0 , b1 , , bkd1 = [bid , bid+1 , , bid+kd1 ],
(i1)
it is directly obtained by a d-bit left-shifting operation on B because
[ ]
(i) (i) (i)
b0 , b1 , , bkd1 = [bid , bid+1 , , bid+kd1 ]
= [b(i1)d+d , b(i1)d+d+1 , , b(i1)d+d+kd1 ] (21)
[ ]
(i1) (i1) (i1)
= bd , bd+1 , , bd+kd1 .
[ ]
(i) (i) (i)
For the latter part, i.e., bkd , bkd+1 , , bkd+d1 = [bid+kd , bid+kd+1 , , bid+kd+d1 ], it is
(i1)
computed from the coordinates of B as follows:
(i)
bkd+j = bid+kd+j = T r(id+kd+j B) = T r(m kdm id+j B)
( )
= T r (n + 1)kdm id+j B = T r(n+kdm+id+j B) + T r(kdm+id+j B)
(22)
= b(i1)d+n+(kdm)+d+j + b(i1)d+(kdm)+d+j
(i1) (i1)
= bn+(kdm)+d+j + b(kdm)+d+j , for 0 j d 1.
For a situation in which d is selected as d m/2 and n satisfies the condition
n m 2d, then n + (kd m) + d + (d 1) kd 1. According to (22), we obtain that
[ ] [( )
(i) (i) (i) (i1) (i1)
bkd , bkd+1 , , bkd+d1 = bn+(kdm)+d + b(kdm)+d ,
( ) ( )] (23)
(i1) (i1) (i1) (i1)
bn+(kdm)+d+1 + b(kdm)+d+1 , , bn+(kdm)+2d1 + b(kdm)+2d1
(i)
Namely, the lattermost d coordinates of B can be computed simultaneously from the
(i1)
coordinates of B with d XOR gates since two bits in GF(2) can be added by a XOR
gate.
(i)
In sum, the following equations of B are obtained from the above derivation:
(0)
[ ]
(0) (0) (0) (0) (0)
B = b0 , b1 , , bm1 , b(0)
m , b m+1 , , b kd+d1
= [b0 , b1 , , bm1 , bm , bm+1 , , bkd+d1 ] (24)

[ ]
= b0 , b1 , , bm1 , (bn + b0 ), (bn+1 + b1 ), , (bn+(k+1)dm1 + b(k+1)dm1 ) .
(i)
The recursive form of B , for 1 i k 1, is
(i)
[ ]
(i) (i) (i) (i) (i) (i)
B = b0 , b1 , , bkd1 , bkd , bkd+1 , , bkd+d1
(i1) (i1)
[ ]
(i1) (i1) (i1)
= d B = (B d) + 0, 0, , 0, bd+kd , bd+kd+1 , , b2d+kd1
[ ]
(i1) (i1) (i1) (i1) (i1) (i1)
= bd , bd+1 , , bd+kd1 , bd+kd , bd+kd+1 , , b2d+kd1 (25)
[ ( )
(i1) (i1) (i1) (i1) (i1)
= bd , bd+1 , , bd+kd1 , bn+(kdm)+d + b(kdm)+d ,
( ) ( )]
(i1) (i1) (i1) (i1)
bn+(kdm)+d+1 + b(kdm)+d+1 , , bn+(kdm)+2d1 + b(kdm)+2d1 .
Based on the above derivation, a scalable dual-basis multiplication algorithm with digit
size d is derived as follows:
Algorithm 2:
Input: A = [a0 , a1 , , am1 ], B = [b0 , b1 , , bm1 ]
Output: C = [c0 , c1 , , cm1 ] = AB
1. Initial step
1.1 Clear each output sub-vector Ci , 0 i k 1, k = m/d
1.2 Construct each sub-vector Ai , 0 i k 1, from A according to (14).

(0)
1.3 Generate B from B according to (24).
2. Multiplication step
2.1 For i = 0 to k 1 do
(i)
2.2 Generate Hankel vectors B i , B i+1 , , B i+k1 from B according to (16)
and (19).
2.3 For j = 0 to k 1 do
2.4 Ci = Ci + B i+j Aj
2.5 Endfor
(i+1) (i)
2.6 Generate B = d B according to (25).
2.7 Endfor
3. Return C = [C0 , C1 , , Ck1 ].
With Algorithm 2, the PB element A is divided into k sub-vectors Ai ; in addition, DB
(0)
element B is transformed into vector B from which the k Hankel vectors B 0 , B 1 , ,
B k1 are generated. The complete product result vector C is obtained after all k rounds
are computed. In each computation round, the sub-vector Ci is obtained by summing
up the iteratively performed k times of Hankel multiplication (Step 2.4). The required k
(i)
Hankel vectors B i , B i+1 , , B i+k1 in each round are generated from vector B which
(i1)
is transformed recursively from B (Step 2.6). An illustrative example is provided as
follows:
Example 3.1. Assume that finite field GF(28 ) is constructed from F (x) = x8 + x2 + 1
and the digit size d is set to 3. Namely, m = 8, n = 2 and k = m/ d = 8/3 = 3. The
original input vectors are A = [a0 , a1 , a2 , a3 , a4 , a5 , a6 , a7 ], B = [b0 , b1 , b2 , b3 , b4 , b5 , b6 , b7 ].
Then, we get
A0 = [a0 , a1 , a2 ], A1 = [a3 , a4 , a5 ], A2 = [a6 , a7 , 0]
and
(0)
[ ]
(0) (0) (0) (0) (0) (0) (0)
B = b0 , b1 , , b7 , b8 , b9 , b10 , b11 = [b0 , b1 , , b7 , b8 , b9 , b10 , b11 ]
= [b0 , b1 , , b7 , (b2 + b0 ), (b3 + b1 ), (b4 + b2 ), (b5 + b3 )].
(0)
In round 0, B 0 , B 1 , B 2 are generated from B :
B 0 = [b0 , b1 , b2 , b3 , b4 ], B 1 = [b3 , b4 , b5 , b6 , b7 ], B 2 = [b6 , b7 , b8 , b9 , b10 ].
The sub-vector C0 is then derived with
C0 = [c0 , c1 , c2 ] = B 0 A0 + B 1 A1 + B 2 A2 .
(1) (0)
In round 1, B is transformed from B :
(1)
[ ] (0)
(1) (1) (1) (1) (1) (1)
B = b0 , b1 , , b8 , b9 , b10 , b11 = 3 B
[ ]
(0) (0) (0) (0) (0) (0) (0) (0) (0)
= b3 , b4 , , b11 , (b6 + b4 ), (b7 + b5 ), (b8 + b6 ) .
(1)
Additionally, B 1 , B 2 , B 3 are generated from B :
[ ] [ ] [ ]
(1) (1) (1) (1) (1) (1) (1) (1) (1) (1) (1) (1) (1) (1) (1)
B 1 = b0 , b1 , b2 , b3 , b4 , B 2 = b3 , b4 , b5 , b6 , b7 , B 3 = b6 , b7 , b8 , b9 , b10 .
C1 = [c3 , c4 , c5 ] = B 1 A0 + B 2 A1 + B 3 A2 .
(2) (1)
In round 2, B is transformed from B :
(2)
[ ] (1)
(2) (2) (2) (2) (2) (2)
B = b0 , b1 , , b8 , b9 , b10 , b11 = 3 B
[ ]
(1) (1) (1) (1) (1) (1) (1) (1) (1)
= b3 , b4 , , b11 , (b6 + b4 ), (b7 + b5 ), (b8 + b6 ) .
(2)
And B 2 , B 3 , B 4 are generated from B :
[ ] [ ] [ ]
(2) (2) (2) (2) (2) (2) (2) (2) (2) (2) (2) (2) (2) (2) (2)
B 2 = b0 , b1 , b2 , b3 , b4 , B 3 = b3 , b4 , b5 , b6 , b7 , B 4 = b6 , b7 , b8 , b9 , b10 .
C2 = [c6 , c7 , c8 ] = B 2 A0 + B 3 A1 + B 4 A2 .
The complete result C = [C0 , C1 , C2 ] is finally obtained. Notably, the redundant bit c8
should be disregarded.
3.3. Proposed architecture. Deriving the proposed scalable architecture requires ad-
ditional equations. The sub-vectors of d-bit length are defined as follows:
(i)
[ ]
(i) (i) (i)
B j = bjd , bjd+1 , , b(j+1)d1 , for 0 i k 1, 0 j k. (26)
(i) (i)
Next, vector B in each round can be expressed as the composition of B j :
(i)
[ (i) (i) (i) (i) (i)
]
B = B 0 , B 1 , B 2 , B k1 , B k , (27)
and the subsequently generated Hankel vectors B i+j expressed as

[ (i) (i) ]
(i)
B i+j = B j , B j+1 \b(j+2)d1 , for 0 j k 1, (28)
(i) (i) (i) (i)
where B j+1 \b(j+2)d1 denotes removing the lattermost bit b(j+2)d1 from B j+1 . Addition-
ally, deriving the recursive equation in (25) leads to
(i) (i1)
B j = B j+1 , for 0 j k 1, (29)
(i)
[ ] (i1) (i1) (i1)
(i) (i) (i)
and B k = bm , bm+1 , , bm+d1 can be calculated from B 1 , B 2 , , B k with
d XOR gates.
Based on Algorithm 2 and the above equations, the proposed scalable architecture for
dual-basis multiplication over GF(2m ) is constructed (Figure 1). The proposed architec-
ture is largely composed of one d d systolic Hankel multiplier, three registers for A,
B and C respectively, one summation circuit () for C and one recursion circuit (d
block) for B. The d d systolic Hankel multiplier (Figure 2) is applied to perform the
Hankel matrix-vector multiplication B i+j Aj , and is composed of d2 U-cell. Each U-cell
(Figure 3) consists of one AND gate, one XOR gate and two 1-bit latches. Here, the AND
gate and XOR gate are used to multiply and add two bits in GF(2), respectively. The
latches are then used for synchronization of computation. This systolic Hankel multiplier
resembles the one presented in [12].
To perform the Hankel multiplication B i+j Aj , the sub-vector Aj (Equation (14)) is
sent into the d d systolic Hankel multiplier through the d bits [a0 , a1 , , ad1 ] on the
upper side. The Hankel vector B i+j (Equation (28)) is sent into the multiplier through
2d 1 bits [b0 , b1 , , bd1 , bd , , b2d2 ] on the upper and right sides. Meanwhile, a d-bit
zero vector is sent into the multiplier from the left side. The d-bit product outcome B i+j
Aj is then obtained from the d bits [c0 , c1 , , cd1 ] on the right side of the multiplier.
Figure 1. Proposed scalable DB multiplier over GF(2m )
Register A consists of k d-bit latches (A0 , A1 , ( , Ad1 ) and functions as ) a circular-

(i) (i) (i) (i)
shift register. Register B comprises k + 1 banks B 0 , B 1 , , B k1 , B k , which are
all d-bit latches. With the assistance of multiplexers (MUXs), when the control signal
ctr1 = 0, register B functions as a circular-shift register. When ctr1 = 1, register B
(i) (i1)
performs the recursive transformation operation B = d B expressed in (25). The
(i)
d block here consists of d XOR gates and generates the lattermost d coordinates of B
(i1)
from B (Equation (23)). As a d-bit latch, register C is responsible for accumulating
and outputting the sub-vector Ci in each computation round. When the control signal
of the switches (SWs) ctr2 = 0, register C accumulates the outputs of the Hankel mul-
tiplier, i.e. B i+j Aj . When ctr2 = 1, register C outputs the accumulated result, i.e.
k1
j=0 B i+j Aj , which is exactly the sub-vector Ci .
Figure 2. Systolic DB Hankel multiplier used to perform B i+j Aj
Figure 3. Detailed circuit of a U-cell

As the computation process begins, register C is initially cleared (Step 1.1). The input
vector A is divided into k sub-vectors Aj and, then, stored into register A (Step 1.2).
(0)
Input vector B is transformed into B (Step 1.3), which is divided into k + 1 sub-vectors
(0)
B j (Equation (26)) and then stored into register B. In the first round (round 0), the
control signals ctr1 and
[ ctr2 are all set to ]0. Register B performs now as a circular-shift
(0) (0) (0)
register. Thus, B j = B j , B j+1 \b(j+2)d1 and Aj , 0 j k 1, are sequentially sent
into the Hankel multiplier to perform the product operations B j Aj . Meanwhile, register
C accumulates the outputs of the Hankel multiplier and thus performs the summation
operation C0 = C0 + B j Aj (Step 2.4). When the multiplier outputs the product of
B k1 Ak1 , the signal ctr2 is transformed to value 1 in order to make register C output
the
[ (0)result C0 . However,
] signal ctr1 is transformed to the value 1 when the data B k1 =
(0) (0)
B k1 , B k \bm+d1 and Ak1 are to be sent into the multiplier. At that time, the content
[ (0) (0) (0) (0) (0) (0)
]
of register B (from bank0 to bankk ) is B k1 , B k , B 0 , B 1 , , B k3 , B k2 . With the
eect
[ (0) of the MUXs when ctr1 = 1, ] the contents are, at the [next clock cycle, changed to
(0) (0) (0) (0) (1) (1) (1) (1)
B 1 , B 2 , B 3 , , B k , B k+1 which is exactly equal to B 0 , B 1 , B 2 , , B k1 ,
(1)
] (0) (1)
B k (according to (29)). Notably, the content of bankk is changed to B k+1 = B k ,
i.e., the output of d block. Simultaneously, the next round of computation (round 1)
begins. The signals ctr1 and ctr2 are all changed back to 0 and the next computation
C1 = C1 + B 1+j Aj is performed and outputted with same scheme as that in round 0.
The remaining sub-vectors, i.e., Ci , 2 i k 1, are then sequentially computed in the
same manner and outputted from register C in the following rounds i.
4. Time and Space Complexity. The proposed scalable DB multiplier architecture

contains one d d systolic Hankel multiplier (Figure 2), which consists of d2 U-cells.
Each U-cell (Figure 3) consists of one AND gate, one XOR gate and two 1-bit latches.
The summation circuit () for C and the d block are all composed of d XOR gates.
Additionally, registers A, B and C are composed of k d-bit latches, k + 1 d-bit latches and
one d-bit latch, respectively. Thus, for the proposed scalable DB multiplier architecture,
totally d2 AND gates, d2 + 2d XOR gates and 2d2 + 2kd + 2d 1-bit latches are required.
Moreover, there are d switches (SWs) for register C and (k + 1)d multiplexers (MUXs)
for register B. Table 1 lists these gate counts.
Table 1. Comparisons between various multipliers over GF(2m )
Proposed
Multiplier Kim et al. [7] Ibrahim et al. [5] Lee et al. [12]
(Figure 1)
Basis Polynomial Dual Dual Dual
Architecture Digit-serial Digit-serial Bit-parallel Scalable
Space complexity
#2-input AND k(2d2 + d) 2kd2 m2 d2
#2-input XOR 2kd2 2kd2 m2 d2 + 2d
#1-bit latch 10kd + k 6kd 2m2 + m 2kd + 2d2 + 2d
#1 2 SW 0 0 m d
#2 1 MUX 2kd d m kd + d
Critical path dTA + dTX TA + 2dTX
TA + TX + TL TA + TX + TL
delay time +(d 1)TM U X + TL +TM U X + TL
Latency
3k 2k 2m k 2 + 2d 2
(unit = cycle)
k = m/d, d: selected digit size
As for the computation latency of the proposed scalable multiplier, Algorithm 2 clearly
indicates that k computation rounds and then k Hankel matrix-vector computations in
each round are required. Hence, totally k 2 Hankel matrix-vector computations are re-
quired to perform a complete m-bit multiplication. Each Hankel matrix-vector compu-
tation performed with the d d systolic Hankel multiplier (Figure 2) requires a latency
of 2d 1 clock cycles [12]. Moreover, due to the feedback structure of the summation
circuit for register C, sub-vector Ci is outputted after k clock cycles in each computation
round. The lattermost sub-vector Ck1 is outputted after k computation rounds. Thus,
the total latency required to obtain the desired complete result of vector C is k 2 + 2d 2
clock cycles. Additionally, critical path delay is the time duration required by each U-cell
in the d d systolic Hankel multiplier. This time duration is TA + TX + TL where TA ,
TX and TL refer to the time delay of a 2-input AND gate, a 2-input XOR gate and an
1-bit latch, respectively. Table 1 summarizes the above space complexity (gate count) and
time complexity (latency and critical path delay) of the proposed scalable multiplier. This
table also compares the proposed multiplier architecture with other non-scalable multipli-
ers (bit-parallel [12] and digit-serial [5,7]). According to this table, the proposed scalable
multiplier architecture has a lower space complexity in terms of total gate count (O(d2 ))
than the non-scalable architectures (O(m2 ) and O(kd2 ) for bit-parallel and digit-serial
ones, respectively).
Next, this more closely examines the trade-o between throughput performance and
hardware complexity for a given word length m, where the time-area complexity in terms
of the product of transistor count and latency in ns is calculated. Here, the 2-input XOR,
2-input AND, 1 2 SW, 2 1 MUX and 1-bit latch are assumed to be composed of
6, 6, 6, 6 and 8 transistors, respectively [6]. Their typical time delays are obtained as
TX = 12ns, TA = 7ns, TM U X = 11ns and TL = 13ns from [22]. Thus, according to Table
1, the total transistor count of the proposed multiplier architecture can be calculated as
28d2 +22kd+40d. Its latency in ns is (7+12+13)(k 2 +2d2)ns, i.e., the product of critical
path delay and latency in cycles. In the same manner, the total transistor counts of the
digit-serial multipliers of Kim [7] and Ibrahim [5] can be calculated as 24kd2 + 98kd + 8k
and 24kd2 + 48kd + 6d, respectively. Their latencies are 3k(7d + 12d + 11(d 1) + 13)
ns and 2k(7 + 24d + 11 + 13)ns, respectively. Figure 4 shows the above total transistor
counts versus selected digit size d for GF(2233 ), indicating that the transistor count of the
proposed scalable multiplier architecture is always lower than that of the other two digit-
serial multipliers. Figure 5 illustrates the time-area complexity, i.e., the product of total
transistor count and latency in ns, versus selected digit size d for GF(2233 ). According to
this figure, the proposed scalable multiplier architecture has a lower time-area complexity
than the other two digit-serial multiplier architectures when digit size d 5. A minor
limitation is that once d is chosen to be too small (such as < 10), the latency increases
rapidly. For the other four finite fields, GF(2163 ), GF(2283 ), GF(2409 ) and GF(2571 ), used
in ECDSA, similar complexity plots as Figures 4 and 5 can be obtained in the same
manner. The superiority of the proposed multiplier architecture over conventional ones
is thus clearly demonstrated. Moreover, according to Figure 5, the optimal digit size d
can be easily selected in practice (d = 17 for GF(2233 )) to achieve minimal time-area
complexity, i.e., the optimal trade-o between throughput performance and hardware
complexity for the proposed scalable multiplier. Table 2 lists the optimal digit size d of
the proposed scalable multiplier architecture to achieve minimal time-area complexity for
the five GF(2m ) used in ECDSA. This data provides a valuable reference for eorts to
implement ECDSA in resource-constrained environments such as embedded systems and
smart cards with the proposed multiplier architecture.
Table 2. Optimal digit size d for GF(2m ) used in ECDSA
m 163 233 283 409 571

dopt 14 17 21 25 35
Figure 4. Transistor count versus digit size d for GF(2233 )
Figure 5. Time-area product versus digit size d for GF(2233 )
5. Conclusions. This work develops a novel scalable architecture for dual basis mul-
tiplication over GF(2m ). By utilizing the block Hankel matrix-vector representation, a
novel low-complexity scalable and systolic dual basis multiplier architecture for GF(2m )
generated by irreducible trinomials is derived and proposed. Analytical results indicate
that both the area and time-area complexities of the proposed scalable architecture are
significantly lower than those of the non-scalable architectures. By selecting an appro-
priate digit size d, the proposed scalable architecture can achieve a satisfactory trade-o
between throughput performance and hardware complexity for implementing ECC cryp-
tosystems such as ECDSA in resource-constrained environments such as smart cards and
embedded systems. Moreover, owing to its regularity, modularity and concurrency, the
proposed scalable multiplier architecture is highly promising for VLSI implementations.
Future eorts should improve upon the core multiplication architecture to further enhance
the total throughput performance.
REFERENCES
[1] D. E. R. Denning, Cryptography and Data Security, Addison-Wesley Longman Publishing Co., Inc.,
1983.
[2] S. T. J. Fenn, M. Benaissa and D. Taylor, GF(2m ) multiplication and division over the dual basis,
IEEE Trans. on Computers, vol.45, no.3, pp.319-327, 1996.
[3] J. H. Guo and C. L. Wang, Digit-serial systolic multiplier for finite fields GF(2m ), IEE Proc. of
Computers and Digital Techniques, vol.145, no.2, pp.143-148, 1998.
[4] I. S. Hsu, T. K. Truong, L. J. Deutsch and I. S. Reed, A comparison of VLSI architecture of finite
field multipliers using dual, normal, or standard bases, IEEE Trans. on Computers, vol.37, no.6,
pp.735-739, 1988.
[5] M. K. Ibrahim and A. Aggoun, Dual basis digit serial GF(2m ) multiplier, International Journal of
Electronics, vol.89, no.7, pp.517-523, 2002.
[6] S. Kang and Y. Leblebici, CMOS Digital Integrated Circuits: Analysis and Design, McGraw-Hill
Professional, 2002.
[7] C. H. Kim, C. P. Hong and S. Kwon, A digit-serial multiplier for finite field GF(2m ), IEEE Trans.
on Very Large Scale Integration Systems, vol.13, no.4, pp.476-483, 2005.
[8] C. Y. Lee, C. C. Chen, W. Y. Lee and E. H. Lu, Scalable and systolic multipliers for gaussian normal
basis of GF(2m ), Proc. of 2006 International Conference on Hybrid Information Technology, Korea,
2006.
[9] C. Y. Lee, C. W. Chiou, J. M. Lin and C. C. Chang, Scalable and systolic montgomery multiplier
over GF(2m ) generated by trinomials, IET Circuits, Devices and Systems, vol.1, no.6, pp.477-484,
2007.
[10] C. Y. Lee, Low complexity bit-parallel systolic multiplier over GF(2m ) using irreducible trinomials,
IEE Proc. of Computers and Digital Techniques, vol.150, no.1, pp.39-42, 2003.
[11] C. Y. Lee and C. W. Chiou, Ecient design of low-complexity bit-parallel systolic Hankel multipliers
to implement multiplication in normal and dual bases of GF(2m ), IEICE Trans. Fundamentals,
vol.E88-A, no.11, pp.3169-3179, 2005.
[12] C. Y. Lee, J. S. Horng and I. C. Jou, Low-complexity bit-parallel multiplier over GF(2m ) using dual
basis representation, Journal of Computer Science and Technology, vol.21, no.6, pp.887-892, 2006.
[13] Y. Li, J. Lam and X. Luo, Hankel norm model reduction of uncertain neutral stochastic time-
delay systems, International Journal of Innovative Computing, Information and Control, vol.5, no.9,
pp.2819-2828, 2009.
[14] A. J. Menezes, P. C. V. Oorschot and S. A. Vanstone, Handbook of Applied Cryptography, CRC
Press, 1996.
[15] M. Morii, M. Kasahara and D. L. Whiting, Ecient bit-serial multiplication and the discrete-time
Wiener-Hopf equation over finite fields, IEEE Trans. on Information Theory, vol.35, no.6, pp.1177-
1183, 1989.
[16] N. Kobliz, Elliptic curve cryptography, Math. Computation, vol.48, pp.203-209, 1987.
[17] National Institute for Standards and Technology, Digital Signature Standard, FIPS Publication 186-
2, 2000.
[18] A. Reyhani-Masoleh and M. A. Hasan, Low complexity bit-parallel architectures for polynomial basis
multiplication over GF(2m ), IEEE Trans. on Computers, vol.53, no.8, pp.945-959, 2004.
[19] M. Y. Rhee, Cryptography and Secure Communications, McGraw-Hill, Singapore, 1994.
[20] A. Skander, M. Nadjim and M. Benslama, A new accurate quantum cryptography control error
reconciliation (QCCER) with XOR operator in BB84 protocol, ICIC Express Letters, vol.2, no.2,
pp.187-192, 2008.
[21] A. Skander, M. Nadjim and B. Malek, Synchronization chaotic communications with closed crypto-
graphic systems, ICIC Express Letters, vol.2, no.3, pp.269-274, 2008.
[22] STMicroelectronics, Logic Selection Guide: STMicroelectronics, http://www.st.com.
[23] A. F. Tenca and C. K. Koc, A scalable architecture for modular multiplication based on Mont-
gomerys algorithm, IEEE Trans. on Computers, vol.52, no.9, pp.1215-1221, 2003.
[24] C. Wei and S. Jiang, Small target detection in FLIR imagery using multi-scale morphological filter
and kernel density estimation, International Journal of Innovative Computing, Information and
Control, vol.5, no.7, pp.1811-1818, 2009.
[25] C. W. Wu and M. K. Chang, Bit-level systolic arrays for finite-field multiplications, The Journal of
VLSI Signal Processing, vol.10, no.1, pp.85-92, 1995.
c ISSN 1349-4198
A FUZZY-BASED ROUGH SETS CLASSIFIER FOR FORECASTING

QUARTERLY PGR IN THE STOCK MARKET (PART II)
Ching-Hsue Cheng1 and Yoush-Yang Chen2,

1
National Yunlin University of Science and Technology
123, Section 3, University Road, Touliu, Yunlin 640, Taiwan
chcheng@yuntech.edu.tw
2
Hwa Hsia Institute of Technology
111, Gong Jhuan Rd., Chung Ho, Taipei 235, Taiwan

Corresponding author: ys chen@cc.hwh.edu.tw
Received September 2009; revised April 2010
Abstract. Three hybrid models are proposed for solving the practical problems of fore-
casting quarterly PGR (profit growth rate) and based on six essential components: expe-
riential knowledge (EK), feature selection method (FSM), discretization method (DM),
fuzzy set theory (FST), rule filter (RF) and rough set theory (RST). Following the Part
I (the previous paper), a called PGR dataset collected from the financial holding stocks
in Taiwans stock market is implemented as the empirical case study in the Part II (this
paper) to evaluate the proposed hybrid models. An external comparison and internal
comparison are conducted; concurrently, some findings and management implications
are disclosed from the experimental results. The results include decision rules of a set to
directly rule the strategy of investment and intelligently oer a powerful explanation for
the investors. They are of value to both academicians and practitioners.
Keywords: Rough set theory, Fuzzy-rule similarity, Profit growth rate (PGR), Feature
selection, Discretization method, Minimize entropy principal approach (MEPA)
1. Introduction. From the Part I (the previous paper) of the study, we had noted that
continually increasing the needs for searching a better way to help investors survive in a
serious environment, such as globally keener, competitive, and dynamic financial climate,
the Asian Economic Crisis in 1997, or a global financial crisis in 2007-2009, is an increased
trend today. Therefore, understanding the financial status and operational performance
of a specific company through an indicator is needed urgently from various perspectives
for them. In practice, one of the most eective ways to assess operating performance of
companies is to conduct a profitability analysis. With the view of stock investors, profit
growth rate (PGR) is one of core financial ratios and is an eective evaluation indicator.
However, academicians rarely explore it in an academic study to forecast stock price for
a short- mid run. The study aims to fill this knowledge gap. As such, the study proposes
three intelligent hybrid models based on rough sets, which include experiential knowledge,
feature selection methods, discretization methods, rule filter, fuzzy set theory and rough
set theory, for classifying quarterly PGR problems faced by investors. Likewise, following
the Part I, the four research objectives are emerged: (1) Conduct three suitable hybrid
models to use fewer selected attributes, fewer generated rules, higher accuracy, and more
stability on forecasting; (2) Examine the determinants of influencing the quarterly PGR;
(3) Evaluate the eects of basic components; and (4) Generate comprehensive decision
rules as knowledge-based systems for investors.
1209
1210 C.-H. CHENG AND Y.-Y. CHEN
The rest of this paper is organized in the following. First, we roughly discuss related
studies of classifying the PGR of a specifically financial industry again. Second, an em-
pirical case study of Taiwans financial holding companies for the proposed hybrid models
is implemented. Verification and comparisons of experimental results are described. Af-
terward, discussion and findings are formulated in the next section, and they are followed
by some conclusions finally.
2. Related Works. This section states profit growth rate, rough set theory, rule filter,
fuzzy set theory, feature selection method, and discretization method. (For a convenience
of a shortened paper, the repeat contents are not presented, and please see the Section 2
of the Part I of the study for the details)
2.1. Profit growth rate (PGR). Simply, PGR refers to the proportion of periodic vari-
ation quarterly of operating profits for a specific company in the study, and there are four
quarters, named first, second, third and fourth quarter in a year. The equation of quarterly
PGR can be calculated in the following: Quarterly PGR (%) = ((Current quarter Profit
Last quarter Profit) 100)/Last quarter Profit. The positive PGR represents good and
optimistic with company for future, whereas negative PGR is bad and pessimistic.
2.2. Rough set theory (RST). RST, first proposed by Pawlak [1], mainly addresses the
continuing problem of vagueness, uncertainty and incompletion by applying the concept
of equivalence classes to partition training instances according to specified criteria. Thus,
the RST approach is based on refusing certain set boundaries, implying that every set
will be roughly defined using a lower and an upper approximation [2].
2.3. Rule filter (RF). Rough set-mediated rule sets extracted by the LEM2 (Learning
from Examples Module, version 2) [3] method usually contain large numbers of distinct
rules [4]. Therefore, a filtering rule solution is used to reduce the number of rules and
based on the computed quality indices of rules in a rule-set. The quality index of each
rule is computed using a specific rule quality function, which determines the strength of
a rule, based on the measure of support, consistency, and coverage.
2.4. Fuzzy set theory (FST). To deal with the uncertain and imprecise data in real-
world applications, Zadeh [5] first proposed FST (also called fuzzy logic). This theory
was developed based on the concept of traditional crisp sets, and it is that an element
belongs to a fuzzy set with a certain degree of membership. He indicated that a linguistic
variable diers from a numerical variable in that its values are not numbers, and they
are words or sentences in a natural or artificial language, such as small, medium, and
large.
2.5. Feature selection method (FSM). One of the critical aspects of any knowledge
discovery process is feature selection [6] that will help to evaluate usefulness of attributes,
to select relevant attributes, and then to reduce the dimensionality of datasets by delet-
ing unsuitable attributes that may degrade the performance of classification, and hence
improves the performance of data mining algorithms [7].
2.6. Discretization method (DM). The study mainly introduces two discretization
methods, minimize entropy principle approach (MEPA) and decision trees (DT) C4.5,
as representative. On the one hand, a key goal of entropy minimization analysis is to
determine the quantity of information in a given data set. The entropy of a probability
distribution is a measure of the uncertainty of the distribution [8]. To subdivide the data
into membership functions, the threshold among classes of data must be established. On
the other hand, DT is a flow-chart-like tree structure where each internal node denotes
A FUZZY-BASED ROUGH SETS CLASSIFIER FOR FORECASTING QUARTERLY PGR 1211
a test on an attribute; each branch represents an outcome of the test, and leaf nodes
represent class or class distributions. The ID3 [9] is a DT algorithm based on information
theory. The basic strategy used by ID3 is to choose splitting attributes with the highest
information gain.
3. An Empirical Case Study for the Proposed Hybrid Models. For solving the
practical problems faced by investors, the study proposes a set of hybrid models for fore-
casting quarterly PGR and dedicating to improve the quality of a rough set classification
system. The proposed hybrid models involve three dierent types of models (named mod-
els A-C), constituted by various combinations of six essential treatments: EK (experiential
knowledge), FSM, DM, FST, RF, and RST. Accordingly, to illustrate and verify the three
proposed hybrid models systematically, a practically collected dataset is selected from 70
publicly traded financial holding stocks as the experimental dataset (called PGR dataset
in the study), which is the quarterly financial reports of them in 2004-2006 in the Taiwan
stock market to demonstrate the proposed hybrid models. There are, in total 636, in-
stances, which are characterized by the following nine (including condition and decision)
attributes that are selected by EK, in the experimental dataset: (i) F assets, (ii) T assets,
(iii) T liabilities, (iv) O expenses, (v) O income, (vi) T salary, (vii) A costs, (viii) C type,
and (ix) PGR (Class); except for the C type, all attributes are continuous data. Based
on the EK, the Class in the experimental dataset is partitioned into three classes: Fair
(PGR > 50%), Good (PGR = 0-50%), and Poor (PGR < 0%). The contented attributes
of this dataset are described in Table 1. For example, the first F-assets attribute in-
cludes continuous data, and its value range is from a minimum of 15990 to a maximum
of 47131292.
Table 1. The attributes information in the PGR dataset
Attributes Attributes Numbers
No. Notes
name information of value
1 F-assets numeric continuous Min: 15990 Max: 47131292
2 T-assets numeric continuous Min: 8390514 Max: 2380547669
3 T-liabilities numeric continuous Min: 5375826 Max: 2297331093
4 O-expenses numeric continuous Min: 1994478 Max: 11852475
5 O-income numeric continuous Min: 34328398 Max: 9318711
6 T-salary numeric continuous Min: 701805 Max: 7535360
7 A-costs numeric continuous Min: 4530602 Max: 188208927
8 C-type symbolic 4 A-D refer to dierent types of company
9 PGR (Class) numeric continuous Min: 1671.79 Max: 247.28
(Unit: thousands in New Taiwan Dollars)
3.1. The proposed hybrid models A-C using the PGR dataset. Initially, the
computing processes of model A using the PGR dataset are implemented, as follows:
Step 1: Collect the attributes of a practical dataset by EK. Based on professional
knowledge in the financial industry, a practiced dataset for 70 publicly traded financial
holding stocks listed on TSEC (Taiwan Securities Exchange Corporation) and OTC (over
the counter) is the experimental dataset. Accordingly, collect the data that are quarterly
financial reports of the 70 financial holding stocks during 2004-2006.
Step 2: Preprocess the dataset. Following Step 1, records with inconsistent values
were firstly deleted, and redundant attributes were eliminated. Accordingly, data are
transformed into EXCEL format. Totally, in the PGR dataset, the F-assets, T-assets,
T-liabilities, O-expenses, O-income, T-salary, A-costs, C-type, and PGR repre-

sent total fixed assets, total assets, total liabilities, operating expenses, operating income,
total salary, agency costs, company type, and profit growth rate of accounting terms on
quarterly financial reports, respectively. Table 2 shows the partial raw data of the PGR
dataset after preprocessing. For example, the values of first record for the nine attributes
are 25071214, 1355605332. . . , and 28.84, respectively.
Table 2. The partial raw data of the PGR dataset after preprocessing
No. F-assets T-assets T-liabilities O-expenses O-income T-salary A-costs C-type PGR
1 25071214 1355605332 1273420783 2826210 3128226 1582198 11149973 A 28.84
2 25119198 1337366451 1258204709 2720560 3975026 1494924 4742858 A 34.75
3 25054427 1296335446 1214666878 2683450 2659381 1489657 4930083 A 25.89
.. .. .. .. .. .. .. .. .. ..
. . . . . . . . . .
634 1889918 167781224 158717933 464557 274609 913035 13935430 D 1.87
635 1872991 154250134 145627009 330814 534620 718302 13725072 D 3.66
636 897360 12464816 5847993 227779 188612 100021 2277396 D 1.25
(Unit: thousands in New Taiwan dollars)
From Table 2, the first eight attributes are condition attributes, the last attribute PGR
refers to a decision attribute, and the study renames it Class. The Class attribute is
classified into three types by EK: F (Fair), G (Good), and P (Poor).
Step 3: Build the membership function by MEPA. Additionally, the continuous at-
tributes will be partitioned. The entropy value of these attributes is computed by em-
ploying MEPA. Repeating this procedure to subdivide the data obtains the cuto points
(thresholds). Table 3 shows thresholds of the seven continuous attributes in the PGR
dataset. Take an example. For the F assets attribute, five thresholds are Min, SEC1,
PRI, SEC2, and Max, which correspond with values of 15990, 130487, 2491896.5,
4565094, and 47131292, respectively. These cuto points of continuous attributes are
used as the midpoint of membership function. The membership degree equals 1 when
the attribute value is lower than SEC1 and the attribute value exceeds SEC2. Hence, the
membership function is generated based on these cuto points. Figure 1 illustrates the
membership function of the F assets attribute by the MEPA. According to the member-
ship function, the maximal degree of membership for the F assets attribute is calculated
to determine its linguistic value; that is, L 1 (small), L 2 (medium), and L 3 (large).
Table 3. Thresholds of the seven continuous attributes in the PGR dataset
Attributes Min SEC1 PRI SEC2 Max
F assets 15990 130487 2491896.5 4565094 47131292
T assets 8390514 254216959.5 288095831 388986088.5 2380547669
T liabilities 5375816 24132180.5 226256240 329154013.5 2297331093
O expenses 1994478 82276 170450 3048754.5 11852475
O income 34328398 25725926 132.5 195489 9318711
T salary 701805 15548.5 96118 1659794 7535360
A costs 4530602 421298.5 1401736.5 2373431 188208927
Take an example of the F assets attribute, and its corresponding value ranges for
L 1, L 2, and L 3 are [15990, 2491896.5), [130487, 4565094), and [2491896.5, 47131292],
respectively.
Step 4: Fuzzify the continuous data into unique corresponding linguistic values. Using
the membership function derived from the previous step, the membership degrees of each
data are recalculated. Table 4 shows the membership degrees for all continuous attributes
Figure 1. The membership function of the F assets attribute by the MEPA
fuzzied in the PGR dataset. Consider an example. The membership degrees of three
linguistic values, L 1, L 2, and L 3, in the F assets attribute on record 1 are 0, 0, and 1,
respectively. Clearly, the maximal degree of membership is 1, which falls into L 3. Hence,
the first record in the F assets attribute belongs to a linguistic value L 3.
Table 4. The membership degrees for continuous attributes fuzzied in the
PGR dataset
F-assets T-assets T-salary A-costs
No.
L1 L2 L3 L1 L2 L3 L1 L2 L3 L1 L2 L3
1 0 0 1 0 0 1 0 0.05 0.95 0 0 1
2 0 0 1 0 0 1 0 0.11 0.89 0 0 1
3 0 0 1 0 0 1 0 0.11 0.89 0 0 1
.. .. .. .. .. .. .. .. .. .. .. .. ..
. . . . . . . . . . . . .
634 0.86 0.14 0 1 0 0 0 0.99 0.01 0 0.48 0.52
635 0.86 0.14 0 1 0 0 0 0.99 0.01 0 0 1
636 0.82 0.18 0 1 0 0 0 0.99 0.01 0 0.19 0.81
Step 5: Select the similarity threshold by FST. To find the closeness between the
decision attribute and each subgroup of condition attributes, the study calculates the
subsethood of them. At first, calculate the subsethood of each continuous attribute by
Equation (14) in the Part I of the study. Based on the foregoing statement, the Class
attribute is classified into three types: Fair (F), Good (G), and Poor (P); there are three
corresponding subgroups: subgroup1, subgroup2, and subgroup3, respectively. Table 5
shows the subsethoods of all continuous attributes based on the three subgroups above
in the PGR dataset. For instance, in the F assets attribute, the S(F; L 1)=0.19 refers
to when the label of Class attribute is F (fair), the linguistic value of the F assets
attribute is L 1 (small), and then the subsethood S equals 0.19. By the same way, the
S(G; L 2)=0.22 refers to when the label of Class attribute is G (good), the linguistic
value of the F assets attribute is L 2 (medium), and then its subsethood S equals 0.22.
Likewise, the S(P; L 3)=0.08 refers to when the label of Class attribute is P (poor),
the linguistic value of the F assets attribute is L 3 (large), and then its subsethood S
equals 0.08.
Accordingly, compute the similarity threshold by Equation (15) in the Part I of the
study, and Table 6 shows the similarity threshold of all continuous attributes in the
the PGR dataset. For example, the similarity threshold of L 1 linguistic value in the
F assets attribute is selected, as follows:
Max{S(F; L 1), S(G; L 1), S(P; L 1)} = Max{0.19, 0.63, 0.73} = 0.73.
Similarly, Max{S(F; L 2), S(G; L 2), S(P; L 2)} = Max{0.30, 0.22, 0.19} = 0.30,
and Max{S(F; L 3), S(G; L 3), S(P; L 3)} = Max{0.51, 0.15, 0.08} = 0.51.
Hence, in the F assets attribute, the similarity thresholds of linguistic values L 1, L 2,
and L 3 are 0.73, 0.30, and 0.51, respectively. By the same computation using Equation
(15) in the Part I of the study, the similarity thresholds of linguistic values L 1, L 2, and
L 3 for the T assets attribute are 0.54, 0.16, and 0.81, respectively.
Table 5. The subsethoods of the seven continuous attributes in the PGR dataset
Decision attribute (class)
Condition attributes
Subgroup1(F) Subgroup2(G) Subgroup3(P)
S(F; L 1)=0.19 S(G;L 1)=0.63 S(P; L 1)=0.73
F assets S(F; L 2)=0.3 S(G; L 2)=0.22 S(P; L 2)=0.19
S(F; L 3)=0.51 S(G; L 3)=0.15 S(P; L 3)=0.08
S(F; L 1)=0.07 S(G; L 1)=0.54 S(P; L 1)=0.53
T assets S(F; L 2)=0.12 S(G; L 2)=0.07 S(P; L 2)=0.16
S(F; L 3)=0.81 S(G; L 3)=0.39 S(P; L 3)=0.3
S(F; L 1)=0.05 S(G; L 1)=0.57 S(P; L 1)=0.64
T liabilities S(F; L 2)=0.05 S(G; L 2)=0.57 S(P; L 2)=0.6
S(F; L 3)=0.51 S(G; L 3)=0.22 S(P; L 3)=0.12
S(F; L 1)=0.12 S(G; L 1)=0.53 S(P; L 1)=0.55
O expenses S(F; L 2)=0.49 S(G; L 2)=0.42 S(P; L 2)=0.43
S(F; L 3)=0.39 S(G; L 3)=0.05 S(P; L 3)=0.02
S(F; L 1)=1 S(G; L 1)=0.9 S(P; L 1)=0.01
O income S(F; L 2)=0 S(G; L 2)=0.1 S(P; L 2)=0.1
S(F; L 3)=0 S(G; L 3)=0 S(P; L 3)=0.11
S(F; L 1)=0.11 S(G; L 1)=0.49 S(P; L 1)=0.47
T salary S(F; L 2)=0.5 S(G; L 2)=0.46 S(P; L 2)=0.5
S(F; L 3)=0.39 S(G; L 3)=0.05 S(P; L 3)=0.03
S(F; L 1)=0.03 S(G; L 1)=0.74 S(P; L 1)=0.88
A costs S(F; L 2)=0.15 S(G; L 2)=0.15 S(P; L 2)=0.06
S(F; L 3)=0.82 S(G; L 3)=0.11 S(P; L 3)=0.05
Table 6. The similarity threshold of the seven continuous attributes in

the PGR dataset
Condition attributes L1 L2 L3
F assets 0.73 0.30 0.51
T assets 0.54 0.16 0.81
T liabilities 0.64 0.64 0.51
O expenses 0.55 0.49 0.39
O income 1.00 0.10 0.11
T salary 0.49 0.50 0.39
A costs 0.88 0.15 0.82
Step 6: Transform the fuzzied attributes. According to the membership degrees de-
rived from Step 4 and the similarity thresholds derived from Step 5 for corresponding
linguistic values of the seven continuous attributes, transform the fuzzied dataset fur-
ther. As such, the transformed values of the seven continuous attributes are calculated
and generated by Equation (16) in the Part I of the study. For example, the mem-
bership degree Ii of three linguistic values in the F assets attribute is [0 0 1] (see Ta-
ble 4), and the similarity threshold STi is [0.73 0.30 0.51] for the same linguistic values
of the same F assets attribute (see Table 6). Consequently, the membership degree
Ii = [0 0 1] [0.73 0.30 0.51] = [0 0 1] is obtained, and the notation denotes a com-
pared operation. These transformed values, in addition to the original values of other
symbolic attributes, build a Table 7. Table 7 shows the partial data of all nine attributes
after the transformation processes, based on similarity threshold in the PGR dataset. For
example, in the F assets attribute, the transformed values of three linguistic values, L 1,
L 2, and L 3, on record 1 are 0, 0, and 1, respectively, and the class of the PGR attribute
is P.
Table 7. The partial data in the PGR dataset after a transformed method
F-assets T-assets T-salary A-costs
No. C-type PGR
L1 L2 L3 L1 L2 L3 L1 L2 L3 L1 L2 L3
1 0 0 1 0 0 1 0 0 1 0 0 1 A P
2 0 0 1 0 0 1 0 0 1 0 0 1 A G
3 0 0 1 0 0 1 0 0 1 0 0 1 A G
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
. . . . . . . . . . . . . . . .
634 1 0 0 1 0 0 0 1 0 0 1 0 D G
635 1 0 0 1 0 0 0 1 0 0 0 1 D P
636 1 0 0 1 0 0 0 1 0 0 1 0 D G
Step 7: Extract the decision rules by the RST LEM2 algorithm. Accordingly, the 636
instances of the PGR dataset are randomly split up into two sub-datasets: the 67% dataset
is used as a training set, and the other 33% is used as a testing set. The experiment of
the PGR dataset with the 67%/33% ratio is run by using the rough sets LEM2 algorithm,
and the decision rules are generated. Table 8 shows the partial rules set in the PGR
dataset. In Table 8, Rules refers to the decision rules generated by the LEM2 algorithm
for forecasting quarterly PGR, and Supports refers to real examples coinciding with
the generated decision rules. Furthermore, for easy presentation, the F assets, T assets,
T liabilities, O expenses, O income, T salary, and A costs attributes are abbreviated as
FA, TA, TL, OE, OI, TS, and AC, respectively. For example, a notation OI L 1 refers to
the L 1 (small) of threshold on O income attribute. For Rule 1, its class is good, and its
support numbers are 164 in total.
In further explanations of Table 8, the generated rules are in total classified into three
classes of rules, which are Fair, Good, and Poor, regarded as knowledge-based invest-
ment systems by investors. The three classes of rules refer to 636 real examples coinciding
with the generated decision rules in the PGR dataset. To simplify, only an example of
each class is further elucidated, as follows:
(1) Fair-class rule:
If OI L 1=0 & OI L 2=0 & OI L 3=1 & TA L 1=1 & TA L 2=0 & TA L 3=0 &
FA L 1=1 & FA L 3=0 & OE L 3=0 & TS L 1=1 & TS L 2=0 & TS L 3=0 & AC L 1=1
& AC L 2=0 & AC L 3=0 Then Class = Fair (supports = 7).
It indicates that when O income / L 1, O income / L 2, O income L 3, T assets
L 1, T assets / L 2, T assets / L 3, F assets L 1, F assets
/ L 3, O expenses
/ L 3, T salary L 1, T salary
/ L 2, T salary
/ L 3, A costs L 1, A costs /
L 2, and A costs / L 3 occur simultaneously, then the class (PGR) is fair. Furthermore,
seven examples totally support this rule. Take an example. The term F assets L 1
refers to the total salaries, which fall into 15990-2491896.5 (unit: NT$) (see Table 3 and
Figure 1).
Table 8. The partial rules set extracted by LEM2 algorithm in the PGR
dataset example
Rules Supports
1 If OI L 1=0&OI L 2=0&OI L 3=1&TA L 2=0&FA L 1=0&FA L 3=1 164
&TS L 1=0&TL L 1=0&OE L 1=0&AC L 1=0 Then Class = Good
2 If OI L 1=0&OI L 3=1&TL L 1=0&FA L 1=0&OE L 1=0&TS L 1=0 149
&AC L 1=0&FA L 3=1&TL L 2=0&FA L 2=0&TA L 1=0&
OE L 2=0 Then Class = Good
3 If OE L 1=0&OI L 1=0&OI L 2=1&OI L 3=0&TS L 1=0&AC L 1=0 49
&FA L 1=0&TL L 1=0&AC L 3=1&AC L 2=0&TA L 1=0&TL L 2=0
&TL L 3=1 Then Class = Poor
4 If OI L 1=0&OI L 3=1&TA L 3=0&TS L 3=0&FA L 3=0&TL L 3=0 48
&OE L 3=0&TA L 1=1&TA L 2=0&AC L 1=0&TL L 1=1&TL L 2=0
&C type=D Then Class = Good
&OE L 3=0&TA L 1=1&TA L 2=0&AC L 1=0&OE L 1=0&OE L 2=1
&TS L 1=0&TS L 2=1&FA L 1=0&FA L 2=1&AC L 2=0 Then Class
= Good
6 If I L 1=0&OI L 3=1&TL L 1=0&FA L 1=0&OE L 1=0&TS L 1=0 19
&AC L 1=0&TA L 3=0&TS L 2=1&OE L 2=1&TS L 3=0&C type=A
&TA L 1=0 Then Class = Good
7 If OI L 1=0&OI L 3=1&TL L 1=0&FA L 1=0&OE L 1=0&TS L 1=0 9
&TA L 3=0&TS L 2=1&OE L 2=1&FA L 2=1&FA L 3=0&TL L 3=0
&OE L 3=0&TS L 3=0&AC L 3=0&TA L 1=1&TA L 2=0&TL L 2=0
Then Class = Good
&OE L 3=0&TA L 1=1&TA L 2=0&AC L 1=0&OE L 1=0&OE L 2=1
&TS L 1=0&TS L 2=1&AC L 2=1&AC L 3=0&FA L 1=0&FA L 2=1
&TL L 1=0&TL L 2=1&OI L 2=0&C type=B Then Class = Poor
9 If TS L 3=0&OI L 1=0&OI L 2=0&OI L 3=1&AC L 3=0&TA L 3=0 8
&OE L 3=0&TS L 3=0&FA L 3=0&TA L 1=1&TA L 2=0&TS L 1=1
&TS L 2=0&AC L 1=1&AC L 2=0&FA L 1=1 Then Class = Fair
10 If FA L 3=0&TA L 1=1&TA L 2=0&TA L 3=0&TL L 1=1&TL L 2=0 7
&TL L 3=0&OE L 1=0&OE L 2=1&OE L 3=0&OI L 1=0&OI L 2=0
&OI L 3=1& TS L 3=0&AC L 3=0&TS L 1=0&C type=A Then
Class = Fair
(2) Good-class rule:

If OI L 1=0 & OI L 2=0 & OI L 3=1 & TA L 2=0 & FA L 1=0 & FA L 3=1 &
TS L 1=0 & TL L 1=0 & OE L 1=0 & AC L 1=0, then Class = Good (supports =
164).
It indicates that when O income / L 1, O income / L 2, O income L 3, T assets
/ L 2, F assets
/ L 1, F assets L 3, T salary
/ L 1, T liabilities
/ L 1, O expenses
/ L 1, and A costs
/ L 1 occur simultaneously, then the class is good. Furthermore,
164 examples in total support this rule.
(3) Poor-class rule:
If OI L 1=0 & OI L 2=1 & OI L 3=0 & TA L 1=0 & FA L 1=0 & TL L 1=0 &
TL L 2=0 & TL L 3=1 & AC L 1=0 & AC L 2=0 & AC L 3=1 & TS L 1=0 OE L 1=0,
then Class = Poor (supports = 49).
It indicates that when O income / L 1, O income L 2, O income / L 3, T assets
/ L 1, F assets / L 1, T liabilities
/ L 1, T liabilities
/ L 2, T liabilities L 3,
A costs / L 1, A costs
/ L 2, A costs L 3, T salary / L 1, and O expenses
/ L 1 occur simultaneously, then the class is poor. Furthermore, 49 examples in total

support this rule.
Step 8: Improve the rule quality by RF. Using the rules set extracted in Step 7, the
filter rule algorithm based on support guides a filtering process. Rules below the support
threshold (< 2) are eliminated. The performance of the refined rules is conclusively
described in next subsection.
Similarly, the computing processes of models B-C using the PGR dataset are not pre-
sented here to give unnecessary details; only the empirical results are shown in the fol-
lowing two subsections.
3.2. Verification and comparison. The final decision rules derived from three dierent
models are used to test data in the PGR dataset, and the performance of these hybrid
models is then evaluated and described conclusively in this section for convenient readings.
Finally, the three proposed hybrid models are compared with other AI techniques. Some
evaluation criteria, such as the accuracy with standard deviation, the number of attributes,
and the number of rules with its standard deviation, are adopted as the standard of
comparisons. Mainly, the comparison is divided into two-fold: internal comparison and
external comparison.
3.2.1. Internal comparison. The PGR dataset is split into two subsets: the 67%/33%
dataset for training/testing, respectively. The experiment for three proposed hybrid mod-
els is run 10 times repeatedly with the 67%/33% random split, and then compute the
average accuracy rate, respectively. The results of the internal comparison for them are
shown in Table 9, and the accuracy for training/testing, number of attributes, and number
of rules are regarded as criteria.
Table 9. Internal comparison results in the PGR dataset
Number of
Methods Training accuracy Testing accuracy Number of rules
attributes
Model A 99.0% (0.005) 96.99% (0.012) 9 24.5 (3.2)
Model B 100.0% (0.000) 94.86% (0.014) 4 3.0 (0.0)
Model C 100.0% (0.000) 97.41% (0.007) 9 63.4 (6.7)
From Table 9, it shows that the model C outperforms other models in terms of both
accuracy and its standard deviation. The results of the experiments oer insights into the
manner in which DT+RST for model C shows superiority in this dataset. Furthermore,
the model B is superior to other models not only in terms of number of rules but also in
terms of its lower standard deviation. Importantly, it is worth it to point out that only
3.0 rules are generated for four attributes in model B. As for the number of attributes,
the model A is better than the other two models due to less-used attributes.
The experiment without RF for the three proposed hybrid models is run 10 times
repeatedly with the 67%/33% random split again, and then compute average accuracy
rate, respectively. Table 10 shows their comparison results, and the criteria are accuracy,
improved accuracy rate, number of rules, and improved rule rate, respectively.
Table 10. Internal comparison of the three hybrid models without/with

RF in the PGR dataset
Improved Improved
Methods Testing accuracy Number of rules
accuracy rate rule rate
Model A (without RF) 95.21% (0.012) 34.7 (5.2)
1.87%1 29.39%2
Model A (with RF) 96.99% (0.012) 24.5 (3.2)
Model B (without RF) 94.86% (0.014) 3.0 (0.0)
0.00% 0.00%
Model B (with RF) 94.86% (0.014) 3.0 (0.0)
Model C (without RF) 95.98% (0.007) 102.0 (6.9)
1.49% 39.27%
Model C (with RF) 97.41% (0.007) 63.4 (6.7)
1
(0.9699-0.9521)/0.9521=1.87%.
2
(34.7-24.5)/34.7=29.39%.
From Table 10, the proposed model C regardless of being without/with RF is the
best one to forecast quarterly PGR in terms of Testing accuracy (95.98%/97.41%) when
they are compared to the other three models, respectively; however, the proposed model
B is superior in terms of Number of rules both with RF (3.0) and without RF (3.0).
Moreover, the higher improved accuracy rate for without/with RF is 1.87% for model A,
and the higher improved rule rate is 39.27% for model C. Particularly, both the improved
accuracy rate and the improved rule rate of model B are zero (0.00%), which represents
that RF processing is not necessary in the PGR dataset. This interesting phenomenon is
worthy of being explored as the root cause.
3.2.2. External comparison for the PGR dataset. The PGR dataset is split into two sub-
sets, the 67% and 33% dataset, for external comparisons. Accordingly, the experiment
is also repeated 10 times with the 67%/33% random split using four dierent AI meth-
ods: DT-C4.5 [10], Bayes Net [11], Multilayer Perceptron [12], and Traditional Rough
Sets [13], respectively. Table 11 compares the performance of the four AI methods above
with proposed hybrid model C (the best performance in the study) in terms of training
accuracy, testing accuracy, and number of rules.
Table 11. External comparison results in the PGR dataset
Training Testing Number

Methods Attributes
accuracy accuracy of rules
Decision Tree-C4.5 [10] 100.0% (0.0) 95.09% (0.020) 9 13
Bayes Net [11] 100.0% (0.0) 89.95% (0.024) 9
Multilayer Perceptron [12] 100.0% (0.0) 87.95% (0.017) 9
Traditional Rough Sets [13] 100.0% (0.0) 82.87% (0.033) 9 204.4 (7.6)
Proposed hybrid Model C 100.0% (0.0) 97.41% (0.007) 9 63.4 (6.7)
The denotes no given answer in this field.
As Table 11 showed, the proposed hybrid model C is the best for forecasting quarterly
PGR in terms of Testing accuracy, and it is superior in the number of rules when com-
pared to the Traditional Rough Sets. The results above demonstrate that the proposed
hybrid model C outperforms the listing methods in terms of both accuracy and its stan-
dard deviation. Therefore, it is obvious that the proposed hybrid model C has the best
performance regardless of in accuracy or its standard deviation.
For further verification of the results in Table 11, a one-way ANOVA test is used to
analyze significant dierences in accuracy among these methods above. It shows that
the DT, Bayes Net, Multilayer Perceptron, Traditional Rough Sets, and model C all
significantly diered at a confidence level of 0.95 (see the fourth column of Table 12).
This shows that significant dierences in accuracy existed in these methods. In further
validation of the significance above, a Schees test is used for a post hoc test among these
five methods. For easy presentation, 1, 2, 3, 4, and 5 are representing the DT, Bayes Net,
Multilayer Perceptron, Traditional Rough Sets, and model C, respectively. Consequently,
the results of one-way ANOVA test and Schees test are deployed in Table 12. From
Table 12, the performance of the DT-C4.5 and proposed hybrid model C significantly
outperforms the Bayes Net, Multilayer Perceptron, and Traditional Rough Sets, but the
dierence between DT-C4.5 and model C is not significance. No significance also occurred
among the remaining three methods.
Table 12. The results of One-way ANOVA test and Schees test for five
dierent methods in the PGR dataset
One-way ANOVA test
Statistical tests Schees test
F-Statistic P-value
Accuracy 1, 5 > 2, 3, 4 109.655 0.0001
1
P < 0.05.
Additionally, in order to compare the performance of Traditional Rough Sets with the
models A-C, the improved accuracy rate and the improved rule rate are calculated, and
the results are shown in Table 13.
Table 13. The improvement results of the three proposed hybrid models
for average accuracy rate with standard deviations in the 10 tests for the
PGR dataset
Improved Number Improved
Methods Accuracy
accuracy rate of rules rule rate
Traditional
82.87% (0.033) 204.4 (7.6)
Rough Sets
Model A 96.99% (0.012) 17.04%1 24.5 (3.2) 88.01%2
Model B 94.86% (0.014) 14.47% 3.0 (0.0) 98.51%
Model C 97.41% (0.007) 17.55% 63.4 (6.7) 68.49%
The denotes no given answer in this field.
1
(0.9699-0.8287)/0.8287=17.04%,
2
(204.4-24.5)/204.4=88.01%.
As Table 13 showed for the rough sets community, the proposed hybrid model C is still
the best method to forecast quarterly PGR in Testing accuracy in the PGR dataset;
moreover, the three proposed hybrid models are also superior to the Traditional Rough
Sets in Number of rules. Particularly, the higher improved accuracy rate is 17.55% for
model C, and the higher improved rule rate is 98.51% for model B when comparing the
proposed hybrid models with the Traditional Rough Sets method. It is obvious that
the fewer rules in the proposed hybrid models are generated. In summary, the results
above demonstrate that there is room to improve the Traditional Rough Sets method
with respect to accuracy and number of generated rules. The results of these models are
summarized in Table 14.
Table 14. The comparison results of seven dierent methods in the PGR dataset
Training Testing Number of Number of

Methods
accuracy accuracy attributes rules
Decision Tree-C4.5 100.0% (0.0) 95.09% (0.020) 9 13
Bayes Net 100.0% (0.0) 89.95% (0.024) 9
Multilayer Perceptron 100.0% (0.0) 87.95% (0.017) 9
Traditional Rough Sets 100.0% (0.0) 82.87% (0.033) 9 204.4 (7.6)
Model A without RF 96.8% (0.012) 95.21% (0.012) 9 34.7 (5.2)
Model A with RF 99.0% (0.005) 96.99% (0.012) 9 24.5 (3.2)
Model B without RF 100.0% (0.0) 94.86% (0.014) 4 3.0 (0.0)
Model B with RF 100.0% (0.0) 94.86% (0.014) 4 3.0 (0.0)
Model C without RF 100.0% (0.0) 95.98% (0.007) 9 102.0 (6.9)
Model C with RF 100.0% (0.0) 97.41% (0.007) 9 63.4 (6.7)
3.3. Analysis results and implications. To explore the implied information of hidden
data in the PGR dataset deeply, the study based on rules set in Figure 1, Tables 1, 3,
and 8-14 presents a summary of bullet points. Meanwhile, some empirical results and
management implications involved in the PGR dataset are aggregated in numbered lists
separately in the following, respectively.
1. The ranking order of accuracy and improved rate are based on the following cases:
In internal comparison: From Table 9, the ranking order of accuracy is 97.41%,
96.99%, and 94.86% for models C, A, and B, respectively.
In without/with RF: From Table 10, the ranking order of accuracy is 97.41%,
96.99%, 95.98%, 95.21%, 94.86%, and 94.86% for model C, model A, model C
without RF, model A without RF, model B, and model B without RF, respec-
tively.
In external comparison: From Table 11, the ranking order of accuracy is 97.41%,
95.09%, 89.95%, 87.95%, and 82.87% for model C, DT-C4.5, Bayes Net, and
Multilayer Perceptron, Traditional Rough Sets, respectively.
In the rough sets community: From Table 13, the ranking order of accuracy is
97.41%, 96.99%, 94.86%, and 82.87% for models C, A, B, and Traditional Rough
Sets, respectively.
In improved accuracy rate for RF: From Table 10, the improved accuracy rate
for RF is 1.87%, 1.49%, and 0% for models A, C, and B, respectively.
In improved rule rate for RF: From Table 10, the improved rule rate for RF is
39.27%, 29.39%, and 0% for models C, A, and B, respectively.
In improved accuracy rate for the rough sets community: From Table 13, the
improved accuracy rate for the rough sets community is 17.55%, 17.04%, and
14.47% for models C, A, and B, respectively.
In the improved rule rate for the rough sets community: From Table 13, the
improved rule rate for the rough sets community is 98.51%, 88.01%, and 68.49%
for models B, A, and C, respectively.
2. The best model is dependent on the following criteria:
In terms of accuracy: From Tables 9 and 10, the experimental results demon-
strate that the proposed hybrid model C is the best one in terms of accuracy in
both internal comparison and external comparison.
In terms of standard deviation on accuracy: From Tables 10 and 11, it is clear

that the proposed hybrid model C is the best one in standard deviation of accu-
racy.
In terms of number of attributes: From Tables 9 and 11, it is clear that the
proposed hybrid models B is the best one in number of attributes.
In terms of number of rules: From Tables 10 and 13, it is clear that the proposed
hybrid model B (without/with RF) is the best one in number of rules.
In terms of standard deviation on number of rules: Similarly, from Tables 10 and
13, it is clear that the proposed hybrid model B (without/with RF) is the best
one in standard deviation of number of rules.
3. The worst model is depended on the following situation:
In internal comparison: From Tables 9 and 10, it is obvious that mode B is the
worst one in terms of accuracy regardless of being without/with RF. Thus, for
the model B, there is still space to improve accuracy, when it is compared with
the other proposed hybrid models.
In external comparison: From Table 11, the worst model is Traditional Rough
Sets in terms of accuracy.
4. The ranking order of importance on condition attributes is dependent on:
A statistical table based on Table 8: The study is further based on a rule set
in Table 8 (in total 20 rules) extracted by the LEM2 algorithm to develop a
statistical table shown in Table 15. In further analyses of Table 15, the first two
high frequencies of condition attributes include O-income (54) and T-liabilities
(51), which is indicated in the shading, which means that they are key determi-
nants for forecasting quarterly PGR by rough sets LEM2 algorithm. Next, the
middle frequency of condition attributes includes F-assets (49), T-salary (48),
O-expenses (47), T-assets (44), and A-costs (35). All are important attributes
for forecasting quarterly PGR, too. The lowest frequency of condition attributes
is C-type (9). It refers to the least importance on forecasting quarterly PGR in
the PGR dataset. As a result, the ranking of the frequency of presentation in this
rule set for eight condition attributes is: O-income T-liabilities F-assets
T-salary O-expenses T-assets A-costs C-type. Particularly in O-
income (L 1 and L 3) and T-liabilities (L 1 and L 3), the two attributes are the
key determinants for forecasting PGR. This situation explicates that the greater
the liabilities of companies is, the greater the influence for quarterly PGR is
restated, so are the operating incomes of companies. The interesting finding is
disclosed, and it will be explored further in the next section. Most attributes
and the orderings in the PGR dataset can provide a reference that is regarded
as evaluation criteria for selecting potential investment targets for investors in
the future. Obviously, these critical variables imply that they will be the most
significant determinants for determining investment portfolios of investors.
5. The eect of components for proposed hybrid models is presented, as follows:
For FSM: From model B, the FSM is an eective way in view of lesser-used
attributes and decreasing the dimensions of attributes in the PGR dataset.
For DM: Based on the literature [14,15], which indicated before the data can
be sent into some rough sets models, they must first be discretized. In the per-
formance comparisons of the proposed hybrid models A-C with the Traditional
Rough Sets method in Table 14, the DM is a eective technique in reducing the
number of rules, improving accuracy rate, and generating fewer rules of a set,
even in lowering standard deviation.
Table 15. The statistical table of a rule set extracted by LEM2 algorithm
in the PGR dataset
1 2
Condition attributes
T-liabilities L 1 18
T-liabilities L 2 15 51
T-liabilities L 3 18
T-assets L 1 16
T-assets L 2 13 44
T-assets L 3 15
F-assets L 1 17
F-assets L 2 14 49
F-assets L 3 18
A-costs L 1 16
A-costs L 2 9 35
A-costs L 3 10
O-income L 1 20
O-income L 2 14 54
O-income L 3 20
O-expenses L 1 18
O-expenses L 2 14 47
O-expenses L 3 15
T-salary L 1 18
T-salary L 2 14 48
T-salary L 3 16
C-type 9 9
1
denotes the frequency of presentation in the rule set.
2
denotes the subtotal frequency of presentation in
same condition attributes.
For FST (Similarity functions): According to Table 9, the FST+RST model (i.e.,
fuzzy-rule similarity + rough sets method) is also an eective way to improve the
accuracy of prediction in the PGR dataset, which also has a high performance
in accuracy with its standard deviation and number of rules with its standard
deviation.
For RF: From Table 10, the improved accuracy rate and the improved rule rate
provide that the RF is a requirement for not only improving the accuracy but
also reducing the number of rules in the PGR dataset except for model B.
6. An interesting phenomenon is disclosed from model B, as follows: From Table 10,
it is well worthy of being indicated that the improved accuracy rate and improved
rule rate are zero (0%) in model B without/with RF. This information transfers a
message that RF that is used to reduce the number of rules is an ineective way in
model B. The RF reduces the number of rules by eliminating rules with low support
(< 2) in the study. Concretely, a real fact for this situation is found that the rule
set of model B has the lowest support, 21; therefore, no rules are omitted, which
results in a zero improved accuracy rate and rule rate. Thus, the RF processing has
no eect for model B, and it is unnecessary.
7. Some important discoveries are described, as follows:
The outstandingly proposed hybrid models: In summarizations of Table 14, it is
clear that DT-C4.5, Multilayer Perceptron, Bayes Net, Traditional Rough Sets,
and the three proposed hybrid models all can be regarded as selected tools for
forecasting the quarterly PGR, because the accuracy of these methods is over
80%, which can be accepted in the forecasting outcomes. More importantly, the
proposed hybrid models outperform the listing methods not only in accuracy
but also in the number of generated rules. Furthermore, the LEM2 algorithm
can generate comprehensive rules as a knowledge-based investment system for
investors.
The outstandingly improved accuracy rate and improved rule rate: From Ta-
ble 13, the proposed hybrid models are superior to the Traditional Rough Sets
method in the PGR dataset in both the improved accuracy rate and the im-
proved rule rate, which are over 14% and 68%, respectively. Also, they have
proven that the Hybrid Rough Sets method surpasses Traditional Rough Sets in
stand-alone models in the PGR dataset. This result is consistent with that of
the literature.
Two significant attributes: From Table 15, the rate of presentation frequency for
the two attributes, the L 1 and L 3 of the O-income and T-liabilities attributes,
is 100%, 100%, 90%, and 90% of all 20 rules in a specific rule set, respectively.
It is clear that the two attributes play a significant role in forecasting quarterly
PGR of financial holding stocks in the stock market.
8. Explanatory power of generated rules set: The proposed hybrid models based on
rough sets can generate a comprehensive decision rules set that contains if ... then ...
rules, assisting market investors to direct the use of knowledge to rule the investment
strategies of the stock market easily. It presents a better explanatory power of
generated rules set.
9. A special characteristic and dierence given to the study: To deserve to be men-
tioned, the major dierences for the problems of forecasting profit growth rate (PGR)
processed between this paper and others are that practitioners usually use the re-
lated issue of PGR to forecast the future stock price or to analyze the potential for a
specific firm, but academicians never see it in the field of research, particularly. The
study aims to fill this knowledge gap. Therefore, the paper can be regarded as the
first article that uses fuzzy set theory in addition to rough set theory to process the
PGR problem until now.
10. A further explanation is given on how to use the empirical results in practice: For
practical applications of the empirical results, the theories used in the study can
be presented systematically and their algorithms presented in step-by-step type,
and then they are further applied to knowledge-based systems as the directions to
distinguish the bad trades from the good ones precisely for investors and help them
learn more about making wise investment decisions. The proposed hybrid models
provide an alternative approach for those who intend to use intelligent technology
in the stock market, because it eectively discovers hidden information from data
and improves some subjective views or manual operations of investors for achieving
better performance to unfold the derivative problems of fast-changing stock market
conditions, contingent with capturing the market opportunity for generating benefits.
These research results can be useful and viable for stock investment, knowledge-
based system development, and subsequent research, and both academicians and
practitioners stand to benefit from them.
4. Discussion and Findings. Through these foregoing introductions and explorations,

some valuable discussions are conducted from the PGR dataset, as follows:
1. About Type I and Type II errors: First, considering misclassification (i.e., Type I
and Type II error) costs is necessary. In the real world, the stock market is a place for
money games, and the stock trading is a highly implied risk investment for expected
earnings. Some error costs are caused by erroneous decision-making (investment).
A best-known error for investment is similar to Type I (a false positive) and Type
II (a false negative) errors, which entails that the forecasting in the next trading
day is going up but reality is going down, or the forecasting in the next trading day
is going down but reality is going up, respectively. In other words, in the study, a
type I error occurred when investors rejected case N class (negative returns) but
it was P class (positive returns), and a type II error occurred when investors did
not reject case P class but it was N class. In a sense, a type II error is twice
or more as bad as a type I error for misclassification costs. For simplicity, the
study assumes the basic assumption of equal costs for both N class and P class
misclassifications. However, there is evidence that the error costs may not be equal
for investors when they consider possible loss functions with respect to Type I and
Type II errors; incorporating them into the forecasting systems can lead to better
and more desirable results. Thus, the evaluation criterion of forecasting methods
should be minimizing the total error costs, not to minimize the accuracy rate only
in the future; that is, join to consider the costs of error rather than just roughly
compare the number of errors (i.e., accuracy). Error costs, a problematic issue, may
result in huge losses to investors; therefore, the study could take into account other
types of models that incorporate error costs in subsequent research.
2. About cause-eect relationships of condition and decision attributes: It is an impor-
tant issue to explore the cause-eect relationships between condition and decision
attributes in the PGR dataset. For the PGR dataset, all 8 attributes, T-assets,
F-assets, T-liabilities, C-type, A-costs, O-income, O-expenses, and T-salary, are the
condition attributes for forecasting the quarterly PGR (decision attribute) problem.
However, to develop a complete cause-eect relationship between condition attributes
and the decision attribute is dicult due to the hardness of discovering all the fac-
tors, in which the former aects the latter. Moreover, it is dicult to distinguish
useful attributes, since some of them may be neglected, hidden, or misstated. Also,
the data of collection are very dicult for attributes. Alternatively, some of them
may be qualitative data that cause a problem of measurement. Therefore, it is a
vital and interesting issue to find out all the core attributes that influence quarterly
PGR in the future.
3. About the performance of automatic DM and expert DM applied to the stock market:
A higher performance for a called automatic DM for condition attributes and another
called expert DM, based on the knowledge of domain experts for decision attributes,
had been proven in improving classification accuracy in this study. Conversely, the
expert DM is used to discretize condition attributes and the automatic DM is used to
discretize decision attributes. Definitely, it is worthy to explore which performance
is better with respect to the two cases. Perhaps, it is hard to say, and only taking a
look to find out the result is relatively dicult.
4. About the agency costs in stock market: Agency costs arise mainly from divergence
of control, separation of ownership and control, and the dierent objectives of the
principal-agent. Core et al. [16] oered empirical evidence showing that firms with
weaker governance structures have greater agency problems and are less eective.
Much literature also suggested that the agents might play an important disciplin-
ing role in managerial employment contracting and incentives [17]. Obviously, in
reference to earlier studies, agency costs are an important component of corporate
governance systems. In this case, it implies that the financial data (agency costs)
and non-financial data (governance structures) are inter-active and inter-influential.
However, the study completely uses financial data to forecast quarterly PGRs of
financial holding stocks and do not consider non-financial data. Similarly, whether
to consider joining the non-financial data, such as macro economic, corporate gover-
nance, and ownership structure, into the forecasting systems will be an alternative
to forecast quarterly PGRs.
From the experimental results of the empirical study from the three proposed hybrid
models in the PGR dataset, the study yields the following findings:
1. A case possibly causing financial crisis (or distress) is found from model A: If the
T-liabilities are higher (i.e., L 3) and the O-income is less (i.e., L 1), this case would
yield a negative (poor) PGR. It implies that companies will cause poor economies of
scale and cost-eectiveness. Moreover, if the negative PGR continually occurs year
after year, then it potentially leads to a financial crisis (or distress) finally. This is
useful reference information to investors for selecting potential target companies into
investment portfolios.
2. Two cases for improved selection of investment portfolios for investors are found:
A negative selection: Based on a rule set extracted by the LEM2 algorithm, it
is found that when the T-liabilities L 3 and O-income L 1 occur simultaneously,
then the class is almost poor (PGR < 0%), which reflects the increased prob-
ability that higher total liabilities and fewer operating incomes yield negative
PGRs. More importantly, it implies that when companies are quoted high loans,
the most common feature of them will be going down to steer high financial lever-
age and to collect a great deal of short-term capital. Because their overweighted
debts produce no eect, the core business will incur severe loss in business; there-
fore, it verifies that the PGR is poor or negative. This is a negative implication
for investing in target companies.
A positive selection: Conversely, when T-liabilities L 1 and O-income L 3, then
class is almost fair (PGR > 50%) or good (PGR = 0-50%), which reflects the
increased probability that fewer total liabilities and higher operating incomes
yield positive PGRs. It usually implies that when companies have slight liabilities
and improved incomes, their quarterly PGR is positive and fair. In fact, in
Taiwan, the Cathay Life Insurance Co. is a well-known representative company
for this example, and her yearly EPS (earnings per share) is always positive.
3. An improved space for the performance of Traditional Rough Sets: For the PGR
dataset, a common thing is found that the performance of the Traditional Rough
Sets is the worst; therefore, there is more room to improve its performance.
4. Finding an alternative applied to the field of the stock market for forecasting quar-
terly PGR: Many previous studies revealed that in comparing the eciency of the
classifiers, they almost positively conclude that the classifiers performances are
strongly dependent on the various contexts. Generally, no tools for AI techniques
can be applied to various applications of the real world, because they always exist
in some uncertain drawbacks themselves. Therefore, the study oers a path to find
out more suitable ways that are applied in the field of the stock market, to classify
quarterly PGR problems.
5. Requirements for data preprocessing: From Tables 9-14 for the PGR dataset, it
has been proven that implementing a DM way into the proposed hybrid models for
improving accuracy and lowering its standard deviation is necessary.
6. Implementing an RF is not essential: From the earlier literature, they supported

that the performance of the RF is an eective method to reduce the number of
rules. However, from the empirical results of model B in Table 10, it provides that
implementing a RF to model B is not essential. This is a vital finding to academicians
and practitioners in the study.
7. Superiority and stability on the proposed hybrid models: According to Tables 11-
14 for the PGR dataset, the proposed hybrid models are superior to the listing
methods in terms of both higher accuracy and fewer attributes. Thus, the results of
experiments oer insight toward the manner in which the proposed hybrid models
show superiority in the study. Similarly, it supports that the proposed hybrid models
are relatively stable in accuracy rate, because they have the best performance of
standard deviation, too.
8. An eective way for the DT+RST and FST+RST hybrid models in the PGR dataset:
From Table 14, the accuracy of the models C and A is 97.41% and 96.99%, and they
are the first two, respectively. Obviously, the hybrid models surpass the stand-alone
models in the PGR dataset. This is matches that of the literature.
9. Validating the importance of hybrid intelligent systems: From Table 13, when the
proposed hybrid models are compared with the Traditional Rough Sets method, the
improved accuracy rate and the improved rule rate in the PGR dataset are 14.47%,
17.04%, and 17.55%, and 98.51%, 88.01%, and 68.49% for models A-C in the PGR
dataset, respectively. It is evident that hybrid intelligent systems are superior to
the techniques in stand-alone models for the PGR dataset, also. Therefore, with the
limitations of intelligent techniques above in stand-alone models, it is worthwhile to
propose a newer hybrid intelligent system in order to amplify the advantages of the
individual models and minimize their limitations. The study oers a validating way
on the importance of hybrid intelligent systems. Similarly, the result is consistent
with that of the literature, too.
10. Dierences with the listing methods: By way of the results of the one-way ANOVA
test in the PGR dataset, they all support an evidence of dierences that are signifi-
cant in accuracy rate at confidence level 0.95 among these listing methods in Table
11. This implies that these dierences are conclusive proof that the proposed hybrid
model C outperforms the listing methods in the PGR dataset, except no significance
is found between proposed hybrid model C and the DT method.
5. Conclusions. The study has proposed three hybrid models (A, B, and C) based on
rough sets for solving the classification problems of quarterly PGRs faced by investors in
the financial industry of the stock market. The proposed hybrid models are constituted
dierently by six basic components: experiential knowledge, feature selection methods,
discretization methods, rule filter, fuzzy-rule similarity, and rough sets. A practiced PGR
dataset collected from financial holding stocks in Taiwans stock market was used in the
empirical case study to verify the proposed hybrid models. The experimental results
show that the proposed hybrid models have higher performance than traditional rough
sets and the listed AI techniques in terms of some criteria, such as accuracy, number
of rules, standard deviation, and number of used attributes. Consequently, it has been
proven that the proposed hybrid models produce higher accuracy, lower standard deviation
(higher stability) on accuracy and generated rules, fewer decision rules (smaller rules set),
and fewer attributes; concurrently, they directly support the if . . . then rules set as
knowledge-based systems to rule the strategy of the investment and intelligently oer
an explanation power for investors. Moreover, the findings based on the PGR dataset,
having important implications, provide a useful insight, in that the key characteristics of
forecasting quarterly PGRs are critical with respect to responsiveness to rapidly changing
market conditions in the financial industry for investors.
To stock market investors, the proposed hybrid models are useful tools, because invest-
ment behaviors consist of making the same buy-sell activities repeatedly, while selecting
investment targets occurs across dierent economic and market environments each time.
The ill selection of a specific company that has negative returns of stock products has led
to a huge loss; therefore, the extent of making more benefits requires this intellectual tech-
nique. In general, the proposed hybrid models surpass the listing methods based on the
following six reasons: (1) The feature selection method firstly discovers core attributes by
Chi-square, InfoGain, Gain Ratio, Consistency, and CFs and deletes redundant attributes
that may decrease accuracy of prediction; (2) Feature selection and the rough sets LEM2
algorithm are used to improve the rough sets classifier; (3) The MEPA method discretizes
attributes based on outcome class using entropy equations to determine thresholds before
the intervals are subdivided by minimum entropy values. This method has demonstrated
the good performance of rule-based classification; (4) The MEPA-based and fuzzy-rule
similarity method are mixed to further improve the performance of rough sets classifiers
and to generate a knowledge-based classification system; (5) The rough sets LEM2 algo-
rithm is superior than traditional methods because it deduces rules set directly from data
with symbolic and numerical attributes; and (6) RF based on support can eliminate rules
with low support and then refine the extracted rules by reducing the number of rules, but
without (or rare) compromising accuracy.
A weakness of the results is that the data of quarterly PGRs used as training were
collected in 2004-2006; therefore, it is only suitable to forecast the stock price of short-
mid runs, and it can not used in the long run. If the proposed hybrid models would be
used to forecast a stock price for a long period in the near future, at least 3 years (in
practice, the period of over 3 years can be regarded as a long run standard in Taiwan)
worth of data are needed. Furthermore, a limitation for the fuzzy logic used in the study
is it relies on an expert to determine a membership function subjectively, but this case
for the opinion of experts is subjective to the individual thoughts, so it can not solve
the problems of the real world objectively. Hence, the MEPA discretization method
incorporated into FST+RST oers an eective and objective technique to construct the
membership function in improving a performance of the fuzzy-based rough sets classifier,
so that the FST+RST method is suitable for working the PGR problems for the stock
investors. Although this fuzzy-based rough sets classifier also presents promising results
in the study, it is still necessary to further conduct more experiments and construct more
models or other methods to improve the study. Generally, six future works to further
improve them can be further explored separately, as follows: (1) More times of data: To
forecast the stock price in the long term, a lot more time periods of data (more than 3
years) will be required; (2) More models: To see dierent types of hybrid models on how
to influence the accuracy rate of classifying quarterly PGRs is a valuable issue; (3) More
datasets: Use other datasets of the same financial markets than stock markets, like the
futures market, to verify the proposed hybrid models; (4) To explore more variables used
as condition attributes related to influence quarterly PGR; (5) Other datasets of dierent
industry sectors, such as the manufacturing industry, services industry, and healthcare
industry, can be considered in further assessing the proposed hybrid models; and (6)
Compare the forecasting performance with other AI models, such as genetic algorithm
(GA) or fuzzy time-series.
Acknowledgment. The authors would like to thank the Editor-in-Chief, associate edi-
tor, and anonymous reviewers for their very useful comments and suggestions, which are
very helpful in improving this study.
REFERENCES
[1] Z. Pawlak, Rough sets, Informational Journal of Computer and Information Sciences, vol.11, no.5,
pp.341-356, 1982.
[2] M. Inuiguchi and M. Tsurumi, Measures based on upper approximations of rough sets for analysis of
attribute importance and interaction, International Journal of Innovative Computing, Information
[3] J. W. Grzymala-Busse, A new version of the rule induction system LERS, Fundamenta Informaticae,
vol.31, no.1, pp.27-39, 1997.
[4] H. Sakai and M. Nakata, On rough sets based rule generation from tables, International Journal of
Innovative Computing, Information and Control, vol.2, no.1, pp.13-31, 2006.
[5] L. A. Zadeh, Fuzzy sets, Information and Control, vol.8, pp.338-353, 1965.
[6] S. Tan, H. Xu and J. Gu, Ecient algorithms for attributes reduction problem, International Journal
[7] M. A. Hall and G. Holmes, Benchmarking feature selection techniques for discrete class data mining,
IEEE Trans. on Knowl. Data Eng., vol.15, no.3, pp.1-16, 2003.
[8] R. Yager and D. Filev, Template-based fuzzy system modeling, Intelligent and Fuzzy System, vol.2,
pp.39-54, 1994.
[9] J. R. Quinlan, Induction of decision trees, Machine Learning, vol.1, no.1, pp.81-106, 1986.
[10] J. R. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann, San Mateo, 1993.
[11] K. P. Murphy, Bayes Net ToolBox, Technical report, MIT Artificial Intelligence Labortary, 2002,
http://www.ai.mit.edu/murphyk/.
[12] R. P. Lippmann, An introduction to computing with neural nets, IEEE ASSP Mag., vol.4, no.2,
pp.4-22, 1987.
[13] J. Bazan and M. Szczuka, RSES and RSESlib A collection of Tools for rough set, Lecture Notes
in Computer Science, Springer-Verlag, Berlin, pp.106-113, 2001.
[14] F. E. H. Tay and L. Shen, Economic and financial prediction using rough sets model, European
Journal of Operational Research, vol.141, pp.641-659, 2002.
[15] Y. S. Chen, J. F. Chang and C. H. Cheng, Forecasting IPO returns using feature selection and
entropy-based rough sets, International Journal of Innovative Computing, Information and Control,
vol.4, no.8, pp.1861-1875, 2008.
[16] J. E. Core, R. W. Holthausen and D. F. Larcker, Corporate governance, chief executive ocer
compensation, and firm performance, Journal of Financial Economics, vol.51, pp.371-406, 1999.
[17] F. Machlup, Theories of the firm: Marginalist, behavioral, managerial, American Economic Review,
vol.57, pp.1-33, 1967.
c ISSN 1349-4198
RE-EVALUATION OF ADAPTIVE X CONTROL CHARTS: A

COST-EFFECTIVENESS PERSPECTIVE
Yan-Kwang Chen1, , Hung-Chang Liao2 and Hsu-Hwa Chang3

1
Department of Logistics Engineering and Management
National Taichung Institute of Technology
No. 129, Sanmin Road, Sec. 3, Taichung, Taiwan

Corresponding author: ykchen@ntit.edu.tw
2
Department of Health Services Administration
Chung-Shan Medical University
No. 110, Sec. 1, Jianguo N. Road, Taichung 40201, Taiwan
hcliao@csmu.edu.tw
3
Department of Business Administration
National Taipei College of Business
No. 321, Chi-Nan Road, Sec. 1, Taipei, Taiwan
hhchang@webmail.ntcb.edu.tw
Received October 2009; revised May 2010
Abstract. Studies have considerably improved the statistical performance of the X con-
trol charts by varying the design parameters as a function of the data from the process.
These design parameters include the sample size, length of sampling interval and con-
trol/warning limits of a control chart. Studies compare adaptive charts with standard
X charts in terms of cost-eectiveness, where the former outperforms the latter. Tradi-
tionally, researchers analyze cost-eectiveness for certain input process/cost parameters,
and the sensitivity to the input parameters examines each factor in turn. This analy-
sis is requires little eort, however, its results depend on the initial input parameters.
Therefore, to achieve an objective analysis, this paper develops a statistical procedure
combining experimental design, regression analysis and covariance analysis to facilitate
cost-eectiveness analysis better. This study presents a numerical example to demon-
strate the use of this new approach in the cost-eectiveness analysis of adaptive charts.
Keywords: Adaptive control chart, Economic design, Cost-eectiveness analysis, Ge-
netic algorithms
1. Introduction. Industries use control charts to observe whether a process is in con-

trol. When there is one quality characteristic to control, process disturbances are typically
monitored with Shewhart control charts [1]. The advantage of Shewhart charts is their
simplicity, however, they may take a long time to detect small disturbances. Thus, re-
searchers have proposed alternatives, such as the Shewhart control chart with supplemen-
tary run rules [2], EWMA charts [3] and CUSUM charts [4,5] to improve the performance
of Shewhart charts.
Recently, adaptive sampling schemes that allow one, two or all of the three design
parameters (i.e., sample size, sampling interval, coecients of control limits) to be variable
during production are applied to the standard Shewhart (SS) chart to improve the average
time to detect small mean shifts. This chart with adaptive sampling schemes is referred to
as an adaptive chart. Several studies have examined adaptive sampling schemes, including
the variable sampling interval (VSI) [6-8], variable sample size (VSS) [9,10], variable
sample size and sampling interval (VSSI) [11-13] and variable parameter (VP) [14,15].
1229
1230 Y.-K. CHEN, H.-C. LIAO AND H.-H. CHANG
Prior to using the control chart, the user must select several design parameters. For an
SS chart, these design parameters are sample size, control limits and sampling interval.
Duncan [16] suggested how to choose these parameters based on the economic criterion,
and called the resulting process economic design. Since Duncans study was introduced,
many researchers have followed up on his work (See, for reviews [17,18]).
The typical approach to the economic design of control chart is to develop a cost
function for the manufacturing process, estimate the functions process/cost parameters,
and find the optimal design parameters that minimize the cost function. Duncan [16]
proposed a cost function in which the process remains in operation during the search for
an assignable cause. Chiu [19] proposed another cost function that allows the process to
shut down during the search for an assignable cause.
For the economic design of adaptive control charts, additional parameters are available:
multiple sampling interval lengths, sample sizes and warning/control limits. Bai and Lee
[20] followed Chius model to study the economic design of VSI charts. Park and Reynolds
[21] considered the economic design of VSSI control charts. The model they adopted is
similar to Duncans model. Numerical results show substantial cost savings when sample
size and interval among samplings are variable. The actual amount of cost savings varies
depending on the process/cost parameters. De Magalhaes et al. [22] and Costa and Rahim
[23] assumed that the length of time the process remains in control follows an exponential
distribution, and thus, the Markov chain approach can be applied to the economic designs
of VP control charts. For a given set of process and cost parameters, they compared the
expected cost of using an SS chart. The results indicated that a scheme for adapting these
design parameters could lead to substantial cost savings.
The aforementioned studies performed a cost-eectiveness analysis for certain input
process/cost parameters; the sensitivity to the input parameters was examined for each
factor in turn. This analysis requires little eort, however, the results depend on the
initial input parameters and will lack objectivity.
The intent of this paper is to examine the cost-eectiveness of adaptive control charts,
however, it diers from previous research in several ways. First, to prepare for the cost-
eectiveness analysis, this study economically designed VP, VSSI, VSI, VSS and SS charts
and compared them in terms of the minimum costs they achieved. Second, instead of
comparing for certain input process/cost parameters, this study compared the control
charts under a wide spectrum of input parameters in the cost-eectiveness analysis to
avoid overlooking some cases.
This paper is organized as follows. Section 2 briefly reviews adaptive control charts.
Section 3 introduces the economic design of the adaptive chart. Here, the cost function
for the economic design follows the model of Costa and Rahim [23], in which only one
assignable cause of variation exists and the time among occurrences is exponentially dis-
tributed. Section 4 proposes a statistical procedure consisting of experimental design,
regression analysis and covariance analysis to examine the cost-eectiveness of adaptive
charts. Section 5 presents a numerical example to illustrate cost-eectiveness analysis.
Finally, concluding remarks are provided in the last section.
2. Adaptive X Charts. Consider a process in which the distribution of observations

of the process is normal with a mean and known variance 2 . The target value of the
process mean is represented by 0 . The process is in control if = 0 and out of control
if = 0 , > 0.
When an SS control chart is employed to monitor the process, a random sample of
fixed size n0 is / taken from the process every h0 hours, and the standard sample means
Zi = (X i 0 ) X (i = 1, 2, 3, ) are plotted on a control chart with fixed control
CONTROL CHARTS
RE-EVALUATION OF ADAPTIVE X 1231
limits k0 until Zi falls outside the range (k0 , k0 ). At this time, the process is stopped
and an assignable cause is sought. When = 0 , this is called a false alarm, and the
process continues. However, if = 0 , this is called a true alarm, and the assignable
cause must be eliminated before the process continues. Here, X represents the standard
deviation of the sample mean.
The VP control chart is a modification of the SS control chart. In the VP control
chart, random samples of dierent sizes are taken at intervals of variable length. Assume
that the sample size and sampling interval length employed in the VP control chart
vary between two pairs: the pair consisting of the minimum sample size and the longest
sampling interval is denoted by (n1 , h1 ), whereas that consisting of maximum sample size
and the shortest sampling interval is denoted by (n2 , h2 ). The pairs are chosen so that
n1 < n0 < n2 and h2 < h0 < h1 . The rule for switching between the pairs depends on the
prior sample point on the control chart. That is, the position of the prior sample point
on the chart determines the size of the current sample and the time of its sampling. If
the prior sample point (i 1) falls in the warning region, (n2 , h2 ) should be used for the
current sample point (i). On the other hand, if the prior sample point (i 1) falls in
the central region, (n1 , h1 ) should be used for the current sample point (i). The warning
region and central region are given by [kj , wj ) (wj , kj ] and [wj , wj ], respectively,
where wj and kj ( wj ) represent the warning limit and control limit, respectively. The
index j = 1 if the sample point comes from the small sample, and j = 2 if it comes from
the large sample. Here, k1 > k0 > k2 .
Because no sample is taken at the start of the process, the first sample size and sampling
interval cannot be determined by the switching rule. Prabhu, et al. [11] and Costa [12]
considered choosing the first sample size and sampling interval that at random. The pair
has a probability of p0 of being (n1 , h1 ) and a probability of (1 p0 ) of being (n2 , h2 ),
where p0 can be simply defined as the conditional probability of a sample mean falling into
the central region, given that it did not fall outside the two control limits. Alternatively,
it may be preferable in practice to use a tightened control (the maximum sample size,
shortest sampling interval, and narrowest control limits), because it provides additional
protection against problems that arise during startup [20].
The VSSI chart is one of the simpler charts and uses two adaptive parameters. In the
VSSI chart, the control and warning limits are both fixed (i.e., k1 = k2 ; w1 = w2 ), but the
sample size and sampling interval length still vary between the pairs (n1 , h1 ) and (n2 , h2 ),
depending on the prior sample point on the control chart.
The VSI and VSS charts are called the adaptive charts with one parameter varying.
The sample size of the former varies between n1 and n2 , and the sampling interval length
of the latter varies between h1 and h2 .
3. Economic Design of Adaptive Charts. The process is assumed to start with an

in-control state ( = 0 ), but after a random time of in-control operation it will be
disturbed by a single assignable cause that shifts the process mean ( = 0 ). After
the shift, the process remains out of control until the assignable cause is eliminated (if
possible). The inter-arrival time of assignable causes that disturb the process is assumed
to follow an exponential distribution with a mean of 1/ hours.
The economic design of adaptive X control charts consists of specifying a cost function,
and searching for the optimal design parameters that minimize the cost function over a
production cycle.
3.1. Cost function. Figure 1 depicts the production cycle, which is divided into four
time intervals: the in-control period, out-of-control period, search period due to a false
alarm, and time period for identifying and correcting an assignable cause. These intervals
are now described before they are grouped together.
Figure 1. Production cycle considered in the cost model
T1. The expected length of the in-control period is 1/.

T2. The expected length of the out-of-control period represents the average time needed
for the control chart to produce a signal after the process mean shift. This average
time is called the adjusted average time to signal (AATS ) and is the most widely used
statistical measure for comparing the eciencies of dierent adaptive control charts. The
memoryless property of the exponential distribution allows the computation of the AATS
using the Markov chain approach. The fundamental concepts used below appeared in
Cnlar [24].
Let M be the average time from the start of the cycle to the time the chart first signals
after the process shift. Then,
AAT S = M 1/. (1)
At each sampling moment during the period M , one of the four transient states presented
in Table 1 is reached according to the status of the process (in or out of control) and
i (central or warning region). If X
the position of X i falls into the action region at some
sampling time while the process status is out-of-control, a true alarm is signaled and the
process enters an absorbing state, State 5.
The transition probability matrix is given by

p11 p12 p13 p14 0
p21 p22 p23 p24 0

TP =
0 0 p33 p34 p35 .
(2)
0 0 p43 p44 p45
0 0 0 0 1
CONTROL CHARTS
The entries of the matrix in (2) are

p11 = Pr|Z| < w1 | |Z| < k1 eh1 , p12 = Prw1 < |Z| < k1 | |Z| < k1 eh2 ,
( )
p13 = Pr|Z| < w1 | |Z| < k1 1 eh1 , p14 = 1 p11 p12 p13 ,
p21 = Pr|Z| < w2 | |Z| < k2 eh1 , p22 = Prw2 < |Z| < k2 | |Z| < k2 eh2 ,
( )
p23 = Pr|Z| < w2 | |Z| < k2 1 eh1 , p24 = 1 p21 p22 p23 ,

p33 = Pr|Y | < w1 | Y N ( n1 , 1), p34 = Prw1 < |Y | < k1 | Y N ( n1 , 1),

p35 = 1 p33 p34 , p43 = Pr|Y | < w2 | Y N ( n2 , 1),

p44 = Prw2 < |Y | < k2 | Y N ( n2 , 1), p45 = 1 p43 p44 , where Z N (0, 1).
Table 1. States of the Markov chain

i Process status at (i + 1)th sampling moment
State Position of X
1 Central region In-control
2 Warning region In-control
3 Central region Out-of-control
4 Warning region Out-of-control
5 Absorbing
Using the elementary properties of Markov chains, the average number of transitions
to each transient state before true alarm signals is given by r (I Q)1 , where r =
(p11 , p12 , p13 , p14 ) is the vector of starting probability; I is the identity matrix of order
4; and Q is the transition matrix, in which elements associated with the absorbing state
have been deleted. The product of the average number of visits to the transient state and
the corresponding sampling interval determines the period M . That is,
M = r (I Q)1 t, (3)
where t = (h1 , h2 , h1 , h2 ) is the vector of the next sampling intervals corresponding to the
four transient states.
T3. Let t0 be the average amount of time wasted searching for an assignable cause when
the process is in-control, and E(F A) be the expected number of false alarms per cycle,
E(F A) = r (I Q)1 f, (4)
where f = (1 , 2 , 0, 0) gives the probability that a false alarm is produced in each
transient state. Thus, the searching period due to false alarms is given by t0 E(F A).
T4. The average time spent identifying and correcting the assignable cause is a constant,
t1 .
The expected length of a product cycle can be represented as
ET = M + t0 E(F A) + t1 . (5)
Let V0 be the hourly profit earned when the process is being operated in the in-control
state, V1 be the hourly profit earned when the process is being operated in the out-of-
control state, C0 be the average search cost if the given signal is false, C1 be the average
cost to discover the assignable cause and adjust the process to the in-control state, and s
be the cost for each inspected measurement. The expected net profit from a production
cycle is given by
EC = V0 (1/) + V1 (M 1/) C0 E(F A) C1 sE(N ) (6)
where E(N ) is the average number of observations that result in a signal during the
production cycle. The cycle is given as
E(N ) = r (I Q)1 (7)
where = (n1 , n2 , n1 , n2 ) is the vector of the sample sizes corresponding to the four
transient states taken for the next sampling.
Finally, the expected loss per hour F is given as
F = V0 EC/ET. (8)
3.2. Solution procedure. The cost function F given by (8) is a function of the pro-
cess parameters (t0 , t1 , , ), cost parameters (s, C0 , C1 , V0 , V1 ) and design parameters
(n1 , n2 , h1 , h2 , w1 , w2 , k1 , k2 ). To design a control chart economically is to derive the design
parameters that minimizes F , given the process and cost parameters. Genetic algorithms
(GAs) can be used for this because the minimization problem has a nonlinear objective
function with the following constraints: n1 and n2 Z+ (positive integers), 0 n1 n2 ,
h1 h2 > 0 (note that the lower bound of h2 must be greater than zero to allow enough
time between the samples to generate the required sample size), 0 w1 k1 , 0 w2 k2
and k1 k2 0. These constraints yield a mix of continuous and discrete decision vari-
ables and a discontinuous, non-convex solution space. Typical nonlinear programming
techniques are insucient and time-consuming at solving this optimization problem.
GAs are global search and optimization techniques modeled on the process of natural se-
lection in biological systems [25]. They have been used successfully to optimize the param-
eters of quality control charts (e.g., [26-29]). The primary distinguishing features of GAs
include an encoding, fitness function, selection mechanism, crossover mechanism, muta-
tion mechanism and culling mechanism. The algorithm for GAs can be formulated accord-
ing to the following steps: the operations of GAs operate as follows: 1) randomly generate
an initial solution population where each candidate solution (n1 , n2 , h1 , h2 , w1 , w2 , k1 , k2 )
in the population is represented as a string of bits; 2) assign each bit string a value ac-
cording to a fitness function (i.e., F ) and select strings from a random sample of the old
population on the basis of their fitness; 3) recombine these strings using crossover and
mutation operators; 4) produce a new generation of strings that are a better fit than the
previous one. The termination condition is satisfied when the number of generations is
large enough or a desired fitness value is obtained.
The quality of solution generated by GAs might depend on the setting for their control
parameters: population size, crossover probability, and mutation rate. Past literature
discusses the optimal setting for GAs control parameters (see for example [28,30-32])
Here, GAs were formulated using a commercial software package, EVOLVER (http://
www.palisade.com.au/evolver), which can function as an add-on to Microsoft Excel.
4. Proposed Procedure. As demonstrated above, process/cost parameters aect the

minimum loss associated with a control chart. Therefore, this study should compare the
minimum loss for dierent control charts using diverse input process/cost parameters. The
proposed procedure for this consists of three phases, which are briefly explained below:
(1) select a wide spectrum of input process/cost parameters for comparison, (2) extract
the critical parameters that aect the minimum loss of control charts and (3) remove the
eects of these parameters on minimum loss for equitable comparison.
Phase I. Select a wide spectrum of input process/cost parameters for compar-
ison.
This phase is realized by an experimental design, in which the process/cost parameters
are assigned to the column of an orthogonal array. Assume the level for each process/cost
CONTROL CHARTS
parameter is that shown in Table 2. This plan has a total of eight parameters with two
levels and one with four levels in this plan. Therefore, a mixed 2 4 levels experimental
design is used to assign the nine parameters to the columns of a modified L16 orthog-
onal array, as shown in Table 3. The 16 assays in this array provide a wide variety of
combinations of the input process/cost parameters.
Phase II. Extract the critical process/cost parameters that aect the minimum
loss of the control chart.
To extract the critical process/cost parameters that influence the minimum loss, we
build a linear regression model for each control chart. In these models, the process and
cost parameters are regarded as possible independent variables and the minimum loss is
considered the dependent variable. To establish the relationship between the independent
variables and the dependent variable, the minimum loss associated with the input pro-
cess/cost parameters selected in Phase I is used, along with their level combination, for
a linear regression analysis. Those independent variables whose coecient diers signifi-
cantly from zero are regarded as critical process/cost parameters aecting the minimum
loss and are retained for the analysis in Phase III.
Table 2. Levels for each process/cost parameter
Independent variables Level 1 Level 2 Level 3 Level 4

s 5 10
C0 250 500
C1 50 500
V0 250 500
V1 50 100
t0 2.5 5
t1 1 10
0.01 0.05
0.5 1.0 1.5 2.0
Table 3. Experimental layout of the modified L16 array
Assay s C0 C1 V0 V1 t0 t1
1 1 1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2 2 2
3 1 2 2 1 1 2 2 2 1
4 1 2 2 2 2 1 1 1 2
5 2 1 1 1 2 1 1 2 3
6 2 1 1 2 1 2 2 1 4
7 2 2 2 1 2 2 2 1 3
8 2 2 2 2 1 1 1 2 4
9 2 1 2 2 2 1 2 1 1
10 2 1 2 1 1 2 1 2 2
11 2 2 1 2 2 2 1 2 1
12 2 2 1 1 1 1 2 1 2
13 1 1 2 2 1 1 2 2 3
14 1 1 2 1 2 2 1 1 4
15 1 2 1 2 1 2 1 1 3
16 1 2 1 1 2 1 2 2 4
Phase III. Remove the eects of critical process/cost parameters on minimum

loss for equitable comparison.
After the critical process/cost parameters are found in Phase II, the minimum loss
for dierent control charts is compared by means of a factorial experiment design using
analysis of covariance (ANCOVA) [31].
In the ANCOVA model, the minimum loss is regarded as the dependent variable, and
the type of control chart and the critical process/cost parameters are regarded as the
treatment and covariates, respectively. ANCOVA involves adjusting the observed mini-
mum loss for the eects of covariates (i.e., the critical process/cost parameters). If this
adjustment is not made, the covariates could inflate the mean square error and make it
harder to detect true dierence or lack of dierences in the minimum loss due to the
treatment (type of control charts). Thus, the ANCOVA is a method of adjusting for the
eects of covariates.
The statistical analysis used in ANCOVA includes the overall test, equal slope test,
equal intercept test/zero slope test, and pair-wise test (see Figure 2). The overall test
is used to examine the null hypothesis: the minimum loss is not aected by choosing a
dierent control chart or critical process/cost parameters, or by their interaction.
Figure 2. Flowchart for ANCOVA analysis

CONTROL CHARTS
When the hypothesis is rejected (i.e., the testing data represent an alternative hypoth-
esis, H1 ), the equal slope test can be further used to determine whether the combination
of a dierent control chart and dierent critical process/cost parameters significantly af-
fected the minimum loss. If the eect of the interaction is insignificant, the equal intercept
and zero slope tests can be used to determine whether the main eects of choosing a dif-
ferent control chart and dierent critical process/cost parameters significantly aected
the minimum loss. If not, an extra pair-wise test is required to determine which critical
process/cost parameter aects the minimum loss depending on the choice of control chart
for further investigations.
5. Numerical Results. The cost-eectiveness of a control chart is typically evaluated

by cost comparison with other control charts given some input parameters (or a level
combination of parameters). In contrast, the proposed method selects a wide range of
input parameters using the experimental orthogonal array for comparison (see Tables 2
and 3). To extract the critical process/cost parameters that influence the minimum loss,
a linear regression model relating the process/cost parameters and the minimum loss was
created for each control chart. The data on the process/cost parameters are the specific
level combinations resulting from the experiment designed. Hence, they are the same for
each regression model. The data on minimum loss were obtained from the economic design
of the corresponding control chart using the given data for the process/cost parameters.
Therefore, data are available for regression analysis (see Table 4).
Table 4. Minimum losses of the 16 assays achieved by the VP, VSSI, VSI,
VSS and SS charts
Assay FV P FV SSI FV SS FV SI
FSS
1 32.63 32.69 34.71 32.91 34.87
2 195.65 196.42 201.11 196.58 201.31
3 134.11 133.07 134.05 133.10 134.11
4 33.25 32.44 38.09 33.71 38.32
5 41.60 41.22 44.91 41.81 45.01
6 62.06 61.41 66.46 62.37 66.33
7 38.75 38.38 40.92 39.03 41.08
8 86.46 83.87 92.86 86.01 93.05
9 100.76 100.92 103.57 101.14 103.77
10 79.03 79.08 83.00 79.52 83.18
11 155.80 150.21 155.54 150.34 155.80
12 42.47 42.53 45.19 42.73 45.34
13 203.39 202.71 207.55 204.11 207.62
14 14.86 14.44 16.84 15.01 16.83
15 22.62 21.75 27.63 23.15 27.79
16 92.93 92.49 94.17 92.95 94.20
Table 5 illustrates the output of linear regression analysis models with a stepwise se-
lection method. The regression analysis was run using the statistical software SPSS 10.0.
The output indicates that significant regression models are properly selected with respect
to the VP (R2 = 0.895; F = 23.324; p-value = 0.000), VSSI (R2 = 0.893; F = 23.033;
p-value = 0.000), VSI (R2 = 0.894; F = 23.174; p-value = 0.000), VSS (R2 = 0.897;
F = 23.977; p-value = 0.000) and SS (R2 = 0.897; F = 24.019; p-value = 0.000) control
charts. The output also shows that the values for the parameters: V0 , t1 , and sig-
nificantly aects the minimum loss regardless of which chart was used. Thus, these four
parameters are the critical factors aecting the minimum loss of the five control charts.
The estimated coecients show that the minimum loss decreases when either V0 , t1 or
decrease, or when increases.
Once the critical parameters are identified, the minimum loss for dierent control charts
can be equitably compared by means of a factorial experiment design using ANCOVA.
Because three of the critical parameters have two levels and one has four levels, a 23 41
factorial experiment design is used to assign the values for the four parameters. That is,
this design has 32 (= 23 41 ) assays. If each assay consists of three replications, a set
of 480 (= 23 41 3 5) data points is forwarded to the ANCOVA model. Note that
Gas are heuristic; thus, the minimum loss obtained from GAs is slightly dierent for each
assay.
Table 6 shows the output of ANCOVA. Table (6-1) is the output of the equal slope
test, in which the model examines the main eects of the dierent control charts used, the
critical parameters (V0 , t1 , and ), and their interaction eect. The indices of R2 = 0.835,
F = 95.992 and p-value = 0.000 in Table (6-1) indicate that the null hypothesis of the
overall test is rejected. The output indicates that the interaction eect terms are not
significant at a level of 0.05. Accordingly, we can adjust the dierence in minimum loss
for dierent control charts by removing the eect of the critical process/cost parameters.
Table (6-2) is the output of the equal intercept and zero slope tests, in which the model
examines only the main eects of the dierent control charts used and the four critical
parameters (R2 = 0.835; F = 297.645; p-value = 0.000). It shows that the dierence in
minimum loss for dierent control charts is insignificant (F = 0.726; p-value = 0.575),
whereas all critical parameters significantly aect the minimum loss.
Table 5. Linear regression analysis output
(5-1) Linear regression analysis for the VP chart

Model F = V0 t1
R2 = 0.895 Model F = 23.324 Pr > F = 0.000
Parameter Estimate Std. error t Pr > |t|

Intercept 45.344 24.762 1.831 0.094
V0 2004.904 285.972 7.011 0.000
t1 5.609 1.271 4.413 0.001
0.192 0.046 4.192 0.002
27.248 10.231 2.663 0.022
(5-2) Linear regression analysis for the VSSI chart

Model F = V0 t1
R2 = 0.893 Model F = 23.033 Pr > F = 0.000

Intercept 44.939 24.772 1.841 0.097
V0 1982.885 286.093 6.931 0.000
t1 5.725 1.272 4.503 0.001
0.188 0.046 4.105 0.002
27.021 10.236 2.640 0.023
CONTROL CHARTS
(5-3) Linear regression analysis for the VSI chart

Model F = V0 t1
R2 = 0.894 Model F = 23.174 Pr > F = 0.000

Intercept 45.726 24.674 1.853 0.091
V0 1982.648 284.955 6.958 0.000
t1 5.689 1.266 4.492 0.001
0.190 0.046 4.172 0.002
26.383 10.195 2.588 0.025
(5-4) Linear regression analysis for the VSS chart

Model F = V0 t1
R2 = 0.897 Model F = 23.977 Pr > F = 0.000

Intercept 46.204 24.433 1.891 0.085
V0 1999.310 282.172 7.085 0.000
t1 5.548 1.254 4.424 0.001
0.200 0.045 4.419 0.001
25.952 10.095 2.571 0.026
(5-5) Linear regression analysis for the SS chart

Model F = V0 t1
R2 = 0.897 Model F = 24.019 Pr > F = 0.000

Intercept 45.999 24.420 1.884 0.086
V0 1999.836 282.021 7.091 0.000
t1 0.200 0.045 4.426 0.001
5.541 1.253 4.420 0.001
26.053 10.090 2.582 0.025
6. Concluding Remarks. This article discussed the cost-eectiveness of four adaptive

charts (VP, VSSI, VSI and VSS charts) as well as the SS chart. Based on the model of
Costa and Rahim [23], these five control charts were economically designed to evaluate
and compare the minimum loss associated with each. Instead of comparing the eect of
certain input parameters, this study used a statistical procedure to re-evaluate the cost-
eectiveness of the five charts under a wide spectrum of input parameters. The results
indicate the following.
(1) The hourly profit earned in the in-control state (V0 ), average time required to detect
and eliminate an assignable cause (t1 ), rate of occurrence of an assignable cause (),
and magnitude of process shift () are the critical process/cost parameters aecting
the minimum loss of the control charts.
(2) The interaction eect of the type of control chart and these four critical parameters
may have a trivial eect on the minimum loss.
(3) The cost savings of using an adaptive chart instead of an SS chart may be trivial.
This result is inconsistent with previous literature in which the cost savings was
explored only for certain input parameters.
(4) There is no significant dierence in minimum loss between the VP, VSSI, VSI and
VSS charts.
These findings were based on the economical model used in Costa and Rahim [23]. This
model assumes the process is disturbed by a single assignable cause and the time between
occurrences is exponentially distributed. Similar to Chius model, this model allows the
process to shut down during the search for an assignable cause. Therefore, it would be
interesting and valuable to conduct further research on other economical models where
some of these imposed conditions failed to hold might lead to dierent conclusions.
Table 6. ANCOVA analysis output
(6-1) Equal slope analysis output

Model F = CC V0 t1 CC V0 CC t1 CC CC
R2 = 0.835 Model F = 95.992 Pr > F = 0.000
Parameter Sum of Squares Degree of Freedom Mean Square F Pr > |F|

Intercept 61644.059 1 61644.059 112.509 0.000
CC 4.866 4 1.216 0.002 1.000
V0 222653.915 1 222653.915 406.373 0.000
t1 3992434.554 1 3992434.554 716.246 0.000
572283.573 1 572283.573 1044.494 0.000
73037.088 1 73037.088 133.302 0.000
CC V0 209.902 4 52.475 0.096 0.984
CC t1 82.857 4 20.714 0.038 0.997
CC 20.484 4 5.121 0.009 1.000
CC 3.253 4 0.813 0.001 1.000
error 249296.752 455 547.905
Total 4190549.704 480
CC: Control charts
(6-2) Equal intercept analysis output

Model F = CC V0 t1
R2 = 0.835 Model F = 297.645 Pr > F = 0.000
Parameter Sum of Squares Degree of Freedom Mean Square F Pr > |F|

Intercept 61644.059 1 61644.059 116.317 0.000
CC 1538.080 4 384.520 0.726 0.575
V0 222653.915 1 222653.915 420.130 0.000
t1 3992434.554 1 3992434.554 740.492 0.000
572283.573 1 572283.573 1079.853 0.000
73037.088 1 73037.088 137.815 0.000
error 249613.247 471 529.964
Total 4190549.704 480
CC: Control charts
CONTROL CHARTS
Acknowledgment. This work is partially supported by National Science Council of

Taiwan. The authors also gratefully acknowledge the helpful comments and suggestions
of the reviewers, which have improved the presentation.
REFERENCES
[1] W. A. Shewhart, Economic Control of Quality of Manufactured Product, Van Nostrand, New York,
1931.
[2] C. W. Champ and W. H. Woodall, Exact results for Shewhart charts with supplementary run rules,
Technometrics, vol.29, pp.393-399, 1987.
[3] J. M. Lucas and M. S. Saccucci, Exponentially weighted moving average control schemes: Properties
and enhancements, Technometrics, vol.32, pp.1-12, 1990.
[4] E. S. Page, Continuous inspection schemes, Biometrika, vol.41, pp.100-115, 1954.
[5] J. M. Lucas, Combined Shewhart-CUSUM quality control schemes, J. Qual. Technol., vol.14, pp.51-
59, 1982.

[6] M. R. Reynolds Jr., R. W. Amin, J. C. Arnold and J. A. Nachlas, X-charts with variable sampling
interval, Technometrics, vol.30, pp.181-192, 1988.
[7] M. R. Reynolds Jr. and J. C. Arnold, Optimal one-side Shewhart control chart with variable sampling
intervals, Sequential Analysis, vol.8, pp.51-77, 1989.
[8] G. C. Runger and J. J. Pignatiello Jr., Adaptive sampling for process control, J. Qual. Technol.,
vol.23, pp.133-155, 1991.
[9] S. S. Prabhu, G. C. Runger and J. B. Keats, An adaptive sample size X chart, Int. J. Prod. Res.,
vol.31, pp.2895-2909, 1993.
[10] A. F. B. Costa, X control chart with variable sample size, J. Qual. Technol., vol.26, pp.155-163,
1994.
[11] S. S. Prabhu, D. C. Montgomery and G. C. Runger, A combined adaptive sample size and sampling
interval X control chart, J. Qual. Technol., vol.26, pp.164-176, 1994.
[12] A. F. B. Costa, X charts with variable sample sizes and sampling intervals, J. Qual. Technol., vol.29,
pp.197-204, 1997.
[13] Y. C. Lin and C. Y. Chou, On the design of variable sample size and sampling intervals X charts
under non-normality, Int. J. Prod. Econ., vol.96, pp.249-261, 2005.
[14] A. F. B. Costa, X charts with variable parameters, J. Qual. Technol., vol.31, pp.408-416, 1999.
[15] M. S. De Magalhaes, A. F. B. Costa and F. D. M. Neto, Adaptive control charts: A Markovian
approach for processes subject to independent disturbances, Int. J. Prod. Econ., vol.99, pp.236-246,
2006.
[16] A. J. Duncan, The economic design of X charts used to maintain current control of a process, J.
Am. Stat. Assoc., vol.51, pp.228-242, 1956.
[17] D. C. Montgomery, The economic design of control charts: A review and literature survey, J. Qual.
Technol., vol.12, pp.75-87, 1980.
[18] C. Ho and K. E. Case, Economic design of control charts: A literature review for 1981-1991, J. Qual.
Technol., vol.26, pp.39-53, 1994.
[19] W. K. Chiu, Economic design of attribute control charts, Technometrics, vol.17, pp.81-87, 1975.
[20] D. S. Bai and K. T. Lee, An economic design of variable sampling interval X-bar control charts, Int.
J. Prod. Econ., vol.54, pp.57-64, 1998.
[21] C. Park and M. R. Reynolds Jr., Economic design of a variable sampling rate X charts, J. Qual.
Technol., vol.31, pp.427-443, 1999.
[22] M. S. De Magalhaes, E. K. Epprecht and A. F. B. Costa, Economic design of a Vp X chart, Int. J.
Prod. Econ., vol.74, pp.191-200, 2001.
[23] A. F. B. Costa and M. A. Rahim, Economic design of X charts with variable parameter: The Markov
chain approach, J. Appl. Stat., vol.28, pp.875-885, 2001.
[24] E. Cnlar, Introduction to Stochastic Processes, Prentice Hall, Englewood Clis, NJ, 1975.
[25] D. E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-
Wesley, Reading, 1989.
[26] D. He and A. Grigoryan, Construction of double sampling S-control charts for agile manufacturing,
Quality Reliab. Eng. Int., vol.18, pp.343-355, 2002.
[27] D. He, A. Grigoryan and M. Sigh, Design of double and triple-sampling X control charts using
genetic algorithms, Int. J. Prod. Res., vol.40, pp.1387-1404, 2002.
[28] F. Aparisi and J. C. Garca-Daz, Optimization of EWMA and MEWMA control charts using genetic
algorithms, Compu. Oper. Res., vol.31, pp.1437-1454, 2003.
[29] Y. K. Chen, Economic design of X control charts for non-normal data using variable sampling policy,
Int. J. Prod. Econ., vol.92, pp.61-74, 2004.
[30] C. Y. Chou, C. H. Chen and C. H. Chen, Economic design of variable sampling intervals T 2 control
charts using genetic algorithms, Expert Sys. Appl., vol.30, pp.233-242, 2006.
[31] Y. Li, Y. Yang, L. Zhou and R. Zhu, Observations on using problem-specific genetic algorithm for
multiprocessor real-time task scheduling, International Journal of Innovative Computing, Informa-
tion and Control, vol.5, no.9, pp.2531-2540, 2009.
[32] H. Li, Y.-C. Jiao, F.-S. Zhang and L. Zhang, An ecient method for linear bilevel programming
problems based on the orthogonal genetic algorithm, International Journal of Innovative Computing,
[33] D. C. Montgomery, Design and Analysis of Experiments, John Wiley & Sons, 1991.
c ISSN 1349-4198
INVENTORY MANAGEMENT IN A (Q, r) INVENTORY MODEL

WITH TWO DEMAND CLASSES AND FLEXIBLE DELIVERY

School of Economics and Management
Tsinghua University
Haidian District, Beijing 100084, P. R. China
{ huangsh; liaob.05; david }@sem.tsinghua.edu.cn
Abstract. This paper considers a single-product inventory system that serves both on-
line and traditional customers. Demands from both channels follow independent station-
ary Poisson processes. Traditional customers have their demands fulfilled upon arrival if
the retailer has stock on hand; otherwise the demand is lost. However, online customers
place their orders in advance, and delivery is flexible in the sense that early shipment is
allowed. In particular, an online order placed at time t should be fulfilled by time t + T ,
where T is the customers demand lead time [1]; late fulfillment incurs a time-dependent
backorder cost. A (Q, r) replenishment policy is used and replenishment lead times are
assumed to be constant. We develop an approximation for the expected annual cost for
the retailer, and compare analytically and simulated results. The optimal parameters of
the system are derived by minimizing the expected annual cost. We illustrate the model
with numerical examples, and discuss the sensitivity of the results to variables such as
demand lead time and the split between online and traditional orders.
Keywords: Inventory model, Multi-channel, Demand lead time
1. Introduction. The Internet has become an important retail channel in the past
decade. In 2004, online sales comprised about 5.5% of all retail sales excluding travel
[2]. In most product categories, a bricks-and-clicks structure, where one company has
both the Internet and a traditional presence, is becoming the dominant form of the In-
ternet participation. Large, traditional retailers such as Wal-Mart, Staples and Best Buy
have websites that sell their products in significant volumes [3]. Meanwhile, some tradi-
tional online-only companies are expanding their business to retail stores. For example,
Dell has installed kiosks in shopping malls and sells its computers through Costco [4].
Such firms now have two dierent channels to reach their customers, and face challenges
in eectively managing inventory given the distinct features of dierent channels.
The objective of this paper is to study the inventory management problems under a
bricks-and-clicks structure. We consider a model where a retailer sells a single product in
both online and traditional channels. The customers in dierent channels exhibit dierent
aversions to waiting. The retailer replenishes its stock according to a continuous review
(Q, r) policy. The objective is to make replenishment decisions to minimize the retailers
long-run average costs. We show how this problem can be solved and present numerical
examples to analyze how certain variables aect the expected costs of the retailer.
There is extensive research on inventory models with multiple demand classes [6-9].
However, few of them consider a combination of backorders and lost sales in multiple
demand classes, especially under a (Q, r) policy. Many papers in this stream do not
deal with inventory issues [10-12], or simply assume a base stock policy and exponential
replenishment lead time [5,13]. One reason for this assumption is that it enables a Markov
1243
1244 S. HUANG, B. LIAO AND D. J. ROBB
process analysis. However, fixed ordering costs do exist and a (Q, r) policy is widely used
in practice. To the best of our knowledge, we are the first to consider such a case in the
prevalent (Q, r) inventory model employed in retailing. We also employ an assumption
of constant lead time; since in a retailing environment, replenishment lead time between
suppliers and retailers is usually specified, and has little variation.
Our model also directly contributes to the stream of research on multiple demand
classes/channels by taking the demand-lead-time of the online customers into considera-
tion. In practice, online customers often place their orders in advance and early shipment
is allowed. Though research on demand-lead-time has traditionally taken the exact deliv-
ery assumption for granted [1,14], the work of [15] argues that flexible delivery arrange-
ments are prevalent in many instances where firms sell directly to consumers.
In this paper, we develop and evaluate a tractable model for bricks-and-clicks retail-
ers, and incorporate more realistic assumptions to assist retailers make better inventory
control decisions. Sensitivity analysis on free waiting time shows its impact on average
inventory cost. We also evaluate how a change in the online demand ratio aects cost
providing additional insights for managers. Although this problem is motivated by in-
ventory problems faced by bricks-and-clicks retailers, we consider our inventory system
structure and solution to be applicable to a wide range of industry settings. For example,
since our model combines two ways to handle shortages under a (Q, r) policy, it is appli-
cable to all three applications described in the work of [5], i.e., a retailer facing both loyal
customers and occasional walk-in customers, a physical-and-online store, and an origi-
nal equipment manufacturer satisfying demand using both emergency and replenishment
demands from local warehouses.
The remainder of the paper is organized as follows. In Section 2 we discuss the relevant
literature. Section 3 introduces the model. Section 4 describes algorithms for deriving
the mathematical expression of the expected total annual cost. Numerical examples and
concluding remarks are presented in Sections 5 and 6.
2. Literature Review. There exists an extensive literature considering inventory sys-

tems with multiple demand classes. The work of [17] provides a high-level categorization
of the previous literature. Our model falls into the research stream of continuous review
inventory control policy, but combines lost sales and backorders.
The work of [6] appears to be the first to study a continuous review (Q, r) inventory
model with two demand classes; they assume a critical-level policy and backordering of
excess demand. The work of [9] extends the model in the work of [6] by relaxing the
assumption of at most one outstanding order. Dierentiating two demand classes by
shortage costs, they introduce a threshold clearing mechanism to fill backorders. Using
this mechanism they develop algorithms to calculate the optimal ordering and rationing
parameters, and demonstrate the eectiveness of their model numerically. The work of
[18] also analyzes a (Q, r) inventory model with two demand classes, but dier from the
work of [6] and [9] by considering a lost sales environment. They also ration inventory
between dierent classes by applying a critical level policy.
All of the above papers dierentiate demand classes based on the priority of the cus-
tomers and apply a critical level policy. Our model diers in that we dierentiate demand
classes based on how excess demands are handled. Furthermore, our focus is on evaluating
system performance given a (Q, r) policy, and selecting optimal parameter values. Wang
et al. [16] use delivery lead-time to dierentiate demand classes. Similar to our model,
they develop system performance evaluations for a given policy, although they consider a
base-stock policy for inventory replenishment. Chen [8] considers multiple demand seg-
ments with diering aversions to shipping delays. This work develops an optimal pricing
INVENTORY MANAGEMENT IN A (Q, r) INVENTORY MODEL 1245
and replenishment strategy for a multi-stage supply chain with backlogging for which
he proves the optimality of an echelon base stock policy.
Related work includes the multiple channel supply chain literature modeling coordi-
nation between traditional and internet channels. Cattani et al. [3] provide a review of
some of the early work. These papers either analyze the dual-channel design problem by
modeling the price and/or service interactions between upstream and downstream eche-
lons [10,11], or analyze a supply chains performance in equilibrium [12]. However, they
do not deal with inventory issues, the theme of our model.
Several studies have examined inventory issues in a dual-channel supply chain. The
work of [19] considers a two-echelon dual-channel inventory model in which stock is held
at both the manufacturers warehouse and a retail store. They assume a one-for-one
policy and discuss insights based on parametric analysis. The work of [13] studies several
rationing policies in which the firm either blocks or backlogs orders for lower priority
customers when inventory falls to or below a certain level. Given that the customers are
willing to wait, they formulate the inventory model as a Markov Process and show that
their rationing policies can improve the overall performance of the firm, compared to FCFS
policies. However, these papers assume base stock policies and exponential production
lead times, which may not be practical in some retailing environments. In contrast,
we assume a (Q, r) inventory policy and a constant supply lead time, which are quite
common in a retailing environment. Seifert et al. [20] model the inventory decisions for a
supply chain comprised of a manufacturer and N retailers and derive optimal base stock
levels. Their model is similar to ours in that the retail customers demand instantaneous
availability and the retailer incurs lost sales if the product is not in stock; whereas internet
customers are willing to wait. However, they assume a base stock policy and demands in
the internet channel are ultimately lost after a fixed delay.
Besides considering a dual-channel structure, we take the characteristics of online cus-
tomers into account when we model the online channel, assuming that they are willing to
accept a reasonable waiting time after placing their order. The concept of free waiting
time of online customers is closely related to the advance demand information (ADI)
literature. Most previous studies on ADI concentrate on the value of ADI in production-
inventory systems, e.g., the work of [21] presents a detailed analysis of a single-stage
make-to-stock queue with ADI. Based on the same basic model, the work of [22] studies
a discrete-time make-to-stock queue, showing that base stock policies are near optimal
for demand lead times below a threshold. For more general inventory systems, the work
of [1] denotes the time from a customers order until the due date as the demand lead
time, and demonstrates that demand lead times directly oset supply lead times. The
work of [14] investigates optimal replenishment policies for a single-stage periodic review
inventory system with ADI. Using numerical results, they show that under the optimal
replenishment policy, ADI can lead to significant cost reductions. This work is later ex-
tended to the multi-stage case in the work of [23]. Marklund [24] extends the work of [1]
to one-warehouse multiple-retailer divergent supply chains.
Most ADI literature assumes exact delivery, early fulfillment is not allowed. In a Just-
in-time (JIT) purchasing environment, this assumption is reasonable. However, as online
customers, we have probably experienced the following situation. After order placement,
we are notified that delivery will take place no later than a particular date. Relatively
few models consider allowing for early delivery what the work of [15] terms flexible
delivery. The work of [25] studies inventory management with a service level constraint
under a flexible time-window fulfillment scheme. They use an (s, S) policy and develop
algorithms for finding the optimal parameters. The work of [15] extends the model in
the work of [14] by allowing for flexible delivery. They consider both homogeneous and
heterogeneous customer demand lead times. In contrast to the work of [15], we consider a
continuous review environment, and allow for two demand classes with online customers
forming one of the classes.
3. The Model. We consider a retailer that carries inventory for a single product to serve
both online and traditional customers. The retailer replenishes inventory with a continu-
ous review (Q, r) policy from an outside supplier with infinite capacity and constant lead
time. Demand from channel i (i = 1 for the traditional channel, i = 2 for the online
channel) follows a stationary independent Poisson process with rate i .
Customers from both channels are served on a first come first served (FCFS) basis as
long as there is inventory on hand. If the inventory level has dropped to zero, demands
from the traditional channel are lost, whereas online customers wait until the next re-
plenishment to be served. Due to the convenience of online purchasing, customers may
accept a reasonable waiting time after they have placed their order, i.e., a suciently
short waiting time will not incur any shortage cost for the retailer. We assume that an
online demand is charged a backorder cost only if its waiting period exceeds the maximum
free waiting time of online demands in a backorder period, denoted by T > 0. Obviously
a meaningful T should not exceed the retailer lead time, L, otherwise the retailer can
meet every online demand as it occurs and need not hold inventory for online customers,
and the system will reduce to a simple system with only one traditional demand class. A
graphical illustration of the system is shown in Figure 1.
Figure 1. An inventory system with both traditional and online demand
We further assume that there is at most one outstanding order. This assumption is
quite common in inventory control literature considering lost sales. The diculties in
analyzing inventory models with lost sales and the usage of this assumption has been
discussed by the work of [26]. Furthermore, the assumption is often satisfied in practice
because order cycles are usually long compared to lead times [27,28].
To derive the expected cost per unit time, we need to describe how the retailer inventory
level, I(t), changes with time. In Figure 2, a graphical illustration of an order cycle is
presented, from the time when the inventory on hand reaches the reorder point r (at
t = 0), to when the replenishment arrives (t = L) and then until the inventory drops to
r again.
The explanation of Figure 2 is as follows: The retailer issues an order of size Q to
replenish its inventory whenever the inventory level drops to r. Since we assume there is
no more than one outstanding order, the inventory level of the retailer immediately before
this replenishment order is equal to the inventory position. Denote this time point as time
0. After time 0, the inventory continues to deplete at rate (= 1 + 2 ) until time T (r),
when all r units have been demanded and the inventory level is 0. From T (r), since there
is no inventory on hand, traditional customers will leave without purchasing, whereas
demands from the online channel will continue to reduce the inventory, i.e., inventory is
depleted at rate 2 after T (r). We denote the time period of lost sales in a cycle as TLS .
Figure 2. Changes in inventory level in a dual-channel inventory system
The backorders of online demands only incur shortage penalty costs after waiting time T ,
i.e., during period TB . The accumulation of backorders that incur shortage costs is shown
in Figure 2 with the bold curve. The replenishment arrives at time L, after which the
inventory decreases again at rate until the reorder point r, completing a typical cycle.
Other notation employed includes: h denotes unit holding cost per unit time at the
retailer, b is unit backorder cost per unit time at the retailer, stands for unit lost sales
cost, i.e., the lost profit (including any lost goodwill) per unit, K and c represent fixed
ordering cost and variable unit purchasing cost, respectively.
Our goal is to obtain Q and r so as to minimize the average cost per unit time, C,
where
C(Q, r) = E[T C]/E[Tc ] (1)
where T C is the total cost accumulated within a cycle. Here
E[T C] = K + cQ + hE[IT ] + E[LS] + bE[BT ] (2)
where IT is the time-weighted inventory in a cycle, LS is the number of lost sales in a
cycle, and BT is the time-weighted backorders in a cycle. The first and second parts in
Equation (2) are easy to obtain. In the next section, we develop an approximate method
to derive the remaining components of the expected cost in a cycle.
We adopt the notation employed in [26] to denote the Poisson probability
(t)j et
p(j; t) = (3)
j!
and its tail probability

P (j; t) = p(i; t) (4)
i=j
4. Analysis. As in Figure 2, denote TLS as the time period of lost sales in a cycle. Since
no online order will be lost, TLS equals the time period for lost sales in the traditional
channel in a cycle. Thus

{
0 when T (r) L,
TLS = (5)
L T (r) when T (r) < L,
where T (r) is the time for r units to be depleted in both channels. Notice that T (r) is
Erlang distributed with parameters r and , its probability density function is p(r1; t).
Thus
L L
r
E[TLS ] = (L t)f (t)dt = (L t)p(r 1, t)dt = LP (r, L) P (r+1, L) (6)
0 0
Then the expected number of lost sales per cycle can be obtained as
1
E[LS] = 1 E[TLS ] = 1 LP (r; L) rP (r + 1; L) (7)

On average all items ordered are consumed in a single cycle, the satisfied demands per
unit time can be denoted as Q/E[Tc ]. This number is also equal to the incoming demands
minus lost sales, i.e., E[LS]/E[Tc ]. By equating these two expressions we obtain
Q + 1 E[TLS ]
E[Tc ] = (8)

Denote the time-weighted backorders per cycle as BT . Due to the free waiting time
period T , backorder costs are only incurred in time interval [T (r), t] and when t falls into
a certain interval, T (r) < t < L T . The expression for BT can thus be formulated as
LT
BT = D2 (t T (r))dt (9)
T (r)
where D2 (t T (r)) denotes online demands arising during the period [T (r), t].
Given that the probability density function of T (r) is f (s) = p(r 1; s), the expected
time-weighted backorders per cycle, E[BT ], is
[ LT ] LT [ LT ]
E[BT ] = E D2 (t T (r))dt = 2 (t s)dt f (s)ds
T (r) 0 s
2 (L T ) 2
2 (L T )r (10)
= P (r; (L T )) P (r + 1; (L T ))
2
2 r(r + 1)
+ P (r + 2, (L T ))
22
We divide the derivation of the expected time-weighted inventory held per cycle, E[IT ],
into two parts: prior to the replenishment time, denoted by E[IB ], and after the replen-
ishment time, denoted by E[IA ]. Defining D(t) as the total demand during (0, t], with
I(t) denoted as the inventory on hand at time t, then
L L
E[IB ] = +
E[I(t)] dt = E[r D(t)]+ dt
L { }
0 0
r1
r1 (11)
(r j)
= (r x)p(x; t) dt = P (j + 1; L)
0 x=0 j=0

In the following steps, we deduce the expected time-weighted inventory held per cycle
after the replenishment time, E[IA ]. Let I(L) be the inventory level just before a replen-
ishment order arrives. According to our assumption, I(L) can only be distributed from
r Q to r. Recall that T (r) is the time for r demands to arrive from time 0. If T (r) > L,
I(L) is greater than 0; otherwise for a given T (r) = t L, the probability that I(L)
equals j (j 0) equals the probability that there are k (k r) demands coming from
both channels in [0, L] and j of the last k r demands come from the online channel.
Thus, we have

p(r j; L), j = 1, 2, . . . , r
( ) ( )k+j ( )j
Pr[I(L) = j] =
k 1 2

p(r + k; L) , j = (Q r), . . . , 1, 0
k=j j
(12)
Now, let Z = Q + I(L) be the inventory level just after a replenishment order arrives.
The state space of Z is {r, r + 1, . . . , r + Q}. Denote Ti as the inter-arrival time of the
demands (including traditional and online demands). Then Ti , i = 1, 2, 3, . . . are i.i.d.
exponential random variables with mean 1/. The time-weighted inventory held from the
replenishment time until the next ordering time is then given by
{
0 if Z = r
IA = (13)
ZT1 + (Z 1)T2 + . . . + (r + 1)Tzr if Z > r
Hence
1
r+Q
E[IA ] = (Pr[Z = z](z + (z 1) + . . . + (r + 1)))

z=r+1
1
r+Q
= (Pr[Z = z](z 2 + z r(r + 1)))
2
z=r+1
( r+Q ) (14)
1
= (z 2 + z) Pr[Z = z] r(r + 1) Pr[Z > r]
2
z=r+1

1
r
= [(j + Q)2 + (j + Q)] Pr[I(L) = j] r(r + 1) Pr[I(L) > r Q]
2
j=r+1Q
Now, we can obtain the total expected holding cost per cycle: hE[IT ] = h(E[IA ]+E[IB ]).
The optimal control parameters Q and r can be determined by solving the following
nonlinear integer optimization problem:
Min C(Q, r),
Q,r
(15)
subject to Q, r > 0 and integer.
5. Numerical Results. Our objective in this section is to draw managerial insights

based on a numerical analysis of our model. We first test the eectiveness of our method
via comparison with simulation results. In the first stage of the experiments, we fix the
financial parameters as follows: K = 100; c = 10; h = 3; b = 4; = 25. We set T = 1 and
study short and long lead times, respectively (L = 2 and 7). Three demand scenarios are
considered in the comparison, i.e., low, medium and high demand rates with = 0.2, 2,
10. In each scenario, we first assume that the demands from two channels are identical.
For each instance, by employing the EOQ as an approximate Q and adjusting r from 0 to
Q1, we compare our analytical results for the expected average cost with the simulation.
The results are shown in Table 1, and Figures 3 and 4.
As shown in the numerical examples, the deviations in our comparisons under low and
medium demand scenarios are quite small, especially when the lead time is short. However,
when the lead time is long, the gap between the simulation and analytical expected costs
is increasing in the reorder point r (see Figures 3 and 4). Thus, the deviations in expected
cost tend to increase when the demand rate is medium or high, and the reorder point is
high. The increased gap associated with increase in L and r results from the assumption
Table 1. Expected total costs for low demand rate cases (1 = 2 = 0.1,
Q = 4)
Short lead time (L = 2) Long lead time (L = 7)

r
Simulated Analytical Dierence (%) Simulated Analytical Dierence (%)
0 13.52 13.50 0.16 11.7802 11.7606 0.17
1 16.31 16.31 0.00 13.7484 13.7367 0.09
2 19.28 19.30 0.09 16.3595 16.3681 0.05
3 22.28 22.30 0.08 19.3207 19.3004 0.11
Figure 3. Expected total costs for medium demand rate cases
Figure 4. Expected total costs for high demand rate cases
of at most one outstanding order. In the worst case (1 = 2 = 5, L = 7, r = 25), that is,
with the highest demand rate, longest lead time, and highest reorder point, the expected
cost derived from the analytical method leads to a 20.51% deviation from the simulation
result. However, considering that the lead time for fast moving products is usually very
short, and the reorder point is not likely to be very high, our analytical method is still
eective in most real situations. Thus, we base the remaining numerical studies on the
analytical model.
In the remaining numerical examples we employ the following parameters: K = 100;
c = 10; h = 3; b = 4; = 3; L = 7. Figure 5 shows the optimal decision variables Q
and r for dierent values of and 1 /. We set 1 / = 0.1, 0.5, and 0.9, respectively.
The upward trend of both Q and r is expected since as the cost of lost sales increases,
it is reasonable to hold more inventory to shorten the period of lost sales. Another
finding is that the increase in both Q and r with is more sudden for larger 1 /,
i.e., both Q and r are more sensitive to changes in lost sale cost when demand from
the traditional channel constitutes the majority of total demand. This finding is intuitive
since only demand losses are only possible from the traditional channel and the lost sales
cost is usually much higher than the backorder cost, and therefore has a more significant
influence on Q and r. However, when the unit lost sales cost reaches an extremely high
level, both Q and r tend to stabilize as the likelihood of a lost sale is extremely low in
such circumstances.
Figure 5. Optimal decision variables
Figures 6 shows the results of sensitivity analysis for dierent values of 1 /. Again,
we set 1 / = 0.1, 0.5 and 0.9. For higher 1 /, the expected cost increases more quickly.
This is understandable since with more demand coming from the traditional channel, the
influence of (which is only incurred when traditional demand is not met) will be higher.
When is very large, the cost dierences among dierent demand scenarios are distinct
and relatively insensitive to changes in . This arises from the optimal inventory policy
for large having a very short period of lost sales, thus the lost sales cost constitutes only
a small portion of the total expected cost and the impact of is thus weak.
Figure 7 shows that the expected cost is more sensitive to T when is small. This is
because when the lost sales cost is small, the dominant parts of the total expected cost
are the holding cost and the backorder cost, which are both significantly aected by the
length of the free waiting time, T . That is, with a longer T , the holding cost and the
backorder cost, constituting the main part of the total expected cost, decrease quickly.
However, when is large (in our example larger than 60), the lost sales cost comprises
the main part of the total cost, hence the impact of T on the total expected cost is much
Figure 6. Expected cost for dierent values of 1 /
less significant. In addition, the high lost sales cost will result in an optimal policy with
higher Q and r , which also weakens the aect of T since the period of stock-out time
is very short in these cases sometimes even shorter than T -rendering T ineective in
reducing cost. This explains why the cost dierences for dierent T values are very small
when is large.
Figure 7. Expected cost for dierent values of T
6. Concluding Remarks. In this paper, a tractable inventory model for bricks-and-

clicks retailers is developed and analyzed. We assume a constant lead time and the
commonly found (Q, r) inventory policy. We allow for a period of free waiting time for
online customers, a practice widely accepted by online customers due to the convenience
of online purchasing. An algorithm is developed to evaluate the system. Numerical
examples validate the eectiveness of the theoretical results when seeking to minimize
expected cost. The algorithm performs particularly well when lead time is not too long.
Inventory systems with multiple channels are becoming more common with the growth
of e-business. How inventories should be managed in such systems is a critical managerial
decision. For given parameter settings our model can be employed to determine the order
size and re-order point for an item, by solving (15), which could be conducted using a
spreadsheet optimization package or a two-dimensional grid search, e.g., with the grid
size reflecting realistic values. Our model can be used to analyze the impacts of dierent
parameters on the optimal inventory policy and the expected cost. For example, when the
cost of lost sales exceeds a particular level, an increase in the online demand proportion
helps to reduce the expected cost, and this cost-reduction can be enhanced if the free
waiting period for online customers is extended. This oers insights for managers to take
measures to make online shopping more appealing, i.e., to encourage traditional customers
to shift to online channels. The savings derived from our analytical results can be used to
determine the price or discount to online customers. Another way to decrease the expected
cost is to increase the free waiting time of online customers. Here the savings derived from
our analysis can be used to determine the compensation for customers waiting longer.
Our work has limitations in not allowing for stochastic leadtimes (and thus potential
for order crossover) nor the possibility of more than one order outstanding at any point in
time. With the relatively short leadtimes found in many e-business/retail environments
we believe these assumptions are not especially problematic.
Our work can be extended in several ways, e.g., by considering pricing strategy [8,30],
which would be especially pertinent were online demand to be modelled as a function of
waiting time. One could also develop heuristic inventory policies for the multiple channel
system, perhaps expanded to include joint retailer-supplier costs, and allow for lost sales
and backorder costs for both traditional and online sales. In this regard it should be noted
that fuzzy models [31,32] are increasingly being developed to address such problems.
Acknowledgment. Shuo Huangs work on this paper was supported by the National
Natural Science Foundation of China, under Grant Nos. 70601018, 71072016 and 7089008
2. David Robbs work on this paper was supported by the National Natural Science Foun-
dation of China, under Grant No. 70890082. The authors also gratefully acknowledge the
helpful comments and suggestions of the reviewers, which have improved the presentation.
REFERENCES
[1] R. Hariharan and P. Zipkin, Customer-order information, leadtimes and inventories, Management
Science, vol.41, no.10, pp.1599-1607, 1995.
[2] M. Mangalindan, Online retail sales are expected to rise to $172 billion this year, The Wall Street
Journal, 2005.
[3] K. D. Cattani, W. G. Gilland and J. M. Swaminathan, Coordinating traditional and internet supply
chains, in Handbook of Quantitative Supply Chain Analysi: Modeling in the E-Business Era, D.
Simchi-Levi, S. D. Wu and Z. J. Shen, (eds.), Springer, New York, 2004.
[4] G. McWilliams and A. Zimmerman, Dell plans to peddle PCs insider sears, other large chains, The
Wall Street Journal, 2003.
[5] P. Enders, I. Adan, A. Scheller-Wolf and G. J. van Houtum, Inventory rationing for a system with
a lost-sales and a backordering customer class, Working Paper, Tepper School of Business, Carnegie
Mellon University, 2008.
[6] S. Nahmias and S. Demmy, Operating characteristics of an inventory system with rationing, Man-
agement Science, vol.27, no.11, pp.1236-1245, 1981.
[7] R. Dekker, M. J. Kleijn and P. J. de Rooij, A spare parts stocking policy based on equipment
criticality, International Journal of Production Economics, vol.56-57, pp.69-77, 1998.
[8] F. Chen, Market segmentation, advanced demand information, and supply chain performance, Man-
ufacturing and Service Operations Management, vol.3, no.4, pp.53-67, 2001.
[9] V. Deshpande, M. A. Cohen and K. Donohue, A threshold inventory rationing policy for service-
dierentiated demand classes, Management Science, vol.49, no.6, pp.683-703, 2003.
[10] W. K. Chiang, D. Chhajed and J. D. Hess, Direct marketing, indirect profits: A strategic analysis
of dual-channel supply-chain design, Management Science, vol.49, no.1, pp.1-20, 2003.
[11] A. Tsay and N. Agrawal, Channel conflict and coordination in the e-commerce age, Production and
Operations Management, vol.13, no.1, pp.93-110, 2004.
[12] A. Dumrongsiri, M. Fan, A. Jain and K. Moinzadeh, A supply chain model with direct and retail
channels, European Journal of Operational Research, vol.187, pp.691-718, 2008.
[13] K. D. Cattani and G. C. Souza, Inventory rationing and shipment flexibility alternatives for direct
market firms, Production and Operations Management, vol.11, no.4, pp.441-457, 2002.
[14] G. Gallego and O. Ozer,
Integrating replenishment decisions with advance demand information,
Management Science, vol.47, no.10, pp.1344-1360, 2001.
[15] T. Wang and B. L. Toktay, Inventory management with advance demand information and flexible
delivery, Management Science, vol.54, no.4, pp.716-732, 2008.
[16] Y. Z. Wang, M. A. Cohen and Y. S. Zheng, Dierentiating customer service on the basis of delivery
lead-times, IIE Transactions, vol.34, no.11, pp.979-989, 2002.
[17] H. Arslan, S. C. Graves and T. A. Roemer, A single-product inventory model for multiple demand
classes, Management Science, vol.53, no.9, pp.1486-1500, 2007.
[18] P. Melchiors, R. Dekker and M. J. Kleijn, Inventory rationing in an (s, Q) inventory model with
lost sales and two demand classes, Journal of Operational Research Society, vol.51, no.1, pp.111-122,
2000.
[19] W. K. Chiang and G. E. Monahan, Managing inventories in a two-echelon dual-channel supply chain,
European Journal of Operational Research, vol.162, pp.325-341, 2005.
[20] R. W. Seifert, U. W. Thonemann, S. A. Rockhold and M. A. Sieke, Integrating direct and indirect
sales channels under decentralized decision making, International Journal of Production Economics,
vol.103, no.1, pp.209-229, 2006.
[21] J. A. Buzacott and J. G. Shanthikumar, Safety stock versus safety time in MRP controlled production
systems, Management Science, vol.40, no.12, pp.1678-1689, 1994.
[22] F. Karaesmen, J. A. Buzacott and Y. Dallery, Integrating advance order information in make-to-stock
production systems, IIE Transactions, vol.34, no.8, pp.649-662, 2002.
[23] G. Gallego and O. Ozer,
Optimal replenishment policies for multi-echelon inventory problems un-
der advance demand information, Manufacturing and Science Operations Management, vol.5, no.2,
pp.157-175, 2003.
[24] J. Marklund, Controlling inventories in divergent supply chains with advance-order information,
Operations Research, vol.54, no.5, pp.988-1010, 2006.
[25] T. Wang, Y. Chen and Y. Feng, On the time-window fulfillment rate in a single-item min-max
inventory control system, IIE Transactions, vol.37, no.7, pp.667-680, 2005.
[26] G. Hadley and T. M. Whitin, Analysis of Inventory Systems, Prentice Hall, Englewood Clis, NJ,
1963.
[27] H. Tempelmeier, Supply chain inventory optimization with two customer classes in discrete time,
European Journal of Operational Research, vol.174, pp.600-621, 2006.
[28] R. H. Teunter and W. K. Klein Haneveld, Dynamic inventory rationing strategies for inventory sys-
tems with two demand classes, Poisson demand and backordering, European Journal of Operational
Research, vol.190, pp.156-178, 2008.
[29] G. P. Zhang, B. E. Patuwu and C. W. Chu, A hybrid inventory system with a time limit on
backorders, IIE Transactions, vol.35, pp.679-687, 2003.
[30] N. R. Smith, J. L. Robles, L. E. Cardenas-Barron and R. G. Gonzalez-Ramirez, Optimal pricing and
production master planning in a multi-period horizon considering capacity and inventory constraints,
[31] C.-C. Chou, Fuzzy economic order quantity inventory model, International Journal of Innovative
[32] T.-S. Shih, J.-S. Su, C.-F. Fuh and H.-M. Lee, Applying fuzzy genetic algorithms to solve a fuzzy
inventory with backorder problem, International Journal of Innovative Computing, Information and
Control, vol.6, no.1, pp.229-237, 2010.
c ISSN 1349-4198
A NOVEL TEXT WATERMARKING ALGORITHM USING

IMAGE WATERMARK

FAST National University of Computer and Emerging Sciences
A. K. Barohi Road, H-11/4, Islamabad, Pakistan
{ Zunera.jalil; arfan.jaar; anwar.m.mirza }@nu.edu.pk
Received October 2009; revised February 2010
Abstract. Authentication and copyright protection of information contents has always

been a concern in print media. The problem has become more critical with the increasing
use of the Internet and digital technologies. Besides image, audio and video, the text
is most extensively used medium travelling over the Internet. Digital libraries oer a
quick and easy access to information such as e-books, archives, images, etc. However,
making the protection of copyright is more complex and dicult. Increasing use of digital
libraries, blogs and electronic commerce has made it necessary to protect digital contents.
Digital watermarking came up as a solution for copyright protection problem. In this
paper, we propose a novel text watermarking algorithm, which embeds the watermark
image in the text (to be protected) logically using embedding algorithm and extract it
later using the extraction algorithm. Experimental results demonstrate the eectiveness
of the algorithm proposed in localized and dispersed attacks on the text. Objective evalu-
ation methods like Normalized Hamming Distances (NHD), Peak-Signal-to-Noise-Ratio
(PSNR), Weighted Peak-Signal-to-Noise-Ratio (WPSNR), Mean Structural Similarity
Index Measures (MSSIM) and Normalized Absolute Error (NAE) metrics are used to
measure the quality and similarity between original and extracted watermark images.
Keywords: Digital watermarking, Information security, Image watermark, Document
authentication, Copyright protection
1. Introduction. Copyright protection, piracy, steganography, watermarking and en-

cryptions are certain terms used frequently in the digital world. The use of the Internet,
e-commerce and digital libraries, mobile phones, iPods; has made the dissemination of
information fast and easy [26]. However, it has become increasingly dicult to han-
dle security, privacy and copyright protection issues of digital contents. Digital contents
mostly comprises of text, image [23], audio, videos and software programs. Authentica-
tion and Copyright protection of image [25], audio, software and video has been given
due to consideration by researchers in the past; however, the amount of work done to
protect text is very scare. Text is the most essential and dominant part of websites,
e-books, newspapers, articles, legal documents, reports and journals; but its protection
has been seriously ignored. The threats of electronic publishing like illegal copying and
re-distribution of copyrighted material, plagiarism and other forms of copyright violations
need to be explicitly addressed [7], particularly for plain text.
Copyright protection of digital content, especially text, is such a need of time that
cannot be neglected. Digital watermarking is one of the solutions for copyright protection
and authentication of text documents and the amount of work done for watermarking
of plain text is very insucient. Digital Watermarking methods identify the original
copyright owner(s) of the information contents. A digital watermark can be described
1255
1256 Z. JALIL, M. A. JAFFAR AND A. M. MIRZA
as a visible or an invisible, preferably the latter, identification code that permanently is

embedded in the data. It means that unlike conventional cryptographic techniques, it
remains present within the data even after the decryption process [1].
Text watermarking is a significant area of research, which might open the new horizon in
digital security world. In past, various text watermarking techniques have been proposed.
These include text watermarking using text image, synonym, pre-supposition, syntactic
tree, noun-verb, word and sentence structure, acronyms, typo errors etc. Generally, we
can classify the previous work on digital text watermarking in the following categories;
an image based approach, a syntactic approach, a semantic approach and the structural
approach.
In image based approach towards text watermarking, watermark is embedded in text
image. Brassil, et al. were the first to propose a few text watermarking methods utilizing
text image [10,11]. Later Maxemchuk, et al. [18-20] analyzed the performance of these
methods. Low, et al. [27,28] further analyzed the eciency of these methods. Huang and
Yan [6] proposed an algorithm based on an average inter-word distance in each line. The
distances are adjusted according to the sine-wave of a specific phase and frequency. The
feature and the pixel level algorithms were also developed which mark the documents by
modifying the stroke features such as width or serif [31].
In syntactic approach, the syntactic structure of text has been used to embed water-
mark. Atallah, et al. first proposed the natural language watermarking scheme by using
syntactic structure of text [12,13] where the syntactic tree is built and transformations
are applied to embed the watermark while keeping all the text properties intact. Hassan
et al. performed morpho-syntactic alterations to the text to watermark it [8]. They also
provided an overview of available syntactic tools for text watermarking [9].
In semantic approach, semantics of text are used to embed the watermark in text.
Atallah et al. were the first to propose the semantic watermarking schemes in the year
2000 [14,15]. Later, the synonym substitution method was proposed, in which water-
mark was embedded by replacing certain words with their synonyms [32]. Sun, et al.
proposed noun-verb based technique for text watermarking [33] which used nouns and
verbs parsed by semantic networks. Later Topkara, et al. proposed an algorithm of the
text watermarking by using typos, acronyms and abbreviation in the text to embed the
watermark [16]. Algorithms were developed to watermark the text using the linguistic
approach of presuppositions [2] in which the discourse structure, meaning, and represen-
tations are observed and utilized to embed watermark bits. The text pruning and the
grafting algorithms were also developed in the past. Another algorithm based on text
meaning representation (TMR) strings has also been proposed [21].
The structural approach is the most recent approach used for copyright protection of
text documents. In this approach, text is not altered, rather it is used to logically embed
watermark in it. A text watermarking algorithm for copyright protection of text using
occurrences of double letters (aa-zz) in text has recently been proposed [34]. Another
zero-watermarking algorithm utilizing prepositions along with double letters has recently
been proposed [35].
Text watermarking algorithms using binary text images are not robust against re-
typing and text reproduction attacks. Text watermarking by using syntactic structure
combined with natural language processing algorithms, is an ecient approach towards
text watermarking but research progress in NLP is very slow. The transformation applied
using NLP algorithms are most of the time non-reversible.
The text watermarking methods based on semantics are language dependent. The
synonym based techniques are not resilient to the random synonym substitution attacks.
Sensitive nature of some text like legal documents, poetry, and quotation do not allow us
A NOVEL TEXT WATERMARKING ALGORITHM USING IMAGE WATERMARK 1257
to make random semantic transformations. The reason behind is the necessity to preserve
the semantic connotation as well the value of text, while performing any transformation.
The structural algorithms are not applicable to all types of text documents and the
algorithms are restricted to alphabetical watermarks.
The problem of Text Watermarking has been studied by researchers in past. However,
major problem with existing methods of Text Watermarking techniques are their lack of
automation and adaptability. The existing Text Watermarking methods are computation-
ally expensive and they can only be applied to specific domains. These methods also make
some assumptions about the type of the requirements. The proposed method addresses
all these problems. It is computationally ecient as well as it needs no assumptions and
advance knowledge about the requirements. It also reduces the emphasis on manual eort
which has its own obvious problems. Text may contain similarity but images dier, so it
is better to use image watermark instead of textual watermark.
The main contributions of this work to the digital community community in general
and text watermarking community in particular are:
A novel text watermarking algorithm, which uses an image as watermark to protect
the text document, has been proposed.
No assumption about watermark image structure, text type and length has been
made. Image can be a logo, fingerprint, face image or signature of the original
author. Text may contain any ASCII character and consist of any number of words.
It is a blind zero-watermarking algorithm which makes no changes to the text, as it
happens in traditional text watermarking approaches. Besides, the original water-
mark is not required for the extraction algorithm.
This paper first analyzes the previous work on text watermarking. The proposed em-
bedding and extraction algorithm are described in detail in Section 2. Section 3 presents
the experimental results for the tampering (insertion, deletion and re-ordering) attacks
with two dierent images. Performance of the proposed algorithm is evaluated by com-
paring it with the most recent relevant work. The last section concludes the paper along
with directions for future work.
2. Proposed Algorithm. The process of integration of a digital watermark into a text

document that carries digital information unique to the owner of copyright or the creator
of the document is called Text Digital Watermarking [35]. Plain text, being the simplest
mode of information exchange, brings many challenges when it comes to the protection
of copyright. Text has a limited capacity for watermark embedding, since any change to
the text need to preserve the meaning, fluency, grammaticality and the value of the text.
The binary nature of text, with clear demarcation between foreground and background,
block/line/word patterning, semantics, structure, style and language rules are some of the
eminent properties of text documents which are needed to be addressed in any text wa-
termarking algorithm. Besides, the inherent properties of a generic watermarking scheme
like imperceptibility, robustness and security also need to be dealt with. We propose a
novel text watermarking algorithm which embeds the watermark image of the original
copyright owner in the text to protect it.
The proposed algorithm embeds the watermark image logically in text which means
watermark is not actually embedded in; rather it is used to generate a key based on text
constituents. We have used text constituents; double letters and most frequently used
words in English text to watermark it. Double letters are very common in the English
language. Almost 7 to 11% words of any text contain double letters [34]. Besides, there
are some frequently used words in English language which are utilized in proposed text
watermarking algorithm. The COBUILD (Collins Birmingham University International
Language Database) provides a list of the ten most frequent words in English [30] as
shown in Table 1. These words are redundant in text and this inherent redundancy is
exploited to protect text in the proposed algorithm. The appearance of double letters
and words that occur frequently in the text is used to embed watermarks. The original
copyright owner of text inputs his/her watermark image and selects a word from Table 1.
Table 1. Top ten most frequently used words in English [22]
Rank Word Occurrence frequency

1 the (309 497) 25.1%
2 of (155 044) 12.6 %
3 and (153 801) 12.5%
4 to (137 056) 11.1%
5 a (1129928) 10.5%
6 in (100 138) 8.1%
7 that (67 042) 5.4%
8 I (64 849) 5.3%
9 it (61 379) 5.0%
10 was (54722) 4.4%
The algorithm generates a unique key as a result of these inputs. This key is used
later for watermark extraction, whenever any copyright conflict arises in future. In the
proposed approach, the original watermark image is not required at the time of watermark
extraction and the embedding algorithm does not make any physical modification in text
to watermark it. Hence it is a blind zero-watermarking algorithm.
A trusted certification authority is a prerequisite for this algorithm with which the
original owner records copyright of its key author. Whenever the content and ownership
of texts are concerned, this trust third party acts as the authority of decision. The trustee
is responsible for implementing detection on behalf of the owner of the content, provided
that the owner has already registered its products watermarked to the third party [34].
The proposed algorithm caters for combined insertion, deletion and re-ordering attack
on the text. A brief description of some key terms used in the proposed algorithm is as
follows:
Definition 2.1. The two consecutive occurrences of an alphabet is called Double letter.
For example, double letter in the word letter is t, and in process is s.
Definition 2.2. An image which acts as a unique identifier of the copyright owner of
the text is called Watermark. This can be a logo, fingerprint, signature, or any other
unique image of the copyright owner.
Definition 2.3. An authorized government or recognized institution which acts as the
neutral facilitator among all entities/stakeholders is called Certifying Authority (CA).
It is just like a registration authority which registers the text documents in the name of
author. Every author who seeks protection from pirates in the digital world embeds a
watermark in the original content [34]. After watermarking the text, the author registers
the key with the Certifying Authority.
Definition 2.4. Any action that a user can perform on a text that can aect the water-
mark, or its usefulness is called Attack [4]. It is the deliberate insertion, deletion and
re-ordering of words and sentences; to and from the original text. The volume of attacks
varies depending on the use of malicious intent. Striker may steal a piece of text or edit
text on a large scale to make it look dierent.
Definition 2.5. The key generated by the embedding algorithm for the original author
based on the watermark and the word given by the author is called Watermark Key.
This key is later registered with the CA along with time as well as date and subsequently
used for the watermark extraction.
The watermarking process involves two stages, watermark embedding and watermark
extraction. Watermark embedding is done by the original author and extraction is done
later by the CA to prove ownership.
2.1. Embedding algorithm. The algorithm which incorporates the watermark in text
is called embedding algorithm. The embedding algorithm embeds the watermark in the
text logically and generates a watermark key. The flowchart of proposed watermark
embedding process is shown in Figure 1.
Figure 1. Flowchart of watermark embedding algorithm
Image acquisition process acquires the watermark image (WI ), which can be a logo,
signature or fingerprint of the original copyright owner. In case of RGB image, RGB to
grayscale conversion is performed and the image is scaled to a standard size (100 100
pixels used in this algorithm). After this step, image pixel values are normalized using
min max normalization method [3] between 97 and 122 to represent small alphabetical
letters (a to z in the ASCII table). Min-max normalization equation is as follows, where
v represents old value and v represents new value:
v min
v = (newmax newmin) + newmin (1)
max min
This watermark image (WI ) is then converted to alphabetical watermark based on
ASCII values. Text acquisition process acquires the plain text and text partitions as well
as groups are formed based on user given Frequently-Used word (F-U-Word). In the next
step, double letters in each group are processed. The average occurring double letter in
each group is identified which later populates the Average Occurring Letter (AOL) list
with corresponding group numbers.
Figure 2. Flowchart of watermark extraction algorithm
The alphabetical watermark (WA ) and AOL list are collectively used to generate a key
using the key generation process. Afterwards, this key is registered with the CA along
with the watermark image, original text, current date and time.
The detailed watermark embedding algorithm is as follows:
1. Input WI , GS, F U Word and T

2. Convert WI to grascale image
3. Normalize WI between 97 and 122 using min max normalization
4. Convert WI to WA
5. Make partitions of T based on F U Word
Number of partitions = (Occurrence count of F U Word) 1
6. Make groups of text based on GS
Number of groups = Number of partitions/GS
7. Count occurrence of each double letter (aa to zz) in each group and populate
AOL list (Average occuring letter) for each group
8. For each letter{of watermark (WA) repeat step 9
key[i] = 0; where j = 1, 2, 3, . . . , wl
9. If wj AOL
key[i
+ 1] = Group Number (AOL)
key[i] = 1;
else if wj AOL key[i + 1] = (wj + k) M OD 26

where k is in Z26
where Z26 represents 26 alphabets (a z)
10. Output WK
WI : Image watermark, GS: Group size, F-U Word: Frequently Used Word, T: text file,
WA : alphabetical watermark, WK: Watermark key, AOL: Average Occuring Letter.
2.2. Extraction algorithm. The algorithm which is used to extract watermark from
the text is known as the extraction algorithm. The flowchart of the watermark extraction
algorithm is shown in Figure 2.
The extraction algorithm takes the watermark key generated by the embedding algo-
rithm as input and extracts the watermark image from the text. The algorithm is stored
with the CA that is used to resolve copyright issues, if any, at a later stage. The detailed
extraction algorithm is as follows:
1. Input WK and T
2. Read F U Word from WK and set counter = 1
3. Make partitions of T based on F U Word
Number of partitions = (Occurrence count of F-U Word)1
4. Make groups of text based on GS
Number of groups = Number of partitions/GS
5. Count occurrence of each double letter (aa to zz) in each group and populate
AOL (Average occuring letter) for each group
6. L1 = length(F U Word, GS), Keyindex = L1+1, L2 = length(WK)
7. While(Keyindex < L2) repeat 9 to 10
8. If (WK(i) equals 0)
W(I) = AOL (groupnumber), goto step 11
else
W(I) = Cipher Letter
9. I = I + 1
10. Convert W to WI
11. De-normalize WI using min max normalization
12. Output WI
WK: Watermark key, T: text file, F-U Word: Frequently Used Word, GS: Group size,
W: alphabetical watermark, AOL: Average Occuring Letter, WI : Image watermark.
In the extraction algorithm, text is first partitioned using Frequently Used Word (F U
Word) obtained from watermark key (WK). These partitions are then combined to form
groups of text as done in the embedding algorithm. In the next step, occurrence of each
double letter (aa to zz) in each group is analyzed and average occurring letter in each group
is identified. The contents of watermark (WK) are then used to obtain watermark from
the text as shown in step 8 to 10 of the extraction algorithm. The obtained watermark W
is denormalized using the reverse min max normalization formula and value of each pixel
is obtained. Hence, alphabetical watermark gets converted to image watermark (WI).
3. Experimental Results and Discussion. To evaluate the performance of the pro-

posed algorithm, we used 20 text samples from the dataset designed and used in [34].
These text samples are divided in four categories based on text length as mentioned in
Table 2. Each text category contains 5 samples. These samples have been randomly col-
lected from the Reuters corpus [24], e-books, newspaper articles, research papers, novels,
web contents, reports etc.
Table 2. Text categories [34]
Text Category Number of Sentences (NS)

1 Small Size Text (SST) < 20
2 Medium Size Text (MST) [21, 100]
3 Large Size Text (LST) [101, 1000]
4 Very Large Size Text (VLST) > 1000
Possible attack on text is termed as tampering. Tampering means random insertion

and deletion of words and sentences; to and from the text. Tampering can be made in two
ways: i) localized tampering and ii) dispersed tampering. Tampering location and volume
in text cannot be predicted. Generally, text encounters both insertion and deletion attack
at the same time. Therefore, we evaluated the proposed algorithm on combined insertion
and deletion attacks. Localized and dispersed tampering was made on all samples of the
text randomly. Insertion and deletion percentages of attacks on all samples of the text
are were kept as in [35].
Experiments were performed to evaluate the resistance of algorithm under localized
tampering attacks (LTA), dispersed tampering (DTA) attacks and in comparison with
other proposed algorithms under dispersed tampering.
The logos of FAST National University of Computer and Emerging Sciences, Islam-
abad, Pakistan and Higher Education Commission of Pakistan have been used as wa-
termark images in our experiments. These images are shown in Figure 3. The group sizes
were kept 2, 3, 10 and 100 for samples belonging to SST, MST, LST and VLST categories
respectively in all experiments.
(a) NU logo (b) HEC logo
Figure 3. Watermark images

3.1. Evaluation criteria. Experiments were conducted to analyze the eect of localized
and dispersed tampering with the text by evaluating the Normalized Hamming Distances
(NHD), Peak Signal to Noise Ratio (PSNR), Weighted Peak Signal to Noise Ratio (WP-
SNR), Mean Structural Similarity Index Measure (MSSIM) and Normalized Average Er-
ror (NAE) between original and extracted watermark images. The description of each
performance metric is as follows:
3.1.1. Normalized hamming distance. In order to evaluate the degree of closeness between
original watermark and the extracted watermark image, we selected the normalized ham-
ming distance (NHD) [5] as the similarity measure. Normalized hamming distance is
defined as:
1
NW
N HD(w, w ) = w(i) w (i) (2)
NW i=1
where, w and w are the original and extracted reference watermarks respectively, NW is
the length of the watermark, and is the exclusive-OR operator. The distance ranges
between 0 and 1. Ideally, two perceptually similar images should have a distance close to
0.
3.1.2. Peak signal to noise ratio (PSNR). Peak signal-to-noise ratio (PSNR) measures
are estimates of the quality of a distorted image compared with an original image. PSNR
gives a single number which reflects the quality of the distorted image [29]. Suppose we
have a source image L(i, j) that contains N by N pixels and a distorted image M (i, j)
where M is corrupted with random noise. Error metrics are computed on the luminance
signal only so the pixel values range between black (0) and white (255). PSNR in decibels
(dB) is computed by using:
{ }
255 255
P SN R = 10 log10 (3)
M SE
where, M SE represents mean squared error as in (4). Typical PSNR values range between
20 and 40 dB.
1
M N
M SE = (L(i, j) M (i, j))2 (4)
M N i=1 j=1
3.1.3. Weighted peak signal to noise ratio (WPSNR). The WPSNR is a modified version
of the PSNR suggested in [30]. This objective quality metric is based on the computation
of a Noise Visibility Function (NVF) [29] which depends on a texture masking function.
The values of NVF are in the range zero (for extremely textured areas) to one (for smooth
areas of an image). The WPSNR can be computed as follows:
{ }
255 255
W P SN R = 10 log10 (5)
N V F M SE
where, MSE represents mean squared error as computed in (4).
3.1.4. Structural similarity index measure (SSIM). The SSIM can be defined as the col-
lection of three important properties of an image [36]. They are Luminance measurement
l(x, y), Contrast measurement c(x, y) and Structure measurement s(x, y) of the image.
SSIM (x, y) = f (l(x, y), c(x, y), s(x, y)) (6)
that is,
(2x y + C1)(2xy + C2)
SSIM (x, y) = (7)
+ 2y + C1)(x2 + y2 + C2)
(2x
where, x , y , are the mean intensity, x , y , are the estimate of the signal contrast. C1
and C2 constants as defined as in [36]. The overall image quality is measured by the mean
SSIM (MSSIM) index defined as
1
M N
M SSIM = SSIM (i, j) (8)
M i=1 j=1
where, M is the total number of local SSIM indexes.
3.1.5. Normalized absolute error (NAE). Normalized absolute error is another important
metric to measure the amount of error in an image [17]. NAE is defined as:
M N
|L(i, j) M (i, j)|
i=1 j=1
N AE = (9)

M N
|L(i, j)|
i=1 j=1
The large value of Normalized Absolute Error (NAE) means that image is of poor quality
and small value represents high quality image.
3.2. Localized tampering attack results. Localized insertion and deletion is the most
frequently occurring form of attack on text. The malicious attacker may be interested in
inserting or removing a portion of the text, particularly in the case of legal documents.
For experiments, insertion and deletion of text was made at a single randomly selected
location in each of the text sample.
The attack location was chosen randomly and the attack volume was kept same as in
[35]. We performed experiments on each sample of small size text (SST), medium size text
(MST), large size text (LST) and very large size text (VLST). The Normalized Hamming
Distances (NHD) between original and extracted watermark images on all text samples
are shown in Table 3 with four dierent words. It can be clearly observed that distances
are nearly zero for all text samples. However, the algorithm performs better with keyword
on as compared with in, to and of.
Table 3. Normalized Hamming Distances between original watermark im-
age (W) and extracted watermark image (W ) (NU logo) after localized
tampering attack on all samples
Normalized Hamming Distance

Text category
in on to of
SST 0.0267 0.0098 0.0035 0.0417
MST 0.0577 0.0139 0.0585 0.0591
LST 0.0315 0.0133 0.0429 0.0662
VLST 0.0165 0.0090 0.0263 0.0326
Figure 4 depicts the performance of proposed algorithm with four most frequently used
words in English language. The hamming distances between original image and extracted
image is always between 0 and 0.1, which is the minimal dierence. To further observe
the impact of tampering on watermark, peak signal to noise ratio, weighted peak signal
to noise ratios, mean structural similarity index, and normalized absolute error measures
of both extracted watermark images were obtained for all text samples under localized
tampering attack. The results can be seen in Table 4.
It can be clearly observed in Table 4 that PSNR and WPSNR values are always greater
than 30 dB for all text samples. For NU logo image, the maximum value of PSNR is
Figure 4. Normalized hamming distance between original and retrieved

watermark under localized tampering attack on all text sample with dier-
ent words
Table 4. PSNR (dB), weighted PSNR (dB), mean SSIM and NAE values
for both extracted watermark images under localized tampering attack on
all text samples
NU logo HEC logo

PSNR WPSNR MSSIM NAE PSNR WPSNR MSSIM NAE
1 32.91 39.60 0.9905 0.0229 32.87 39.85 0.9862 0.0227
2 32.82 39.19 0.9893 0.0232 32.09 38.94 0.9840 0.0245
SST 3 33.04 39.34 0.9331 0.0317 33.05 37.42 0.9397 0.0344
4 32.28 28.35 0.8660 0.0229 32.87 39.85 0.9862 0.0227
5 31.92 27.68 0.7854 0.0646 31.17 27.02 0.8149 0.0707
1 31.85 29.28 0.8588 0.0517 31.50 30.67 0.8815 0.0505
2 32.21 35.38 0.9407 0.0332 31.88 34.11 0.9517 0.0362
MST 3 32.04 32.43 0.8800 0.0459 32.08 25.55 0.8673 0.0598
4 32.94 37.96 0.9773 0.0263 32.89 37.00 0.9690 0.0253
5 32.91 39.60 0.9905 0.0229 32.87 39.85 0.9862 0.0227
1 31.30 30.66 0.8687 0.0521 31.13 30.38 0.9041 0.0524
2 32.21 35.97 0.9523 0.0317 31.88 34.87 0.9597 0.0343
LST 3 32.91 39.60 0.9905 0.0229 32.87 39.85 0.9862 0.0227
4 32.28 34.52 0.9634 0.0314 32.37 36.57 0.9633 0.0292
5 31.96 33.43 0.9466 0.0371 31.96 35.25 0.9498 0.0335
1 32.94 39.30 0.9868 0.0246 32.89 39.03 0.9782 0.0240
2 32.91 39.60 0.9905 0.0229 32.89 39.03 0.9782 0.0227
VLST 3 32.30 33.93 0.9472 0.0351 32.39 34.82 0.9441 0.0321
4 32.39 34.79 0.9049 0.0402 32.53 36.03 0.9142 0.0410
5 32.91 39.60 0.9905 0.0229 32.87 39.85 0.9862 0.0227
33.04 for SST3 and WPSNR is 36.90 for SST1, MST5, LST3 and VLST2. For HEC logo
image, the maximum value of PSNR is 33.05 for SST3 and WPSNR is 39.85 for SST1,
SST4, MST5, LST3 and VLST5. The values of MSSIM are close to 1 for all text samples
as desired. The maximum values of MSSIM are 0.9905 and 0.9862 for NU and HEC
logo image respectively. Normalized absolute error of watermark images under localized
tampering attacks on all text samples is shown in Figure 5. The values are close for 0
for all text samples and minimum values of NAE for NU and HEC logo image are 0.0229
and 0.0227 respectively.
Table 5. Extracted watermark (We ) and error images (E) for some sam-
ples under attack
Sample Attack Extracted watermark (We ) Error image (E)
Localized
SST5 Insertion=25%
Deletion=13%
Localized
MST1 Insertion=26%
Deletion=11%
Dispersed
LST2 Insertion=7%
Deletion=5%
Dispersed
SST5 Insertion=35%
Deletion=39%
The extracted watermark image (We ) and error images (E) of four randomly selected
text samples under localized and dispersed tampering attacks are shown in Table 5. The
extracted algorithm successfully retrieves the watermark image for each text sample and
the original copyright owner can be recognized. The error images (E) are almost invisible
which represents very small error in the extracted image.
3.3. Dispersed tampering attack results. Dispersed tampering is a form of attack
where the attacker inserts and deletes, words and phrases, and from several places in the
text. In general, the intention is to make the text look dierent; destroying author writing
style, change meaning, and/or value of the text. To conduct experiments, insertion and
deletion of variable size text was made at multiple random locations in each of the text
samples.
The peak signal to noise ratio (PSNR), weighted peak signal to noise ratios (WPSNR)
and mean structural similarity index measures (MSSIM) of extracted NU logo and HEC
logo images for all text samples under dispersed tampering attack are shown in Table 6.
It can be clearly observed in Table 6 that PSNR and WPSNR values are always greater
than 30 dB for all text samples.
For NU logo image, the maximum value of PSNR is 32.89 and WPSNR is 39.64. The
maximum values of PSNR are 32.87 and WPSNR is 39.85 for HEC logo image. The
values of MSSIM are close to 1 for all text samples as desired. The maximum values
of MSSIM are 0.9925 and 0.9862 for NU and HEC logo image respectively. Normalized
Figure 5. Normalized absolute error of watermark images under localized

tampering attacks on all text samples
Figure 6. Normalized absolute error of watermark images under dispersed

tampering attacks on all text samples
absolute error of watermark images under dispersed tampering attacks on all text samples
is shown in Figure 6. The values are close for 0 for all text samples and the minimum
values of NAE for NU and HEC logo image are 0.0229 and 0.0227 respectively.
3.4. Comparative results. The performance of proposed algorithm is compared with a
text watermarking algorithm proposed by Lu et al. [20] (which is named here as OTMR)
Table 6. PSNR (dB), weighted PSNR (dB), mean SSIM and NAE values
for both extracted watermark images under dispersed tampering attack on
all text samples
NU logo HEC logo

PSNR WPSNR MSSIM NAE PSNR WPSNR MSSIM NAE
1 32.18 36.51 0.9680 0.0303 31.88 35.66 0.9669 0.0324
2 32.84 39.64 0.9893 0.0247 32.85 32.53 0.9662 0.0330
SST 3 32.18 34.74 0.9369 0.0346 31.88 33.40 0.9430 0.0381
4 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
5 31.88 28.29 0.8110 0.0590 32.02 24.28 0.7977 0.0747
1 32.27 34.75 0.9664 0.0314 32.37 36.57 0.9633 0.0292
2 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
MST 3 32.33 32.61 0.8948 0.0411 32.52 30.79 0.9007 0.0435
4 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
5 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
1 32.18 35.31 0.9480 0.0332 31.88 34.11 0.9517 0.0362
2 31.04 30.26 0.8487 0.0615 30.87 24.34 0.8290 0.0834
LST 3 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
4 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
5 31.42 31.19 0.9213 0.0465 31.58 32.94 0.9263 0.0410
1 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
2 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
VLST 3 32.29 34.14 0.9511 0.0351 32.39 34.82 0.9441 0.0321
4 32.89 39.62 0.9925 0.0229 32.87 39.85 0.9862 0.0227
5 33.03 39.53 0.9413 0.0317 33.05 37.42 0.9397 0.0344
and the preposition based algorithm [35]. 9 TMR samples from MIKRKOSMOS were used
for comparison as used by Lu et al. [20] in his natural language watermarking algorithm.
It was impractical to use TMR samples in their present form in our algorithm (which
requires plain text). Therefore, we used the original text sample, from which the TMR
sample was created for our experiments. All the text samples belong to Small Size Text
(SST) category. The optimum group size (2) for SST category and the NU and HEC logo
are used as watermark images.
The original TMR samples were attacked by performing dispersed insertion and dele-
tion of words and sentences. Same attacked samples were used for experiments in OTMR,
Preposition based and proposed algorithms. The embedding algorithm takes TMR sam-
ples in the original form and the extraction algorithm use attacked TMR samples. In
the OTMR algorithm, the watermark is embedded in the sentence; hence the probability
of deletion of the sentence containing watermark bits is also positive. However, in the
proposed algorithm, since the watermark image is not actually embedded, so even if text
is tampered, the watermark can still be detected.
Table 7 shows the ratios of successfully detected watermark by the proposed algorithm
with both images as compared with OTMR algorithm.
Figure 7 shows the comparative statistics of the algorithms, where results of OTMR
algorithm as well as the preposition based algorithm are inferior to the results of proposed
algorithm. The proposed algorithm has better robustness against dispersed tampering
attacks for all 9 samples. The watermark is resilient under combined insertion and deletion
attack. It is more secure since the existence of watermark remains unknown to the world
except the original author and the certifying authority. It is also computationally ecient
Table 7. Comparison of proposed algorithm with previous algorithms
Ratio of successfully detected watermark

Preposition Proposed Proposed
OTMR
Sample No. based Algorithm Algorithm
20 bit
Algorithm (NU logo) (HEC logo)
TMR0011 92% 99.7% 100% 100%
TMR0020 90% 99.9% 100% 100%
TMR0046 86% 97.8% 98.67% 99.12%
TMR0070 84% 99.0% 99.48% 99.55%
TMR0589 86% 95.1% 97.27% 96.12%
TMR0592 78% 100.0% 97.27% 96.12%
TMR2099 90% 99.1% 99.67% 99.10%
TMR2103 84% 94.2% 98.23% 98.76%
TMR2301 88% 97.4% 99.48% 99.61%
Figure 7. Performance comparison of proposed algorithm with OTMR

and Preposition based algorithms
with worst case complexity of O(n) and gives better accuracy. Also, the probability of
false alarm is very low. In the proposed algorithm since we are taking user input along
with watermark to generate key in embedding algorithm. It is nearly impossible to get
the same key for two dierent documents even with the same watermark image.
Text watermarking is one of the neglected research areas in information security domain
and there is no benchmark text available to facilitate comparisons of dierent algorithms.
It is a great obstacle towards making further comparison with other proposed approaches.
However, it can be claimed that none of the formerly proposed algorithms using image as
watermark give near zero error in extracted watermark image under dispersed tampering
attacks on text.
The proposed algorithm can be used to protect all open digital contents from forgery.
Even, the algorithm does not prevent copying, but it makes the attacker cautious about
it. It will be obvious to everyone that the original copyright holder can claim ownership
at any time and certification authority to detect tampering of copyright. In addition,
cyber-laws may be made and implemented to prevent illegal copying and counterfeiting.
4. Conclusions. In this paper, we have proposed a novel idea to embed image watermark
in text logically utilizing the inherent properties of the text. The algorithm with a blind
zero watermarking approach gives a robust solution for the text watermarking problem.
Image such as logo, signature or fingerprint is first converted in alphabets and a key
is populated using the occurrence frequencies of double letters. The image watermark
can be extracted later for copyright protection and text authentication. We tested the
performance of the algorithm for tampering; both in localized and dispersed forms on 20
texts. Normalized hamming distances, PSNR, WPSNR, MSSIM and NAE are used as
performance measures to compare original and extracted watermarks. We also compared
the performance of the algorithm with two recently proposed algorithms on two dierent
images. The results show that our algorithm is more robust, secure and ecient. In
addition, it has minimal computational requirements.
Acknowledgment. One of the authors, Zunera Jalil, 041-101673-Cu-014 would like to

acknowledge the Higher Education Commission of Pakistan for providing funding and
resources to complete this work under Indigenous Fellowship Program.
REFERENCES
[1] A. Khan, A. M. Mirza and A. Majid, Optimizing perceptual shaping of a digital watermark using
genetic programming, Iranian Journal of Electrical and Computer Engineering, vol.3, pp.144-150,
2004.
[2] B. Macq and O. Vybornova, A method of text watermarking using presuppositions, Proc. of the SPIE
International Conference on Security, Steganography, and Watermarking of Multimedia Contents,
2007.
[3] R. M. Bolle, J. H. Connell, S. Pankanti, N. K. Ratha and A. W. Senior, Guide to Biometrics,
Springer-Verlag, New York, 2004.
[4] C. Lopez, Watermarking of digital geospatial datasets: A review of technical, legal and copyright
issues, International Journal of Geographical Information Science, vol.16, no.6, pp.589-607, 2002.
[5] S. Counsell, F. Mendes, S. Swift and A. Tucker, Evaluation of an object-oriented cohesion metric
through Hamming distances, Tech. Rep., BBKCS-02-10, Birkbeck College, University of London,
UK, 2002.
[6] D. Huang and H. Yan, Interword distance changes represented by sine waves for watermarking text
images, IEEE Trans. Circuits and Systems for Video Technology, vol.11, no.12, pp.1237-1245, 2001.
[7] H.-X. Wang, Z.-M. Lu, Y.-N. Li and S.-H. Sun, A compressed domain multipurpose video wa-
termarking algorithm using vector quantization, International Journal of Innovative Computing,
[8] H. M. Meral, et al., Natural language watermarking via morphosyntactic alterations, Computer
Speech and Language, vol.23, no.1, pp.107-125, 2009.

[9] H. M. Meral, E. Sevinc, E. Unkar,
B. Sankur, A. S. Ozsoy and T. G ungor, Syntactic tools for text
watermarking, Proc. of the 19th SPIE Electronic Imaging Conference on Security, Steganography,
and Watermarking of Multimedia Contents, San Jose, CA, 2007.
[10] J. T. Brassil, S. Low, N. F. Maxemchuk and L. OGorman, Electronic marking and identification
techniques to discourage document copying, IEEE Journal on Selected Areas in Communications,
vol.13, no.8, pp.1495-1504, 1995.
[11] J. T. Brassil, S. Low and N. F. Maxemchuk, Copyright protection for the electronic distribution of
text documents, Proc. of the IEEE, vol.87, no.7, pp.1181-1196, 1999.
[12] M. J. Atallah, C. McDonough, S. Nirenburg and V. Raskin, Natural language processing for infor-
mation assurance and security: An overview and implementations, Proc. of the 9th ACM/SIGSAC
New Security Paradigms Workshop, Cork, Ireland, pp.51-65, 2000.
[13] M. J. Atallah, V. Raskin, M. C. Crogan, C. F. Hempelmann, F. Kerschbaum, D. Mohamed and
S. Naik, Natural language watermarking: Design, analysis, and a proof-of-concept implementation,
Proc. of the 4th Information Hiding Workshop, LNCS, Pittsburgh, PA, vol.2137, 2001.
[14] M. Atallah, V. Raskin, C. F. Hempelmann, M. Karahan, R. Sion, U. Topkara and K. E. Triezenberg,
Natural language watermarking and tamperproofing, Proc. of the 5th Information Hiding Workshop,
LNCS, Noordwijkerhout, The Netherlands, vol.2578, 2002.
[15] M. Topkara, C. M. Taskiran and E. Delp, Natural language watermarking, Proc. of the SPIE In-
ternational Conference on Security, Steganography, and Watermarking of Multimedia Contents VII,
2005.
[16] M. Topkara, U. Topraka and M. J. Atallah, Information hiding through errors: A confusing ap-
proach, Proc. of the SPIE International Conference on Security, Steganography, and Watermarking
of Multimedia Contents, San Jose, CA, 2007.
[17] M. Mrak, S. Grgic and M. Grgic, Picture quality measures in image compression systems, Proc. of
the Eurocon 2003 Conference, Ljubljana, Slovenia, pp.233-237, 2003.
[18] N. F. Maxemchuk ans S. H. Low, Performance comparison of two text marking methods, IEEE
Journal of Selected Areas in Communications, vol.16, no.4, pp.561-572, 1998.
[19] N. F. Maxemchuk, Electronic document distribution, AT&T Technical Journal, vol.73, no.5, pp.73-
80, 1994.
[20] N. F. Maxemchuk and S. Low, Marking text documents, Proc. of the IEEE International Conference
on Image Processing, Washington, DC, pp.13-16, 1997.
[21] P. Lu, et al., An optimized natural language watermarking algorithm based on TMR, Proc. of the
9th International Conference for Young Computer Scientists, 2009.
[22] Preposition, The Columbia Encyclopedia, 6th Edition, 2008, Encyclopedia.com, 2009.
[23] Q. Gu and T. Gao, A novel rreversible watermarking aarithm based on wavelet lifting scheme, ICIC
[24] Reuters Corpus, http://about.reuters.com/researchandstandards/corpus/index.asp.
[25] S. Zhou, Q. Zhang and X. Wei, An image scrambling method based on chaos in low intensity laser-
DNA molecule systems, ICIC Express Letters, vol.3, no.3(B), pp.645-650, 2009.
[26] S. D. Lin and Y.-H. Huang, An integrated watermarking technique with tamper detection and
recovery, International Journal of Innovative Computing, Information and Control, vol.5, no.11(B),
pp.4309-4316, 2009.
[27] S. H. Low, N. F. Maxemchuk and A. M. Lapone, Document identification for copyright protection
using centroid detection, IEEE Transactions on Communications, vol.46, no.3, pp.372-381, 1998.
[28] S. H. Low and N. F. Maxemchuk, Capacity of text marking channel, IEEE Signal Processing Letters,
vol.7, no.12, pp.345-347, 2000.
[29] S. Voloshynovskiy, A. Herrigel, N. Baumgaertner and T. Pun, A stochastic approach to content adap-
tive digital image watermarking, Proc. of the 3rd International Workshop on Information Hiding,
1999.
[30] J. Sinclair, Corpus, Concordance, Collocation, Oxford University Press, 1991.
[31] T. Amano and D. Misaki, A feature calibration method for watermarking of document images, Proc.
of ICDAR, pp.91-94, 1999.
[32] U. Topkara, M. Topkara and M. J. Atallah, The hiding virtues of ambiguity: Quantifiably resilient
watermarking of natural language text through synonym substitutions, Proc. of ACM Multimedia
and Security Conference, Geneva, 2006.
[33] X. Sun and A. J. Asiimwe, Noun-verb based technique of text watermarking using recursive decent
semantic net parsers, Lecture Notes in Computer Science, vol.3612, pp.958-961, 2005.
[34] Z. Jalil and A. M. Mirza, An invisible text watermarking algorithm using image watermark, Inno-
vations in Computing Sciences and Software Engineering, Springer, ISBN: 978-90-481-9111-6.
[35] Z. Jalil and A. M. Mirza, A robust zero-watermarking algorithm for copyright protection of text
documents, Journal of the Chinese Institute of Engineers, Accepted for publication.
[36] Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli, Image quality assessment: From error
visibility to structural similarity, IEEE Transactions on Image Processing, vol.13, no.4, pp.600-612,
2004.
c ISSN 1349-4198
AN EFFICIENT AND SECURE THREE-PASS AUTHENTICATED KEY

AGREEMENT ELLIPTIC CURVE BASED PROTOCOL
Zeyad Mohammad1 , Chien-Lung Hsu2 , Yaw-Chung Chen3 and Chi-Chun Lo1

1
Institute of Information Management
3
National Chiao Tung University
No. 1001, University Road, Hsinchu 300, Taiwan
zeyad.cs94g@nctu.edu.tw; cclo@faculty.nctu.edu.tw; ycchen@cs.nctu.edu.tw
2
Chang-Gung University
No. 259, Wen-Hwa 1st Road, Kwei-Shan, Taoyuan 333, Taiwan
clhsu@mail.cgu.edu.tw
Abstract. Key agreement protocols are a fundamental building block of cryptography

to establish a common secret key over public network. We propose an ecient and secure
three-pass authenticated key agreement protocol based on elliptic curve where three-pass
protocols have significant advantages over two-pass in terms of security properties and
applications. The three-pass protocols can prevent denial of service attacks in complex
and unpredictable communication environments such as wireless networks and Internet.
We show the proposed protocol can withstand a stronger adversary under eCK security
model by using a trick in its block of hashing a static secret key with an ephemeral secret
key. Furthermore, it can provide an assurance of the identity authentication of its part-
ner, thus it can withstand non-repudiation attacks. Therefore, it is suitable for electronic
commerce to provide non-repudiation services. By comparing the security and computa-
tional complexity of the proposed protocol with other existing protocols in our study, we
show that the proposed protocol not only satisfies all security attributes but also obtains
computational eciency with a cost of 3 point multiplications.
Keywords: Cryptography, Key agreement, Elliptic curve cryptosystem, Extended Cane-
tti-Krawczyk
1. Introduction. With the rapid development of Internet-based applications which are

using wireless communications, there are diverse communication devices used in daily life
to transform the trends and style of traditional environment into a technology based one.
Due to the exposure of transmitted data over wireless mediums, several security issues
and requirements must be considered in the developments [9,34]. Therefore, a secret key
distribution and user authentication become the most important security services for open
networks. Nowadays, secure and ecient user authentication and key exchange schemes
are important for successful Internet based commerce [10,12,24,28,29,34,35].
A key agreement protocol is a fundamental building block of cryptography over an
open network. In two-party authenticated key agreement asymmetric scheme based, the
two parties are exchanging static and ephemeral public keys between them, thereafter,
they combine their secret keys with public keys to compute a common secret key. The
common secret key can provide data privacy, data integrity, non-repudiation and entity
authentication over an open communication channel.
1273
1274 Z. MOHAMMAD, C.-L. HSU, Y.-C. CHEN AND C.-C. LO
A key agreement protocol providing implicit key authentication to both participating

parties is called an Authenticated Key Agreement (AKA) protocol which involves two
or one message exchanges, respectively. A key agreement protocol providing explicit key
authentication to both participating parties is called AKA with key Confirmation (AKC)
protocol, which involves three-message exchange [22]. Normally, three-pass AKA proto-
cols are extended from the original two-pass AKA protocols where message authentication
algorithms and two independent hash functions in the protocol block are used to provide
AKC. Since these protocols cannot support an assurance of the entity authentication of
its partner, they cannot withstand non-repudiation attacks. However, non-repudiation is
an important issue in electronic commerce to support billing of non-repudiation services
[10].
Canetti-Krawczyk [6-8,16] showed that the AKC protocol has significant advantage over
AKA protocol in terms of security properties and applications. Since the AKC protocol
can provide key confirmation, Universally Composable (UC) security and full Perfect
Forward Secrecy Resilience (fPFS-R), they are leaked in AKA protocol. UC security can
ensure the security of cryptographic protocol in complex and unpredictable environments
such as Internet and wireless communication. In many applications of key agreement,
lack of key confirmation may lead to a form of Denial of Service (DoS) attack in which
the initiator party starts using the session key to send protected information to responder
party, while the responder party is unable to process this information because it has not
yet established the session key. A DoS attack or Distributed DoS (DDoS) attack is an
attempt to disrupt, destroy or render services unavailable. Nowadays, reliable, fast and
secure communication is essential for commercial success on the internet. Unavailable,
slow and insecure web pages could motivate clients to switch to an alternative business,
leading to a loss in customer base for the service provider [29].
In an Authenticated Key Exchange (AKE) protocol, two parties exchange information
and compute a secret key, whereby they use a function that has at least four secret
information parameters. LaMacchia et al. presented a strong security definition for AKE
protocol by introducing the idea of hashing a static private key with an ephemeral private
key to compute an ephemeral public key that is extended the well-known CK [6] (eCK)
definition [17]. This idea prevents an adversary from extracting the discrete logarithm of
the ephemeral public key unless both leakage of the static and ephemeral secret keys.
Cryptosystems in elliptic curve groups have a significant potential advantage over tradi-
tional cryptosystems in Galois finite field, since they have higher security, fewer key length
and transmission overheads. An elliptic curve cryptosystem has a lot of advantages over
other cryptosystems in finite field in key size and computational complexity since 224
bits have the same crack years of 2048 bits key size of RSA [19,21,27]. A key agreement
protocol based on Elliptic Curve Die-Hellman [11] (ECDH) is suitable for those devices
that have limited computation, power, memory capacity and network bandwidth. Many
protocols have been designed with elliptic curve implementation. Law et al. proposed
MQV protocols (ECMQV) based on ECDH as an ecient AKA protocols [18]. ECMQV
protocols are standardized in [1,2,13]. Kaliski discovered that two-pass ECMQV protocol
is vulnerable to an Unknown Key Share Resilience (UKS-R) attack [14].
Starngio proposed an ecient elliptic curve Die-Hellman AKA protocol (ECKE-1)
[30]. However, Wang et al. improved ECKE-1 protocol (ECKE-1N) to overcome Key
Compromise Impersonation Resilience (KCI-R) attacks in ECKE-1 [33]. Strangio pro-
posed a three-pass ECKE-1R as a revised version of his original protocol to overcome
KCI-R attack [31]. Moreover, Mohammad and Lo [20] showed that ECKE-1N is vul-
nerable to both leakage of ephemeral private keys and KCI-R attacks under stronger
THREE-PASS AUTHENTICATED KEY AGREEMENT ELLIPTIC CURVE 1275
adversary in eCK security model. They enhanced ECKE-1N (EECKE-1N) which can
withstand KCI-R and PFS-R attacks under eCK [24].
Popescu proposed an ecient and secure AKA protocol based on the elliptic curve
Die-Hellman [25,26], Yoon and Yoo improved Popescus protocol to overcome KCI-
R attacks, reflection attacks and replay attacks [36]. However, the Yoon-Yoo protocol
cannot cope with a superior adversary and has higher computational complexity than
Popescus protocol. Therefore, Lim et al. enhanced Popescus protocol version claiming
their protocol provided stronger and more ecient security features than the Yoon-Yoo
and Popescus protocols [20].
The design of a secure and ecient key agreement protocol is to solve an optimization
problem and minimize the computational costs and the numbers of messages transmitted
during a run of protocol as well as maximizing the security attributes. Both eciency
and security are a fundamental in designing a new protocol. The eciency of a designed
protocol makes the protocol suitable to diverse devices in their capabilities.
The aim of this paper is to propose an ecient and secure three-pass AKA protocol
(3P-AKAP) based on ECDH. The 3P-AKAP uses a trick of hashing the ephemeral se-
cret key with the static secret key that it makes the proposed protocol can withstand
stronger adversary attacks. Furthermore, our proposed protocol can provide an assurance
of the identity authentication of its partner and can withstand non-repudiation attacks.
The proposed 3P-AKAP is ecient due to the fact that each party performs one point
multiplication oine and two point multiplications online.
The rest of the paper is organized as follows: In Section 2, we propose the 3P-AKAP,
and we analyze its security. In Section 3, we present the comparisons the 3P-AKAP with
known existing protocols in terms of security attributes and computational complexity.
We present our conclusion in the last section.
2. The Proposed Three-Pass Authenticated Key Agreement Protocol. The no-

tation of Elliptic Curve Cryptography (ECC) was first introduced independently by Miller
[23] and Koblitz [15]. The domain parameters of ECC are ECC = (q, F R, S, a, b, P, n, h),
where q is a prime power (in practice, either q = p, an odd prime, or q = 2m ), F R a valid
field representation, S a seed for randomly generated elliptic curve, and two field elements
a, b Fq , which define the equation of the elliptic curve E over Fq (i.e, y 2 = x3 + ax + b in
the case b > 3, where 4a3 + 27b2 = 0). The base point P (P.x, P.y) satisfies the defin-
ing equation and verifies nP is the point at infinity, n is a large prime, and the cofactor
h = #E(Fq )/n (where #E(Fq ) denotes the number of points in the elliptic curve E(Fq ))
[18,22,33].
2.1. Notation and assumptions. Notation and assumptions used in this paper are
summarized as follows.
A/B: Two communicating stations, initiator A and responder B.
M: Malicious (adversary) station. M has full control of communications
on the open network, such that M is capable of intercepting, deleting,
injecting, altering and replaying messages in any online instance of the
key agreement protocol.
IDA , IDB : Identities of two communicating parties.
P: The generating point of ECC large prime order in E(Fq ).
P : A cyclic subgroup of P .
P : The set of nonidentity elements P .
P : The identity point in E(Fq ).
wA , wB : Static private keys of A and B, where wA , wB R [1, n 1].
WA , WB : Static public keys of A and B, where WA = wa P and WB =

wb P . The Certificate Authority (CA) should verify that the
static public key WA of a party A belongs to P as defined
in [15].
rA , rB : Ephemeral private keys of A and B.
RA , RB : Ephemeral public keys of A and B.
H1 , H2 : {0, 1} F|q| : A collision of resistant hash functions which the output
length is |q| bits.
H : {0, 1} {0, 1} : Hash functions of modeled random oracle, where is a se-
curity parameter.
k: The computed ephemeral session key by two-party.
kAB : The derived session key by two-party.
Assumption 2.1. Let E be an elliptic curve defined over a finite field Fq and P be a
generator with prime order n. Given R E(Fq ). The Elliptic Curve Discrete Logarithm
Problem (ECDLP) is hard to find the integer r [1, n1] such that R = rP in polynomial
time.
Assumption 2.2. Let E be an elliptic curve defined over a finite field Fq and let P
E(Fq ) be a point of prime order n. Given two points S E(Fq ) and T E(Fq ), the
Elliptic Curve Computational Die-Hellman Problem (ECCDHP) is hard to find stP
from the two given points, where s = ECDLP (S) and t = ECDLP (T ) in polynomial
time.
The domain parameters ECC should be suitably selected so that no polynomial al-
gorithms exists that can solve ECDLP or ECCDHP in the subgroup P . The domain
parameters must be validated as defined in [18].
The security of the 3P-AKAP depends on the ECDLP and ECCDHP. In addition, we
assume digital certificates (denoted by certA and certB respectively) are issued by CA.
2.2. Protocol description. The 3P-AKAP is based on ECDH which verifies the session
identifier of two participating intended parties. The session identifier verification can pro-
vide entity authentication to key agreement protocol as shown by Bellare and Rogaway
(BR) [3,4] security model for public key two-entity AKA setting. Moreover, the 3P-AKAP
hashes the static private key with the ephemeral private key to generate the ephemeral
public key that makes our proposed protocol can withstand a revelation of secret infor-
mation from two-party participants according to eCK security model. Furthermore, the
proposed protocol contains the verification steps in its block to verify partner identity
from computing its static public key. In addition, the 3P-AKAP uses the hash function
to derive the session key to thwart the guessing session key attacks, key replicating attacks
and UKS-R attacks as defined in [32].
The 3P-AKAP is depicted in Figure 1 which consists of five steps.
Step 1: Initially, A performs the following operations to establish a session key with B.
Select a random number rA R [1, n 1], where RA = H1 (rA wA )P .
Send (IDB , IDA , RA ) to B.
Step 2: Upon receiving As message, B performs the following operations:
Verify RA P by using an embedded key validation of RA . If RA P
then B terminates the protocol run with a failure output.
Select a random number rB R [1, n 1], where RB = H1 (rB wB )P .
Compute: k = hH1 (rB wB )RA = hH1 (rA wA )H1 (rB wB )P , eB = H2 (ID A
IDB xRA xRB xk ), cB = H1 (rB wB ) + wB mod q and ZB = cB eB ,
A(wA ) Public information: B(wB )

P , WA , WB
rA R [1, n 1]
RA = H1 (rA wA )P
(RA ) -
rB R [1, n 1]
RB = H1 (rB wB )P
k = hH1 (rB wB )RA
eB = H2 (IDA IDB xRA xRB xk )
cB = H1 (rB wB ) + wB mod q

ZB = c B eB
(RB , ZB )
k = hH1 (rA wA )RB
eB = H2 (IDA IDBxRA xRB xk )
cB = ZB eB
?
WB = cB P RB
eA = H2 (IDA IDB yRB yRA xk )
cA = H1 (rA wA )
+ wA mod q
zA = cA eA
(ZA ) -
eA = H2 (IDA IDB yRB yRA xk )

c A = ZA eA
?
WA = c A P R A
kAB = H(IDA IDB xRA xRB cA cB xK ) kAB = H(IDA IDB xRA xRB cA cB xK )
Figure 1. The proposed of an ecient and secure authenticated key agree-

ment elliptic curve based protocol
where xRA denotes the x-coordinate of RA , xRB denotes the x-coordinate of

RB , xk denotes the x-coordinate of k and h denotes the cofactor.
Send (IDA , IDB , RA , RB , ZB ) to A.
Destroys rB and eB .
Step 3: Upon receiving Bs message, A performs the following operations:
Verify the tuple session identifier of (IDA , IDB , RA , RB , ZB ) with his own
the tuple session identifier of (IDA , IDB , RA ). If they dont match, then A
terminates the protocol run with a failure output.
Verify RB P by using an embedded key validation of RB . If RB P
then A terminates the protocol run with a failure output.
Compute k = hH1 (rA wA )RB = hH 1 (rA wA )H wB )P , eB = H2 (IDA
1 (rB

IDB xRA xRB xk ) and cB = ZB eB = cB eB eB .
Verify whether WB = cB P RB holds. If the verification fails, A terminates
the execution; otherwise, A proceeds to compute eA = H2 (IDA IDB yRB yRA
xk ), cA = H1 (rA wA ) + wA mod q and ZA = cA eA , where yRA denotes
the y-coordinate of RA , xk denotes the x-coordinate of k, yRB denotes the
y-coordinate of RB .
Send (IDB , IDA , RA , RB , ZB , ZA ) to B.
Destroy rA , eA and eB .
Step 4: B performs the following operations:
Verify (IDB , IDA , RA , RB , ZB , ZA ) with (IDB , IDA , RA , RB , ZB ). If they
dont match then B terminates the protocol run with a failure output.

Compute eA = H2 (IDA IDB yRB yRA xk ) and cA = ZA eA = cA eA
eA .
Verify whether WA = cA P RA holds. If the verification fails, B terminates
the execution. Otherwise, B can be assured of As identity.
Destroy eA .
Step 5: Both A and B terminate and agree on the common session key which is computed
as follows: kAB = H(IDA IDB xRA xRB cA cB xk ) then they destroy cA , cB
and k.
Correctness: The correctness of the protocol is derived from the following equation
k = hH1 (rA wA )H1 (rB wB )P ; in this case, only the honest parties A and B will compute
the same session key from the elliptic curve. The scalar multiplication using the cofactor
h prevents the small subgroup attacks and invalid elliptic curve equation [18].
2.3. Security analysis.

Theorem 2.1. If the verification equation WB = (ZB eB )P RB = wB P holds. A
may verify the message (RB , ZB ) sent from B.

Proof: Since cB = ZB eB = (cB eB ) eB = cB 0 = H1 (rB wB ) + wB mod q;

hence eB = eB , in this case, both the initiator and responder parties are able to compute
k because eB = H2 (IDA IDB xRA xRB xk ) depends on its value. We have the following:
WB = cB P RB = (H1 (rB wB ) + wB mod q)P H1 (rB wB )P
WB = H1 (rB wB )P + wB P H1 (rB wB )P = wB P
Hence, the 3P-AKAP is a symmetric protocol that means the responder B will verify
initiators static public key from the message (ZA ) sent by the initiator party A.
Theorem 2.2. The 3P-AKAP provides the following security attributes: fPFS-R, Key
Control Resilience (KC-R), Known Key Security Resilience (KKS-R), UKS-R, KCI-R,
ReFlection Resilience (RF-R), Replay Resilience (R-R), Key Replicating Resilience (KR-
R).
Proof:
1- fPFS-R: Assume one of the following pairs of keys: (wA , wB ), (rA , rB ), (wA , rB )
or (wB , rA ) is compromised at one time and is known to Eavesdropper (E). Since the
ephemeral session key equation is k = hH1 (rA wA )H1 (rB wB )P which contains the four
secret keys from the two intended parties, E cant compute the previous k. Thus, the
secret information needed to compute k is (wA , rA ) or (wB , rB ) randomly. Hence, H1 is a
random oracle, (RA , WA , wA ) and (H1 (rA wA )P , WA , wA ) are indistinguishable. If E is
able to distinguish (RA , WA , wA ) from (H1 (rA wA )P , WA , wA ) in a random oracle model
then E must have queried H1 (rA wA ), thus E must have computed wA = ECDLP (WA ).
In the same way, (RB , WB , wB ) and (H1 (rB wB )P , WB , wB ) are indistinguishable. If
E is able to distinguish (RB , WB , wB ) from (H1 (rB wB )P , WB , wB ) then E must have
queried H1 (rB wB ), thus E must have computed wB = ECDLP (WB ). However, ECDLP
is an intractable problem.
In case E tries to construct the ephemeral session equation hH1 (rA wA )H1 (rB wB )P
from transmitted messages (RA , RB ), he needs to find ECCDHP (RA , RB ). However,
ECCDHP is an intractable problem. The CK security model not showing a two-pass
AKA protocol can provide fPFS-R. Instead, a two-pass can provide weak PFS-R. The
completed session in the 3P-AKAP already provides assurance that the ECDH values
were chosen and explicitly authenticated by the sessions responder (B); in this case,
the session key is called a clean session. On the other hand, if an uncorrupted party A
establishes the 3P-AKAP session with uncorrupted responder B, then the session key
kAB remains secure even though the adversary corrupts A after kAB expires at A, or it
corrupts B after kAB expires at B [7,8,16]. Therefore, the 3P-AKAP can satisfy fPFS-R.
2- KC-R: Since k = hH1 (rA wA )H1 (rB wB )P , it ensures the computed session key is
generated using input from the two intended participating parties executing the protocol,
confirming that no single party will be able to force the session key to be a preselected
value.
3- KKS-R: From the messages exchange during a run of the 3P-AKAP, it is trivial the
3P-AKAP satisfies KKS-R. Since k has the two ephemeral secret keys in its equation, it
ensures the computed session key is unique for each of the established session keys and
also the session key is dependent on the ephemeral and static secret information which is
generated by the two intended participating parties. The uniqueness of the session key
insures that compromise of a session key does not expose keys from other sessions.
4- UKS-R: Assume that C makes A believed that it is sharing a secret with C = B,
while B believes that the session key is shared with A. To do this, (a) C needs to obtain
the static private key of B (to impersonate B to A) from Bs static public key WB . (b)
C should obtain the value of H1 (rB wB ) between them using this message RB . (c) Once
the session key is obtained, C should establish a session key with B by forcing the session
key to the pre-selected value k. Tasks (a) and (b) are equivalent to solving an instance
of ECDLP as explained in proving fPFS-R and KKS-R. Task (c) is not possible because
of KC-R. Moreover, the verification steps in second and third messages flow confirm that
the party produced the verification equation because it contains the partys static private
key.
Furthermore, the session key derivation includes identities of two intended parties that
prevent UKS-R attacks and also includes RA , RB , cA and cB in kAB assures non-matching
sessions between two parties A and B that have dierent session keys [6,32].
5- KCI-R: From the verification steps in the block of 3P-AKAP and the construction
of k; the 3P-AKAP is shown to withstand KCI-R in the following cases:
Assume that the static private key (wA ) of A is compromised and it is known to
M . Clearly, M can impersonate A to the other entities using wA . However, to
impersonate any other entity (B) to A during the execution of the 3P-AKAP, M
would need the static secret key (wB ) of B to compute cB . It is clear that M needs
to solve wB = ECDLP (WB ). On the other hand, M cannot be able to compute
H1 (rA wA ) = ECDLP (RA ) by knowledge of wA ; hence M needs rA that is randomly
generated by A.
Assume that the ephemeral private key (rA ) of A is compromised and it is known to
M . Clearly, M can impersonate A to other entities using rA . However, to imperson-
ate any other entity (B) to A during the execution of the 3P-AKAP, M would need
the static secret key (wA ) of A. It is clear that M needs to solve wA = ECDLP (WA );
Hence M needs to compute H1 (rA wA ).
6- RF-R: The 3P-AKAP can detect a party by establishing a session key with itself since
the 3P-AKAP has three-pass in its block. The initiator party is able to verify its partner
identity in the second pass transmitted message (RB , ZB ). In similar way, the responder
can verify the initiator identity from the third message flow (ZA ). It is straightforward
that Theorem 2.1 showed the key reflection resilience because the verification equation
(ZA /ZB ) contains the private static key of the initiator/responder party, which provides
a signature of that party.
7- R-R: A replay attack occurs in any protocol run when an adversary copies a stream
of messages between two parties and replays the stream to one or more parties that it
is purposed to deceive a legitimate participant party. However, our assumption in the

Section 2.1 that the prime order n of point P is arbitrarily large such that the probability
of a protocol party selecting the same ephemeral key rA , rB R [1, n 1] in two dierent
sessions is negligible. Suppose the adversary masquerades as A by replaying As first old
message RA from a previous protocol run between A and B. After B has sent its a fresh
RB , ZB in the second message flow, the adversary would abort since it could not produce
ZA corresponding to RB . On the other hand, the adversary cannot masquerade as B to
A by using Bs old message; hence 3P-AKAP is a symmetric protocol. Therefore, the
3P-AKAP can withstand replay attacks.
8- KR-R: In this attack, an adversary forces the establishment of a session (other than
the test session or its matching session) that has the same key as the test session [16].
However, the session key derivation of the 3P-AKAP is kAB = H(IDA IDB xRA
xRB cA cB xk ) that contains the computed ephemeral session key, public ephemeral
keys, identities of the two participating intended parties and two decrypted verification
equations; therefore, the KR-R attacks are thwarted, hence their probability is negligible
[32]. On the other hand, The verification steps in the 3P-AKAP can provide message
integrity of RA and RB because the two parties have been safe guaranteed by having
each party to calculate the hash function of eA and eB which is used to decrypt ZA and
ZB , respectively. Therefore, the adversary cannot be able to force the establishment of
non-matching sessions to possess a common session key.
Claim 2.1. The 3P-AKAP can withstand static private key leakage to malicious party.
Proof: Assume Alice is willing to obtain Bobs static private key; therefore the 3P-
AKAP is shown to withstand that attack as follows:
1- Alice selects wA [1, n1] and gets a certificate for her public static key WA = wA P
as legitimate party through CA; hence it allows in eCK security model [17,32].
2- She selects rA [1, n 1], computes RA = rA P and sends (RA ) to Bob.
3- Upon receiving (RA ), Bob computes k = hH1 (rB wB )RA = hrA H1 (rB wB )P , eB =
H2 (IDA IDB xRA xRB xk ), cB = H1 (rB wB ) + wB mod q and ZB = cB eB
while Bob is computing the session key, Alice reveals an ephemeral private key of
Bob (rB ).
4- On receiving Bobs message (RB , ZB ), Alice cannot obtain Bobs static private key
from computing the following: k = hrA RB = hrA H1 (rB wB )P , eB = H2 (IDA IDB
xRA xRB xk ) then Alice decipher the cipher message (ZB ) as follows:

cB = ZB eB = cB eB eB = cB 0 = cB
cB = H1 (rB wB ) + wB mod q;
hence, Alice needs to know H1 (rB wB ) to extract the static key of B.
However, (RB , WB , wB ) and (H1 (rB wB )P , WB , wB ) are indistinguishable because
H1 is modeled as a random oracle. If Alice is able to distinguish (RB , WB , rB ) from
(H1 (rB wB )P , WB , rB ) then she must have queried H1 (rB wB ), thus she must have
computed wB = ECDLP (WB ). However, ECDLP is an intractable problem.
Therefore, 3P-AKAP can prevent a malicious party from recovering the victims static
private key under strong adversary revelation attacks.
2.4. Computational complexity analysis. From the run of the 3P-AKAP it is clear
that each party needs to generate ephemeral public key which takes one point multiplica-
tion oine and computes the session key from the following equation k = hH1 (rA wA )RB .
This equation takes 1 point multiplication and to verify identity takes another one point
multiplication. Thus the total number of point multiplication is 3. However, the total time
cost of point multiplication is important for communication devices having constraints in

computation, power and memory such as PDA and mobile station. Since, the protocols
use a speeding up algorithm to reduce the total time cost that occupy more memory than
protocols do not use it, thus it may not be suitable for devices having limited memory.
In addition the 3P-AKAP computes six hash functions and two exclusive OR (XOR) op-
erations. Hence, the time complexity of hash functions and XOR operations is negligible
when compared with point multiplications in public key cryptosystems [12,35].
3. Security and Computational Complexity Comparisons. Table 1 shows the se-

curity attributes for each protocol in this paper to identify the flaws of the protocols. The
first column presents KCI-R, the second column shows UKS-R, the third column displays
KF-R, the fourth column identifies R-R, the fifth column indicates whether the protocol
can withstand the ephemeral and the static key, one of each from the two dierent par-
ties is exposed (ES-R), and the last column insights in detail whether the protocol can
withstand both leakages of ephemeral secret keys (EE-R).
Table 1. Security attributes comparison for various protocols
Protocol KCI-R UKS-R KF-R R-R ES-R EE-R

Popescu [20, 36] No Yes No No No No
ECMQV [14] Yes No No No No No
ECKE-1 [33] No Yes No No Yes Yes
ECKE-1N [24] No Yes No No Yes No
EECKE-1N [24] Yes Yes No No Yes Yes
ECKE-1R [20,31] Yes Yes Yes Yes Yes Yes
Yoon-Yoo [20] Yes Yes Yes Yes Yes No
Lim et al. [20] Yes Yes Yes Yes Yes Yes
3P-AKAP Yes Yes Yes Yes Yes Yes
We can see from Table 1 that Popescus protocol does not provide KCI-R, KF-R, R-R,
EE-R and ES-R [36], ECMQV protocol has only KCI-R from the security attributes listed
in Table 1 [14]. A protocol of ECKE-1 cannot withstand KCI-R, KF-R and R-R attacks
[20,33]. ECKE-1N protocol is vulnerable under stronger adversary revelation attacks if an
adversary reveals the ephemeral secret key of the initiator party and the test session does
not have a matching session; then an adversary is able to lunch KCI-R attacks in this
protocol. In addition, ECKE-1N cannot withstand EE-R attacks in a completed session
[24]. ECKE-1N and EECKE-1N protocols cannot withstand R-R and KF-R attacks.
However, all of the two-pass AKA protocols in Table 1 cannot withstand KF-R because
these protocols dont have a verification mechanism of its partner identity in their block
design. Table 1 shows that Yoon-Yoo protocol does not have EE-R [20].
In addition to other security attributes of AKA such as KC-R, KKS-R and PFS-R
[5,22], all protocols in Table 1 can provide KC-R. Nevertheless KKS-R and PFS-R have
dierent degree of security in each protocol such as the case of analyzing an ephemeral
session key equation when disclosing both ephemeral secret keys. This is significant to
obtain the shared static key between the two participating parties, thus, it can subvert the
KKS-R between these two parties. Furthermore, Lim et al.s protocol cannot withstand
KKS-R if an adversary reveals both ephemeral secret keys plus the ephemeral session key
equation since an adversary can compute the shared static key between the two-party par-
ticipants. The ephemeral session key equation of 3P-AKAP is H1 (rA wA )H1 (rB wB )P ;
totally ephemeral randomly, therefore; the 3P-AKAP can provide KKS-R and PFS-R
without any impacts in establishing a session key in the future or expose a previous ses-
sion key under eCK. We can observe from Table 1 that the ECKE-1R and 3P-AKAP can
provide all the security attributes.
We showed in Theorem 2.1 that the 3P-AKAP can provide an assurance of the identity
authentication of its partner therefore it can withstand non-repudiation attacks. Non-
repudiation is an important issue to the service provider and customer in internet-based
applications and it needs a strong security protocol. The three-pass AKA protocol can
provide key confirmation, UC security and fPFS-R security attributes, therefore, our pro-
posed protocol can withstand DoS attacks and other strong adversary attacks, thus, the
3P-AKAP is suitable for diverse environments such as internet and wireless communica-
tion. However, the protocols of Popescu, ECMQV, ECKE-1, ECKE-1N and EECKE-1N
do not have AKC, UC security and fPFS-R security properties. The extended variants
protocols cannot withstand non-repudiation attacks because they do not have a verifica-
tion mechanism of its partner identity in their block design.
Table 2 shows the communication and computational complexities required by each
party for protocol in our study. First column counts the number of messages transmitted
during a protocol run, the second column counts the number of point multiplications while
column three displays the number of field inversions. Fourth column shows hash function
calculations, and the last column enumerates the number of field multiplications.
Table 2. Communication and computational complexity comparisons for

various protocols
Protocol Pass PM FI H FM
Propescu [25,26] 2 3 0 0 1
ECMQV [18] 2 2.5 0 1 1
ECKE-1 [30] 2 3 0 3 4
ECKE-1N [33] 2 2.5 1 3 2
EECKE-1N [24] 2 2.5 1 4 2
ECKE-1R [31] 3 4 0 3 3
Yoon-Yoo [36] 3 6 2 3 5
Lim et al. [20] 3 4 0 3 3
3P-AKAP 3 3 0 6 3
We can observe from Table 2 that Popescu, ECMQV, ECKE-1, ECKE-1N and EECKE-
1N protocols have less communication cost than ECKE-1R, Yoon-Yoo, Lim et al. and
the 3P-AKAP protocols. However, the two-pass AKA protocols lack in providing key
confirmation, UC security and fPFS-R when compared to three-pass AKA protocols.
We can see from Table 2, the computational cost of Popescu, ECKE-1 and 3P-AKAP
are 3 point multiplications but Popescu and ECKE-1 have security flaws. Moreover,
ECKE-1N and EECKE-1N perform 2.5 point multiplications and 1 field inversion, there-
fore, they have higher computational costs than 3P-AKAP. The 3P-AKAP has slightly
higher computational cost than ECMQV; however, ECMQV cannot withstand stronger
revelation adversary attacks since the three-pass of ECMQV protocol cannot provide an
assurance of identity authentication of its partner and this property is important in in-
ternet and network usage. Moreover, the protocols of ECKE-1R, Yoon-Yoo and Lim et
al. have higher computational complexity than the 3P-AKAP.
Table 2 shows the eciency of the 3P-AKAP due to its computational cost of 3 point
multiplications. In addition, the proposed protocol has higher hash function calculations
than the other three-pass AKA protocols examined in our study. However, the time
complexity of hash function compared to point multiplications in public key cryptography

is negligible [35]. The eciency of the 3P-AKAP has a significant potential advantage in
diverse communication devices used in dierent applications.
4. Conclusions. The proposed protocol is based on ECDH where ECDH has a significant
potential advantage over DSA-type groups in wireless communications. Therefore, ECDH
allows our proposed protocol to use a small key size, which saves network resources such
as bandwidth and reduces the computation complexity; hence, the validation for verifying
points on elliptic curves of prime order is free. Due to eciency of the proposed protocol,
we claim it is suitable for mobile and PDA communication devices.
Moreover, we have proved that the proposed protocol meet all necessary security at-
tributes and we claim the proposed protocol can withstand stronger adversary revelation
attacks as defined in the eCK security model, showing that our proposed protocol provides
a set of desirable security properties which arms the proposed protocol can fit complex
communication environment such as internet and wireless network.
We also have thoroughly investigated security attributes for other protocols mentioned
in our study in addition to our proposed protocol.
Acknowledgement. We would like to thank anonymous referees for their valuable sug-
gestions. This work was supported in part by the Chang Gung University Grants UARPD
390111 and UARPD390121, Chang Gung Memorial Hospital Grant CMRPD390031, and
in part by National Science Council under the grants NSC 98-2410-H-182-007-MY2.
REFERENCES
[1] ANSI X9.42, Public Key Cryptography for the Financial Services Industry: Agreement of Symmetric
Keys Using Discrete Logarithm Cryptography, American National Standards Institute, 2003.
[2] ANSI X9.63, Public Key Cryptography for the Financial Services Industry: Key Agreement and Key
Transport Using Elliptic Curve Cryptography, American National Standards Institute, 2001.
[3] M. Bellare and A. Palacio, The knowledge-of-exponent assumptions and 3-round zero-knowledge
protocols, CRYPTO2004, LNCS, vol.3152, pp.273-289, 2004.
[4] M. Bellare and P. Rogaway, Entity authentication and key distribution, CRYPTO1993, LNCS,
vol.773, pp.232-249, 1994.
[5] C. Boyd and A. Mathuria, Protocols for Authentication and Key Establishment, Springer, 2003.
[6] R. Canetti and H. Krawczyk, Analysis of key-exchange protocols and their use for building secure
channels, EUROCRYPT2001, LNCS, vol.2045, pp.453-474, 2001.
[7] R. Canetti and H. Krawczyk, Universally composable notions of key exchange and secure channels,
EUROCRYPT2002, LNCS, vol.2332, pp.337-351, 2002.
[8] R. Canetti, Universally composable security: A new paradigm for cryptographic protocol, Proc. of
FOCS2001, pp.136-145, 2001.
[9] Y.-F. Chang, A practical three-party key exchange protocol with round eciency, International
[10] T.-H. Chen, An authentication protocol with billing non-repudiation to personal communication sys-
tems, International Journal of Innovative Computing, Information and Control, vol.5, no.9, pp.2657-
2664, 2009.
[11] W. Die and M. E. Hellman, New directions in cryptography, IEEE Transaction Information The-
ory, vol.22, no.6, pp.644-654, 1976.
[12] H.-F. Huang and W.-C. Wei, A new ecient and complete remote user authentication protocol with
smart cards, International Journal of Innovative Computing, Information and Control, vol.4, no.11,
pp.2803-2808, 2008.
[13] IEEE Std 1363-2000: Standard Specifications for Public-Key Cryptography, 2000.
[14] B. Kaliski, An unknown key-share attack on the MQV key agreement protocol, ACM Transaction
on Information and Systems Security, vol.4, no.3, pp.275-288, 2001.
[15] N. Koblitz, Elliptic curve cryptosystems, Mathematics of Computation, vol.48, pp.203-209, 1987.
[16] H. Krawczyk, HMQV in IEEE P1363, submission to the IEEE P1363 working group, http://grouper.
ieee.org/groups/1363/P1363-Rearm/submission/krawczyk-hmqv-spec.pdf, 2006.
[17] B. LaMacchia, K. Lauter and A. Mityagin, Stronger security of authenticated key exchange, Proc.
of the 1st International Conference on Provable Security, LNCS, vol.4784, pp.1-16, 2007.
[18] L. Law, A. Menezes, M. Qu, J. Solinas and S. Vanstone, An ecient protocol for authenticated key
agreement, Design and Codes Cryptography, vol.28, no.2, pp.119-134, 2003.
[19] C.-T. Li, M.-S. Hwang and Y.-P. Chu, An ecient sensor-to-sensor authentication path-key estab-
lishment scheme for secure communications in wireless sensor networks, International Journal of
[20] M.-H. Lim, C.-M. Yeoh, S. Lee, H. Lim and H. Lee, A secure and ecient three-pass authenticated
key agreement protocol based on elliptic curves, Networking2008, LNCS, vol.4982, pp.170-182, 2008.
[21] F. Liu and L. Lu, A WPKI-based security mechanism for IEEE 802.16e, Proc. of Int. Conference on
Wireless Comm., Networking and Mobile Computing, pp.1-4, 2006.
[22] A. J. Menezes, P. Oorschot and S. Vanstone, Handbook of Applied Cryptography, CRC Press, 1997.
[23] V. Miller, Uses of elliptic curves in cryptography, Proc. of CRYPTO1985, pp.417-426, 1986.
[24] Z. Mohammad and C.-C. Lo, Vulnerability of an improved elliptic curve Die-Hellman key agreement
and its enhancement, Proc. of EBISS2009, pp.1-5, 2009.
[25] C. Popescu, A secure authenticated key agreement protocol, Proc. of MELECON2004, vol.2, pp.783-
786, 2004.
[26] C. Popescu, A secure key agreement protocol using elliptic curves, International Journal of Com-
puters and Applications, vol.27, no.3, pp.147-152, 2005.
[27] G. V. S. Raju and R. Akbani, Elliptic curve cryptosystem and its applications, Proc. of IEEE-
SMC2003, vol.2, pp.1540-1543, 2003.
[28] W.-G. Shieh and M.-T. Wang, An improvement to Kim-Chungs authentication scheme, ICIC Ex-
press Letters, vol.3, no.4(B), pp.1215-1220, 2009.
[29] D. Stebila and B. Ustaoglu, Towards denial-of-service-resilient key agreement protocols, ACISP2009,
LNCS, vol.5594, pp.384-406, 2009.
[30] M. A. Strangio, Ecient Die-Hellmann two-party key agreement protocols based on elliptic curves,
Proc. of the 20th ACM Symposium on Applied Computing, pp.324-331, 2005.
[31] M. A. Strangio, Revisiting an ecient elliptic curve key agreement protocol, Cryptology ePrint
Archive: Report 081, 2007.
[32] B. Ustaoglu, Obtaining a secure and ecient key agreement protocol from (H)MQV and NAXOS,
Design and Codes Cryptography, vol.46, no.3, pp.329-342, 2008.
[33] S. Wang, Z. Cao, M. A. Strangio and L. Wang, Cryptanalysis and improvement of an elliptic curve
Die-Hellman key agreement protocol, IEEE Communication Letters, vol.12, no.2, pp.149-151, 2008.
[34] R.-C. Wang, W.-S. Juang and C.-L. Lei, A robust authentication scheme with user anonymity for
wireless environments, International Journal of Innovative Computing, Information and Control,
vol.5, no.4, pp.1069-1080, 2009.
[35] C.-C. Wu, C.-C. Chang and I.-C. Lin, An authenticated payword scheme without public key cryp-
tosystems, International Journal of Innovative Computing, Information and Control, vol.5, no.9,
pp.2881-2891, 2009.
[36] E.-J. Yoon and K.-Y. Yoo, An improved Popescus authenticated key agreement protocol,
ICCSA2006, LNCS, vol.3984, pp.276-283, 2006.

!!
" # $%&'
()* +*, -.
/&

!&01200

!
" # $ "% &
$
$ $" '$

" ( " $
)
* * "
( +
,
! $ $
* - " $ "
$

3 !& &&! + &!

'&&
'&&
4 /5
' ! ' 4&(5 +
& & 4+5

! "# $ $
% & ' "#

(
) *
+
,
- .$ /

$* ( *

0
- 1"#2
, 3 13!!2
/
$
- %4 5 6 7 8 9 &
3!!
)$
,

:
; * $
* 3!!

%7 8 4 5 6 7&
1<02 %8 9 &

/ 3!!

0 3!!

=
=
<0 $
-

> , -:$ - -
/

> /
*

,

-
>

/
"# , ,
+ / $
: "# "#
;

$ ! $

$
<0 ; -
! 4 ! 5
! 6 - ! 7 -
*

! "#$% ! <0
$
? <0

<0 *
$
; <0

+ /

,
;
*

$

$ !
$
<0 1!?<02 - <0

) /
%4&
!

<0
*
@$
$

$
1 ' 2A

12 B 12 C 12 C 12 12 12

!
/ /

@
$ - A
$ - - 12
$ - - 12 ,
$

$

$

DE

DE
-
B 12

F -

$F 1 2 %4& ;

$ 1 / -2
- - 1 $

$
2
* $/

$
%4&

/
$

$
- - '

- <0

;

$ ,
$
<0 ; /
"#
$
<0 ) "# ; *

"#

&!" "'"#()!$*( @ * ,

10.2

-
*
%& 0 0.
*

,
!
,
1G2 $
1G0.2 ; G $

;
,
6 4 1 2 ;

( ; - "#

- ) % 4 5 6 7 8 9&
! "# $ "# $
* "# /* $

G0. ! "#

(
$$

"# )
D$*E

* $
,
, <0
F "#
+ "#

1"<#2

; <0

4 +$"#"! ! "*')( < - $

( * ) '<0

$

4 ") !
( 1- 2
6 1 B 62 8 1 B 7672
* 8 1 B 852 ; 12
$ 12
0
$ "
- /

4
' / ! "#
/
? 13 0 G!02 "
) 1 2
- ' 0
" 3
(H *

* - 8 I :

. -
4
; )
" 5 :
*

; -

; ) 1 ' 42
A
@ J-
' J-
' J-
;

)

# -

4 %)!' $"%%) ; -

$ '
"# ;
- ;
G
5 4
1 ;
2
; F
$

<0
-
< '
3 '-
'- < G1'<G2 0
J-
4 '- < 1'< 2 0
J-
5 '- ? ' J-
!, 1'?!2 J-
6 . @J-
7 . .
? 4 5

, *!"'% *"" *!"'% *!"'%
< < < <
1 2 ? 1 62 ? 1 6 82 ?
14 52 ? 14 5 82 ? 14 5 7 92 ?
18 92 ? 16 7 2 ? 16 7 4 82 ?
1 62 1 4 62 1 4 6 82
15 92 15 7 92 1 5 7 92

44 !'(%% & #+) #!+ /

- $
- ; "# 1$
2
$

A 12 B % 12 C 12 C
12 12& 12 12 12 6

12 12 12 <0 ;
-
"/ 12 4 5 "# 0 -

$
"# ' 5
0 -

$

"#
44

*
1. -
2
* -
*
- * -
- F
*
F -

H
> -

? -

, 0
0
$
-
* F @

/ - /
, <0
/

> <0 #
-
G* * ) 12 $
3 -

- ;
-
A
; -
- ; - 12 '

12 ; 12 -
'
- ;

'
- A

B B

B

142

(
; '

$

;
) >
'

; #
- /
"#
- *

5 "%'% ! %"-!% ;

) 4 5
* -
; * - ;
4 ; $
;
5 "- 4 5

* -
4 5

<
*
?
;
; 4 4
' 5 5
? '

1' 2 = -$
4 5

, *!"'% *"" *!"'% *!"'%
? ? ?
./00/ ....12/ .2034 ...34 .3356 ..135
6 9 7 77 77 7786 86
58 87 7 75 75 76 47
556 8 6974 74 748 88
5757 79 687 7 789 5
5796 7 6769 69 778 56
5 7 675 75 79 57
5975 69 7 7 7 7 78
588 67 7 6 6 79 5
6 6 95 666 66 8 57

; < ?

B
8695 4
77 8775
? B ./00/

B 8 9 95
74
? B ....12
; < ?
96 8 75

B 745 454 56

949 9 49
? B .2034
68 7 8 894

B 49 9 66 4

5 89
? B ...3.
' < ?

8666 6 7 9 76

5
B
9
9 7
87 4 5 649
764 57 956 65
? B .3356

5 4 6 4 49

B
49
55
766
4 79
5 67

45 9456 54 65
? B ..135
; 1;
5 ' 6 72
>
"# 1 2 F

4 "# 14 42 5 "# 15 52
; 1;
62 "#
1 2
14 "#
14 42 5 "# 15 52 2 ! (

0.8
2 Channel Independency
3 Channel Independency
0.75
Normalised determinant values 4 Channel Independency
0.7
0.65
0.6
0.55
0.5
0.45
0.4
1 2 3 4 5 6 7 8 9 10
Number of samples
)
0.025
2 Channel Dependency
0.02
Normalised determinant values
0.015
0.01
0.005
0
1 2 3 4 5 6 7 8 9 10
Number of samples
? )

5 !%!' !!'(%% & %"% %"$% + 0+K0

-

; $
1 )
2 $ F

$

5 *!"' ! 7 *!"' ! / *!"' !
88 9 5
885 49
99 5 59
986 4 4
876 5 58
8 57
4 79 4 58
4 89 8 5
4 74 59
44 9 9 59
"! 50// 026 //5
140/010 22073/4 /124001
, )
, $ 6
;
0+K0 A
0 4 5
0 4 5
4 ; : 4 5
0+K0 4 5
! !! ! ' .
'
< 998 55 47
" 77 8 6
; 967
0+K0 4 5

! !! ! ' .
'
< 8 6 84869
" 8
;
0+K0 $

4 5
! !! ! ' .
'
< 66 869 8646875 67 67
" 44 8 8664
; 6647657
; 0+K0 , ; , , 0+K0

1 ;
7 = 92
A
K
) 1

2

K
)

1
2
; 0+K0
- A
; ,

; ! !/ 1!!2
; 12
; !/ 1!2 !!:
; , !H
; - $
$
12 0 $
; , 1 ' 8 = 2
- L
)
- $
$ '
)
0.7
0.65
0.6
0.55
0.5
2 Channels 3 Channels 4 Channels

Independency levels
0+K0 4 5

6 %%% ; -
)
' ;

4 * -
' ;
5

#
-

)
' ;
6

:
5
: ;
*
;
F A
@ 4 5
1 9 2

G
$
<0

3
x 10
18
16
Normalised determinant values 14
12
10
0
2 channels 3 channels 4 channels
Dependency levels
0+K0 4 5
3000
2500
2000
1500
1000
500
0
2 channels 3 channels 4 channels
Ratio of Independency and dependency levels
0+K0

4 5
; *

- -

/
4 5 >
*

?
-

$
1 ;
52

;

- ;

"#
7 '%% - J
"# ;

)
J

/
; $
-
) ) $
; -$
)

; )

/

/ * -
"#
/

'
$
<0 )
) ' <0

/
$

-

@ "#

* ;

"# -

67 0 80 9 +0 0 :

90 + "0 0 3& ( %
%

;' &&& ! .$ %
0 0 0 <0
67 (0 "0 1 =0 90 9 0 )0 ! 0 )
&> 3 !&& 3 &! /
( / +0 -( 1 2 +
/&
0.-- .0
67 /0 1 0 0 / * 3 /% 4 *
* ? :
' & > @ 1 0
67 +0 /AA 0 B0 C B & B0 0
/ >%
&
& %
& ( /
0
0"( %
0 0 0
6<7 ?0
0 0 +
?0 + A 3 /
)&& $/ <0
6.7 0 ?0 ?& 0 :0 3&&
'&&
&!
& /
%
0. 0 <0
6-7 (0 "0 1 =0 90 9 0 )
&>
/ & ( & &!
&&! &'& !&! !&& /
! '' /&
0- - < 0
67 #0 #' 80 1 ,0 3 @0:0 & " !&
'

& & 1

'&&

%
0< 0- 0<-.. 0

67 0 9 &%' 0 "0 /& 90 @& / )=B &

'

& &

%
0< 04/5 0.--. 0
6 7 /0 ( =0 & B0 0 0 & /
>%

&&D E
D
&! ;

/
0
0"( 0..- 0
67 30 1 0 @ & 0 9 90 /1A> #0 #

'&&

&
' ! &!
& &

&F ) G &! E&
5 %
)
5 %
0 0 0
0
67 /0 =0 0 90 +0 !
& '

> &&& &!
1 %
& ! 3 %
0< 0 0
<0
67 @0 3 ?0 1 30 C 90 =0 90 C1 !
&
' ! ' &
&!
'&& /
0< -0
67 0 B 0 :! :080 B& A &&
&!&

%
0< 04/5 0-
0
6<7 @0#0 :! 00 C "00 ? @0:0 3! !
A &
%!
H 4 %
0 04/5 0-- 0
6.7 ,0 8! ,0 ( =0 , +
H !
E' & > '

%
0< 0. 0..
0
6-7 )0
'&& > I
/ %
0. 0
0- 0
67 /0 ?0 +
#0 ?0 H >&1 / ;A

&

%

0 %
0- 0. 0< <0
67 /0 3'% B& & D;
! &
'&&
! 0 0"( %
0 0 0... 0
6 7 /0 3'% ?0 9 0 *H * :
' & >
@ 1 0
67 ?0 0 +&H 0 =
* % ! 6 . $ <
:
& J :
1& +
<0
67 #0 )0 ?! 0 1! #0 :0 C 0 ?0 1 > (0 + > /0 ?0 +
#0 ?0 H >&1

'&&
&!
& / -(
* %
0 0.. 0
67 /0 +0 + 0 =0 !
0 /H /
(&
&& > &
&
.$ . & %
0- 0 0<. 0
67 ?0 9 & 0 % / / *; $
%&' )&& + & 0
6<7 +0 (0 C 1 =0 B0 ! 0 0 ? & / & (
&
&
&%

&
&
0 %
0 0 0- 0
6.7 30 /0 3> #0 ! $% & ! &! ( G &
! / + 78 B C
BC 0--< 0
6-7 0 *1 @0 & 90 #0 #&H (
&&D &!
> 1 & &!

%
0< 04+5 0< 0
67 30 # #0 ( =0 *1 30 #1 90 # / &' &(
! &!
> 1

%
0< 04+5 0-- 0
c ISSN 1349-4198
H FILTERING FOR MARKOVIAN JUMP SYSTEMS WITH

TIME-VARYING DELAYS
Jinliang Liu1 , Zhou Gu2 and Songlin Hu3

1
College of Information Science and Technology
Donghua University
No. 2999, Renmin North Road, Songjiang District, Shanghai 201620, P. R. China
liujinliang@vip.163.com
2
College of Power Engineering
Nanjing Normal University
No. 78, Bancang Road, Nanjing 210042, P. R. China
guzhouok@yahoo.com.cn
3
Department of Control Science and Engineering
Huazhong University of Science and Technology
No. 1037, Luoyu Road, Wuhan 430074, P. R. China
songlin621@126.com
Abstract. This paper proposes a class of H filter design for Markovian jump systems
with time-varying delays. Firstly, by exploiting a new Lyapunov function and using the
convexity property of the matrix inequality, new criteria is derived for the H perfor-
mance analysis of the filtering-error systems, which can lead to much less conservative
analysis results. Secondly, based on the obtained conditions, the gain of filter can be
obtained in terms of linear matrix inequalities (LMIs). Finally, numerical examples are
given to demonstrate the eectiveness and the merit of the proposed method.
Keywords: Time-delay systems, H filter, Markovian jump systems
1. Introduction. During the past few decades, Markovian Jump Systems (MJSs) have
been attracted much attention [1, 2, 3, 4, 5, 6], which can be regarded as a special class
of hybrid systems with finite operation modes whose structures are subject to random
abrupt changes. The system parameters usually jump among finite modes, and the mode
switching is governed by a Markov process. MJSs have many applications, such as failure
prone manufacturing systems, power systems and economics, etc. A great number of
results on estimation and control problems related to such systems have been reported in
the literature [7, 8, 9, 10, 11].
Recently, the problem of H filtering of linear/nonlinear time-delay systems has also
received much attention due to the fact that for many practical filtering applications, time-
delays cannot be neglected in the procedure of filter design and their existence usually
results in a poor performance [12, 13, 14]. Some nice results on H filtering for time-delay
systems [15, 16] have been reported in the literature and there are two kinds of results,
namely delay-independent filtering [17] and delay-dependent [18, 19, 20, 21, 22, 23]. The
delay-dependent results are usually less conservative, especially when the time-delay is
small. The main objective of the delay-dependent H filtering is to obtain a filter such
that the filtering error system allows a maximum delay bound for a fixed H performance
or achieves a minimum H performance for a given delay bound.
1299
1300 J. LIU, Z. GU AND S. HU
This paper addresses the problem of H filter design for MJSs with interval time-
varying delay. To obtain less conservative results, a new Lyapunov function is constructed,
which includes the lowerand upper delay bound of interval time-varying delay. Based on
t
this, one splits the item t (t) eT (s)Q2 (t )e(s)ds into two parts to deal with, respectively,
and uses the convexity of the matrix functions to avoid the conservative caused by en-
larging (t) to M in the deriving results. Compared with the existing method [24, 25],
the conservativeness of the derived H performance analysis results is further reduced.
Novel H filter design criteria are obtained. Examples used in [24, 25] are employed to
show the eectiveness and less conservativeness of the proposed methods.
Notation: Rn and Rnm denote the n-dimensional Eculidean space, and the set of nm
real matrices; the superscript T stands for matrix transposition; I is the identity matrix
of appropriate dimension; stands for the Euclidean vector norm or the induced matrix
2-norm as appropriate; the notation X > 0 (respectively, X 0), for X Rnn means
that the matrix X is real symmetric positive definite (respectively, positive semi-definite).
When x is a stochastic variable, E{x} [ stands ] for the expectation of x. For a matrix B and
A
two symmetric matrices A and C, denote a symmetric matrix, where * denotes
B C
the entries implied by symmetry.
2. Problem Statement and Preliminaries. Fix a probability space (, F, P) and

consider the following class of uncertain linear stochastic systems with markovian jump
parameters and time-varying delays ()

x(t)
= A(t )x(t) + Ad (t )x(t (t)) + A (t )(t)

y(t) = C(t )x(t) + Cd (t )x(t (t)) + C (t )(t)
(1)

z(t) = L(t )x(t) + Ld (t )x(t (t)) + L (t )(t)
x(t) = (t), t [ , ]
M m
where x(t) Rn is the state vector, y(t) Rr is the measurement vector, (t) L2 [0, )
is the exogenous disturbance signal, z(t) Rp is the signal to be estimated, {t } is a
continuous-time Markovian process with right continuous trajectories and taking values
in a finite set S = {1, 2, , N } with stationary transition probabilities:
{
ij h + o(h), i = j
P rob{t+h = j|t = i} = (2)
1 + ii h + o(h), i = j
where h > 0, lim o(h)

= 0 and ij 0 for j = i is the transition rate from mode i at time
h0 h
t to the mode j at time t + h and

N
ii = ij (3)
j=1, j=i
In the system (1), the time delay (t) is a time-varying continuous function satisfying
the following assumption.
0 m (t) M < , (t) , t > 0 (4)
where m is the lower bound and M is the upper bound of the communication delay.
In this paper, we consider the following filter for system (1)

x (t) = A(t ) x(t (t)) + G(t )(
x(t) + Ad (t ) y (t) y(t))
y(t) = C(t ) x(t (t))
x(t) + Cd (t ) (5)

z(t) = L(t ) x(t (t))
x(t) + Ld (t )
H FILTERING FOR MARKOVIAN JUMP SYSTEMS 1301
The set S comprises the various operation modes of system (1) and for each possible
value of t = i, i S, the matrices associated with i-th mode will be denoted by
Ai := A(t = i), Adi := Ad (t = i), Ai := A (t = i)
Ci := C(t = i), Cdi := Cd (t = i), Ci := C (t = i)
Li := L(t = i), Ldi := Ld (t = i), Li := L (t = i)
where Ai , Adi , Ai , Ci , Cdi , Ci , Li , Ldi , Li are constant matrices for any i S. It is
assumed that the jumping process {t } is accessible, i.e., the operation mode of system
() is known for every t 0.
Let e(t) = x(t) x(t) and z(t) = z(t) z(t). Then we have the following filtering error
system:
{
= Ai e(t) + Adi e(t (t)) + Ai (t)
e(t)
(6)
z(t) = Li e(t) + Ldi e(t (t)) Li (t)
where Ai = Ai + Gi Ci , Adi = Adi + Gi Cdi , Ai = Ai Gi Ci .
The H filtering problem addressed in this paper is to design a filter of form (6) such
that
The filtering error systems (6) with (t) = 0 is exponentially stable.
The H performance z (t)2 < (t)2 is guaranteed for all nonzero (t)
L2 [0, ) and a prescribed > 0 under the condition e(t) = 0, t [M , m ].
The following lemmas and definitions are needed in the proof of our main results.
Lemma 2.1. [26] For any constant matrix R R, R = RT > 0, constant M > 0 and
vector function x : [M , 0] Rn such that the following integration is well defined, it
holds that
t [ ]T [ ][ ]
x(t) R R x(t)
M T
x (s)Rx(s)ds
(7)
tM
x(t M ) R R x(t M )
Lemma 2.2. [27] Suppose 0 m (t) M , 1 , 2 and are constant matrices of

appropriate dimensions, then
( (t) m )1 + (M (t))2 + < 0 (8)
if and only if (M m )1 + < 0 and (M m )2 + < 0 hold.
Definition 2.1. The system (6) is said to be exponentially stable in the mean-square
sense (EMSS), if there exist constants > 0, > 0, such that t > 0
E{x(t)2 } et sup {(s)2 } (9)
M <s<0
Definition 2.2. For a given function V : CFb 0 ([M , 0], Rn ) S R, its infinitesimal
operator L [28] is defined as
1
LV (xt ) = lim+ [E(V (xt+ |xt ) V (xt ))] (10)
0
3. Main Results.
Theorem 3.1. For some given constants 0 m M and , the systems (6) is EMSS
with a prescribed H performance , if there exist Pi > 0, Q0 > 0, Q1 > 0, Q2i > 0,
R0 > 0, R1 > 0, Z1 > 0, Z2 > 0, Mik and Nik (i S, k = 1, 2, , 5) with appropriate

dimensions such that the following matrix inequalities hold

11
21 22
= 31
< 0, s = 1, 2 (11)
32 33
41 (s) 42 (s) 0 R1

N
ij Q2j Zk , k = 1, 2 (12)
j=1
where

N
11 = Pi Ai + ATi Pi + Q0 + Q1 + Q2i R0 + m Z1 + Z2 + ij Pj
j=1
T
Adi Pi Mi1T + Ni1T
R0 + Mi1T
21 =
Ni1 T
T
Ai Pi

Mi3 + Mi2T + Ni3 Q0 R0 + Mi3 + Mi3T
22 =
Mi4 + Ni4 Ni2 T
Mi4 Ni3T
Q1 Ni4 Ni4
T

Mi5 + Ni5 Mi5 Ni5 2 I

R Ai R Adi 0 0 m R0 Ai
m 0 m 0
31 = R1 Ai , 32= R1 Adi 0 0 R1 Ai
Li Ldi 0 0 Li
33 = diag{R0 , R1 , I}

41 (1) = Mi1T , 41 (2) = Ni1T
[ ]
42 (1) = Mi2T Mi3T Mi4T Mi5T
[ T T T T ]
42 (2) = Ni2 Ni3 Ni4 Ni5
= (1 )Q2i Mi2 Mi2T + Ni2 + Ni2T
= M m .
Proof: Introduce a new vector
[ ]
T (t) = eT (t) eT (t (t)) eT (t m ) eT (t M ) T (t)
[ ] [ ]
and two matrices 1 = Ai Adi 0 0 Ai , 2 = Li Ldi 0 0 Li .
Rewrite (6) as
{
e(t)
= 1 (t)
(13)
z(t) = 2 (t)
Let xt (s) = x(t + s) ( (t) s 0). Then, similar to [29], {(xt , t ), t 0} is a Markov
process. Construct a Lyapunov functional candidate as

4
V (xt , t ) = Vi (xt , t ) (14)
i=1
where
V1 (xt , t ) = eT (t)P (t )e(t)
t t t
V2 (xt , t ) = eT (s)Q0 e(s)ds + eT (s)Q1 e(s)ds + eT (s)Q2 (t )e(s)ds
tm tM t (t)
t t tm t
V3 (xt , t ) = m e T (v)R0 e(v)dvds
+ e T (v)R1 e(v)dvds

tm s tM s
t t tm t
T T
V4 (xt , t ) = e (v)Z1 e(v)dvds + e (v)Z2 e(v)dvds
tm s tM s
Let L be the weak infinite generator of the random process {xt , t }. Then, for each
t = i (i S), we have
( )

N
L[V (xt , t )] e (t) 2Pi Ai +
T
ij Pj + Q0 + Q1 + Q2i + m Z1 + Z2 e(t)
j=1
+2eT (t)Pi Adi e(t (t)) + 2eT (t)Pi Ai (t) eT (t m )Q0 e(t m )
(1 )eT (t (t))Q2i e(t (t))
+ e T (t)R1 e(t)
t ( N )

+ eT (s) ij Q2j e(s)ds + m 2 T
e (t)R0 e(t)

t (t) j=1
t
e (t M )Q1 e(t M ) m
T
e T (s)R0 e(s)ds

tm
tm t tm
T
e (s)R1 e(s)ds
e (s)Z1 e(s)ds
T
eT (s)Z2 e(s)ds (15)
tM tm tM
Note that
( )
t
N
eT (s) ij Q2j e(s)ds
t (t) j=1
( N ) ( )
tm t
N
= eT (s) ij Q2j e(s)ds + eT (s) ij Q2j e(s)ds (16)
t (t) j=1 tm j=1
From (12) and (16), we have

( )
t
N t tm
T
e (s) ij Q2j e(s)ds e (s)Z1 e(s)ds
T
eT (s)Z2 e(s)ds
t (t) j=1 tm tM
[ ] ( N )
t
N tm
= eT (s) ij Q2j Z1 e(s)ds + eT (s) ij Q2j e(s)ds
tm j=1 t (t) j=1
tm
eT (s)Z2 e(s)ds
tM
t [ N ]

eT (s) ij Q2j Z1 e(s)ds
tm j=1
[ ]
tm
N
+ eT (s) ij Q2j Z2 e(s)ds 0 (17)
t (t) j=1
Applying Lemma 2.1, we have

t [ ]T [ ][ ]
e(t) R0 R0 e(t)
m T
e (s)R0 e(s)ds
(18)
tm
e(t m ) R0 R0 e(t m )
Combining (15), (17), (18) and introducing slack matrices Mi , Ni , i S, we obtain
L[V (xt , t )] 2 T (t)(t) + zT (t)
z (t)
( )
N
eT (t) 2Pi Ai + ij Pj + Q0 + Q1 + Q2i + m Z1 + Z2 e(t)
j=1
+2e (t)Pi Ai (t) eT (t m )Q0 e(t m ) eT (t M )Q1 e(t M )

T
(1 )eT (t (t))Q2i e(t (t)) + m 2 T

+ e T (t)R1 e(t)
e (t)R0 e(t)
[ ]T [ ][ ] tm
e(t) R0 R0 e(t)
+ e T (s)R1 e(s)ds

e(t m ) R0 R0 e(t m ) tM
+2eT (t)Pi Adi e(t (t)) 2 T (t)(t) + T (t)T2 2 (t)

[ tm ]
+2 (t)Mi e(t m ) e(t (t))
T
e(s)ds

t (t)
[ ]
t (t)
+2 T (t)Ni e(t (t) e(t M ) e(s)ds
(19)
tM
where
[ ]
MiT = Mi1T Mi2T Mi3T Mi4T Mi5T
[ ]
NiT = Ni1T Ni2T Ni3T Ni4T Ni5T
Note that
tm tm
2 (t)Mi
T
e(s)ds
e T (s)R1 e(s)ds
+ ( (t) m ) T (t)Mi R11 MiT (t) (20)
t (t) t (t)
t (t) t (t)
2 (t)Ni
T
e(s)ds
e T (s)R1 e(s)ds
+ (M (t)) T (t)Ni R11 NiT (t) (21)
tM tM
From (19) (21), we can obtain

L[V (xt , t )] 2 T (t)(t) + zT (t)
z (t)
( )
N
e (t) 2Pi Ai +
T
ij Pj + Q0 + Q1 + Q2i + m Z1 + Z2 e(t)
j=1
+2e (t)Pi Adi e(t (t)) 2 T (t)(t) + T (t)T2 2 (t)

T
+2eT (t)Pi Ai (t) eT (t m )Q0 e(t m ) eT (t M )Q1 e(t M )

(1 )eT (t (t))Q2i e(t (t)) + m 2 T
+ e T (t)R1 e(t)
e (t)R0 e(t)
[ ]T [ ][ ]
e(t) R0 R0 e(t)
+
e(t m ) R0 R0 e(t m )
+2 T (t)Mi [e(t m ) e(t (t))] + 2 T (t)Ni [e(t (t) e(t M )]
+( (t) m ) T (t)Mi R11 MiT (t) + (M (t)) T (t)Ni R11 NiT (t) (22)
Using Lemma 2.2 and Schur complement, it is easy to see that (11) with s = 1, 2 are
sucient conditions to guarantee
L[V (xt , t )] 2 T (t)(t) + zT (t)
z (t) < 0 (23)
Then, the following inequality can be concluded

E{LV (xt , i, t)} < min ()E{ T (t)(t)} (24)
Define a new function as W (xt , i, t) = et V (xt , i, t), Its infinitesimal operator L is given
by
W(xt , i, t) = et V (xt , i, t) + et LV (xt , i, t) (25)
By the generalized Ito formula [28], we can obtain from (25) that
t t
E{W (xt , i, t)} E{W (x0 , i)} = s
e E{V (xs , i)}ds + es E{LV (xs , i)}ds (26)
0 0
Then, using the similar method of [2], we can see that there exists a positive number
such that for t > 0
{ }
E{V (xt , i, t)} sup (s)2 et (27)
M s0
since V (xt , i, t) {min (Pi )}x (t)x(t), it can be shown from (27) that for t 0
T
{ }
E{xT (t)x(t)} t sup (s)2 (28)
M s0
where
= /(min Pi ). Recalling Definition 2.1, the proof can be completed.
Remark 3.1. In the above proof, ( it should
) be noted that a new Lyapunov function is
t
N
constructed, and t (t) xT (s) ij Q2j x(s)ds in (17) are separated into two parts.
j=1
Moreover, examples below show that this method has less conservative than the existing
ones [6, 25].
Remark 3.2. Theorem 3.1 provides a delay-dependent stochastic stability condition for
MJS with interval time-varying delays. Throughout the proof of Theorem 3.1, it can be
seen that the convexity property of the matrix inequality is treated in terms of Lemma 2.2,
which need not enlarge (t) to M , therefore, the common existed conservatism caused by
this kind of enlargement in [30, 31, 32, 33] can be avoided, which will reduce the conser-
vative of the result.
In the following, we are seeking to design the H filtering based on Theorem 3.1.
Theorem 3.2. For some given constants 0 m M and , the augmented systems
(6) is EMSS with a prescribed H performance if there exist Pi > 0, Q0 > 0, Q1 > 0,
Q2i > 0, R0 > 0, R1 > 0, Z1 > 0, Z2 > 0, G i , Mik and Nik (i S, k = 1, 2 , 5) with
appropriate dimensions such that the following LMIs hold for a given > 0

11

= 21

22 < 0, s = 1, 2 (29)

31 32 33
41 (s) 42 (s) 0 R1

N
ij Q2j Zk , k = 1, 2 (30)
j=1
where

N
11 = Pi Ai + ATi Pi + G
Ti + Q0 + Q1 + Q2i R0 + m Z1 + Z2 +
i Ci + CiT G ij Pj
j=1
T
Adi Pi + Cdi Gi Mi1T + Ni1T
T T
i Ci
m Pi Ai + m G
R0 + Mi1 T
21
=
,

31 = Pi Ai + G i Ci
Ni1T
Li
ATi Pi Ci T T
Gi

P A + i Cdi 0 0 m Pi Ai m G
G i Ci
m i di m
32
= Pi Adi + G i Cdi 0 0 Pi Ai G i Ci
Ldi 0 0 Li
33 = diag{2Pi + 2 R0 , 2Pi + 2 R1 , I}

and 22 , 41 (s), 42 (s) and are as defined in Theorem 3.1.
Moreover, the filter gain in the form of (5) as following:
Gi = P 1 Gi
i (31)
Proof: Defining G i = Pi Gi , from (11) and using schur complement, the matrix in-
equality (11) holds if and only if

11

= 21

22 < 0, s = 1, 2 (32)
31 32 33
41 (s) 42 (s) 0 R1
where
33 = diag{Pi R1 Pi , Pi R1 Pi , I}
0 1
Due to
(Rk 1 Pi )1 (Rk 1 Pi ) 0, k = 0, 1, iS (33)
which gives
Pi Rk1 Pi 2Pi + 2 Rk , k = 0, 1, iS (34)
Substituting Pi Rk1 Pi
with 2Pi + Rk (k = 0, 1) in (32), we obtain (29), so, if (29)
2
holds, we have (11) holds, and from above proof, we have Gi = Pi1 G
i . This completes
the proof.
Remark 3.3. The inequality (34) is used to bound the term Pi Rk1 Pi . This step also can
be improved by adopting the cone complementary algorithm [34], which is popular in recent
control designs. Here the scaling parameter > 0 can be used to improve conservatism in
Theorem 3.2.
4. Numerical Example. In this section, well-studied examples are used to illustrate

the the eectiveness of the approaches proposed in this paper.
Example 4.1. Consider a Markovian jump system in (6) with two modes and the follow-
ing parameters [25]:
[ ] [ ] [ ]
2.2460 1.4410 1.8999 0.8156 0.7098 1.1908
A1 = , A2 = , Ad1 = ,
1.5937 2.9289 0.6900 0.7881 0.6686 3.2025
[ ] [ ] [ ]
1.5198 1.6041 0.0403 0.5689
Ad2 = , A1 = , A2 = ,
0.1567 1.2427 0.6771 0.2556
[ ] [ ] [ ]
L1 = 0.3775 0.2959 , L2 = 1.4751 0.2340 , Ld1 = 0 0 ,
[ ]
Ld2 = 0 0 , L1 = 0.1184, L2 = 0.3148.
Suppose the transition probability matrix is given by 11 = 3.
For several values of and 22 , the computation results of M are listed in Tables 1 3 .
Tables 1 3 show the maximum allowable values of M for = 0.5 and 22 = 0.6,
= 1 and 22 = 0.6, = 1 and 22 = 1, respectively. From Tables 1 3, we can see
our method can lead to less conservative results.
Table 1. Maximum allowable values of M for = 0.5 and 22 = 0.6
0.5 1 1.5 2
M by [25] 0.3657 0.4062 0.4168 0.4219
M by Theorem 3.1 0.3968 0.4388 0.4510 0.4569
Table 2. Maximum allowable values of M for = 1 and 22 = 0.6
0.5 1 1.5 2
M by [25] 0.2772 0.2846 0.2857 0.2861
M by Theorem 3.1 0.3966 0.4355 0.4468 0.4520
Table 3. Maximum allowable values of M for = 1 and 22 = 1
0.5 1 1.5 2
M by [25] 0.2773 0.2846 0.2857 0.2861
M by Theorem 3.1 0.3896 0.4296 0.4408 0.4463
To illustrate the proposed method on filtering design, another example is considered as

follows.
Example 4.2. Consider linear Markovian jump systems in the form (1) with two mode.
For modes 1 and 2, the dynamics of system are described as

3 1 0 0.2 0.1 0.6 1

A1 = 0.3 2.5 1 , Ad1 = 0.5 1 0.8 , A1 = 0 ,

0.1 0.3 3.8 0 1 2.5 1
[ ] [ ]
C1 = 0.8 0.3 0 , Cd1 = 0.2 0.3 0.6 , C1 = 0.2,
[ ] [ ]
L1 = 0.5 0.1 1 , Ld1 = 0 0 0 , L1 = 0,

2.5 0.5 0.1 0 0.3 0.6 0.6
A2 = 0.1 3.5 0.3 , Ad2 = 0.1 0.5 0 , A2 = 0.5 ,
0.1 1 2 0.6 1 0.8 0
[ ] [ ]
C2 = 0.5 0.2 0.3 , Cd2 = 0 0.6 0.2 , C2 = 0.5,
[ ] [ ]
L2 = 0 1 0.6 , Ld2 = 0 0 0 , L2 = 0.
[ ]
0.5 0.5
Suppose the transition probability matrix is given by = and the initial
0.3 0.3
[ ]T [ ]T
conditions x(0) = 0.8 0.2 0.9 , x(0) = 0 0.2 0 . This system is nominal one
consider in [24]. By Theorem 3.2, we get the maximum time delay M = 6.4763 for
m = 0, = 0, = 10 and = 1.2. This upper bound is much larger than the one
M = 1.9195 given by [24], which shows our method has less conservative than that of
[24].
The corresponding filter are given by

0.5399 0.4684
G1 = 1.1596 , G2 = 0.3923
3.5158 1.3604
To illustrate the performance of the designed filter, choose the disturbance function as
follows
0.5, 5 < t < 10
(t) = 0.5, 15 < t < 20

0, otherwise
With this filter, Figures 1 4 show the operation modes of the MJS, interval time-
varying delay, estimated signal z(t), z(t) and estimated signals error (t) = z(t) z(t),
respectively.
3 7
6
2.5
5
2
4
Mode
(t)
1.5
3
1
2
0.5 1
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time t(sec) Time t(sec)
Figure 1. Operation modes Figure 2. Interval time-

of MJS varying delay
0.6 0.4
0.4 0.2
0
Estimated signals
0.2
0.2
0
(t)
0.4
0.2
0.6
0.4
0.8
0.6 1
0.8 1.2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time t(sec) Time t(sec)
Figure 3. Estimated signal Figure 4. Estimated signals

z(t) and z(t) error (t) = z(t) z(t)
5. Conclusions. In this paper, we have studied a class of H filter design for Markovian
jump systems with time-varying delays via manipulating a new Lyapunov function and
using the convexity property of the matrix inequality. With the proposed method, an
LMI-based sucient condition for the existence of the desired H filter has been derived,
which can lead to much less conservative analysis results. Finally, Numerical examples
have been carried out to demonstrate the eectiveness of the proposed method.
Acknowledgment. The authors would like to acknowledge Natural Science Founda-

tion of China (NSFC) for its support under grant Nos. 61074025, 61074024, 60834002,
60904013 and 60774060.
REFERENCES
[1] H. Morita, E. Uchino and T. Yamakawa, Markov model based adaptive web advertisement system
by tracking a users taste, International Journal of Innovative Computing, Information and Control,
vol.5, no.3, pp.811-819, 2009.
[2] D. Yue and Q. Han, Delay-dependent exponential stability of stochastic systems with time-varying
delay, nonlinearity, and Markovian switching, IEEE Trans. on Automatic Control, vol.50, no.2,
pp.217-222, 2005.
[3] L. Zhang and E. Boukas, Mode-dependent H filtering for discrete-time Markovian jump linear
systems with partly unknown transition probabilities, Automatica, vol.45, no.6, pp.1462-1467, 2009.
[4] H. Shao, Delay-range-dependent robust H filtering for uncertain stochastic systems with mode-
dependent time delays and Markovian jump parameters, Journal of Mathematical Analysis and
Applications, vol.342, no.2, pp.1084-1095, 2008.
[5] O. Costa and M. Fragoso, A separation principle for the H2 control of continuous-time infinite Markov
jump linear systems with partial observations, Journal of Mathematical Analysis and Applications,
vol.331, no.1, pp.97-120, 2007.
[6] S. Xu, T. Chen and J. Lam, Robust filtering for uncertain Markovian jump systems with mode-
dependent time delays, IEEE Trans. on Automatic Control, vol.48, no.5, pp.900-907, 2003.
[7] J. Qiu and K. Lu, New robust passive stability criteria for uncertain singularly Markov jump systems
with time delays, ICIC Express Letters, vol.3, no.3(B), pp.651-656, 2009.
[8] H. Zhang, H. Yan, J. Liu and Q. Chen, Robust H filters for Markovian jump linear systems
under sampled measurements, Journal of Mathematical Analysis and Applications, vol.356, no.1,
pp.382-392, 2009.
[9] Z. Wang, Y. Liu, L. Yu and X. Liu, Exponential stability of delayed recurrent neural networks with
Markovian jumping parameters, Physics Letters A, vol.356, no.4-5, pp.346-352, 2006.
[10] Z. Wang, J. Lam and X. Liu, Exponential filtering for uncertain Markovian jump time-delay systems
with nonlinear disturbances, IEEE Trans. on Circuits and Systems II: Express Briefs, vol.51, no.5,
pp.262-268, 2004.
[11] J. Xiong and J. Lam, Robust H2 control of Markovian jump systems with uncertain switching
probabilities, International Journal of Systems Science, vol.40, no.3, pp.255-265, 2009.
[12] Z. Wang and D. Ho, Filtering on nonlinear time-delay stochastic systems, Automatica, vol.39, no.1,
pp.101-109, 2003.
[13] S. Nguang and P. Shi, Delay-dependent H filtering for uncertain time delay nonlinear systems: An
lmi approach, Control Theory & Applications, IET, vol.1, no.1, pp.133-140, 2007.
[14] Z. Wang, D. Ho and X. Liu, Robust filtering under randomly varying sensor delay with variance
constraints, IEEE Trans. on Circuits and Systems II: Express Briefs, vol.51, no.6, pp.320-326, 2004.
[15] J. Liu, W. Yu, Z. Gu and S. Hu, H filtering for time-delay systems with markovian jumping
parameters: Delay partitioning approach, Journal of the Chinese Institute of Engineers, vol.33,
no.3, pp.357-365, 2010.
[16] J. Liu, S. Hu and E. Tian, A novel method of H filter design for time-varying delay systems, ICIC
[17] C. de Souza, R. Palhares and P. Peres, Robust filter design for uncertain linear systems with multiple
time-varying state delays, IEEE Trans. on Signal Processing, vol.49, no.3, pp.569-576, 2001.
[18] C. Peng, D. Yue, E. Tian and Y. Zhang, Improved network-based robust H filtering for uncertain
linear systems, International Journal of Innovative Computing, Information and Control, vol.5, no.4,
pp.961-970, 2009.
[19] Y. Liu and W. Wang, Fuzzy H filtering for nonlinear stochastic systems with missing measurements,
[20] F. Wang, Q. Zhang and B. Yao, LMI-based reliable H filtering with sensor failure, International
[21] M. Basin, E. Sanchez and R. Martinez-Zuniga, Optimal linear filtering for systems with multiple state
and observation delays, International Journal of Innovative Computing, Information and Control,
vol.3, no.5, pp.1309-1320, 2007.
[22] C. Gong and B. Su, Robust L2 -L filtering of convex polyhedral uncertain time-delay fuzzy systems,
International Journal of Innovative Computing Information and Control, vol.4, no.4, pp.793-802,
2008.
[23] D. Yue and Q. Han, Network based robust H filtering for uncertain linear systems, IEEE Trans.
on Signal Processing, vol.11, pp.4293-4301, 2006.
[24] L. Hu, P. Shi and Y. Cao, Delay-dependent filtering design for time-delay systems with Markovian
jumping parameters, International Journal of Adaptive Control and Signal Processing, vol.21, no.5,
pp.434-448, 2007.
[25] S. Xu, J. Lam and X. Mao, Delay-dependent H control and filtering for uncertain Markovian jump
systems with time-varying delays, IEEE Trans. on Circuits and Systems I, vol.54, no.9, pp.2070-2077,
2007.
[26] K. Gu, V. Kharitonov and J. Chen, Stability of Time-Delay Systems, Birkhauser, 2003.
[27] D. Yue, E. Tian, Y. Zhang and C. Peng, Delay-distribution-dependent stability and stabilization of
T-S fuzzy systems with probabilistic interval delay, IEEE Trans. on Systems, Man, and Cybernetics,
Part B, vol.39, no.2, pp.503-516, 2009.
[28] X. Mao, Exponential stability of stochastic delay interval systems with Markovian switching, IEEE
Trans. on Automatic Control, vol.47, no.10, pp.1604-1612, 2002.
[29] E. Boukas, Z. Liu and G. Liu, Delay-dependent robust stability and H control of jump linear
systems, International Journal of Control, vol.74, no.4, pp.329-340, 2001.
[30] B. Chen, X. P. Liu and S. C. Tong, Delay-dependent stability analysis and control synthesis of fuzzy
dynamic systems with time delay, Fuzzy Sets and Systems, vol.157, pp.2224-2240, 2006.
[31] X. F. Jiang and Q. L. Han, Robust H control for uncertain Takagi-Sugeno fuzzy system with
interval time-varying delay, IEEE Trans. on Fuzzy Systems, vol.15, pp.321-331, 2007.
[32] E. Tian and C. Peng, Delay-dependent stability analysis and synthesis of uncertain T-S fuzzy systems
with time-varying delay, Fuzzy Sets and Systems, vol.157, pp.544-559, 2006.
[33] H. N. Wu and H. X. Li, New approach to delay-dependent stability analysis and stabilization for
continuous-time fuzzy systems with time-varying delay, IEEE Trans. on Fuzzy Systems, vol.15,
pp.482-493, 2007.
[34] L. El Ghaoui, F. Oustry and M. AitRami, A cone complementarity linearization algorithm for static
output-feedback and related problems, IEEE Trans. on Automatic Control, vol.42, no.8, pp.1171-
1176, 1997.
c ISSN 1349-4198
SURFACE TEXTURE CHARACTERIZATION OF FIBERS USING

FRACTIONAL BROWNIAN MOTION MODEL
Jiun-Jian Liaw1 , Chuan-Pin Lu2, and Lin-Huang Chang3

1
Department of Information and Communication Engineering
Chaoyang University of Technology
No. 168, Jifong E. Rd., Wufong Township, Taichung County, Taiwan
jjliaw@cyut.edu.tw
2
Department of Information Technology
Meiho Institute of Technology
No. 23, Pingguang Rd., Neipu, Pingtung, Taiwan

Corresponding author: chuan.pin.lu@gmail.com
3
Department of Computer and Information Science
National Taichung University
No. 140, Min-Sheng Rd., Taichung, Taiwan
lchang@mail.ntcu.edu.tw
Received November 2009; revised April 2010
Abstract. The texture of the surface of individual fiber is an important characteristic.

The fibers can be classified and recognized using the surface texture by a specific parame-
ter. To describe the texture, the fractal parameter (or Hurst coecient) is a proper value.
In this study, we use the Fractional Brownian Motion (FBM) to model the texture of the
fiber surface. And we apply a Fourier-domain Maximum Likelihood Estimator (FDMLE)
to calculate the fractal parameter of FBM. According to the experimental results, we can
objectively classify dierent types of fibers.
Keywords: Texture characterization, Fiber classification, Fractional Brownian motion,
Fourier-domain maximum likelihood estimator
1. Introduction. In recent years, the developments in sciences and technology, such as

Internet, digital image processing and micro-manufacture, etc, is drastically changing the
life of the mankind. Science and technology have been used in radar systems, robots,
networks and communications [1-4]. Using science and technology to investigate crime
scenes is the major aspect of fighting crime in democracies. They are necessary not
only to fight crime eciently, but also to protect human rights and maintain the social
order. Evidence collected from the crime scene includes fingerprints, DNA, blood and
fibers. The evidence can be used to identity the criminal or to reduce the number of
suspects. Many classification schemes were proposed for fingerprints, DNA and blood, but
nothing for fabrics. However, fiber identification has not only been an important object
in the forensic science, but also in many applications such as wool manufacture, textile
processing, archeology and zoology [5]. In the textile industry, accurate classification of
animal fibers proved very dicult. In the past, physical and chemical properties of fibers
have been used to solve the classification problem. Recently, some techniques have been
proposed to identify animal fibers by the cuticle on the surface of the fiber, such as image
processing methods and artificial intelligence techniques [6-11]. As seen from the proposed
literature, classification of fibers is actually important and the texture of fiber surface is
a characteristic useful in classifying fibers.
1311
1312 J.-J. LIAW, C.-P. LU AND L.-H. CHANG
The properties of texture have been analyzed more than 30 years ago. Chetverikov
discussed the texture analysis methods and proposed some schemes of texture inspections
[12]. The texture can be segmented in the spatial domain using the sub-band domain
co-occurrence matrix and the neighboring gray level dependent matrix. Some researchers
also extract the texture characterizations from the frequency domain. Wavelet transform
is used to inspect the defects of texture with segmentation algorithm [13]. Several blocks
are segmented by using wavelet technology, and the features of texture can be obtained
by computing the co-occurrence matrix. In this way, the small block selection may lose
some features. However, this scheme can not exactly recognize the location of the defect.
To improve eciency of the texture characterization, the advanced wavelet transforms
were developed. Many researchers use them to compute the detail values for determining
the texture features. To obtain the homogeneous feature, the roughness index and the
homogeneous index are built [14]. The technologies of texture characterization can be
summarized as following.
(1) The stochastic model is used to analyze the statistical properties of textures [15,16].
(2) The matrix methods are used to extract the texture features for classification [17,18].
(3) The transform-based methods are applied to characterizing textures and locating
defects [19,20].
However, those previous works concentrated upon the feature extraction of textures, not
the description of textures. The natural texture is very dicult to describe by conventional
Euclidean geometry (such as lines, circles, triangles, rectangles, . . . , etc). For modeling
the natural texture, the fractal geometry was proposed [21]. The kernels of the fractal
theory are the statistical self-similarity concept and the fractal dimension is not necessary
an integer. Fractal geometry had been successfully used in many applications [22-29].
In the use of the fractal models, the most important procedure is to measure the
fractal parameter (or the Hurst coecient). The fractal parameter is related to the
fractal dimension and characterizes the surface roughness of natural textures. A fractal
analysis method for estimating the fractal parameter is called box-counting [30], but some
literatures mention that box-counting does not perform very well [31,32].
Among the fractal models, a generalized form of ordinary Brownian motion, which is
called Fractional Brownian Motion (FBM), is proposed [33,34]. FBM is one of the most
useful fractal models for the random fractals found in nature. And FBM is widely used to
analyze data in many applications, such as random fractals in nature [27], target detection
of radar techniques [35] and bone radiographic images [24]. It has also been shown that
FBM is suitable for digital images. Since the probability density function of FBM is
well known, a Maximum Likelihood Estimator (MLE) has been used to estimate the
fractal parameter on a self-similar texture image [32]. Later a Fourier-domain Maximum
Likelihood Estimator (FDMLE) is developed to save extensive computations based on
MLE [36].
FBM model and FDMLE technique have been applied to inspect the surface defects of
textile fabrics (defect segmentation) [37]. However, FBM and FDMLE methods do not
deal the problem of the surface texture characterization of fiber.
The surface texture of an individual fiber can be classified and recognized by using a
specific parameter. The fractal parameter (Hurst coecient) is a proper value to describe
the texture of a fiber surface. In this paper, we use FBM to model the texture of the fiber
surface. And we apply a FDMLE technique to calculate the fractal parameter of FBM.
The fractal parameter can be used to characterize the surface texture of the fiber.
2. Fractional Brownian Motion Model. Fractional Brownian Motion (FBM), BH (t),

was first proposed by Mandelbrot [33]. In the FBM model, the most important parameter
SURFACE TEXTURE CHARACTERIZATION OF FIBERS 1313
is the fractal parameter, H, which is also called Hurst coecient and it ranges from zero
to one. FBM is a non-stationary zero-mean Gaussian random function, and it is a single
variable Brownian motion. FBM is an ordinary Brownian motion when H = 0.5. The
path of BH (t) is smoother when H is larger than 0.5 and it is rougher when H is smaller
than 0.5. Three paths of BH (t) with dierent H are shown in Figures 1(a), 1(b) and 1(c),
respectively.
FBM is a non-stationary process and it is not convenient to analyze. The increment of
FBM is called Fractional Gaussian Noises (FGN), and it is a stationary process, therefore,
is more suitable for mathematical analysis than FBM.
The sampled data (the image of texture) is a discrete FBM model and it can be denoted
as BH [n] (n means the nth data), and the discrete FGN (DFGN) [32] can be written as
xH [n] = BH [n] BH [n 1]. (1)
And the autocorrelation of DFGN can be derived as
2
rxH [n] = (|n + 1|2H + |n 1|2H 2|n|2H ), (2)
2
where n means the nth data and 2 is the variance of xH .
(a) (b) (c)
Figure 1. The example signals of FBM with dierent value of H: (a)

H = 0.3, BH (t) is rougher; (b) H = 0.5, BH (t) is an ordinary Brownian
motion; (c) H = 0.7, BH (t) is smoother
3. Fourier-domain Maximum Likelihood Estimator. Since the probability density

function of FBM (and FGN) is well known, the Hurst coecient (H) can be obtained by
a Maximum Likelihood Estimator (MLE) [32]. However, the computation time of MLE
is huge and it is not suitable for practical applications. In this section, we briefly sum-
marize a faster method that is called the Fourier-domain Maximum Likelihood Estimator
(FDMLE) [36,37] for computing the Hurst coecient.
According to Equation (1), the discrete-time Fourier transform of DFGN is denoted as
XH , and each element in XH is written as following:

N 1
XH (u) = xH [n]ej2un/N , (3)
n=0
where N is the number of the data and u = 0, 1, . . . , N 1. The auto-covariance of XH
can be written as

N
1 N 1
T
E[XH (u)XH (v) ] = rxH [m n]ej2(um/N vn/N ) , (4)
m=0 n=0
where N is the number of the data and u,v = 0, 1, . . . , N 1. The value of E[XH (u)XH (v)T ]
is real because rxH [n] = rxH [n]. Then we can define a matrix, H , and each element in
H is written as
apq = E[XH (p)XH (q)T ], (5)
where p, q = 0, 1, . . . , N 1, and H can be approximated a diagonal matrix [36]. Because
of linearity, we can assume that the discrete-time Fourier transform of DFGN still has
a Gaussian distribution, and the probability density function, pro(), of DFGN can be
written as { }
1 1 1
pro(XH ; H) = N 1 exp XH H XH ,
T
(6)
(2) 2 |H | 2 2
where H is the auto-covariance matrix given in Equation (5). After log-likelihood cal-
culation of Equation (6), we can obtain
( )
N N XH 1 T
H XH N 1
log pro(XH ; H) = log 2 log log | H | , (7)
2 2 N 2 2
where H = 12 H . We obtain H , which has to be inverted, and its determinant has
to be computed. We can approximate H as a non-zero diagonal matrix and obtain its
inverse and determinant directly.
We use the gray level variation of image to compute values of the left side of Equation
(7), log pro(XH ; H), with H values varying from 0.01 to 0.99, then we can obtain the
estimated H which is with the maximum value of log pro(XH ; H).
In the practical application, we select a size of a window in the image of the fiber surface.
We assume that the selected size is M M , in the other words, there are M vertical lines
and M horizontal lines in the window. Each line is used to estimate a fractal parameter,
H. The H value of the selected window is represented by the mean of H values of all
lines on the window. We take Figure 2 for an example: Figure 2(a) is an Acrylic fiber
image. We select a window on a fiber surface. The selected window is shown in Figure
2(b) and the size of selected window is 25 25. In the other words, there are 25 vertical
lines and 25 horizontal lines in the window. Vertical lines are denoted as v1, v2, . . . , v25
and horizontal lines are denoted as h1, h2, . . . , h25. We extract two vertical lines (the
1st and 17th) and two horizontal lines (the 2nd and 19th) to show the gray level variation
(as shown in Figures 2(c) 2(f)). The estimated H of each line and the mean H of the
selected window are shown in Table 1. The mean H of the selected window is used to
characterize the surface texture of the fiber.
Table 1. The estimated H of each line and the mean value H of the
selected window (see Figure 2(b))
the vertical line v1 v2 v3 v4 v5 v6 v7 v8 v9 v10 v11 v12 v13

estimated H 0.87 0.85 0.86 0.85 0.87 0.88 0.91 0.87 0.89 0.88 0.87 0.88 0.89
the vertical line v14 v15 v16 v17 v18 v19 v20 v21 v22 v23 v24 v25
estimated H 0.86 0.91 0.89 0.86 0.91 0.91 0.89 0.89 0.91 0.86 0.91 0.91
the horizontal line h1 h2 h3 h4 h5 h6 h7 h8 h9 h10 h11 h12 h13
estimated H 0.68 0.83 0.86 0.81 0.87 0.87 0.77 0.74 0.67 0.63 0.65 0.89 0.86
the horizontal line h14 h15 h16 h17 h18 h19 h20 h21 h22 h23 h24 h25
estimated H 0.86 0.89 0.75 0.77 0.86 0.82 0.64 0.82 0.72 0.65 0.72 0.86
mean value H of the selected window: 0.8314
(a) (b)
(c) (d)
(e) (f)
Figure 2. The example for selected window and each line of the selected
windows: (a) the original image; (b) the selected window (the black rectan-
gle is the selected window and the image is lighter for showing the selected
windows); (c) the 1st vertical line (v1) of the selected window; (d) the 17th
vertical line (v17) of the selected window; (e) the 2nd horizontal line (h2)
of the selected window; (f) the 19th horizontal line (h19) of the selected
window
4. Experiments.
4.1. Computer generated fractal images. In order to demonstrate the performance of
the Fractional Brownian motion model with Fourier-domain maximum likelihood estima-
tor (FDMLE), we apply the Cholesky decomposition method to generate two-dimensional
synthetic fractal images [38] and use those images to test the FDMLE method. Nine
32 32 images with dierent H values (0.1 0.9) are obtained (see Figure 3). We choose
64 lines (1D data, 32 rows, 32 columns) for each image, and apply the FDMLE method
(Section 2.3) to compute their mean H values. The results are shown in Table 2. We can
see that the FDMLE method provides a good estimation of those synthetic images.
H = 0.1 H = 0.2 H = 0.3 H = 0.4 H = 0.5
H = 0.6 H = 0.7 H = 0.8 H = 0.9

Figure 3. The nine 32 32 synthetic fractal images generated by the
Cholesky decomposition method, H = 0.1 0.9
Table 2. Results from the application of the FDMLE method to the syn-
thetic fractal images (see Figure 3)
True H values 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Estimated H 0.095 0.192 0.291 0.390 0.490 0.587 0.685 0.778 0.875
4.2. Classification of synthetic fibers. Two dierent Acrylic fiber images are taken
with a microscope (with 250): Sample A1 is the Acrylic with two-component, 3 denier
per filament and semi-dull luster (Figure 4(a)); Sample A2 is the Acrylic with modified wet
spun, 3 denier per filament and semi-dull luster (Figure 4(b)). We select seven windows
on the surface of each fiber. The selected windows of Figures 4(a) and 4(b) are shown
in Figures 4(c) and 4(d), respectively. We apply the FDMLE technique to estimate the
H value of each window, the results are shown in Table 3. The H range of Sample A1
and Sample A2 are from 0.82 to 0.86 and 0.70 to 0.75, respectively. From the experiment
results, the dierently processed Acrylic fibers can be classified by the H value.
Table 3. The results of Acrylic fibers (Figure 4)
Sample selected window estimated H range of H

A1W1 0.8200002
A1W2 0.8310000
A1W3 0.8604000 0.8200002
Sample A1 A1W4 0.8447999 |
A1W5 0.8502001 0.8604000
A1W6 0.8482001
A1W7 0.8487999
A2W1 0.7011254
A2W2 0.7402000
A2W3 0.7480000 0.7011254
Sample A2 A2W4 0.7270001 |
A2W5 0.7204000 0.7513999h
A2W6 0.7513999
A2W7 0.7316000
(a) (b)
(c) (d)
Figure 4. The experiment of Acrylic fibers: (a) Sample A1: two-

component, 3 denier per filament and semi-dull luster; (b) Sample A2:
modified wet spun, 3 denier per filament and semi-dull luster; (c) and (d)
are the selected windows of (a) and (b), respectively (the black rectangle is
the selected window and the image is lighter for showing the selected win-
dows)
4.3. Classification of animal fibers. There are two animal fibers: Sample W1 is mo-
hair (Figure 5(a)) and Sample W2 is merino (Figure 5(b)). The selected windows of
Figures 5(a) and 5(b) are shown in Figures 5(c) and 5(d), respectively. The H value of
each window and the H range of each fiber are shown in Table 4. From the experimental
results, we can see that the H value is dierent with dierent fiber of wool, in other words,
we can classify dierent types of animal fibers using the FDMLE technique.
Table 4. The results of animal fibers (Figure 5)
Sample Selected window estimated H range of H

W1W1 0.3985000
W1W2 0.3696186
0.3696186
W1W3 0.3756666
Sample W1 |
W1W4 0.3754167
0.3985000
W1W5 0.3837629
W1W6 0.3915001
W2W1 0.6399423 0.6147701
Sample W2 W2W2 0.6259195 |
W2W3 0.6147701 0.6399423
5. Conclusions. In this paper, we first summarized the definitions of FBM (modeling the
surface) and the FDMLE technique (estimating the H value). Then we have applied the
(a) (b)
(c) (d)
Figure 5. The experiment of animal fibers: (a) Sample W1 (mohair);

(b) Sample W2 (merino); (c) the selected windows of (a); (d) the selected
windows of (b) (the black rectangle is the selected window and the image
is lighter for showing the selected windows)
FDMLE technique to calculate the H value (fractal parameter) of the fiber surface. Based
on experimental results, now we can objectively classify dierent types of fibers. This
method can be applied not only in animal fibers, but also to synthetic fibers. The fiber
identification can be used not only in the forensic science, but also in wool manufacture,
textile processing, archeology and zoology.
Acknowledgment. This research was partially supported by the National Science Coun-
cil of Taiwan under contracts NSC 96-2221-E-324-018 and NSC 97-2221-E-324-041. The
authors would like to thank Taiwan Textile Research Institute for providing the fiber
image samples.
REFERENCES
[1] Y.-N. Chung, T.-C. Hsu, M.-L. Li, T.-S. Pan and C.-H. Hsu, A dynamic multiple-model estimator
and neural algorithm for radar system, International Journal of Innovative Computing, Information
and Control, vol.5, no.12(B), pp.4809-4818, 2009.
[2] T. Matsuo, S. Wada and H. Suemitsu, Model-based and non-model-based velocity estimators for
mobile robots, International Journal of Innovative Computing, Information and Control, vol.4, no.12,
pp.3123-3135, 2008.
[3] C.-W. Liao, C.-Y. Lu, K.-Y. Zheng and C.-C. Ting, A delay-dependent approach to design state
estimator for discrete stochastic recurrent neural network with interval time-varying delays, ICIC
[4] A. Hussain, M. A. Jaar, F. Jabeen and A. M. Mirza, Impulse noise removal using robust statistical
estimators and fuzzy logic, ICIC Express Letters, vol.3, no.4(A), pp.1119-1124, 2009.
[5] L. X. Kong, F. H. She, S. Nahavandi and A. Z. Kouzani, Feature extraction for animal fiber identi-
fication, Proc. of SPIE, vol.4875, pp.699-704, 2002.
[6] S. Fujishige and Y. Koshiba, Identifying cashmere and merino wool fibers, Textile Research Journal,
vol.67, pp.619-620, 1997.
[7] L. X. Kong and F. H. She, Image processing and pattern recognition in textiles, Proc. of SPIE,
vol.4552, pp.1-6, 2001.
[8] L. X. Kong, F. H. She, S. Nahavandi and A. Z. Kouzani, Fuzzy pattern recognition and classification
of animal fibers, IFSA World Congress and 20th NAFIPS International Conference, vol.2, pp.1050-
1055, 2001.
[9] D. Robson, Animal fiber analysis using imaging techniques, part I: Scale pattern data, Textile Re-
search Journal, vol.67, pp.747-752, 1997.
[10] F. H. She, L. X. Kong, S. Nahavandi and A. Z. Kouzani, Intelligent animal fiber classification with
artificial neural networks, Textile Research Journal, vol.72, pp.594-600, 2002.
[11] X.-J. Shi, Z.-H. Yuan, G. Liu and W.-D. Yu, Animal fiber classification based on fuzzy neural
network, Proc. of the 5th International Conference on Fuzzy Systems and Knowledge Discovery,
vol.1, pp.195-199, 2008.
[12] D. Chetverikov, Texture imperfections, Pattern Recognition Letters, vol.6, no.1, pp.45-50, 1987.
[13] A. L. Amet, A. Ert uz
un and A. Ercil, Texture defect detection using domain co-occurrence matrix,
IEEE Southwest Symposium on Image Analysis and Interpretation, pp.205-210, 1998.
[14] J. J. Warren, J. G. Stephen and H. Potlapalli, Texture characterization and defect detection using
adaptive wavelets, Optical Engineering, vol.35, no.11, pp.3140-3149, 1996.
[15] P. B. Chapple, D. C. Bertilone, R. S. Caprari and G. N. Newsam, Stochastic model-based processing
for detection of small targets in non-gaussian natural imagery, IEEE Trans. on Image Processing,
vol.10, no.4, pp.554-564, 2001.
[16] R. Paget and I. D. Longsta, Texture synthesis via a noncausal nonparametric multiscale Markov
random field, IEEE Trans. on Image Processing, vol.7, no.6, pp.925-931, 1998.
[17] H. S. Don, K. S. Fu, C. R. Liu and W. C. Lin, Metal surface inspection using image processing
techniques, IEEE Trans. on System, Man, and Cybernetic, vol.14, no.1, pp.139-146, 1984.
[18] J. S. Weszka, C. R. Dyer and A. Rosenfeld, A comparative study of texture measures for terrain
classification, IEEE Trans. on Systems, Man, and Cybernetics, vol.6, no.4, pp.269-285, 1976.
[19] W. J. Jasper, S. J. Garnier and H. Potlapalli, Texture characterization and defect detection using
adaptive wavelets, Optical Engineering, vol.35, no.11, pp.3140-3148, 1996.
[20] C.-Y. Wen, S.-H. Chiu, W.-S. Hsu and G.-H. Hsu, Defect segmentation of texture images with wavelet
transform and co-occurrence matrix, Textile Research Journal, vol.71, no.8, pp.743-749, 2001.
[21] B. B. Mandelbrot, The Fractal Geometry of Nature, W. H. Freeman, New York, 1983.
[22] L. O. Chua, V. I. Sbitnev and S. Yoon, A nonlinear dynamics perspective of wolframs new kind
of science part V: Fractals everywhere, International Journal of Pattern Recognition and Artificial
Intelligence, vol.15, no.12, pp.3701-3849, 2005.
[23] B. Daudert and M. L. Lapidus, Localization on snowflake domains, World Scientific Journals, vol.15,
no.3, pp.255-272, 2007.
[24] C. Fortin, R. Kumaresan, W. Ohley and S. Hoefer, Fractal dimension in the analysis of medical
images, IEEE Engineering in Medicine and Biology, vol.6, pp.65-71, 1992.
[25] Y. Muraoda, K. Inoue, H. Tagaya and K. Nishizawa, Fiber crimp analysis by fractal dimension,
Textile Research Journal, vol.65, pp.454-460, 1995.
[26] G. Di Martino, A. Iodice, D. Riccio and G. Ruello, A novel approach for disaster monitoring: Fractal
models and tools, IEEE Transactions on Geoscience and Remote Sensing, vol.45, no.6, pp.1559-1570,
2007.
[27] A. P. Pentland, Fractal-based description of natural scenes, IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol.6, pp.422-436, 1984.
[28] Q. Zhang and L. Gao, Medical image retrieval based on nonsubsampled contourlet transform and
fractal dimension, Proc. of the 3rd International Conference on Bioinformatics and Biomedical En-
gineering, pp.1-4, 2009.
[29] D. Zhang, A. Samal and J. R. Brandle, A method for estimating fractal dimension of tree crowns
from digital images, International Journal of Pattern Recognition and Artificial Intelligence, vol.21,
no.3, pp.561-572, 2007.
[30] K. Falconer, Fractal Geometry, John Wiley & Sons, New York, 1990.
[31] Q. Huang, J. R. Lorch and R. C. Dubes, Can the fractal dimension of images be measured? Pattern
Recognition, vol.27, no.3, pp.339-349, 1994.
[32] T. Lundahl, W. J. Ohley, S. M. Kay and R. Siert, Fractional Brownian motion: A maximum
likelihood estimator and its application to image texture, IEEE Trans. on Medical Imaging, vol.5,
no.3, pp.152-161, 1986.
[33] B. B. Mandelbrot and J. W. Ness, Fractional Brownian motion, fractional noises and applications,
SIAM Rev., vol.10, no.4, pp.422-436, 1968.
[34] R. F. Voss, Characterization and measurement of random fractals, Physica Scripta, vol.13, pp.27-32,
1986.
[35] C. V. Stewart, B. Moghaddam, K. J. Hintz and L. M. Novak, Fractional Brownian motion models
for synthetic aperture radar imagery scene segmentation, Proc. of the IEEE, vol.81, pp.1511-1522,
1993.
[36] C.-Y. Wen and R. Acharya, Self-similar texture characterization using a Fourier-domain maximum
likelihood estimation method, Pattern Recognition Letters, vol.19, pp.735-739, 1998.
[37] S.-H. Chiu, S. Chou, J.-J. Liaw and C.-Y. Wen, Textural defect segmentation using Fourier-domain
maximum likelihood estimation method, Textile Research Journal, vol.72, pp.253-258, 2002.
c ISSN 1349-4198
POWER GRID NODE AND LINE DELTA CENTRALITY MEASURES

FOR SELECTION OF CRITICAL LINES IN TERMS OF BLACKOUTS
WITH CASCADING FAILURES
Hwachang Song1 , Rodel D. Dosano2 and Byongjun Lee3

1
Seoul National University of Science and Technology
Gongreung-Dong, Nowon-Gu, Seoul 139-743, Republic of Korea
hcsong@seoultech.ac.kr
2
School of Electronic Engineering and Information Engineering
Kunsan National University
Miryong-Dong, Kunsan, Jeonbuk 573-701, Republic of Korea
rodel dosano@kunsan.ac.kr
3
School of Electrical Engineering
Korea University
Anam-Dong, Sungbuk-Gu, Seoul 136-701, Republic of Korea
leeb@korea.ac.kr
Abstract. To investigate and evaluate the mechanism of how cascading failures spread
in power grids, this paper presents the concept of network centrality for power systems
and related two local indices. These indices can provide look-ahead situation awareness
for operators concerning impending large power grid blackouts. These indices are also
easy to interpret and can be used to identify the location of critical edge components in
terms of blackouts with cascading failures. In this paper, results from the application of
the indices to 118-bus test power system are presented.
Keywords: Cascading failures, Complex network theory, Delta centrality index, Multi-
ple contingency, Power grid blackouts
1. Introduction. Blackout history in [1,2] and descriptions of the 2003 blackout in the
U.S. and Canada [3] indicated a crucial warning in power system vulnerability that led
to even bigger and more damaging outages. This is a major crisis arising from those
phenomena, probably forcing catastrophic events to occur that would have otherwise
been unexpected if adequate control schemes had been applied in advance. In the last
few years, the level of allowable capacity in a power system has been reduced; mostly due
to the deregulation of the power industry, which focused more on economic features of
power system operation and planning. Currently, the system is already at the point where
the need to install more generating facilities for sustainable energy resources is highlighted.
Thus, system planning and related implementations that take into consideration possible
catastrophic events due to structural and energy resource deficits are very important.
In the literature, several works on widespread blackouts in power systems have been
investigated using various models. One type is based on the cascading failure model, most
of which are based on artificial systems with the intent of understanding the blackout
process, such as the OPA model [4], CASCADE model [5], branching process model [6]
and hidden failure model [7]. These models have previously gained a great deal of interest.
Due to the continued growth of the importance of the topic, many other aspects of network
1321
1322 H. SONG, R. D. DOSANO AND B. LEE
structure have begun to attract the attention in its applications. By employing complex
network theory in critical structures such as in [8,9], the authors have found out that
there are substantial non-trivial topological features and interdependency among several
components in the systems, which are neither purely regular nor purely random. However,
the blackout models were proposed to better understand the process of cascading tripping
events and the eects on power system security.
To study how failures spread in the given power grids, in this paper, network centrality
based on both system responses and topological features is formulated. For this purpose,
two local indices for network centrality on the possible power system network evolutions
are introduced, which can be concluded through system topological changes such as cas-
cading events. The objective of this paper is to provide the properties of local node and
line centrality measures on system vulnerability, which make it relatively simple to inter-
pret the mechanism of cascading events that threaten system security. These measures
can be further employed to select critical lines in terms of grid blackouts with cascading
failures and to generate sets of multiple contingencies which result in catastrophic failures
in a system. The generated sets of critical multiple contingencies are essential to system
planning and assessment, and in operations; they could be used for alternative decision
support for the establishment of adequate countermeasures.
2. Contingency Selection Method with Respect to Multiple Contingencies. In

power system security assessment, contingency selection and evaluation plays an impor-
tant role [10]. Usually it is a two-stage process in which the primary objective is to
separate critical contingencies from all possible contingencies and to determine their im-
pact on the system. Existing methods proposed in the literature [11,12] for contingency
selections suer from the curse of having substantial number of combinations included
in the computations. These methods work well for most single contingencies but may
fail when dealing with multiple contingencies and compounded contingencies. For fast
contingency selection, basically considering only N -1 contingency, the use of performance
indices have proven to be a promising, accurate and ecient method [13].
The generation of N -k contingencies in a shorter time that include additional opera-
tional constraints and controls is a big challenge in this work. In [14], to find potentially
harmful lines with high edge betweenness centrality in the power grids, a Cray XMT
system is applied for parallel computation because of its computational burden. For the
purpose, in this paper, two local indices are proposed to locate the weak areas of a power
system and to identify lines which are becoming critical when applying cascading tripping
events of them. These indices employ the performance-index technique together with the
concept of complex network theory. Furthering, to overcome the problem of dimensional-
ity, pre-selections are performed to generate sets of multiple contingencies employing the
properties of the indices variation in an evolving system.
3. Power System Network Centrality Indices. Mapping the power grid to a topol-
ogy graph is the initial step which can be analyzed using the complex network theory
[15-17]. In the references related to graph theory [7-9,18], an electric power network can
be represented as a valued undirected/directed graph G, with N nodes and K edges.
Network connection points such as substations for generators and loads are represented
as nodes or vertices while the zero potential points (the ground) are not directly repre-
sented. Branch-paths of transmission lines and transformers are considered as the edges.
The graph G can be further described by the N by N adjacent matrix, whose (i, j) el-
ement is eij . If a transmission line exists between bus i and bus j, eij has a value of (0,
POWER GRID NODE AND LINE DELTA CENTRALITY MEASURES 1323
1]; otherwise, eij = 0. In this mode, the adjacent matrix includes the information of the
network topology at the given state.
By a slight modification of the adjacent matrix, further information of line loading
ratios or eciency of transmission lines can be incorporated. If the value of eij is 1, it is
said to be 100% ecient, and if less than 1, the line experiences overload. Adapting the
model proposed in [9,18] considering the power grid as a dynamic mode, the topology of
which may vary by such tripping events, the graph is then regarded as being weighted
and its magnitude changes by the system response. The dynamic evolution of the system
can be viewed from the values of eij . Initially at time t = 0, value of eij = 1 for all active
lines, means all lines are perfectly working and there is no overload. Then the index of
line eciency for the (i, j) component can be computed as follows:
{
eij (0) LCi (t)
i
, if Li (t) > Ci
eij (t + 1) = (1)
eij (0), if Li (t) Ci
where Li (t) is the actual line power flow and Ci is the rated capacity of the transmission
line at time t. In the opposite sense, the reciprocal of eij can be used as a measure of line
severity.
To precisely model the behavior of the power grid in terms of cascading events, power
system structure alone is not enough to give a complete overview of the process that
is caused in the system when a disturbance occurs. Some other important and unique
features of the power grid must be included, such as redistribution and direction of line
flows. These changes in the information flow between each component and the system
structures can easily be viewed through the concept of centrality. The idea of centrality
was first introduced in [19] to study social behavior in a small group of people. Since
then, various measures of centrality have been proposed to evaluate the influence of an
individual in a social network. Standard centrality can be divided into two classes: (i)
those based on the relevance of distance between individuals and (ii) those based on how
the node acts as intermediary among others. Degree and closeness are measures of the
first kind, flow betweenness is a measure of the second kind. One of the contributions of
this study is to introduce measure on power grid buses and transmission lines based on
the idea of structural and information centrality.
This paper employs the concept of delta centrality [20], which is a generic measure of
classical centralities. Its main concept is based on the idea of how important a node (an
individual) is when the network (social network) responds to the deactivation of one of its
components. Given a graph G(V, E) representing an electric network, the delta centrality
( centrality) of node i can be defined as:
P [G] P [G ] (P )i
Ci = = (2)
P [G] P
where P is a quantity for the relationship of the whole graph as a system index and P
is the variation of P under the deactivation of the i-th component in the graph. Delta
centrality can be a measure of any of the standard centrality measures. The eectiveness
of the centrality measure depends on the choice of the system parameters to evaluate the
above equation. As an example, taking P as a measure of eciency E of the network, this
quantity represents how well the node on the network transfers power with other nodes.
The network eciency E can be defined by the equation:
1
E[G] = eij (3)
n(n 1) i=jG
where n is the total number of branches in the graph. It is noted that E is a system
index summing the eciencies of the whole branches. The delta centrality measure on
each node can be calculated as a measure of eciency of each line as:
E[G] E[G ] (E)i
CiE = = (4)
E[G] E
Other centrality measures [15] that can be used for power systems are degree centrality
(refers to the number of lines incident to a given bus) and closeness centrality (refers to
the distance between buses).
To view the spread of severe nodes during a cascading process, this paper adopts the
following local node centrality index:
1
N Cm = ejm (5)
Km (Km 1) j=Sm
where m is the node of interest; Km indicates the number of edges linked to node m; Sm
stands for the set of neighboring nodes from node m; and ejm denotes the line eciency
of line j-m.
During the deactivation of important lines from the given set of cascading events one
by one, the local node index of (5) for the total number of buses is calculated. Then, they
are compared with the previous values in order to evaluate node delta centrality. From
the rigorous simulation of all system components on these indices, we could obtain certain
properties, like locating which areas/parts of a system are becoming critical and identify
lines which are essential during the blackout process. These properties of node index will
serve as guides to distinguish sets of sequential outages resulting in system blackout.
The property for critical lines using the local node index, however, is too complicated
to generate dierent blackout scenarios from the seed cascading events. By taking the
advantages of combining structural properties of an evolving system (topological infor-
mation), magnitude of line flows and the role of intermediary to other components, a
comprehensive line index is formulated. The local line centrality index is further applied,
on this paper, and the formulation of it is as follows, together with Figure 1:
1 ejm Nk,Le
LC u = ekj (6)
Nj ej,Le kSjLe Nk
where the superscript u represents the evaluated line j-m; j is for the sending end of the
line of interest; Nj is the number of the neighboring lines from node j; ejm and ekj stands
for the eciency of line j-m and k-j from the adjacent matrix, respectively; ej,Le is the
sum of the line eciencies of the neighboring lines with entering power flows into node j;
Nk is the number of neighboring lines from node k; Nk,Le is the number of the neighboring
lines whose flows enter node k; SjLe denotes the set of lines with entering flows to node
j.
Figure 1 illustrates a node (node j in region I) with incoming edges from its neighboring
nodes (in region II) carrying the information of how they are connected to other nodes
(in region III). To evaluate the line criticality of system, node j is visited first and only
edges that have outgoing direction are considered for evaluation. After visiting all nodes,
the line index of all edges are also completely evaluated because incoming edges of the
given nodes are outgoing edges of their neighboring nodes.
In (6), the term ekj Nk,Le /Nk is used to emphasize the lines in region II directly
connected to node k which have more entering branches, and 1/ej,Le is to normalize
the impact of the last term. This conjecture is based on the idea that the removal of a
line whose sending node with many entering branches to the total neighboring lines may
Figure 1. Sub-graph for the evaluation of the line centrality index
not have a severe impact because power flows are easily distributed to the remaining active
lines. Also, the last term of (6) gives a higher score for neighboring lines with a larger
ratio of the number of entering lines to node k to the total number of the neighboring
lines. This term can be interpreted as the bridging coecient of node j with other nodes.
The term 1/Nj is included in the formulation to consider node j as important if it has
lesser branches connected.
From the authors experience, the outage of a system component can impact its neigh-
borhood in a way that the disturbance causes power transfer disturbances mainly up
to 3-level distance from its occurrence. To include the important concept of spreading
regions, the line centrality index of (6) can further be expressed as:
1 ejm 1 ekj Nl,Le
LC u = elk (7)
Nj ej,Le Nk l=1 k=1 ek,Le Nl
where elk is the eciency index of line l-k; ek,Le is the sum of the line loading indices
of all the edges with entering flows to node k; Nl is the number of edges connected to l;
Nk,Le is the number of edges entering k from neighboring nodes.
Then the line delta centrality with (7) can be defined, when the i-th system component
is tripped, as:
LC[G] LC[G ]
CiLC = . (8)
LC[G]
In this paper, the property of the line centrality index was also obtained in terms of
system blackouts. System blackouts are identified, in this paper, by checking out whether
power system operating point exists using the developed optimal power flow (OPF) [21].
The property of the line delta centrality index in selecting critical lines is much simpler
than that of the node delta centrality index using (5). The details of the properties will
be explained in the next section.
4. Simulation Results. The properties of the proposed two local indices with respect
to power system blackouts were obtained by the rigorous simulation and investigation on
the IEEE 118 bus test system [22]. For every case, it is assumed that all the line flows
are calculated with a conventional power flow for the base case operating solution. If no
overload is identified during the first run, an arbitrarily selected contingency is applied
as the initiating outage. For a contingency to be considered, the line to be tripped can
be selected from the following candidates: (1) lines which carry the largest amount of
load, (2) lines with low eciency, (3) lines which are frequently tripped and (4) randomly
selected lines. For the operating point evaluation for contingent cases, OPF is applied
to consider the control strategies available. Basically, all the o-diagonal elements eij in
the adjacent matrix are set to 1, unless the corresponding line experiences overload, so,
overloaded lines can be identified using eij after refreshing the adjacent matrix using the
condition (1). Then, lines with lowest eij are the next candidate lines to be deactivated.
As the number of line outages increases, the number of elements whose values are lower
than 1 in the adjacent matrix also gradually increases. For a case applying sequential
outages in the simulation, the line eciencies at each outage are shown in Table 1.
To view how disturbances spread in power systems, the local node centrality index of
(5) can be used to identify the parts which are becoming critical as the set of sequential
outages are applied. By monitoring the values of the node centrality index, the areas of
the power system infected as each line component is deactivated can be visualized. First
an initiating line outage occurs. The line is one of the double lines (49 54). The buses
with high node centrality index values are clustered based on the areas they are located.
The buses which are connected directly to these buses with a high node centrality index
are then considered as being in the same cluster. Line 34 36 is the next candidate to
be tripped, which carries the highest carrying current at the condition. Then new sets
of clustered buses are identified, and most of them are added to the existing clustered
buses or a generation of new clusters. Lines 16 17 and 38 37 are the next lines to be
deactivated.
Table 1. Line eciencies of severe lines due to a set of multiple contingencies
Line eij
from to Outage 1 Outage 2 Outage 3 Outage 4
16 17 0.841731
38 37 1 0.530291
26 25 1 0.924437 0.774494 0.773054
25 27 1 1 1 0.777011
63 59 1 1 0.989217 0.970031
68 116 1 0.841731 0.924437 0.91393
89 92 1 1 0.841731 0.841816
Figure 2 shows the clusters of the buses after applying the set of outages. Based on
the results and after further verification of the authors, the following summary can be
deduced in selecting high-risk line outages using the node centrality index: (i) lines linking
two separate clusters are mostly critical lines, the sequential outage of which could lead to
a system collapse; (ii) lines whose outage could divide the pre-existing cluster into parts
are severe; and (iii) lines directly connecting a cluster which are reluctant to shift their
line flows to neighboring lines are severe.
Examples of condition (i) mentioned above are lines 70 71, 24 71, 43 44, 44
45 and 16 17. These lines connect newly form clusters (smaller one) after an outage.
Examples of (ii) are lines 49 69, 94 96 and 89 92. These lines, if they are tripped, will
separate large cluster into smaller clusters. Both condition (i) and (ii) stem from the fact
that if clusters of nodes are separated, the transfer of power from one area to the other
area which lack generation will be limited. Examples of condition (iii) refer to those lines
connected to clusters of nodes with constant line flows. Mostly, these lines have already
reached or almost within their line capacity.
Detecting sets of critical outages from an evolved power system can be forecasted using
the line delta centrality of (8), CiLC . For the line delta centrality as the sequential outages
occur in the system, the line centrality of each line is computed using Equation (7). The
Figure 2. Clusters of weak buses identified with node centrality index
values of the line delta centrality are then plotted with respect to its time of occurrence.
Figure 3 shows examples of line delta centrality variation with respect to the case of
applying a sequence of events.
Figure 3. Examples of line delta centrality variation
As shown in Figure 3, the critical lines are line ID 6, 7, 8, 9, 10, 11, 12, 13 and 15. All
the critical lines except line 15 have a gradual increase in their line delta centrality and
quite a large variation right before blackout. By the observation of the authors simulation
results, the properties of the line delta centrality can be summarized as follows: i) lines
with increasing values of line centrality index are definitely critical components whose
tripping would cause blackouts; and ii) lines are critical which are located on the areas
of a cluster with a line centrality index which is dormant to the change with neighboring
disturbances.
An initiating outage and consecutive line tripping events serve as the basis for generat-
ing the list of multiple contingency. Critical lines are chosen by observing the variation of
the values of the line delta centrality index. Members of the N -k contingency generation
basin are composed of the initiating line outage, consecutive line outages and combination
with one from the selected critical lines. All high-risk N -1 contingencies, resulting in the
reduction of the system eciency, are also included in the reduced contingency basin.
Consider an example with the initiating outage of line 49-54 and two consecutive events
(lines 16 17 and 38 37). From this seed contingency, 65 critical lines are found through
the variation of line delta centrality.
By simulating other power system evolving scenarios from dierent initiating events,
other sets high risk N -k contingencies can be generated. Then, the list of high risk N -
k contingencies is updated. Important factors aecting system evolution scenarios may
include line tripping, load shedding and generation dispatch, and they can be taken as
priorities in order to create worst case system conditions.
A verification of the results of the proposed algorithm mentioned above is evaluated
further. Using the same procedures, and by individually tripping each line to initiate an
outage for each blackout scenario generated, dierent sets of critical lines are generated
in each case. The purpose of the verification is to observe the correlations of critical lines
with dierent initiating line outage.
Table 2 shows the top 30 lines with high scores as critical lines based from dierent
initiating outages and with the same sequence of line tripping. The correlation of lines
with high scores can be observed in Figure 4. The selected critical lines from dierent
scenarios are closely correlated, and most of the critical lines determined from other
initiating outages are also included in the sets of critical lines which were previously
determined. The combination of line outages are made up of lines from critical lines
previously determined and the initiating outage used. The degree of severity in each line
varies depending on the combinations of critical lines chosen.
Table 2. Line eciencies of severe lines due to a set of multiple contingencies
Line ID from to Counts Line ID from to Counts

114 68 69 161 127 75 118 153
151 89 90 161 111 65 66 151
118 69 75 160 115 68 81 151
119 69 77 160 44 26 30 149
139 80 98 160 122 70 75 146
137 80 96 159 167 98 100 145
138 80 97 159 68 42 49 144
140 80 99 159 81 49 54 143
117 69 70 158 129 76 118 139
148 85 89 157 112 65 68 134
155 90 91 157 80 49 51 124
15 9 10 156 43 25 27 120
150 88 89 156 107 62 66 116
77 47 69 155 39 23 32 112
85 49 69 155 36 22 23 97
For comparison, the results with an edge betweenness centrality index, slightly modified
from the one in [14,23], was obtained. In this simulation, source vertices are only from
the set of generating nodes and sink vertices are from that of load nodes, to reduce the
number of source-sink sets. In Table 2, the lines with the underlined ID are critical edges
identified using the proposed method and the modified edge centrality index. Dierence
in the results of the two algorithms can be summarized as follows:
In the viewpoint of implementing the edge betweenness centrality index, power sys-
tem network is assumed to have unweighted and undirected graphs. In the proposed
algorithm, line flows with respect to individual line capacities serve as weights of each
line. Power flow direction which is inherent in system evolution is being considered.
The proposed algorithm considers the system states after applying sequential line
trips to the system, while the edge centrality index only takes into account a single
sate. That is, the proposed algorithm is based on the pattern of line centrality index
of each line including system responses when applying a set of line outages.
Figure 4. Correlation of critical lines with high scores generated from

dierent initiating outages
In summary, identifying critical lines in terms of system blackout by cascading events

for bulk power systems becomes more prominent as the number of line outages increases.
This means that without pre-selecting critical lines, generating N -k contingencies, which
are needed for countermeasure determination in security assessment, imposes a severe
computation burden because of its combinatorial feature as described in [14]. By adopting
two local indices based on network centrality in the paper, dierent from system indices
in [8,9], it is possible to find such a line outage group that have the coherency from a
seed set of cascading events leading the system to blackout. Thus critical lines in terms
of cascading events can be more easily identified using the proposed local indices. From
the authors experience on identification of critical lines from other seed sets of cascading
events, applications of the line delta centrality index to critical line outage detections are
promising in fast multiple contingencies selection.
5. Conclusions. In this paper, local node and line centrality indices are proposed to
obtain information on how a blackout can spread through a network by a given set of
sequential multiple contingencies and which lines may dominantly participate in the seed
blackout scenario. Also, the properties of the two centrality indices were obtained, which
could be the basis for the generation of multiple contingency listing. The methodology is
capable of reducing the highly dimensional combination of higher order contingencies by
considering the characteristics of node and line centrality indices.
Acknowledgment. This work was supported by the Ministry of Knowledge Economy

(MKE) through the EIRC program at Korea University.
REFERENCES
[1] M. Schlapfer, Comparative case study on recent blackouts, Proc. of the 3rd EAPC/PfP Workshop
on Critical Infrastructure Protection and Civil Emergency Planning, Zurich, Swiss, 2005.
[2] CIGRE Working Group 34.08, Protection Against Voltage Collapse, CIGRE Brochure, 1998.
[3] Joint U.S.-Canada Power System Outage Task Force, Final Report on the August 14, 2003 Blackout
in the United States and Canada: Causes and Recommendations, 2004.
[4] I. Dobson, B. A. Carreras, V. E. Lynch and D. E. Newman, Complex system analysis of series
blackouts: Cascading failure, criticality and self-organization, Proc. of Bulk Power System, Dynamics
and Control-VI, Cortina dAmpezzo, Italy, 2004.
[5] I. Dobson, J. Chen, B. A. Carreras and D. E. Newman, Examining criticality of blackouts in power
system models with cascading events, Proc. of 2001 Hawaii Int. Conference on System Science,
Hawaii, 2001.
[6] I. Dobson, B. A. Carreras and D. E. Newman, A branching process approximation to cascading
load-dependent system failure, Proc. of 2004 Hawaii Int. Conference on System Science, Hawaii,
2004.
[7] S. Tamaronglak, Analysis of Power System Disturbances Due to Relay Hidden Failures, Ph.D. Thesis,
Virginia Polytechnic and State University, Blacksburg, 1994.
[8] P. Crucitti, V. Latora and M. Marchiori, Model for cascading failures in complex networks, Phys.
Rev. E, vol.69, 2004.
[9] P. Crucitti, V. Latora, M. Marchiori and A. Rapisarda, Eciency of scale-free networks: Error and
attack tolerance, Physica A, vol.320, pp.622-642, 2003.
[10] A. O. Ekwue, A review of automatic contingency selection algorithms for on-line security analysis,
IEEE Conf. Publ., vol.336, pp.152-155, 1991.
[11] R. A. Schlueter, J. E. Sekerke and A. G. Costi, Improved contingency measures for operation and
planning applications, IEEE Trans. on Power Systems, vol.4, no.4, pp.1430-1437, 1989.
[12] C. Fu and A. Bose, Contingency ranking based on severity indices in dynamic security analysis,
IEEE Trans. on Power Systems, vol.14, no.3, pp.980-986, 1999.
[13] J. Zhiao and H. D. Chiang, An enhanced contingency selection method with respect to multiple
contingencies for online voltage stability assessment, Proc. of 2006 International Conference on
Power System Technology, pp.1-6, 2006.
[14] S. Jin, Z. Huang, Y. Chen, D. Charvarria-Miranda, J. Feo and P. C. Wong, A novel application
of parallel betweenness centrality to power grid contingency analysis, Proc. of the 24th IEEE Int.
Parallel and Distributed Processing Symposium, Atlanta, 2010.
[15] A. Barrat, M. Barthelemy and A. Vespignari, Dynamical Processes on Complex Networks, Cambridge
University Press, New York, 2008.
[16] Y. Liu, J. Zhang and S. Yang, Research on consumers cross-site transfer behavior based on complex
network theory, ICIC Express Letters, vol.3, no.4(B), pp.1359-1364, 2009.
[17] D. Zhang, F. Long and S. Cai, The dynamic behavior of SIRS epidemic model in complex networks,
[18] V. Latora and M. Marchiori, Ecient behavior of small-world networks, Phys. Rev. Lett., vol.87,
2001.
[19] A. Bavelas, A mathematical group structure, Human Organization, vol.7, pp.16-30, 1948.
[20] V. Latora and M. Marchiori, A measure of centrality based on network eciency, New Journal of
Physics, vol.9, no.6, 2007.
[21] H. Song, B. Lee and Y.-H. Moon, Reactive optimal power flow incorporating margin enhancement
constraints with nonlinear interior point method, IEE Proc. of Gener. Trans. Dist., vol.152, no.6,
pp.961-968, 2005.
[22] http://www.ee.washington.edu/research/pstca/pf118/pg tca118bus.htm.
[23] M. Girvan and M. E. Newman, Community structure in social and biological networks, Proc. of the
National Academy of Sciences of the United States of America, vol.99, no.12, pp.7821-7826, 2002.
c ISSN 1349-4198
OPTIMIZING MIMO-SDMA SMART ANTENNAS BY USING LINEAR

ARRAY PHASE PERTURBATIONS BASED ON PARTICLE SWARM
OPTIMIZATION
Chao-Hsing Hsu
Department of Computer and Communication Engineering
Chienkuo Technology University
No. 1, Chieh-Shou N. Road., Changhua 500, Taiwan
chaohsinghsu@yahoo.com
Abstract. In this paper, based on phase-only perturbation method, an innovative up-

link Multiple-Intput Multiple-Output Spatial Division Multiple Access (MIMO-SDMA)
optimization technique of smart antennas by is proposed. Particle swarm optimization
(PSO) algorithm is used to search the optimal weighting vector of the phase shift pertur-
bations for array factor. The deduced radiation pattern formulas available for searching
optimal solutions are used to search the optimal weighting vector of the array factor of
a smart antenna. The design for an optimal radiation pattern of a smart antenna can
not only adjustably suppress interferers by placing nulls at the directions of the interfer-
ing sources but also provide maximum main lobes in the directions of the desired signals
at the same time, i.e., to maximize the Signal to Interference Ratio (SIR). In order to
achieve this goal, a new convergent method referred as the two-way convergent method for
particle swarm optimization is proposed. The optimal radiation pattern design of smart
antennas is studied by phase-only perturbation method to achieve uplink MIMO-SDM
optimization.
Keywords: Multiple-intput multiple-output, Spatial division multiple access, Smart
antennas, Linear phase array, Particle swarm optimization
1. Introduction. The advent of fast and low-cost digital signal processors has made
smart antennas practical for cellular land or satellite mobile communication systems.
Therefore, the study of smart antenna systems is becoming the interest of engineers and
researchers in the communication field. The goal is to improve the system performance by
increasing channel capacity and spectrum eciency, extend range coverage, steer multiple
beams to track more mobiles, suppress multipath and co-channel interferences for signal
propagations [1].
PSO algorithm combines the advantages of ecient heuristics incorporating domain
knowledge and population-based search approaches to solve the optimization problems.
In PSO algorithm, each single solution is a bird in the search space. It is called parti-
cle. All of particles have fitness values which are evaluated by the fitness function to be
optimized, and have velocities which direct the flying of the particles. The particles fly
through the problem space by the current optimum particles. Particle swarm optimiza-
tions are used to search the optimal weighting vector of the phase-only perturbations of
array factor [2]. Compared with nulling or main lobe designs, the optimal radiation pat-
tern design brings dierent convergent criteria for optimization problems [3-6]. It must
not only suppress the interferences in their directions but also enhance the main lobe
towards the desired signals direction. The simulation results show the eectiveness of
the particle swarm optimization for this kind of optimization problem. The optimization
1331
1332 C.-H. HSU
design of a smart antenna has been limited to adjust the main lobe of radiation pattern
towards the users. But, it is not good enough for the practical mobile communication
system.
In the modern wireless communication, it asks for high quality and good eciency. Base
station antennas use omnidirectional or sectored pattern, which could cause the power
waste in unexpected direction and interference for the others. A single antenna cannot
change the beam pattern without mechanism. The beam pattern can be changed if the
communication system has two or more antennas combined together. In order to have
high quality and good eciency, the practical mobile communication system must not
only adjustably maximize the main lobe toward the direction of the desired signals, but
also suppress interferences by placing nulls in the directions of interfering signals in the
smart antenna radiation pattern. Thus, Multiple-Input Multiple-Output Spatial Division
Multiple Access (MIMO-SDMA) can be really come true. In a smart antenna system,
there will be several signals use a co-channel at the same time. Therefore, these signals
will interfere with one another [7,8]. The uplink SDMA technique can oer the allocations
of multiple mobile users to a common communication channel with its spatial separation
ability and can avoid making interferences one another. Thereby, the system capacity
can be increased. The optimal radiation pattern method can make uplink MIMO-SDMA
optimization come true and is able to be satisfied with the practical mobile communication
system. In order to achieve this goal, a new convergent method referred as the two-way
convergent method for particle swarm optimizations is proposed. The system can be is
optimized via putting pattern nulls in interfering directions and maximizing the main
lobes power toward users so that the signal to interference ratio (SIR) is maximized. So,
the uplink smart antenna system possesses the optimal SIR and achieves Spatial Division
Multiple Access (SDMA) optimization [9,10].
2. Deduction of Radiation Pattern Array Factor Formula. Uplink system diagram

of a smart antenna is designed by phase perturbations in a linear array using particle
swarm optimization (PSO) as shown in Figure 1. For a linear array of 2N equispaced
sensor elements, a signal of wavelength from direction with respect to the array normal
impinges on any two adjacent element n and element n + 1 by the distance d as shown in
Figure 2. Obviously, the signal will reach element n first, then element n+1. The time
dierence can be expressed as follows [11]:
The is the propagation speed of radio wave. Then, there is a time delay as follows:
d sin
= (1)

The corresponds to a phase shift of 2
d sin .
2
= d sin = kd sin (2)

Then, for far field, the array factor of the adaptive linear array for receiver m of a smart
antenna can be written as
1
2N
AFm () = wmn ej(n1) , m = 1, 2, 3, . . ., M (3)
M n=1
where M is the number of users. wmn is the complex array weight at element n of receiver
m. It can be expressed as:
wmn = ejmn (4)
where : amplitude weight at element n for receiver m, mn : phase shift weight at element
n for receiver m.
OPTIMIZING MIMO-SDMA SMART ANTENNAS 1333
Figure 1. Uplink smart antenna system designed by phase-shift pertur-

bations using a particle swarm optimization
Figure 2. The incident signal reaching any two adjacent elements

1334 C.-H. HSU
If the reference point is set at the physical center of the array, the array factor can be
rewritten as:
1 1 j[(nN 0.5)+mn ]
2N 2N
j(nN 0.5)
AFm () = wmn e = e , m = 1, 2, 3, . . ., M (5)
M n=1 M n=1
Furthermore, assumed, amplitude weights are constant and phase shift weights are in odd
symmetry, Equation (5) can be simplified to
1
N
AFm () = 2 cos[(n 0.5) + mn ], m = 1, 2, 3, . . ., M (6)
M n=1
The array factor, given by (6), describes the model of the radiation pattern and is suitable
for optimal solution search. As only the real part is left, it is available for searching the
optimal solutions using optimization technique. For using the optimization technique, the
fitness function cannot include the imaginary part.
3. Particle Swarm Optimization for Optimal Pattern Design. In PSO algorithm,

each particle (individual) adjusts its flying according to its own flying experience and its
flying experience of companions. Each particle is treated as a point in a D-dimensional
space. The ith particle is represented as XI = (xi1 , xi2 , . . ., xiD ). The best previous
position (the position giving the best fitness value) of the ith particle is recorded and
represented as PI = (pi1 , pi2 , . . ., piD ). The index of the best particle among all the
particles in the population is represented by the symbol g. The rate of the position
change (velocity) for particle i is represented as VI = (vi1 , vi2 , . . ., viD ). The particles are
manipulated according to the following equations:
vid = w vid + c1 Rand() (pid xid ) + c2 Rand() (pgd xid ) (7)
xid = xid + vid (8)
where c1 and c2 are two positive constants, c1 and c2 are usually c1 = c2 = 2, Rand() is
two random functions in the range [0,1], and w is the inertia weight.
Equation (7) is used to calculate the particles new velocity according to its previous
velocity and the distances of its current position from its own best experience (position)
and the groups best experience. Then the particle flies toward a new position according
to Equation (8). The performance of each particle is measured according to a pre-defined
fitness function, which is related to the problem to be solved. The inertia weight w is
employed to control the impact of the previous history of velocities on the current velocity,
thus to influence the trade-o between global (wide-ranging) and local (nearby) explo-
ration abilities of the flying points. A larger inertia weight facilitates global exploration
(searching new areas) while a smaller inertia weight tends to facilitate local exploration
to fine-tune the current search area. Suitable selection of the inertia weight can provide
a balance between global and local exploration abilities and thus require less iteration on
average to find the optimum. In this paper, an analysis of the impact of this inertia weight
together with the maximum velocity allowed on the performance of particle swarm opti-
mization is given, followed by experiments that illustrate the analysis and provide some
insights into optimal selection of the inertia weight and maximum velocity allowed. The
utility of a particle swarm optimization is shown for global search [12-14].
A particle swarm optimization is used to adjust the phase shift weights based on the
power of the array in the interfering directions. The goals are to minimize the total output
power of the interfering signals and maximize output power of the desired signals to the
receiver. During the process of PSO algorithm iteration, the weighting vector kept for the
next step iteration should make the output power of the desired signal to be increased
and the output power of the interfering signal to be decreased monotonically. So, the
PSO algorithm is applied to find the optimal radiation pattern of the proposed smart
antenna by using the two-way convergent method. Obviously, this technique can increase
the signal to interference ratio (SIR). Then, the MIMO-SDMA optimization of smart
antennas can be achieved.
So, the fitness function is the square of AFm () in Equation (6). Obviously, this tech-
nique can increase the signal to interference ratio (SIR).
The flow chart of particle swarm optimization is given in Figure 3. Detailed steps of
particle swarm optimization approach are as follows:
Step 1. Initialization:
The first step toward implementation of the PSO algorithm is to pick the parameters
that need to be optimized and give them a reasonable range in which to search for the
optimal solution. This requires specification of a minimum value for each dimensional
optimization.
Step 2. Initialize random swarm location and velocities:
To begin search for the optimal position in the solution space, each particle begins
at its own random location with a velocity. They are random both in its direction and
magnitude. Since its initial position is the only location encountered by each particle at
the runs start, this position becomes each particles respective individual best. The first
global best is then selected from among these initial positions.
Step 3. Evaluate particle fitness:
The fitness function, using the coordinates of the particle in solution space, returns a
fitness value to be assigned to the current location. If that value is greater than the value
at the respective individual best for that particle, or the global best,then the appropriate
locations are replaced with the current location.
Step 4. Update the individual best and global best:
As a particle moves through the search space, it compares its fitness value at the current
position with the best fitness value which has ever attained at any time up to the current
time. The best position that is associated with the best fitness encountered so far is called
the individual best. Each particle individual best is evaluated according to the updated
position. It is the best position among all of the individual best positions achieved so far.
Hence, the global best can be determined
Step 5. Update the velocity and position:
It is the velocity of the moving particles represented by a real-valued vector. The
manipulation of a particles velocity is the core element of the entire optimization. Careful
understanding of the equation used to determine the velocity is the key to understanding
the optimization as a whole. The velocity of the particle is changed according to the
relative locations of individual best and global best. Based on the updated velocities, it
is simple to move the particle to its next location.
Step 6. Termination criteria:
This is the condition under which the search process will terminate. In this study, the
search will terminate if the following criteria is satisfied: The number of iterations reaches
the maximum allowable number. If one of the termination criteria is satisfied, then stop,
or else go to step 3.
4. Computer Simulations and Results. In this design, the necessary parameters of

the particle swarm optimization are defined as follows:
the population size P equals 300;
the maximum number of generation equals 500;
the maximum value of inertia weight w is 0.8;
1336 C.-H. HSU
Figure 3. The flow chart of particle swarm optimization
the minimum value of inertia weight w is 0.3;

the acceleration constantsC1 and C2 are 2;
the maximum speed of particle is 1000.
The population size, maximum number of generation, the maximum value of inertia
weight, and the minimum value of inertia weight, the acceleration constants and the
maximum speed of particle are specified before the implementation of the algorithm.
Their values definitely aect the process of optimal solution search and results.
In this problem, a linear antenna array is composed of 20 isotropic elements. So,
N = 10. N is variable number. The distance d of two adjacent elements is a half of .
The technique features are by phase-only perturbations. Amplitude weights are constant
and phase shifter weights are in odd symmetry. The value of is constant set between
0.1 and 1. The value of n is set between and . The unit of n is rad.
Example: In this uplink smart antenna system, the signals of user 1 and user 2 come
from 40 and 60 respectively. As this example, a two-user system is studied. The
users number is 2 in this smart antenna system. So, M equals 2. If the number of the
current generation is equal to 600, the particle swarm optimization iteration will stop
automatically.
Radiation pattern of user 1: For user 1, the signal of user 2 is considered as its
interfering signal. After particle swarm optimization implementation, the iterations stop
in 500 generations automatically. The weight vector [mn ], m = 1, n = 1, 2, 3, . . ., N , for
the optimal radiation pattern of user 1 after 500 generations, is derived as listed in Table
1. The optimal radiation pattern is derived as shown in Figure 4. In user 1s radiation
pattern, the null is derived in the 60 interfering direction and the maximum main lobe
is derived at 40 . Thus, the optimal radiation pattern of user 1 has been derived. The
SIR is 59 dB. So, the interference on user 1 due to user 2 can be ignored.
Radiation pattern of user 2: For user 2, the signal of user 1 is considered as its
interfering signal. After particle swarm optimization implementation, the iterations stop
in 500 generations automatically. The weight vector [mn ], m = 2, n = 1, 2, 3, . . ., N , for
the optimal radiation pattern of user 2 after 500 generations, is derived as listed in Table
1. The optimal radiation pattern is derived as shown in Figure 5. In user 2s radiation
pattern, the null is derived in the 40 interfering direction and the maximum main lobe
is derived at 60 . Thus, the optimal radiation pattern of user 2 has been derived. The
SIR is 60 dB. So, the interference on user 2 due to user 1 can be ignored.
For user 1s radiation pattern, the main lobe is toward the user 1 according to Figure
4. At the same time, the null of the radiation pattern of user 1 is also toward user 2. So,
the optimal radiation pattern of user 1 possesses the interference nulling and the desired
signal maximizing. Likewise, for user 2, it is the same. The eciency of particle swarm
optimization approach is high, and the interference can be ignored in real time signal
processing so that it is suitable for the practical mobile communication. The radiation
pattern of the smart antenna in Cartesian coordinate is shown in Figure 6. The radiation
patterns of the smart antenna in polar coordinate are shown in Figure 7 and Figure 8.
Table 1. The weight vector [mn ] for the optimal radiation pattern of two users
Radiation pattern of user 1: Radiation pattern of user 2:

desired signal direction at 40 desired signal direction at 60
and interfering source at 60 and interfering source at 40
11 = 2.424 21 = 2.584
12 = 0.959 22 = 0.729
13 = 2.169 23 = 2.877
14 = 3.136 24 = 0.255
15 = 0.317 25 = 3.140
16 = 0.928 26 = 2.990
17 = 2.096 27 = 0.928
18 = 1.282 28 = 1.317
19 = 2.001 29 = 0.436
110 = 2.902 210 = 2.430
5. Conclusion. The Multiple-Input Multiple-Output Spatial Division Multiple Access

(MIMO-SDMA) optimization design of a smart antenna based on phase-only perturba-
tions using the particle swarm optimization is proposed and achieved. In the paper, first,
the array factor formula suitable for optimal solution search based on signal phase shift
for a smart antenna is deduced. In order to be able to adopt the particle swarm optimiza-
tion to search the optimal radiation pattern, the formula is reformed through amplitude
weights are constant and phase shift weights are in odd symmetry.
1338 C.-H. HSU
Figure 4. Optimal radia- Figure 5. Optimal radia-

tion pattern for user 1 at tion pattern for user 2 at 60
40 and interfering source and interfering source at
at 60 40
Figure 6. Radiation Figure 7. Radiation

pattern of two-user up- pattern of two-user uplink
link smart antenna in smart antenna in polar
Cartesian coordinate coordinate (sacled in dB)
Particle swarm optimizations are applied to find the optimal weighting vectors of array
factor of the proposed smart antenna. The optimal radiation pattern means that the
power of desired signal has to be highest and the power of interfering signals has to be
lowest at the same time. So, in this proposed smart antenna, for each users optimal
radiation pattern, there are one main lobe formed towards one desired user and nulls
formed in the directions of other users whose signals are considered as its interfering
sources. Thus, this smart antenna system can not only derive maximized main lobes in
the directions of all the desired signals but also suppress interferences by placing nulls
in its radiation pattern adjustably. In this way, the MIMO-SDMA eect is achieved and
optimized. Then, the uplink MIMO-SDMA optimization of a smart antenna has been
achieved.
Figure 8. Radiation pattern of two-user uplink smart antenna in polar

coordinate (sacled in decimal)
REFERENCES
[1] A. U. Bhobe and P. L. Perini, An overview of smart antenna technology for wireless communication,
IEEE Aerospace Conference Proceedings, vol.2, pp.2875-2883, 2001.
[2] J. Robinson and R. S. Yahya, Particle swarm optimization in electromagnetics, IEEE Transaction
on Antenna and Propagation, vol.52, no.2, pp.397-407, 2004.
[3] C.-H. Hsu and W.-J. Shyr, Adaptive pattern nulling design of linear array antenna by phase-only
perturbations using memetic algorithms, International Journal of Communications in Numerical
Methods in Engineering, vol.24, pp.1121-1133, 2008.
[4] C.-H. Hsu, T. Y. Chen and J. S. Pan, Designing adaptive linear array antenna to achieve pattern
steering optimization by phase-amplitude perturbations using genetic algorithms, Chinese Journal
Electronics, CJE, vol.14, no.1, pp.125-129, 2005.
[5] C. H. Hsu, P. H. Chou, W. J. Shyr and Y. N. Chung, Optimal radiation pattern design of adaptive
linear array antenna by phase and amplitude perturbations using memetic algorithms, International
[6] C. H. Hsu and W. J. Shyr, Memetic algorithms for optimizing adaptive linear array pattern by phase-
position perturbations, Circuits, Systems and Signal Processing, vol.24, no.4, pp.327-341, 2005.
[7] C.-H. Hsu and C.-H. Chen, Optimizing beam pattern of adaptively linear array antenna by phase
perturbations using genetic algorithms, Lecture Notes in Computer Science, LNCS/LNAI, vol.3213,
pp.929-936, 2004.
[8] M. P. Lotter and P. V. Rooyen, An overview of space division multiple access techniques in cellular
systems, Proc. of 1998 IEEE South African Symposium on Communications and Signal Processing,
COMSIG98, University of Cape Town, Rondebosch, pp.161-164, 1998.
[9] C. H. Hsu, Downlink MIMO-SDMA optimization of smart antennas by phase-amplitude perturba-
[10] C. H. Hsu, Uplink MIMO-SDMA optimization of smart antennas by phase-amplitude perturbations
based on memetic algorithms for wireless and mobile communication systems, IET Communications,
vol.1, no.3, pp.520-525, 2007.
[11] R. A. Monzingo and T. W. Miller, Introduction to Adaptive Arrays, Wiley, New York, 1968.
[12] M. A. Abido, Optimal design of powerSystem stabilizers using particle swarm optimization, IEEE
Transaction on Energy Conversion, vol.17, no.3, pp.406-413, 2002.
1340 C.-H. HSU
[13] J.-F. Chang and K.-L. Chen, Applying new investment satisfied capability index and particle swarm
optimization to stock portfolio selection, ICIC Express Letters, vol.3, no.3(A), pp.349-354, 2009.
[14] J. Nagashima, A. Utani and H. Yamamoto, Ecient flooding method using discrete particle swarm
optimization for long-term operation of sensor networks, ICIC Express Letters, vol.3, no.3(B), pp.833-
840, 2009.

!!
"# $ "%#$

& ' "# $# ( !) ** +,

- #.)- /0#0# $001,- $ ) )/0# $001,

!
"

#

%! 2
. 3 #
#
#

!
!

"

! #

$%&'%

$%&%('% )

!
% $*'

! +

" ,

!

!

-

!

,% ! "

.

/ 0

! !
!
#

1%

!

2

%
/ !

!
-
!

!
)

)
! .

.
- % ! )

"

) ! 2 3 % !
!

) 2 3 *% !

#

2 3 &% ! # ! 2
3 4% !
5
! !
0
#

3 ( 6% !
% !

! 7
)

)

2

% )
% !
)

! .

3

% 326" % 8
9 :89;%
!

<

$'
/
! )

1
$ '

)

!

) !

=! $' "

)
!

# 2 %

)
"

)
>
$' 2
% > 89

? #

$' !

#
!
89

!

) @!
%

%
!

$4'
%

+

" !

A $B' <
$C' A

% !

% % %
% #
%
D+

! !

% ! <

3>7 :3

>
7;
"
#

+

A < 1

!
#

!

#
!
2
% !

#

2 % A

2
% ! , !
! +
+
$4' "

!
!

!
:;% !

+

%

"#

.

) .

!
7

! !

% !

$&%(%B'
E
$&'%

"

8
9 :89; 2

%

#

!

#
#
! # @!
%

! !

"
+

,

!
! F

2

!

> $B'%

!

!

" !

% !

: ;

"

"
!

!

! .

" 33 : .
+

;

2

!
%

2 %
.
.

)

"

E $(' !

"

! !

2

!
%

#
/ .

#

!

"

! 2

!

%
@!
%

.

< % !

/
.%

.

-

!
"% !
%

!
,

#
- !

.

)

% #
!
0

1
.

.

%

!
6

%
.
.

% ! !

" !

!
! )

* "$ 2

%

, @ !

/ ,

!

,

6

% ,
+ ! +

1
% !
)

! #

* % & 9

+

" #

!

#
"% !

! )

%

/ #

/

2 92? GHA
%
#

1
92? GHA
! 6
:;% !

!

I !

!J
K L 4 B L & :;
/ !

!

K * K * 92? GHA %

%

! (
9

:
;
% ! *

.

* '( & ? #

#

2

% #

"

)

! $ 44'
9 #

#
% % # ! 6

: ; 6
#
% !
.

:;

: ;
:; 9

: ;
8
#

& " ) $ 2

!

% !

% !
+ 2 !
!

& * ?%
!
:;

! !

"
!

!
% %

) )

!

.%
%

-
%

- &
!

!

%

* 4 "
,
"

!

"

, !
!
3

. ) .

0

%

!

#

"

& $ & 2

/

%

#

%

2

!

! /

2

%

!

! ) !

%

#

! 6
:;% !

C
4 *

% %

"
%

- #

!

) # 2 #
% !
%

%
4 ! C C
% % #
! 6
: ; "

4

! ( (
6
:; "

6

*
:; : ; :;
:; 8 !

.
: ; :;

% % % #

"
6
*:; !
!

-
!

% !
.

6
*: ; !
!

!
:
;
!

!

:; : ;
:; 0 !
: ; 0 !

@!
% !

"

6
# !

% !
!

!

! 6

%

!
%

4 "# 0

% !

% !

2

.

2

%

/

%

!
!

# 3 &% !

!

M

F

:; ! !

%
!

" #

!

7!%

E
$&'%
,
: ; !

. 3
% E $(' > $B'

.
: ;
! ! ?

.

!
!
,

%
.

2

% !

%
!

-

%
! /

!

1 !
!
%
!

#
"
!

# 2 !

! %
" !

%

?#

!

!

?#

.

$&%(%B'
0

2

E $('%

!

2 !
E%

)

!

"
! #

2

) %
!

%

/

"

%

/
!
!

/ !

# 3 * 6

%

!
2

% !

!

!
)
!

1
% !

! !

!

!
"
/ !
# 3 4
4 (# $(

,
!

! !

.

!

)
- / !

2

%

% !

,
! # !
6
4
"

0

! !

"

!
! .
3

!

@!
%
%
! ! /

% !
.

4 % # - F

"
!

"

!
! !

7
$C'
2

%
!
!

%

!

#
3 &%

+

%
!

" !

1
% ! !

!

! %
!

!

2

/
!
% ! !

! +

-

% !

4
+
/#
% ! 6
(:;

! #
4

2 4
# /
%
!

0
!%

!
!
!

! 6
(:*; (:&; "

!

# "
) ! /
%

!

1
/

" #
!
6
B #

! !

:; :; :*; :&; : ;

:; 8
: ; 9

!

9

%

"

! !

6
%

%
/

!

!%

%
E
$&' > $B' 2 EF

!
%

/

$(' "

!

)

@!
%

.

"

/

% ! 4 !

0
/
E
>

.

( +($
(
! +($ ( !

?

% #
! 4
( 4

! 6
C
2 #
% /
% !

&

%
3
!

!

! !
" ?#

& #% !

"

#

"
! 6

" /

! 6
!

!

%

%

( "
.
! !

(

2
! B %

) %

"
!
B

0
!%

!

6
C%
%
! , !

"

!

:
; ! %

4%

: ; " (

*&

# I)
! !

(
% ! C4
%

% !

) +
"
) #

+

.
% !

6

!

.

/%
)

( "# +($ !
# 3 4%

!
!

%

4
"

4
! &4 &4
: 6
(: ;;
4 "
!
!

!

% "

)

!

!

B 4

- 4 :
6
C; #

"

#

! ) #

"! % !

% !

#
"
! 6
:; "

# !
#
B "
!

3 &
2 % ! #

% !

!

!
" #

!

!

!

"

#
! 6
: ;
:; : ;
$
!
% :; 9

: ;

1

#
6
% !

+
"

!

!
@!
% /

!

# %

#

"

3 &

!%

% 6
?
! #
%
!
!
)

F !

" !

! !
,

2

% !
&

%

:6
: ;; " !

4

! 4%
(% B C%

! 6
:; !%
!

:
!

;
! 6
#% !
4%
# B !
*

:; : ;
:; )

: ;
! .

!

6

% ! ! #

" #
#
2 #
% !

!

4

!

"
#

6
:; ! /

% ! #

6

% ! #
% !

* !% #
#%

4

!
6
: ; "
4 ! 6

:;
B % "

% !

2 )
% !

%

!
1 #

)

!
#
"

#

)
!

!

.% !
!

7

%

:; : ; :;
:;
!

: ;

:;
! !

.%

!

"

0

!
!
! !

45 0 0 () 2
. %! ## $ ,
6
#

%
0 0 ,,07*& 0
45 0 "
0 )# 3,,# ,
! ,
!

. $
%%% &' ,,0 0
45 80 0 2 9 8#% ! # #
% )$, #

( %
07 0 ,,0 0
45 +0 #, 0 :# +0 # # # %! %
,#
# 9
%%% ) & ' %
07 07
,,0 0
4*5 0 ; 0 )1 (0 ( < %## , !# # ,
,
,
$## !# ) $ ,

#

%
0* 0 ,,0&*& 0
475 (0 (0 ($
# 20 0
0 ) 0 =
#,# %#
,
9! 9
# # )! $ * +, ( -
( *0
4&5 0 > '0 : (0 (
#

. # !

%#
,,
# $ %%% &' ,,0&& 0
45 (0 " 0 .) ;0 # 2
. #! #,, %
!
$, # $
-,% & , 8 3 ? +,#@0
45 A0 B
A0 3
( ,
!

. #! # %# C# $
%%% &' ,,0 7 0

4 5 (0 D9 (0 3 '0 (0 $ ,

# !# %!
' - . ,,07*& 70
45 0
### '0 0 ; ;0 3# (0 # !# #
!#

. $ %%% &' ,,0*&* 0
45 0 :# +0 #, 0 2$ +0 # =
%# %!
$ %%% + ( ,,0 0
45 20 +! 0 8! 30 =#
%
9 9

#! E C

%
0* 0 ,,0*7 0
45 0 () E ,
6
# #
#

. ,,! ,
#

%
0* 0 ,,0 0
4*5 F0 30 + '0 .) D0 ' ) (0 # ) 3 #
#

.
# %
%#
# $ / %%% ' &
' ,,0** 0
475 F0 30 + '0 .) '0 ( D0 ' ) (0 # ) 3 #
%!
# %
%#
# %! 0 %
0 0?B@ ,,07*&
0
4&5 E0 F0 #
;0 +0
+0 ## #
%! G ! %#
#% !
,, # # $ * - & - 0
45 '0
# 30
0 () 30 3 (0 (# 3 # %!

,# 9
)9$#
& %
07 0 ,,07 7* 0
c ISSN 1349-4198
OBSERVER-BASED ITERATIVE LEARNING CONTROL WITH

EVOLUTIONARY PROGRAMMING ALGORITHM
FOR MIMO NONLINEAR SYSTEMS
Yan-Yi Du1 , Jason Sheng-Hong Tsai1, , Shu-Mei Guo2, , Te-Jen Su3,

and Chia-Wei Chen4
1
2
Department of Computer Science and Information Engineering
National Cheng Kung University
No. 1, University Road, Tainan 70101, Taiwan
n2896109@mail.ncku.edu.tw

Corresponding authors: { shtsai; guosm }@mail.ncku.edu.tw; tejensu@mail.ksu.edu.tw
3
College of Information Technology
Kun Shan University
No.949, Dawan Road, Yongkang, Tainan 710, Taiwan
tejensu@mail.ksu.edu.tw
4
Department of Mechanical and Automation Engineering
Kao Yuan University
No. 1821, Jhongshan Road, Lujhu Township, Kaohsiung County 821, Taiwan
cwchen@cc.kyu.edu.tw
Received September 2009; revised January 2010
Abstract. In this paper, the observer-based iterative learning control with/without evo-
lutionary programming algorithm is proposed for MIMO nonlinear systems. While the
learning gain involves some unmeasurable states, this paper proposes the observer-based
iterative learning control (ILC) for nonlinear systems and guarantees the tracking error
convergences to zero via continual learning. Moreover, a sucient condition has been
presented to alleviate the traditional constraint, i.e., identical initial state, in the conver-
gence analysis. Then, an idea of feasible reference based on polynomial approximation
is proposed to overcome the limitation of ILC initial state error. To speed up the con-
vergence of the iterative learning control, evolutionary programming is applied to search
for the optimal and feasible learning gain to reduce the training time. In addition, two
improved issues of ILC, an appropriate selection of the initial control input and the im-
proved learning rule for the system whose product matrix of output matrix C and input
matrix B is not full rank, are presented in this paper. Three multi-input multi-output
(MIMO) illustrative examples are presented to demonstrate the eectiveness of the pro-
posed methodology.
Keywords: Iterative learning control, D-type ILC, Observer-based ILC, Evolutionary
programming
1. Introduction. For repeated tracking control tasks, iterative learning control (ILC)
which improves the performance of a system via continual learning is one of the ecient
control algorithms [1]. Furthermore, the controller can exhibit the perfect control perfor-
mance according to a less priori knowledge of the system. Due to its advantages, higher
eciency and more convenience, many researchers ardently discussed the subjects of ILC
and usually applied the ILC-based controllers in several physical systems, such as indus-
trial robot, computer numerical control (CNC) machine and antilock braking system, etc
[1].
1357
1358 Y.-Y. DU, J. S.-H. TSAI, S.-M. GUO, T.-J. SU AND C.-W. CHEN
The major issues of iterative learning control focus on learning convergence and con-
vergence speed. For the subject of convergence, Arimoto et al. proposed the first order
D-type ILC to train industrial robotic systems via the derivative of output error [2]. Then,
P-, D- and PD-type iterative learning controls were proposed for nonlinear systems [1].
Besides, based on Lyapunov theory, a robust iterative learning control was presented for
systems with time-varying parametric uncertainties and local Lipschitz nonlinearities [3].
For the monotonic convergence of ILC, Meng et al. gave sucient conditions in terms of
linear matrix inequalities (LMIs) based on the bounded real lemma [4].
Recently, some researchers focus on designing observer-based ILCs. When the structure
of ILC algorithm contains some unknown/unmeasurable terms, such as unmeasurable sys-
tem states and unknown parameters, some corresponding learning controllers are dicult
to be decided without constructing an observer to estimate these significants but unknown
factors [5]. For example, Mi et al. established a vehicle speed observer to develop the ro-
bust antilock braking control based on ILC [6]. Gunnarsson et al. proposed the ILC-based
robot control by applying Kalman filter to estimate the arm angle [7]. When the input
matrix of a system B contains some unmeasurable states, Taybei and Xu proposed an
observer-based iterative learning control for a time-varying nonlinear system and proved
that the tracking error convergences to zero via increasing learning epoch [8]. However,
Taybei and Xu did not consider the constraint, initial state error, into the algorithm.
Initial state error which indicates the dierence of states between the real system and
ideal reference model at the initial time instant is a fundamental issue for iterative learning
control. The inability to resolve the problem leads the tracking error does not converge.
Hence, in order to overcome this situation, Chen et al. proposed the initial state learning
scheme by D-type ILC updating law for time-varying linear and nonlinear uncertain sys-
tems [9]. Lee and Bien adopted the PD-type ILC algorithm to relieve this restriction [10],
and Park improved the algorithm based on the average operator for linear time invariant
systems [11]. Applying the finite initial rectifying action, Literature [12] presented the
uniform convergence of the output tracking error in a fixed initial shift. Recently, adopt-
ing the concept of adaptive control, adaptive ILC (AILC) has been presented to overcome
the constraint of ILC in literature due to randomly initial states [13].
Although above literatures guarantee the convergence condition for the worst-case sce-
nario, they might lead to a slower convergence speed due to conservative designs [14].
Therefore, recently, some researchers have investigated the convergence speed of iterative
learning control. Literature [14] introduces the Q-factor, a learning performance index,
and minimizes the factor to increase the convergence rate of iterative learning control.
In addition, Ye et al. presented a unit-gain compensator in iterative learning control to
improve the learning speed from the view point of frequency domain [15]. Sugie and Sakai
applied the H2 optimization to specify the convergence speed of the ILC for uncertain
systems [16]. For constraint multi-objective optimization problems [17-19], evolutionary
programming (EP) can easily attain the optimal solution. Evolutionary programming is
a parallel stochastic search technique that mimics the natural evolution, survival of the
fittest, by an eective multi-agent searching strategy [20]. The algorithm can eciently
resolve the global optimization problem with various constraints for various applications
[21-23].
In this paper, the observer-based iterative learning control combined with evolutionary
programming is proposed for MIMO nonlinear systems. While some unmeasurable states
inject into a system input matrix, the observer based on optimal linearization algorithm
[24] and linear quadratic regulator accurately estimates the system states for the learning
gain design. Then, a sucient condition is proposed in this paper (Remark 2.1) to alle-
viate the conservative assumption of D-type ILC adopted by [8] so that the condition of
OBSERVER-BASED ITERATIVE LEARNING CONTROL 1359
identical initial state can be released as one of identical initial outputs. Besides, a resolu-
tion based on polynomial approximation is presented to overcome the constraint induced
by the initial state error for the ILC. Finally, evolutionary programming enhances the
convergence speed of ILC for the complicated nonlinear system such as chaotic systems.
The paper is organized as follows: A proposed observer-based iterative learning control
is presented in Section 2. Then, the improved ILC via evolutionary programming is
proposed in Section 3, and some other improved issues of ILC are presented in Section 4.
The design procedure of the proposed algorithm is summarized in Section 5. Illustrative
examples are given to demonstrate the eectiveness of the proposed method in Section 6,
and conclusions are presented in Section 7.
2. Observer-based Iterative Learning Control.

2.1. Problem description. The basic idea of ILC represents that a system performs the
same task repeatedly, and the tracking performance of the system can be improved by
training the control input via recursive iterations. Consider the class of nonlinear systems
x i (t) = f (xi (t)) + B(xi (t))ui (t), (1a)
yi (t) = Cxi (t), (1b)
where xi (t) n , ui (t) m and yi (t) p are the state vector, the control input and the
system output, respectively. The subscript i represents the iteration number of learning,
and f () and B() are nonlinear functions. The control problem is precisely formulated
as follows: For the nonlinear system, suppose that the desired trajectory r(t) over the
time interval [0, T ] is dierentiable. The main objective aims for designing an iterative
updating rule of input such that the output yi (t) converges to the desired trajectory
r(t) for t [0, T ] when the learning epoch i approaches to infinity. Due to the specific
situation, the input signal needs to be stored via a memory unit in the ILC scheme. For
designing the memory-based ILC, the following assumptions are considered.
A1. There exists an input upper bound umax which denotes the physical limitation of a
system, and the control input ui (t) satisfies the rule ui (t) umax .
A2. There exists a bounded control rule ud (t) umax over the time interval [0, T ] such
that r(t) = Cxd (t) and x d (t) = f (xd (t)) + B(xd (t))ud (t).
A3. The matrix B has B() Bm and satisfies the Lipschitz condition B(x(t1 ))
B(x(t2 )) KB x(t1 ) x(t2 ) with respect to x(t1 ) n and x(t2 ) n over
the time interval [0, T], where Bm and KB are positive constants. Furthermore, the
function f () satisfies the condition, too.
A4. rank(CB()) = m and CB() gm , where gm is a positive constant.
A5. The resetting initial condition yi (0) = r(0) is satisfied at each iteration, where r(0)
is the initial desired trajectory.
Lemma 2.1. Since matrix B() and function f () in (1) satisfy the Lipschitz condition,
respectively, there exists
C[B(x(t1 )) B(x(t2 ))] KB y(t1 ) y(t2 ) (2)
and
C[f (x(t1 )) f (x(t2 ))] Kf y(t1 ) y(t2 ) , (3)
where KB and Kf are positive constants.
Proof: At first, the Lipschitz condition represents
B(x(t1 )) B(x(t2 )) KB x(t1 ) x(t2 ) (4)
from Assumption 3. Premultiply both sides of (4) by C to obtain

C B(x(t1 )) B(x(t2 )) C KB x(t1 ) x(t2 ) . (5)
According to Literature [26], one has
C[B(x(t1 )) B(x(t2 ))] C B(x(t1 )) B(x(t2 )) and (6)
KB y(t1 ) y(t2 ) = KB Cx(t1 ) Cx(t2 ) KB C x(t1 ) x(t2 ) . (7)
From (7), there exists a constant KB to satisfy
KB y(t1 ) y(t2 ) = KB C x(t1 ) x(t2 ) , (8)
where KB KB . Then, combining (5) to (8), one obtains the following inequality
C[B(x(t1 )) B(x(t2 ))] KB y(t1 ) y(t2 ) . (9)
Similarly, the inequality C[f (x(t1 )) f (x(t2 ))] Kf y(t1 ) y(t2 ) is verified by
using the same derivation procedure.
Remark 2.1. The phenomena C[B(x(t1 ) B(x(t2 )] KB y(t1 ) y(t2 ) and C
[f (x(t1 )) f (x(t2 ))] Kf y(t1 ) y(t2 ) of Lemma 2.1 induces the condition yi (0) =
r(0), which helpfully alleviates the limitation xi (0) = xd (0) in Literature [8].
In this paper, some
norms are used as follows:
M = max { nk=1 |mjk |} for a given matrix M = [mjk ] pn , V = max |vj | for
1jp 1jp
a given vector V = [v1 , , vp ]T , the sup-norm denoted by (t) = sup (t), and
t t[0,T ]

the -norm denoted by (t) = sup e (t) . Moreover, the time argument tof the
t[0,T ]
following equations are omitted in order to present a concise form.
2.2. Iterative learning control with LQR-based observer. In many practical prob-
lems, some states of (1) are unmeasurable due to technical or economic reasons and
aect the learning gain design for ILC. Therefore, adopting LQR-based observer [25], the
observer-based iterative learning control (OB-ILC) is proposed and depicted in Figure 1.
Consider the state observer as follows:
x i = A(
xi )
xi + B( xi )(yi yi ),
xi )ui + Lc ( (10a)
yi = C xi , (10b)
where A(xi )
xi is derived from f (xi ) by the optimal linearization algorithm [24], Lc ( xi ) is
the observer gain, and A( xi ) xi ) at any operating state xi with xi (0) = C + yi (0).
xi = f (
According to the observer design in [25], the observer gain Lc ( xi ) can be obtained by
xi ) = Po,i C T Ro1 ,
Lc ( (11)
A(xi )Po,i + Po,i AT ( xi ) Po,i C T Ro1 CPo,i + Qo = 0, (12)
where Qo , Ro , and Po,i is the symmetric and positive definite solution of
nn pp
Riccati Equation (12). Notice that if the weighting matrices Qo and Ro satisfy the high
gain property (i.e., qo ro ), where Qo = qo In and Ro = ro Ip ; qo and ro are scalars, the

estimated state error ei = yi yi is bounded and asymptotically converges to zero.
Now, consider the following iterative learning controller
( )
vi+1 = ui + G ( xi ) r y i
(13a)
= ui + G ( xi ) (r CA ( xi ) xi CB (xi ) ui CLc (
xi ) ei ) ,
ui+1 = sat(vi+1 ), (13b)
Figure 1. Observer-based ILC for a class of nonlinear systems
where r and yi are the time derivatives of r and yi , respectively, xi is the estimated state
by the observer (10), and G( xi ) is the learning gain. The initial control input satisfies the
condition u0 umax . Moreover, the saturation function of control input is defined as
[ ]T
sat(U ) = sat(u1i ), . . . , sat(um
i ) , in which (14)

uki , if uki umax
k
sat(ui ) = umax , if uki > umax , for k {1, . . . , m} , (15)

umax , if ui < umax
k
T
i ] , and the superscript k means the k
where U = [u1i , . . . , um m th
element of the vector
U.
2.3. Convergence analysis of OB-ILC. Define the input error as ui = ud ui and

vi = ud vi such that
vi+1 = ud vi+1 = ui (vi+1 ui ) (16a)
= ui G(
xi )(r CA( xi CB(
xi ) xi )ui CLc (
xi )ei ) (16b)
= (I G( xi )CB( ui G(
xi )) xi )C(B(xd ) B( xi ))ud
(16c)
G( xi )C(f (xd ) f (
xi )) + G(xi )CLc (
xi )ei ,
where xd is the desired state vector generated by the desired control input ud . According
to Assumption A2 and the definition of ui+1 in (13), ui+1 vi+1 is a reasonable
inequality. In view of Assumptions A2 and A3, applying the previous fact with Lemma
2.1 and taking the norm of (16) obtains the following inequality

ui+1 I G( xi )
xi )CB( ui + ( + 2 ) G(
xi )
yi + 2 G(
xi ) yi , (17)

where = KB ud + Kf , 2 = CLc ( xi ), yi = r yi and yi = r yi . Adopting
Assumption A2 and the state observer (10), the output error yi is derived as
t
yi = yi (0)+ C[f (xd ( ))f (
xi ( ))+B(xd ( ))ud ( )B(xi ( ))ui ( )Lc (
xi ( ))ei ( )]d.
0
(18)
According to Assumption A3 and the Bellman-Gronwall Lemma [8], (18) can be derived
as t

yi t
yi (0) e + (1
ui ( ) + 2 ei ( ))e(t ) d . (19)
0
where 1 = CB( xi ) = gm . In addition, by Assumption A2 and (1), one attains

t
yi = yi (0) + C[f (xd ( )) f (xi ( )) + B(xd ( ))ud ( ) B(xi ( ))ui ( )]d . (20)
0
Then, taking the norm of (20) and applying the Bellman-Gronwall Lemma attains
t
yi yi (0) e +
t
1
ui ( ) e(t ) d. (21)
0
From (19) and (21), multiplying (17) by et , > and applying the -norm, one
easily obtains

ui+1 (1 + 2 )
ui + 3 , (22)
1 = I G(
xi )CB(
xi ) , (23)
( + 22 )1
2 = (1 e()T ) G(
xi ) , (24)

( + 2 )2
3 = G( xi ) ei +(+2 ) G( xi )
yi (0)+2 G( xi ) yi (0) . (25)

As the learning control gain G( xi ) is bounded, there exists a suciently large
making 2 arbitrarily small and 1 + 2 < 1 by assigning 1 < 1. Hence, one has
{ }
3
lim ui with = min 2umax , . (26)
i 1 1 2
Then, multiplying (21) by et , > , and applying the -norm, one obtains
{ t }
{ }
yi sup 1 e ()(t )
d
ui + sup yi (0) e()t
t[0,T ] 0 t[0,T ]
(27)
1 (1 e ()T
)

ui + yi (0) .

Applying (26) and Assumption A5, inequality (28) can be derived from (27) as
1 (1 e()T )
lim r yi . (28)
i
Due to the designed observer (10), ei converges to zero. For a suciently large ,
3 ( + 2 ) G(
xi )
yi (0) + 2 G(
xi ) yi (0) for i . Moreover, if yi (0) = 0
and Assumption A5 are satisfied, then 0 implies the tracking error converges to zero
when i . From the above derivation, two observations are summarized as follows:
Theorem 2.1. Consider system (1) with observer-based iterative learning control (13),
where xi is given by the state observer (10). If all Assumptions A1 A5 are satisfied,
then the following statements hold.
(i) If 1 < 1, there exists > such that
( )
1 1 e()T
lim r yi . (29)
i
(ii) If 1 < 1 and yi (0) = r(0), then r yi is bounded and converges to zero when
i .
Figure 2. Observer-based EP-ILC with tuning matrix for a class of

nonlinear systems
3. Improved OB-ILC Tracker Via Evolutionary Programming. According to the

previous convergence analysis, if the learning gain G(
xi ) satisfies the convergence condition
Im G(
xi )CB(
xi ) < 1 (30)
for the ILC tracker, the tracking error will approach to zero when i . In general,
the learning gain is assigned as (CB(xi ))1 for simplicity and convenience. However, the
learning convergence rate of ILC might be not desirable for complicate and nonlinear sys-
tems. In order to increase the speed of the converging rate, choosing a good learning gain
must be considered in the ILC scheme. Hence, applying an optimal searching algorithm
called evolutionary programming (EP) constructs the EP-based ILC (EP-ILC).
In order to apply the EP algorithm to search for an optimal learning gain GEP (xi ), let
the convergence condition (30) be equal to a tuning matrix mm as
= Im GEP (
xi ) CB (
xi ) . (31)
From (30) and (31), the tuning matrix satisfies
< 1; (32)
the modified learning gain GEP (
xi ) can be easily derived according to (31) and (32) as
GEP ( xi ))1 = (Im )G(
xi ) = (Im )(CB( xi ), (33)
where G( xi ) = (CB(xi ))1 . The observer-based EP-ILC with the tuning matrix is
proposed and depicted in Figure 2.
After that, performing evolutionary programming searches for the optimal tuning ma-
trix to enhance the learning convergence rate of iterative learning control. The detailed
procedure of evolutionary programming is presented in [23]. Notice that only the ini-
tial population and the objective function are redefined for the proposed OB-ILC in this
paper.
The initial population is generated from
ij (k) = lij + ij (k) (uij lij ) for 1 i m and 1 j m, (34)
where ij (k) is the element of tuning matrix for k = 1, 2, . . . , N , and N is the total
size of the initial population; lij and uij are corresponding lower bound and upper bound,
respectively; ij (k) is a random number decided by applying the quasi-random
[ ] sequences
(QRS) [23]. According to the quasi-random sequence, R (n) 0 1 is obtained for
any real number n. Let us assign 11 = 2 (k), 12 = 3 (k), 13 = 5 (k),. . ., mm = (k),

where is the (m m)th prime. Then, each initial tuning matrix k is generated by
{ }

k = k = [k,ij ] [ k,ij k,ij ]|1 i m, 1 j m , (35)
where k = 1, 2, . . . , N . Moreover, k,ij and k,ij denote the upper bound and lower bound
of k,j , and k satisfies the convergence condition (30). Therefore, the initial population
P = {[k ]|k = 1, 2, . . . , N } is given by (35) (e.g., N = 30). In addition, a spare population
SP = {[k ]|k = N + 1, N + 2, . . . , N + N } of size N is generated by (35), too.
For evaluating the performance index of ILC tracker via EP, each objective function
score OF (k ), k = 1, . . . , N is obtained from the individual k and defined as

OF (k ) = Ok = if inal for k = 1, . . . , N, (36)
where if inal is the final iteration epoch of the EP-ILC tracker when the sup-norm of the
tracker error sup[yif inal r] satisfies the tolerant error bound ILC ; evolutionary program-
ming searches for the optimal learning gain by (36). Notice that the individual k must
be replaced with some one from SP if the corresponding objective function score OF (k )
reaches the tolerant learning epoch iILC .
4. Other Improved Issues of the ILC Tracker.
4.1. Initialization of the control input for the ILC tracker. In the field of it-
erative learning control, one of the objectives focuses on decreasing the number of the
learning epoch as large as possible, and the decision for initial control input is a critical
factor to enhance the learning convergence rate. However, Literature [8] does not discuss
how to set it. Hence, the control input of the high performance linear quadratic analog
tracker (LQAT) [24] is proposed in this paper to act as the initial control input of ILC to
significantly speed up the convergence rate of ILC.
4.2. Modified learning rule for robot manipulator. Robot manipulator is an ef-
ficient instrument in industry. Many kinds of control algorithms usually exhibit their
tracking performances through the device. Particularly, ILC adequately controls a robot
manipulator due to the property of the industry robot which repeatedly executes the same
task. However, the dynamic equation of the robot manipulator usually does not satisfy
Assumption A4. Hence, the modified learning rule of ILC is presented as follows:
Consider the dynamic model of an n rigid-link robot manipulator as:
M (q)
q + V (q, q)
+ G(q) = , (37)
q Rn1 are the joint position, velocity, and acceleration vectors, respectively;
where q, q,
M (q) Rnn is the inertia matrix, which is symmetric and positive definite; V (q, q)
R n1
is the centripetal-Coriolis vector; G(q) R n1
is the gravity vector; R n1
is the
control torque. According to (37), the dynamic equation of the robot manipulator can be
derived as
[ ] [ ][ ] [ ]
q 0nn Inn q 0nn
= + , (38a)
q M 1 (q)G(q)
M 1 (q)V (q, q)
q M 1 (q)
[ ]
[ ] q
y = cq cq , (38b)
q
[ ] [ ]
[ ] bq 0nn
where C = cq cq and B = = ; G(q)q and V (q, q)
q are derived from
bq M 1 (q)
G(q) and V (q, q) by the optimal linearization algorithm [24], respectively. Furthermore,
the elements of C denote the coecient matrices of states q and q, respectively; the
elements of B(q) are bq and bq. In general, cq and cq are the constant matrix and the zero
matrix, respectively, because each output of the robot usually represents the joint position
or its linear combination. Due to the pre-mentioned phenomenon, CB(q) is a null matrix.
Based on the convergence analysis of ILC shown in Section 2.3, the convergence condition
(30) does not hold even if the learning gain is appropriately adjusted.
To fit the convergence condition (30), the learning rule of ILC must be modified as
vi+1 = ui + G( r yi ),
xi )(
(39)
ui+1 = sat(vi+1 ),
where r and yi are the double derivatives of r and yi , respectively; yi is derived as

[ ]
[ ]
X
y(t) = cX1 cX2
1
= cX1 X ,
1 = Cx (40)
X2
[ ]T [ []T ] [ ]T
where x = X1T X2T
= q q T T
, cX2 = 0, C = 0nn CX1 , x = X1 X2
T
T .
According to the convergence analysis of OB-ILC presented in Section 2.3, the convergence
condition (30) can be rewritten as

In G(
xi )CB(xi ) = In G( xi )cX1 bX 2 < 1, (41)

[ ]T
i) =
where B(X bT bT . Hence, the proposed learning gain G(
xi ) can be assigned
X X 2
( 1
)1
as non-singular cX1 bX 2 for simplicity and convenience.
Remark 4.1. The proposed modified learning rule can be applied to the class of the
Newton-Euler equations of motion, whose CB is not full rank.
4.3. Resolution of initial shift problem for OB-ILC. In practical control engineer-
ing, the desired reference may not equal the system output at initial time instant. How-
ever, the situation apparently conflicts with Assumption 5 mentioned in Section 2.1.
Hence, this paper resolves the problem by constructing a feasible reference for OB-ILC.
The main idea of the resolution for the initial shift problem is depicted in Figure 3, where
the desired reference r(t) is dierent from the system output y(t) at initial time instant;
applying any control algorithm, the system output unavoidably needs some settling time
to approach the desired reference. Hence, constructing a feasible reference rf (t) to satisfy
Assumption 5 for OB-ILC is allowable during the settling time interval.
In order to create a feasible reference, two conditions are required as follows:
I. The initial value of the feasible reference is set as that of the system output (i.e.,
rf (t0 ) = y(t0 )).
II. At the settling time instant, the value and the corresponding derivative of the feasible
reference are respectively assigned as ones of the desired reference (i.e., rf (tk ) = r(tk )
and rf (tk ) = r( tk ), where rf (t) dt and r( t) dr(t)
drf (t)
dt
).
Then, the feasible reference is constructed via the following equation
{
a1,j t2 + a2,j t + a3,j t0 t tk,j
rf,j (t) = for j = 1, . . . , p, (42)
rj (t) tk,j t tf,j
where t0 , tk and tf represent the initial time instant, the settling time instant, and the final
]rf (t) consists of a second
p
simulation time instant, respectively; the feasible reference
[
order polynomial at the settling time interval t t0 tk and the desired reference r(t)
Figure 3. The generation of the feasible reference

[ ]
at the other interval t tk tf . Then, according to above two conditions and (42),
polynomial equations in matrix form for the j th system output can be attained as
T A = R, (43)

t20,j t0,j 1 [ ]T [ ]T
where T = t2k,j tk,j 1 , A = aT1,j aT2,j aT3,j and R = yjT (t0 ) rjT (tk ) r Tj (tk ) .
2tk,j 1 0 [ ]
Solving (43) can obtain the coecients of rf,j (t), t t0,j tk,j , as A = T 1 R.
5. Design Procedure of Observer-based EP-ILC for Nonlinear Systems. In this

section, an observer-based EP-ILC is presented for a class of the nonlinear systems. Notice
that the analog control system design is simulated under the fast sampling time t = Tf ,
and the design procedure is illustrated as follows:
Step 1: Formulate the optimal linear model
Apply the optimal linearization algorithm to obtain the optimal linear model
of the nonlinear system (1) at each sampling time t = kTf as
x i = f (xi ) + B(xi )ui Ai,k xi + B
i,k ui , for kTf t < (k + 1)Tf , (44)
where Ai,k A( xi (kTf )) is derived from f (xi ) by assigning the estimated state
xi (kTf ) with xi (0) C + yi (0) in Step 2 and B i,k B( xi (kTf )).
Step 2: Formulate the LQR-based observer
According to (10) and (44), the state observer of ILC based on LQR is given
by
x i = Ai,k xi + B
i,k ui + Lc,k (yi yi ), (45a)
yi = C xi , (45b)
where xi is the estimated state of xi , xi (0) C yi (0) and Li,k Lc (
+
x(kTf ))
np
is the observer gain at the fast sampling time t = kTf . Furthermore, from
(11), the observer gain Li,k is designed as follows:
Li,k = Po,i,k C T Ro1 , (46)
Ai,k Po,i,k + Po,i,k ATi,k Po,i,k C T Ro1 CPo,i,k + Qo = 0, (47)

where Qo , Ro
nn pp
and Po,i,k is the symmetric and positive definite
solution of Riccati Equation (47). Assign the weighting matrices Qo and Ro
satisfies the high gain property (i.e., qo o ) where Qo = qo In and Ro = o Ip ,
then the estimated state xi is close to the system state xi .
Step 3: Realize the optimal linear model and initialization (of the control )input
Realize the formulation of optimal linear model Ai,k and B i,k based on the
estimated xi (kTf ) in Steps 1 and 2. Then, determine the control input of the
observer-based high-performance linear quadratic analog tracker [24] u(t) to act
as the initial control input of ILC u0 (t).
Step 4: Design the observer-based iterative learning control
The learning update rule of input ui for OB-ILC is given by
vi+1 = ui + Gi,k (r y ) = ui + Gi,k (r C Ai,k xi C B i,k ui CLi,k Cei ), (48a)
ui+1 = sat(vi+1 ), (48b)
where r is the time derivative of desire reference input r at t = kTf , ei xi xi
and Gi,k G( xi (kTf )) = (C B i,k )1 is the learning gain of ILC.
Step 5: Go to Step 1 and repeat the process of ILC until either the sup-norm of the
tracker error sup[yif inal r] satisfies the tolerant error bound ILC or it reaches
the specified learning epoch.
Remark 5.1. Resolution of EP-based ILC with initial shift for nonlinear systems. The
observer-based iterative learning tracker can be constructed according to the above proce-
dure. To speed up the learning convergence of the ILC, the o-line evolutionary pro-
gramming is applied to tune the weighting gain Gi,k . Firstly, let the tuning matrix
= Im GEP (xi ) CB (
xi ). Then, the new learning gain GEP,k is obtained as:
GEP,k = (Im )(C B i,k )1 , where (C B
i,k )1 exists. (49)
The observer-based EP-ILC is developed according to the procedure of the evolutionary
programming in Section 3. For the initial shift problem of ILC, the feasible reference is
constructed according to (42) presented in Section 4.3.
Some observations of the proposed methodology are given as follows:
1. Based on the above design procedure, the proposed observer-based EP-ILC can deal
with some repeatable tasks for the class of the Newton-Euler equations of motion, such
as robot tracking control and high precision CNC. Although the proposed ILC scheme
requires mathematical models of real systems (i.e., model-based ILC) to construct the
learning structure, the system identification algorithm can be applied to overcome this
restriction. For example, one can adopt either the o-line or on-line identification methods
[27,28] to have the identified time-invariant/varying linear model for unknown/known
time-invariant nonlinear systems. System approximation techniques via fuzzy and neural
networks are also the popular estimation tools. Furthermore, the o-line observer/Kalman
filter identification (OKID) is able to formulate the unknown system to its corresponding
low-order linear model, through an extension of the eigensystem realization algorithm
(ERA) [29,30].
2. Although performing evolutionary programming could obtain a global optimal learn-
ing gain for ILC, the algorithm is time-consuming in the evolutional process and parameter
decision. To alleviate the situation, adopting the fast evolutionary programming can re-
duce the searching time for the o-line optimization problem; for the on-line case, combing
some feedback control theorems [24,31] with ILC, the so-called plug-in type ILC [6,13]
significantly improves the training convergence speed.
3. Considering the practical implementation for the proposed ILC, one can apply
the techniques of the optimal linearization and prediction-based digital redesign [24] to
develop the corresponding time-invariant/time-varying linear digital controller for the
sampled-data linear/nonlinear system without altering the whole ILC scheme; therefore,
the digital controller is easily established by microprocessor or digital signal processor
(DSP). All above comments can be considered as highly potential research topics.
6. Illustrative Examples.
6.1. Observer-based ILC for a class of nonlinear systems. Consider the nonlinear
system
x = f (x) + B(x)u,
y = Cx,

cos(x2 ) 3 1
where f (x) = x1 + sin(x2 ) + cos(x3 ) , B(x) = 0 0 and C =
x1 + sin(2x2 ) + 3x3 x1 1 2(2 + 0.5 sin(x2 ))
[ ]
1 0 0
. The nonlinear function f (x) can be partitioned into Al x and (x) based on
0 0 1

0 0 0 cos(x2 )
[8], where Al = 1 0 0 and (x) = sin(x2 ) + cos(x3 ) . Obviously, matrix B(x)
1 0 0 sin(2x2 ) + 3x3 x1
includes the state x2 which can not be acquired from the system output, and it aects the
decision of learning gain of ILC. Hence, designing an observer must be considered for this
case. However, according to Reference [8], (Al , C) is not an observable pair, so the original
observer design could not work. Nevertheless, the proposed LQR-based observer can avoid
the restriction and accurately estimate the system state, since the re-constructed (Ak , C)
pair is observable at every operating state xk .
Let the parameters of the observer-based ILC for the system be given as follows: The
[ ]T [ ]
desired reference is r = sin(2t) cos(2t) for time interval t 0 . The tolerant
[ ]T
error bound is assigned as ILC = 8 104 8 104 , and the specified learning
epoch is given as iILC = 30. Moreover, in the observer scheme (11) and (12), constant
weighting matrices are given as Qo = 3 104 I3 and Ro = I2 , respectively. Then, the
simulation result of observer-based ILC is shown in Figure 4. Figure 4(b) indicates that
the sup-norm errors sup[ri yi ] for i = 1, 2 converge to the tolerant error bound ILC at
the 19th iteration; Figure 4(a) shows that systems outputs yi for i = 1, 2 all match the
corresponding references at the final iteration. Furthermore, Figure 4(c) shows the LQR-
based observer can accurately estimate the unmeasurable state x2 even though (Al , C) is
not observable.
6.2. Alterative learning update rule of ILC for robot manipulator. Consider the
dynamic model of a two rigid-link robot manipulator as follows:
M (q)
q + V (q, q)
+ G(q) = ,
where
[ ]
m2 l22 + (m1 + m2 )l12 + 2m2 l1 l2 cos(q2 ) m2 l22 + m2 l1 l2 cos(q2 )
M (q) = ,
m2 l22 + m2 l1 l2 cos(q2 ) m2 l22
[ ]
m2 l1 l2 sin(q2 )(2q1 q2 + q22 )
V (q, q)
= ,
m2 l1 l2 sin(q2 )q12
(a) (b)
(c)
Figure 4. (a) The responses ri and yi by the OB-ILC at the 19th iteration
for i = 1, 2, (b) Sup-norm errors sup[ri yi ] by the OB-ILC at each iteration
for i = 1, 2 and (c) The estimation errors of x2 at the 1st and 19th epoches,
respectively
[ ]
(m1 + m2 )gl1 cos(q1 ) + m2 l2 cos(q1 + q2 )
G(q) = ,
m2 gl2 cos(q1 + q2 )
[ ]T
q = q1 q2 , q1 and q2 are the angular positions. The related parameters of the
robot represent the link masses, m1 = 4.6Kg, m2 = 2.3Kg, the lengths l1 = l2 = 0.5m,
and the gravitational acceleration g = 9.8m/s2 , respectively. According to the[ optimal
]
bq
linearization algorithm, this model can be rearranged as (38), where B(q) = =
bq
[ ]
022 [ ] [ ]
1 and C = cq cq = I22 022 . Obviously, the product matrix CB is a
M (q)
null matrix, which violates Assumption 4. To overcome this situation, apply the proposed
learning update rule (39) to replace the original one (13) via assigning the learning gain
( )1 [ ]T

xi ) as the non-singular cq bq
Gi ( . When parameters r = 0.25 cos(4t) 0.5 cos(5t)
[ ] [ ]T
for t 0 3 , ILC = 2.5 104 5 104 , iILC = 30, Qo = 104 I4 and Ro = I2
are given, simulation results of the observer-based ILC for robot manipulator are depicted
(a)
(b)
Figure 5. (a) The responses ri and qi by the OB-ILC at the 23th iteration
for i = 1, 2, (b) Sup-norm errors sup[ri qi ] by the OB-ILC at each iteration
for i = 1, 2
in Figure 5, which demonstrates the high-speed convergence of the proposed ILC and the
high-performance tracking purpose.
6.3. Observer-based EP-ILC tracker for chaotic system with initial state error.
In the previous simulations, the observer-based ILC exhibits a good control performance
after finite epochs. However, slowing convergence learning rate may exist in the ILC for
complicate and nonlinear systems. Therefore, the o-line evolutionary programming is
applied to improve the ILC tracker. Now, consider the chaotic system

x 1 35(x2 x1 ) 0 0 [ ]
x 2 = 7x1 x1 x3 + 28x2 + 4.5 + sin(x1 ) 0 u1
u2
x 3 x1 x2 3x3 0 4.85

[ ] [ ] x1
y1 0.18 0.6 0
= x2 .
y2 0.05 0 0.7
x3
[ ]
3 + 5.65 sin(11.45t)
The related parameters of the observer-based ILC tracker are r =
20 + 5 cos(12t)
for t [0 1], ILC = [0.01 0.01] , iILC = 80, Qo = 10 I3 and Ro = I2 . Moreover, in or-
T 5
der to overcome the constraint of initial state error for OB-ILC, the feasible reference
[ T T
]T
rf (t) = rf,1 (t) rf,2 (t) according to Section 4.3 is given as follows:
{
a1,1 t2 + a2,1 t + a3,1 t0 t tk,1
rf,1 (t) = and
r1 (t) tk,1 t 1
{
a1,2 t2 + a2,2 t + a3,2 t0 t tk,2
rf,2 (t) = ,
r2 (t) tk,2 t 1
where a1,1 31227.402, a2,1 435.2171, a3,1 = 5 and tk,1 = 8 103 s; a1,2
20358.705, a2,2 399.9914, a3,2 = 23 and tk,2 = 0.01s. In addition, according to
Section 3, the parameters of evolutionary programming are given as = 10, = 1,

= 0.0001, = 1, generation
[ = 20, N = 30, N = 5, and
] the interval of the turning
[ 0.8 0.8 ] [ 0.8 0.8 ]
matrix is assigned as I = . By applying the o-line EP
[ 0.8 0.8 ] [ 0.8 0.8 ]
to the observer-based ILC, the final tuning matrix and optimal leaning gain GEP ( xi )
are obtained as follows:
[ ]
0.605 0.23309
= , GEP ( xi ) = (I2 )G(xi ).
0.09154 0.001878
Figure 6 depicts simulation results by the proposed OB-ILC tracker and the proposed
EP-ILC tracker with the proposed observer. From Figure 6(a), one can see the convergence
speed of the proposed EP-ILC tracker is extremely faster than the one of the proposed
OB-ILC tracker. Moreover, applying the proposed initialization of input for the EP-
ILC, according to Section 4.1, significantly increases the learning rate of convergence. In
addition, Figure 6(b) indicates the proposed EP-ILC overcomes the restriction of initial
state error, and the system output accurately traces the desired reference.
7. Conclusions. In this paper, observer-based iterative learning control with evolution-

ary programming is proposed for a class of nonlinear systems. By applying linear quadratic
regulator and optimal linearization algorithm, the LQR-based observer can accurately es-
timate some unmeasurable states to design the learning gain of ILC for a class of nonlinear
systems. Furthermore, when the observability of the pair (Al , C) in the literature is not
full rank, the proposed observer can still perform the eective work. In terms of conver-
gence analysis, a sucient condition alleviates the traditional constraint, identical initial
state; the proposed OB-ILC guarantees the tracking error converges to zero via repeated
learning. Then, the feasible reference based on polynomial approximation is constructed
to overcome the problem of initial state error. In addition, an alterative learning rule is
proposed to the ILC tracker for the class of the Newton-Euler equations of motion where
CB is not full rank, such as robot manipulator. Finally, in order to increase the learning
convergence speed of OB-ILC as fast as possible, evolutionary programming is applied to
(a)
(b)
Figure 6. (a) Sup-norm errors sup[ri yi ] by the OB-ILC, EP-ILC and

EP-ILC with initial input decision at each iteration for i = 1, 2, (b) Track-
ing results ri and yi by the EP-ILC with initial input decision at the 20th
iteration for i = 1, 2
identify the optimal learning gain in some specific search domain. The simulation result
indicates that the number of learning iterations for observer-based EP-ILC is extremely
less than that of the original OB-ILC. Moreover, adopting the input signal of LQAT as the
initial control input of ILC upgrades the learning convergence rate of the observer-based
EP-ILC significantly.
Acknowledgement. This work was supported by the National Science Council of Tai-
wan under contract NSC99-2221-E-006-206-MY3, NSC98-2221-E-006-159-MY3. The au-
thors also gratefully acknowledge the helpful comments and suggestions of reviewers,
which have improved the presentation.
REFERENCES
[1] D. A. Bristow, M. Tharayil and A. G. Alleyne, A survey of iterative learning control, IEEE Control
Systems Magazine, vol.26, no.3, pp.96-114, 2006.
[2] S. Arimoto, S. Kawamura and F. Miyazaki, Bettering operation of robots by learning, J. of Robotic
Systems, vol.1, no.2, pp.123-140, 1984.
[3] J. X. Xu and R. Yan, On initial conditions in iterative learning control, IEEE Trans. on Automatic
Control, vol.50, no.9, pp.1349-1354, 2005.
[4] D. Meng, Y. Jia, J. Du and F. Yu, Robust design of a class of time-delay iterative learning control
systems with initial shifts, IEEE Trans. on Circuits and Systems I: Regular Papers, vol.56, no.8,
pp.1744-1757, 2009.
[5] Y. Wang, F. Gao and F. J. Dole III, Survey on iterative learning control, repetitive control, and
run-to-run control, J. of Process Control, vol.19, no.10, pp.1589-1600, 2009.
[6] C. Mi, H. Lin and Y. Zhang, Iterative learning control of antilock braking of electric and hybrid
vehicles, IEEE Trans. on Vehicular Technology, vol.54, no.2, pp.486-494, 2005.
[7] S. Gunnarsson, M. Norrlof, E. Rahic and M. Ozbek, On the use of accelerometers in iterative learning
control of a flexible robot arm, International J. Control, vol.80, no.3, pp.363-373, 2007.
[8] A. Taybei and J. X. Xu, Observer-based iterative learning control for a class of time-varying nonlinear
systems, IEEE Trans. on Circuts and Systems I, Fundamental Theory and Application, vol.50, no.3,
pp.452-455, 2003.
[9] Y. Chen, C. Wen, Z. Gong and M. Sun, An iterative learning controller with initial state learning,
IEEE Trans. on Automatic Control, vol.44, no.2, pp.371-376, 1999.
[10] H. S. Lee and Z. Bien, Study on robustness of iterative learning control with nonzero initial error,
International J. of Control, vol.64, no.3, pp.345-359, 1996.
[11] K. H. Park, An average operator-based PD-type iterative learning control for variable initial state
error, IEEE Trans. on Automatic Control, vol.50, no.6, pp.865-869, 2005.
[12] M. Sun and D. Wang, Iterative learning control with initial rectifying action, Automatica, vol.38,
no.7, pp.1177-1182, 2002.
[13] R. Chi, Z. Hou, S. Sui, L. Yu and W. Yao, A new adaptive iterative learning control motivated
by discrete-time adaptive control, International Journal of Innovative Computing, Information and
Control, vol.4, no.6, pp.1267-1274, 2008.
[14] J. X. Xu and Y. Tan, Robust optimal design and convergence properties analysis of iterative learning
control approaches, Automatica, vol.38, no.11, pp.1867-1880, 2002.
[15] Y. Ye, A. Tayebi and X. Liu, All-pass filtering in iterative learning control, Automatica, vol.45, no.1,
pp.257-264, 2009.
[16] T. Sugie and F. Sakai, Noise tolerant iterative learning control for a class of continuous-time systems,
Automatica, vol.43, no.10, pp.1766-1771, 2007.
[17] C. A. Liu and Y. Wang, A new dynamic multi-objective optimization evolutionary algorithm, In-
ternational Journal of Innovative Computing, Information and Control, vol.4, no.8, pp.2087-2096,
2008.
[18] C. Liu, New evolutionary algorithm for multi-objective constrained optimization, ICIC Express Let-
ters, vol.2, no.4, pp.339-344, 2008.
[19] C. Liu and Y. Wang, A new evolutionary algorithm for multi-objective optimization problems, ICIC
Express Letters, vol.1, no.1, pp.93-98, 2007.
[20] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolutionary Programs, Springer-Verlag,
1996.
[21] J. F. Chang, Y. S. Yang, T. L. Liao and J. J. Yan, Parameter identification of chaotic systems using
evolutionary programming approach, Expert Systems with Applications, vol.35, no.4, pp.2074-2079,
2008.
[22] M. L. Hung, J. S. Lin, J. J. Yan and T. L. Liao, Optimal PID control design for synchronization of
delayed discrete chaotic systems, Chaos, Solitons and Fractals, vol.35, no.4, pp.781-785, 2008.
[23] J. S. H. Tsai, Y. Y. Du, A. C. Dunn and L. S. Shieh, EP-based robustness analysis of optimal
load allocation for boiler systems, IMA J. of Mathematical Control and Information, vol.26, no.2,
pp.197-211, 2009.
[24] S. M. Guo, L. S. Shieh, G. Chen and C. F. Lin, Eective chaotic orbit tracker: A prediction-
based digital redesign approach, IEEE Trans. on Circuits and Systems I, Fundamental Theory
and Applications, vol.47, no.11, pp.1557-1570, 2000.
[25] J. S. H. Tsai, C. L. Wei, S. M. Guo, L. S. Shieh and C. R. Liu, EP-based adaptive tracker with ob-
server and fault estimator for nonlinear time-varying sampled-data systems against actuator failures,
J. of the Franklin Institute, vol.345, no.5, pp.508-535, 2008.
[26] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas with Application to Linear
Systems Theory, Princeton University Press, 2005.
[27] T. Soderstrom and P. Stoica, System Identification, Prentice Hall, NJ, 1989.
[28] S. M. Guo, L. S. Shieh, C. F. Lin and J. Chandra, State-space self-tuning control for nonlinear
stochastic and chaotic hybrid systems, International J. Bifurcation and Chaos, vol.11, no.4, pp.1079-
1113, 2001.
[29] J. N. Juang, Applied System Identification, Prentice Hall, NJ, 1994.
[30] J. S. H. Tsai, T. H. Chien, S. M. Guo, Y. P. Chang and L. S. Shieh, State-space self-tuning control
for stochastic chaotic fractional-order chaotic systems, IEEE Trans. on Circuits and Systems Part
I: Regular Papers, vol.54, no.3, pp.632-642, 2007.
[31] Y. Morita and H. Wakuya, Online feedback error learning control for an inverted pendulum, Inter-
national Journal of Innovative Computing, Information and Control, vol.5, no.10(B), pp.3565-3572,
2009.

!"# $
%#

&' ('
%) *+# ',
%+# '
- .%% / -
0% !"#
' 1 (/' $% * $
% !+%
'
"- /% 2 +"''/'

!"# $
% !"

!" $ %

% %-# 3) &-
-
#""

! " # "
$ "

# %
& " ' %( !

) *
+
, "
# $

* " %

!
( % &

'
%( %( ! "

"

"
- (
"# ./

"
) .01/ "
"

* # % ) 2

.3/ 2

.4(/ 2

% % (
2 2

* 2 (

5

" " #
(
"
2
'

"
"
6

"

%

"

% $
2
2

7 8+ 978:
2
" )

78(
;
78( (
. /

78( , "

"
$ (

< ./ " 78(
-
78 96=78: 2 , "

" $
< ' ; = ./ " 2

# , "

2 < ; .0/

78( %

"
78(
% 78 $
% "

$ + " $ 78(
= ' "
> ) ) 0
$ % " ) 1
) 3
!" 2 "
#$ 7 2+ 978: "(# " (
2 78

# (
78 78
# (

#
;&? .1/ $ (

# ( " 9 @ : "

% "

# 78 78
-
(
#

# " " @ 9 :

@ 9
: $ "

# 9 :

"

( "
"

A " A2 9:

9 : @ 9

:
9:

" @ %

# "
@ @
#
%

% 9 : 9

:(+ % $

9 :
$ 78
#
" % $
$ 78 (
"
$
#

( #
%& '" "%( B = .3/

78(

# " " "
"
#
" @ 0 "
"

78 % $
"

" C%
#
" + @ ) "

78 % $

"
$
# % - %
% # %
D % E )
DE
" " " %

" % = ' - (

% $
$
# 4
" 9 @ 4: -
# 3 " %
" " % "

%
% "

%
%

% "
%
%

$ % % $

% % $ % "
DE
% $

%
% , " % "
% +

" +

! " - % % $
0 "

$ -

%
0
% 0 % 0 "
% 0
% $
$ % % $

"

$

%
% B " % '
%
% B
% $
$ % % 3

%

% $
%

%
D E % $
" % ' %
% % $

"

% $ % 3 % $

A% = '

(

$

"
%
" (
%

# + @ 0 -
# + @ 0 " " "
% @ % @ 0 $ "
# % $ +
2 3 %
"

" %

$

" %
# %
# "
" $ - %
"

" (
9 : 9: $ %

"
" (
9 : 9 :
0
% )" %( 78( (
- " - % $

$

)
% $

78( % " $

% 0

A% = '

" (

$ > "
0 )""& )%" "

- "
% $

" ) +
+
# + "
#
$ %

# % % $
" -
2 % % " % #
+ "
#
" 2
) "
# F=! .4/ " (
# F=!

#
# " * "
78( " "

# % $
= 2 G
" "

" *
#

C% " "
% $
$
A2 9:
H
@

9:

" * %

* % % " 1 $ G

"
% 9 @ :

; *
0 (*& )%" ! % $

-
(

78(
%

- A2 90:I
9 9 :1: 90:
" (-
9 :1 $

78( %

% "

! %

%
I - " " I
= I @ @
= I @ @
5

+ ,
@

@
I - 9 : 9 : " " I
= I @ @
= I @ @
5

4
5

''' '''

5 ''' '''
0I - " I

@
$ 0

"
) I % -
. 0 /

"

( $
> -
"

% &
+

%
$ " %

9: 9
:
9: % $

F=!
#
9
: A % $

) ! % 00 % " %
$
) @
00 @ ! @ @ 34
@ 9 :
) A% " 00 " " 33 $ "
#
" 31
F=! !
% $

F=!
# " 39:

) 0 $ - % @ J =
I
H 0
@ @ 3
@ @ 3 0 @ 3
) 1 9 : 9 : @ = I @
% @
) 3 ) 0 ) 1 % -
" - "
% $
39:
" 39
:
$ % @ 14 "

@ 3 0
@
14 , " 14 . 0 /
9 .1 31/: $ @ 14
93314:
C

2
00 -& "& )%"

% $
9 :(+
#
!

$ %
"
! %

%
%
A% I

@ @

A% I

@ @ H 9 :

A% 0I

@ @ H 9 :

A% 1I 0

@ @ H 9 :

A% 3I

@ 9 : @ % %
A% 4I ( -
@ %
, " %
%
$ % $
"
49:
9: 9
:
9: A % $
9
: % $

) @
00 @ @ 34 & @
00
@
)
% @ "
=
A2 9:
H 0
@ @ 3
) 0 =# % % ! %
1I
0 9 1 00:
%
@ %
@ H 9 : 9 @ 0 :

) 1 ) ) 0 % -
" - "
% $
49: "
49
: @ 9 : %
1 -)( "" % % &
K ;
F ? , & 3 3 % + " "
B $ "
# " 1 1 % + !
#
34 " " (
" ) +
78 % $
2 40J1 9 93 3:L91 1::
9: &
9
: ! 9: ;
9: F 9: ? , 9: &

! )%
J 78 % $
$
F)C %
% 78 $ %
"

$
0 "

$

78 % $
$
A2
91:I
$
@ $ )+ = 9
:
$ )+ 9%: 91:

$
0

$

"
$
"
"
$
0
" % 40
$
1 "

9 + :
$
>

"

9: &
9
: ! 9: ;
F)C @1 B & F)C @0 3J & F)C @00BJ &
9: F 9: ? , 9: &

F)C @0 BB & F)C @0 & F)C @0J44 &
" 5 78
,

6 6 6 6 6 6
-# -# -# -# -# -#

78 '8 '8 *7* '8 8 '8 8 '8 7 '8
8 '8 8 '8 7 '8 '8 7 '8 '8
* '8 7 '8 '8 '8 * '8 *8 '8
! 8 '8 8 87 '8 8* '8 88 '8 8* '8 8 '8
! * '8 8 '8 8 '8 8 '8 88* '8 8 '8
C

(&" * . ))
(* / " B J 4 J
$
3 "
< ;' =
' $ "
=

- $ M < ;' " "
"

3 "" "
78
$ $
% "

<
;' = ' 6

< ;' =
' $

= .0/ .3/ (

"
# $ "
%" $ "

&'"
%' () %' () %' ()
*
87 '8 7 '8 8* '8
8 '8 7 '8 8 '8
7* '8 7 '8 8 '8
7 '8 '8 8 '8

'88 '8* 88* '8
+ 8*8 '8 '8 8** '8
, 8 '8 7 '8 88 '8
" "

" , "
%
$
%
)
" "
2 " %
(
"
-./.-.0%.&
9: ' ';' < / '' (!"
= !" "/ / /%
"
!
'8
'* --' '
9: >' ? / $' ? @ !
!"
=%
% "/ !

% "
&% ' !
' 'A@B --'* '
9: 'C' $" D'0' % E'' 0 / '' % (!"
/ /% /
""
"" "
# %" "% -"
%
- " ( !
'
' --' 7 8'
9: <' $ (!"
/ //% "% /F G-" &&& $
) ! $ !
' ' --' 7 '
98: H' I% / ' 1% . "% %- //% # G-
% /J /
&&& ' !
' ' --' 7'
97: 1'C' $ / '' % /% "/ > -""/ %" "% % /"
/ /
"%
!
'8
'* --' 8 '
9*: ;'' C 'I' ( / '<' C @--
#% "
"%
"%
% 3)
&% ' !
' 'A6B --' '
9: ' < / D' E @ /%
% =% " "/ ! 3) & &
$ ) !
'.8 '7 --' 8 87 '
9: '' % '' C / $'' @ - !/ > / = "
% "% --
-
#""
! !
' ' --'*
* *'
9 : ;'' C / '' 1% %
%" =% # ! 3) ( *
) $- --'*7* '
9: 6' ;% I'' C / 'D' (!"
=% > -""/ / (
+ ,)$&- ! (
6/ --' 8'
9: ';' C / '' % D/% / > -""/ %" "% /""
-"

!
'* ' --' 7'
9: 'D' ;% / ;'' C (!"
/ /% > /G
"/ / "

! !
' '7 --' * '
9: ;' C/ @' 6) / (' ' ?# @
% ! 3) /"% &&& $
!
' ' --'8 '
98: '' % 1'' 1 / ;'' D C ""
"" !# > /G
/// "
/
! !
' ' --' *7 *'
97: ' $' <
F / ' $' <
F ( , -% ; = 7'

c ISSN 1349-4198
A NOVEL TECHNIQUE FOR DESIGNING DECENTRALIZED

STABILIZERS FOR ROBUST CONTROL IN POWER
SYSTEMS USING AN H CRITERION
Alexandre C. Castro1 , Jose M. Araujo2 , Eduardo T. F. Santos2

Fabrcio G. S. Silva2 and Clivaldo S. de Arau
jo1
1
Centro de Tecnologia, Departamento de Engenharia Eletrica
Universidade Federal da Paraba
Campus Universitario I, Joao Pessoa 58051-900, Brazil
{ castro; clivaldo }@ct.ufpb.br
2
Departamento de Tecnologia em Eletro-Eletronica, Grupo de Pesquisa em Sinais e Sistemas
Instituto Federal da Bahia
Rua Emdio dos Santos, S/N, Barbalho, Salvador 40301-015, Brazil
{ jomario; eduardot; fsimoes }@ifba.edu.br
Received October 2009; revised April 2010
Abstract. This paper addresses the problem of excitation control for damping the elec-
tromechanical oscillation modes (EOM) in power systems with decentralized stabilizers
(PSS) of reduced order to ensure robust stability and good performance of a power sys-
tem under disturbances and uncertainties. Good stabilizer coordination is achieved by si-
multaneously designing all PSS. A frequency domain H mixed sensitivity optimization
approach is proposed. The proposed technique is not aected by the pole-zero cancellation
problem associated with the conventional mixed sensitivity techniques.
Keywords: Power systems stabilizers, H control, Decentralized control
1. Introduction. Large scale, interconnected power systems are ubiquous in present

days. As a consequence, quality indexes based on reliability [1], frequency and magnitude
of voltage [2] among others are strictly required. Low frequency electromechanical oscil-
lations have been a problem of great concern and continuous interest for interconnected
power systems. These oscillations are mainly associated with local and interarea EOM.
It is known that the operating conditions of the system changes continuously. Thus,
for secure operation of power systems, it is required that the EOM must be suciently
dampened at all common operating conditions. This means that the PSS must be ro-
bust for these changes [3, 4]. Conventional lead-lag PSS have been successfully applied
to dampened EOM over decades. Its use is consolidated in power industry all over the
world. However, its application in multimachine power systems have been carried out
by independent designs or by a sequential design, which, due to interactions, may not
provide the optimal settings. Moreover, the conventional PSS design does not take into
account the disturbances, measurement noises and uncertainties due to nonlinearities, ne-
glected dynamics in the model, changes in operating conditions, etc [4]. Recently, some
researchers have been concerned in providing sucient damping for the EOM in spite of
uncertainties. The techniques most used for robust design of controllers in power systems
are: linear matrix inequality [4, 5, 6] and mixed sensitivity optimization [7, 8, 9]. All
these techniques have a common limitation of resulting into large centralized controllers
when directly designed in a multimachine power system model. In order to minimize
this problem, the model order is reduced and then controller order is also reduced [8].
1387

1388 A. C. CASTRO, J. M. ARAUJO, E. T. F. SANTOS, F. G. S. SILVA AND C. S. DE ARAUJO
The mixed sensitivity approach deserves a special consideration due to its attractive fre-
quency response shape where robust stability and robust performance improvements can
be easily visualized. However, its conventional approach has two additional drawbacks:
the pole-zero cancellation and the needing of selection of three weighting matrices [10].
These drawbacks make it dicult to apply this approach in large systems. The basic
concern of this paper consists of exploring improvements for the control structure already
existing in the power industry, leaving for the future the replacement of current PSS for
controllers by new structures. In this paper, a new procedure is proposed for PSS tuning,
aiming H robust control of a power system, with decentralized stabilizers with reduced
order. The use of decentralized stabilizers is a practical requirement in order to avoid the
need for remote feedback over long distances, as expected for decentralized power systems.
The frequency domain propositions of H robust control together with an optimization
procedure are applied in order to determine the PSS gains and time constants. The PSS
systems connected to the set of selected generators are tuned simultaneously in order to
take into account the interactions among generators and controllers. The optimization
procedure applied to tune PSS consists on minimizing a weighted transfer function ma-
trix, yielding a mixed-sensitivity objective function [8] suitable for the application under
study. The cancellation problem related to EOM complex zeros is avoided by constraining
the weights, poles and zeros of PSS to be in the real domain.
2. Signal Analysis for Decentralized Control. Only a limited number of PSS require
the damping of the critical EOM. Thus, the most suitable generators for application of
PSS should be previously selected. The PSS function is the EOM damping, aiming to
reduce the undesired eects of disturbances and noise measurements and also to achieve
a good response despite of model uncertainties. Tracking is not of concern in PSS design.
Frequency domain techniques applied to damp EOM are closely related with the phys-
ical phenomenon and with the eld practice in power systems. Besides, frequency do-
main multivariable techniques may be successfully applied for interactions analysis, input-
output selection and for designing robust controllers.
A general power system with n generators is described by
y(j) = G(j)u(j) (1)
where G(j) is the transfer function matrix, y(j) and u(j) are respectively the output
and the input vector.
The EOM controllability may be dened as the ability of the system to damp the mode
in order to achieve acceptable performance with bounded inputs and outputs. Likewise,
the observability of an EOM may be dened as the contribution of the mode on system
response. For analysis of controllability and observability of MIMO power systems, the
singular values of the matrix G(j) are required.
The maximum singular value of G is denoted by (G), the minimum singular value is
denoted by (G) and the condition number is given by (G) = (G)/(G).
The singular values
and in the frequency of an EOM are, respectively, measurements
of observability and controllability of the mode. EOM with small damping and strongly
observable in the output signals display large peaks on the graph of (j). Furthermore,
a large in the frequency of an EOM means that the mode is strongly controllable. A dip
on the graph of (j) represents a complex system zero with signicant eect on response
and possibly on the control.
The eect of on system performance with respect to set-point changes, R, distur-
bances, d, and measurement noises, , is introduced, considering the power system G(s),
with feedback controllers H(s), as usually found in practice, as shown in Figure 1.
A NOVEL TECHNIQUE FOR DESIGNING DECENTRALIZED STABILIZERS 1389
d
+
R + m + y
G ( s)
-
+
+ h
H (s )
Figure 1. The power system with feedback control
The following relations are obtained from Figure 1, omitting s = j

y = T R + Sd T H (2)
u = (I + HG)1 R (I + HG)1 H( + d) (3)
where S = (I + GH)1 is the sensitivity transfer function matrix and T = SG is the
closed-loop transfer function matrix.
Consider a change on setpoints, assuming d = 0 and = 0. Using singular value prop-
erties [11] from (2), it is obtained that y / R = (T )
(G)/(GH + I), resulting
(T )
(G)/((G)(H) 1).
Similarly, by considering only the eect of disturbances on y, it is veried that y / d
1/((G)(H) 1). Then, (G) should be large enough in order to reduce (T ) which is
usually large at the frequencies of the EOM, and also to reduce the eects of disturbances
and measurement noises. Obviously, the controller should be designed to contribute on
both cases.
It is also important to impose bounds in order to avoid input saturation, which con-
strains the control action. Considering only the eect of setpoints in (3), results that
u R /((G)(H) 1). Again, (G) must be large.
Therefore, the most suitable generators set which should be selected for applications of
PSS is the set with the largest (G) in the frequency range of the EOM.
3. Robust H Damping Control. The eects of unstructured uncertainties are now

taken into account. Two common types of uncertainties are presented in Figure 2 [11],
where A and M are respectively the additive and the multiplicative uncertainty ma-
trices and, WA , WM and W1 are weighting function diagonal matrices.
The goal is to design stabilizers in order to damp the EOM, yielding robust stability
and satisfactory system performance, not only for the nominal plant G(s), but for the set
of all plants.
WM (s )DM W1 (s )
+
u(s) + y(s)
G ( s)
+
WA (s )DAW1 (s )
Figure 2. System model with uncertainties

M ( s)
Figure 3. M structure of the system
This set is dened by the expression G (s) = (I+WM (s)M W1 (s))G(s)+WA (s)A W1 (s).
Introducing the uncertainties into the model of Figure 1, considering one uncertainty at
a time, and arranging the blocks, results the model M of Figure 3, where M (s) =
WM T HW1 for = M and M (s) = WA HSW1 for = A .
Assume that the nominal system M (s) is stable and that the uncertainties are stable.
Thus, the M systems are stable for all uncertainties, satisfying (T ) 1, if and
only if [11]:
(M (j)) 1
(4)
For simplicity sake, it is assumed round weights, that is, WA (s) = a (s)I, WM (s) =
m (s)I and W1 (s) = 1 (s)I. Then, from (4), results:
1 (s)m (s) (T ) 1
(H) (5)
for multiplicative uncertainties, and
1 (s)a (s) (S) 1
(H) (6)
for additive uncertainties.
The weight 1 (s) helps in scaling T and S, since, with feedback control, as in Figure 1,
T + S = I, it becomes dicult to satisfy both inequalities of (5) and (6) simultaneously.
Considering, from (2), that S should be small for disturbance rejection and T should
be small enough for noise modulation, then, for robust stability and robust performance,
(T ) and
(S) should be minimized. To satisfy these requirements, the following mixed
sensitivity optimization problem is proposed:
[ ( )]
1 m T H
min sup (7)
1 a HS
3.1. Justification for using identical controller structures. For analysis of this
case, the uncertainties of Figure 2 are lumped into a single multiplicative output un-
certainty matrix given by E = 1 m M + 1 a A G1 . The true power system (with
uncertainties) is given by G (s) = [I + E ] G and the true sensitivity matrix of the system
aected by uncertainties is given by S = [I + G H]1 = [I + (I + E )GH]1 (I + GH)S.
Assuming that in the frequency range of main concern (frequency range of EOM),
occurs that (GH) 1, results:
(H)(G)
(S )

(S) (8)
(I + E )
From (8) it can be seen that the sensitivity of the system may be signicantly deteri-
orated by uncertainties, mostly in the frequencies of the critical EOM, where (G) 1.
Also, the controller may constrain the deterioration on sensitivity if (H) = 1, which
means identical decentralized controllers on all units.
Assuming that a (s) = m (s) = 0 (s), the optimization problem of (7), satisfying
( )
T
(5) and (6), consists on adjusting the parameters of H in order to minimize ,
S
Figure 4. Conguration of the power system
resulting: )
(
T 1

< (9)
S 1 (s)0 (s)
(H)
That allows to leave T and S separated from the weights, allowing the application of
the proposed technique to larger systems, since the use of MATLAB is limited by the
dimensions of the matrix when obtaining singular values. Besides, 0 (s) and H(s) should
have only rst order terms in order to avoid the pole-zero cancellations of the EOM and
complex zeros.
If the last inequality is met (9), the robustness of stability with good performance is
attained. If the inequality is not met, the order
( of)the controller should be increased. It
T
is obvious that, with the minimization of , the peaks of T and S are attened,
S
meaning that the EOM are dampened.
The objective function to be minimized in (9) is not an explicit expression. Then, an
optimization technique to determine the minimum of a function without calculation of
derivatives seems attractive. The direct pattern search optimization technique of Hooke
and Jeeves [12] is used due to its good adaptation to this problem. The limitation of the
technique is that it converges more slowly than the derivative methods.
4. Application to a Multimachine Power System. A multimachine power system is

shown for illustration of proposed technique application [13]. Its transmission congura-
tion is shown in Figure 4. The generator #10 is a large equivalent generator, representing
the rest of the system. The data set used for tests, which may be obtained from the
authors, was omitted. It is assumed that speed signals are used as outputs and input
voltages of the exciters are the control inputs.
In Figure 5, the singular values (T ) and (T ) of the complete system are displayed.
The peaks on the graph of (T ) reveal that the system has four EOM insuciently
damped. These EOM are critical and they were named: Mode 1 (0.46 Hz), Mode 2 (0.86
Hz), Mode 3 (0.89 Hz) and Mode 4 (0.92 Hz). The two former ones are interarea modes
and the latter ones are local modes.
It is known that modern exciter systems with fast responses and high gains are more
appropriate for application of PSS and that small generators may not provide sucient

Table 1. Values of inertia constants and exciter gains of the generators
Generator 1 2 3 4 5 6 7 8 9
H 11.0 11.5 10.0 6.4 9.9 9.9 9.0 9.0 15.0
KA 50.0 50.0 20.0 15.0 100.0 100.0 10.0 10.0 50.0
damping for EOM [13, 14]. Then, only the largest generators with high gains, KA , will
be considered for application of PSS.
Table 1 shows the values of inertia constants, H, and exciter gains, KA , of the gen-
erators. From Table 1, the generators #3, 4, 7 and 8 do not satisfy the requirements,
and are consequently discarded, because these are generators with low gains KA . Then,
generators #1, 2, 5, 6 and 9 are selected for further analysis. Usually, a number of PSS
equal to the number of critical EOM are usually enough to damp these modes. Thus, it
is assumed that PSS ( should
) be applied, as an attempt, on four generators.
5
Thus, there are = 5 sets of generators for nal analysis. The graphic analysis
4
of these sets shows that the set of generators #1, 2, 5 and 6 presents the largest (T ) in
the range of the EOM. The graphs of (T ) and (T ) of this set are presented in Figure
6. Since this set presents the largest controllability with good observability of all critical
EOM, the generators #1, 2, 5 and 6 are selected for application of PSS. The graph of
Figure 6 shows that the selected set in this paper provides better controllability of the
critical EOM than the complete set of 9 generators in Figure 5, because they present higher
values of (T ). This is due to the interactions among generators, which may constrain
the control of EOM, mainly Mode 4. It means that PSS applied on the complete system
may not provide results as good as the ones shown for the selected set.
As can see the graph of (T ) in Figure 6, there is a zero complex (dip on the graph).
However, the graph analysis leads to the conclusion that the zero does not aect the
controllability of any EOM, because it hasnt the same frequency of the modes.
80
dB
60
40
20
- 20
-40
-60
-80
-100 2 0
-1 1 2 3 4
10- 10 10 10 10 10 10
rad/s
Figure 5. Graphs of
(T ) and (T ) of the complete system
In [13], by using eigenanalysis techniques, a dierent set of generators was selected in

order to damp the same EOM. In the modal technique of that paper, the one at a time
selection was applied. Although acceptable results were obtained, the set obtained by
simultaneous selection in this paper yields better controllability and observability of the
EOM, which are required for robust control.
For PSS design, it is considered 1 (s) = 0.13 0.95s+1
0.05s+1
and 0 (s) = 0.25s+0.15
0.5s+1
. The mixed
weight 1 (s)0 (s) is a unique upper bound of uncertainty associated with all control
channels.
The design methodology consists of applying
( )identical PSS, with the conventional struc-
2
Tw s 1+T1 s
ture of reduced order, h(s) = Ks 1+T , on the selected generators and tuning
ws
( 1+T2)s
T
Ks , T1 and T2 in order to minimize , as shown in (9) Assuming that Tw = 20s
S
and using optimization technique, results( that K )s = 0.7, T1 = 0.05s and T2 = 0.009s.
T
In Figure 7, it is shown the graph of for the set of generators #1, 2, 5 and 6
S
with PSS and, also, the same graph with the PSS gain of generator #6 decreased by 50%,
representing an uncertainty in the generator, when compared with the graph of | 1 1 0 h |.
The Equation (9) is satised, thus the robustness of stability with good performance is
achieved. The minimization of the peaks resulted in at graphs, as required for robustness.
In Table 2, the damping ratios of critical EOM with and without the four PSS applied
are compared. The damping ratio provides a mathematical means of expressing the
damping level of a system [14]. A higher ratio leads to a higher damping of the oscillations.
It is observed that all critical EOM were damped with the PSS. It is interesting to see in
Table 2 that for obtaining a robust control, only small increases in damping of the EOM
were required. The small displacements of the eigenvalues were due to the low gains Ks .
With the low gains of the PSS, good dampings of the exciter modes are preserved. This
is an important result, since exciter modes usually limit the robustness of PSS.
In Figure 8, it is shown the generator #6 time responses of the system with and without
PSS, considering the input uncertainty on machine #6. In both cases, it was assumed
50
dB
40
30
20
10
0
-10
-20
-30
-40
-50 -2 -1 0 1 2
10 10 10 10 10
rad/s
Figure 6. Singular values

and of the set of generators 1, 2, 5 and 6

an impulsive disturbance on the mechanical torque of the generator. It is shown that the
system returns to the conditions before the disturbance. A good performance is achieved
under disturbance and uncertainty.
5. Conclusions. A new design technique was proposed aiming to satisfy robust control
requirements with good performance, by using decentralized reduced order stabilizers
rad/sec
50
dB
45
40
35
30
25
20
15
___ 1/ w1 w0 h
10
* PSS on 1,2 ,5 and 6
5
PSS on 1,2,5 and 6 and
0
50% u ncertainty on 6
-5 -3 -2 -1 0 1 2 3 4
10 10 10 10 10 10 10 10
rad/s
( )
T
Figure 7. Graphs of for generators #1, 2, 5 and 6 with and without 50%
S
uncertainty in the input channel of generator #6 and | 1 1 0 h |
20
Amplitude
_____ PSS on 1,2,5 and 6
15
and 50% uncertainty on 6
10 * System without PSS
-5
-10
-150 0.5 1 1.5 2 2.5 3

Time (sec)
Figure 8. Time responses of 6 for the system with and without PSS under an
impulsive disturbance on machine #6
Table 2. Damping ratios of critical EOM
Modes without PSS with 4 PSS

1 0.0395 0.0767
2 0.0405 0.0758
3 0.0394 0.0808
4 0.0296 0.0779
in power systems. A simple mixed sensitivity approach with simple weights is applied
in order to reach this goal. Previous reduction of the system model and reduction of
the controller are not required. PSS with the conventional structure is applied and its
parameters are tuned by an optimization technique. In order to achieve robust control,
the best pole placement for PSS application is previously selected by using a multivariable
frequency domain technique. All pole placements are simultaneously selected, providing a
complete visualization of observability and controllability of all critical oscillation modes.
This paper contributes to show that the conventional lead compensation type of PSS
needs to be tuned in order to provide H robust control in power systems, if the best
placement for application of PSS is previously selected. This paper also contributes
in proposing ecient techniques which can be applied in the design of reduced order
decentralized controllers in general MIMO systems. The limitation of this analysis is
the representation of unstructured uncertainties with a single bound for all generators.
It represents a conservative measure for all procedure. However, the frequency domain
technique provides an unique way of providing a decentralized reduced order with known
structure as robust controller.
Acknowledgment. The authors would like thanks to their Institutions, Universidade

Federal da Paraiba and Instituto Federal da Bahia for supporting. They are also gratefully
acknowledge the helpful comments and suggestions of the reviewers, which have improved
the presentation.
REFERENCES
[1] J. Park, S. Jeong, J. Choi, J. Cha and A.(Rahim)A. El-Keib, A probabilistic reliability evaluation
of Korea power system, ICIC Express Letters, vol.2, no.2, pp.137-142, 2008.
[2] H. Lee, C. Lim, S. Lee, K. Kim, T. Kim, J. Shin and S. Nam, Development of an intelligent voltage
control system for Korean power systems, ICIC Express Letters, vol.3, no.4(B), pp.1315-1320, 2009.
[3] G. E. Boukarim, S. Wang, J. H. Chow, G. Taranto and N. Martins, A comparison of classical, robust
and decentralized control designs for multiple power system stabilizers, IEEE Trans. on Power
Systems, vol.15, pp.1287-1292, 2000.
[4] B. Chaudhuri, B. C. Pal, A. C. Zolotas, I. M. Jaimoukha and T. C. Green, Mixed-sensitivity approach
to H control of power system oscillations employing multiple FACTS devices, IEEE Trans. on
Power Systems, vol.18, pp.1149-1156, 2003.
[5] N. Chen, G. Zhai and W. Gui, Robust decentralized H control of multi-channel discrete-time
descriptor systems with time-delay, International Journal of Innovative Computing, Information
[6] K. A. Folly, N. Yorino and H. Sasaki, Improving the robustness of H PSS using the polynomial
approach, IEEE Trans. on Power Systems, vol.13, pp.1359-1364, 1998.
[7] B. C. Pal, A. H. Coonick, I. M. Jaimoukha and H. El-Zobaidi, A linear matrix inequality approach
to robust damping control design in power systems with superconducting magnetic energy storage
device, IEEE Trans. on Power Systems, vol.15, pp.356-362, 2000.
[8] S. S. Ahmed, L. Chen and A. Petroianu, Design of suboptimal H excitation controllers, IEEE
Trans. on Power Systems, vol.11, pp.199-205, 1996.
[9] P. S. Rao and I. Sen, Robust pole placement stabilizer design using linear matrix inequalities, IEEE
Trans. on Power Systems, vol.15, pp.313-319, 2000.

[10] M. C. Tsai, E. J. M. Geddes and I. Postlethwaite, Pole-zero cancellations and closed-loop properties
of an H mixed sensitivity design problem, Automatica, vol.28, pp.519-530, 1992.
[11] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control Analysis and Design, Wiley, New
York, 2005.
[12] B. S. Gottfried and J. Weisman, Introduction to Optimization Theory, Prentice Hall, Englewood
Clis, 1973.
[13] C. S. Araujo and J. C. Castro, Analysis of input-output couplings for decentralized control of oscil-
lations in power systems, Proc. of the IASTED International Conference, Cancun, 1997.
[14] G. Rogers, Power System Oscillations, Kluwer Academic Publishers, 2000.

!
" # $ % & '
" (') $ *

+ , - .*+ " &+ .+
/0)+
$ *

1
(')
+ , 2/ &+ .+

$ $
" (') $

) * " 34 1
) * 5 +6)07 ++/

! " # $$
% $
& % '( $ %
" '( $
$ )#
#
& %
% $ )#
*
%$ " +
$

&)* $ .*
8)
)

!
" # # $ ! #
%&' " # " & ( " ) ( #
( *+
,

-
,

- ! . . +
( ( #/
-( ( "

-
((( (

( ( (( "

!
. "
,

-
# ( ( #

- ! ( " ( ! " . #
& ! ( # ( ( (
# " . ( ##( 01 .

,

- ( 2
,

- 3
#" # (

4 # ,

* %
,,!# 5,( #,

(,!,#' ,( * " #,
" %

& 6'
#
,( #,
,
" "
" ( " " #,

4 (( . " * ,
*

" #,
" ( 5,(
# # " !( . # .
(( * ) ( #

*
# # -# ,(
. #- (,( .
"
" ( 7 " *
" . ! 8
" ( " ! . -#
* # !
* ! . ! " ( ( !
& # "
,,!# 5,(

,

-+ %'
,,!#+ ((( -

,
! !# ( %
' 5,
(+ !( . ( ( %' (
" (

(
4 -# #
,,!# 5,(
,

- 2 -# . ! 9 " #

( %(' ! , ( " : " (

! (
2 ## ! ! ! . " !
"
,
* ! (

! "#$
,
#,
" " *

# # ,( * ! "
( # ,( + #" ( %'
# . ;<=> #" ! #/ .
( ;>

,

- * " ( # (
4 ( - " ( # #/ : "
%(' # (#
" 2 # (

# ( ( ! ( ( 4
! # ( ( ? * #
. % &" .'
( # (
):
@ % A' -# * "
##( ( #,
;=> -# ! ,
( ( " : " (
##( 8 (/ #( : #(
" ( ( 8 ( 8 " (
# #- ## # " ##(
! ( #,
" ( * ( .
" # ( "
0 1 ## " 8 "
,

- ( ( # " ( 2

( "

9 " ( # $. % B' # ( %
#' " .( ;<>
%' # ,
#,

* #
" . . ;> !
#, * # #
*
* " (( ,

* !
# * ! ! " ( #

,,!# -#
,
#"

0!#(1
,,!# ( (,

# ! .
,
!
(#

C % 6' )
% ' * ;6A>
C % 6' : !

,
# " ;6> ) % '
# # *

,
# " #(
# ;A>
# # %( ##( .('
#" ,
( . ) 3 % ' 0& 1 ;> C3
% A' 0 (
1 ;> # . ( ( !
- ,( * 5,&( . 0(1
# (
6 % !& '($ $& 2 (( ,
( * ! -# ( !
. "
,

-
,

- # #" ,(
DE2 ! ( ( ( (
!
,

- * " (
* - # & " #
( "
,

- ! #
,

. ,( "
,
( % (
'

+ 2 :
,
(
,

# (# ( " !

( 4 ! #" ,( (
,,
!# ? / #,# -
,

( !#

,

,,!# 0!#1

" 2
,,!# :
,
#"

0!#(1 / ( #

" (( ( ( !# (

,
% ( %''

+ 5,( -# . (
(( (
% ( %
''
(( ( . ! ( (
((

+
,,!# 5,( ! # #,

( 2 #

! *
# ( ! ## #,

,

( # !
! ! * (
# ! # *
#(,!,# " ( " ,
!
! * 2" ! * .
! # ( "
"
( ! ! *
(
# ! # # # ! #
-# !
9 ! #" . ! ! " . %
,
( '
( (# % ( '
- " F# %.9' (

. # ,# # ! G #(
# H,/ # " " # " ! < (
! # " < !
" #(,!,#

"

. " #
I "
. ! # #(,!,
# % A' 3 ; > ! ( "
. "
#,
E2 ( 2#
# (
J.,
;> 2 #(,!,# # ! # !(,# (#

!( % ( ' # ( #
#
# ! 4 #
#

* ( "
,,!# 5,(

9
" #"
8 ! #
# # " #
: "
A )*+ ,
,,!# :
,
(

!#( # # "

,,!# # ,

" #- " ( # " ( E( 5,( ( (
( ((
"
" ( (
( 9 " ! . ( 5,( #
,
,!# 5,( ( ( # # ! 4 -#
#" " " ,( * 8 ,
#- " " (
" ( !

" ( * : ,* ,
! , %= =' (
A **!!$ ! 7-# ! 4 2,B
5 (# #
! ! #
# # # # B " #

-# "! 6D$: A ! 4 ! K "
#( -# "! ! # ( & )
F
A -!*!$ ! % ! " ' ## 7-#,
( ( ! A6 %(( " 6' ! (,
A6 !$# ! $+ D ( ( (

##( ! 9 ,( #"
,(

,

- * & 9 ,(
. #
7-# ,- * ! !
= = ( ( * ! ( % ,(' * !
! !
"
4 #
" ( . # " .+ " ,
( . ,( . -,( . 4 " (
! # #- ( " .+ ! #- . ,
#- . ( #- . ! !

9 %

& AA' ,- 8 ( ! 7-# #,
! ! 8 # ( ! (

" * % ( A'
#
( -# # . -#,
! !
9 #
"
9 ! " # %( ,
" 8 '
# -# .

9 ! ( *.
#
! #

.

#
!
" # " #

4 / ## . .
/ .
! % #
! (
.
'
4 ! #

# ! #
!

-# ( . "*
. ,
# # # 6 AB ( * !
# (
9 #
AA $& ( " -# ) %
,,!# 5,
(
,

- 5,(' 5 " ( %&H &4 JH' #- "
( %5 ! ) $(' " !
. ! : 2
-# "+ ## A 6 "
( 6= . L B<A (
#- M ( ( " " ( !

8 ! ;B=> 2 -# #- " (
!

9 " -# ## (
-# #- " # !
# # "
## ## ! . # " !
( !
# ( " !( ( % L
L 6 L A L B L ' .(
9( ## ! ! 6= # ( ( #
"
! ( . 9( #- "
# ## ! ! # (
#
" - ! ! " # " # !
: ( ## ! (
" #" #- ! # ! ( "
#- ! ( : 6 % !+ #- !
! B
+ #- ! " B 6 (+ #- !
6

AB
$ $ HI % " ' ! # !
: #" "
9 #"
! #" ! .3 (/ 8 %$&'
AB & ! # # ,
" ! (/ 8
!
" %6 AA' L B #

! 6 "
,,!# " 5,( 66 "
#(,!,# B< "
,

- # . $&
!
,,!# ,( #(,!,# ! "

,

- %# B' ! (/ 8

! 8
! : ,
% 66' L B #
,,!# 5,( #(,!,

#
,

- 8 (/
!
( :, %
,,!#+ % 66' L A B # 5,(+

% 66' L G # #(,!,#+ % 66' L G #
,

-+ % 66' L < 6 # ' ( B !

( :, J " :, (/
8 ! (( #" " "
* :,
)
( :,

( =
#- "
( 8
" ( " JH

+ ! #- ! (/ 8

! " #- ! (/
8
! " %6 AA' L G < #

# . $& !
,,!# 5,( #(,!,

# ! "
,

- %# ' ! (/

8 ( #- !
(/ 8
! " %6 AA' L = #

# . $& !
,,!# 5,( #(,

!,# ! "
,

- %# ' !

(/ 8

+ ! #- ! (/ 8

! " ( #- !
(/ 8
! " %6 AA' L A #
# . $& !
,,!# 5,(

#(,!,# ! "
,

- %# ' "

! (/ 8

+ " = ( " ! " (
! ( = %= ('
! ( = " . ! #-
,,!# 5,
( #(,!,#
,
,
- !
# # *
( "
( !
,

- # #"
. ! 4
" ( ( "
# # *

)
#-
%5 ! ) $(' " ( 8
" (
" JH
( G ! " &H

+ ! #- ! (/ 8

! " ( #- !
(/ 8
! " %6 AA' L < 6 #
# . $& !
,,!# 5,(

#(,!,# ! "
,

- %# ' !

(/ 8

+ ! #- ! (/ 8

! " %6 AA' L G # # . $&

!
,

- ! "
,,!# 5,(

#(,!,# %# B' "
,

-

!
# !

- #( !
" " ! #- #
(
" (
,

-
( ! (/ 8 ,
#- ! (/ 8

! " ( #- ! (/
8
! " %6 AA' L B #

# . $& !
,,!# 5,( #(,!,

# ! "
,

- %# ' ! (/

8

+ " = ( " ! " (
! ( < %= ('
) %! &

'

#- %5 ! ) $(' " ( 8 ,

" ( " &H
" &4 " ! " &H
2 &H ( : " JH
( G # # * ! ( " ( ! (
#- &H $ !
,
,
-
( #"
# # * " ( ! #- ( : "
&H " JH ( :
8 #" " * 2 # ! (
: * ! ! 8 #" N 4
-# 7-#
AB " ! ! (,
/ 8
!
,,!# 5,( #(,!,#

,

- ! 6O "
,,!# =O " 5,

( AGO " #(,!,# BBO "
,

-
"

AB6 &

9 ! * *
" , ,G %G "
" ! ' % ( <' ! (/ 8
!
,,!# 5,( #(,!,#
,

-

9 #" %6 AA' L G # # . %$&'

! #(,!,# !

#
,,!# 5,( !

,

- #

! (/ 8 &

9 #(,!,# ! M-

# # %
!
,,!# 5,(' #(

" . #

9 ( " " *
B )*+ . " 7-# !
,,!# 5,(

#(,!,# !

,

- " ( #- .

,,!# 5,( ! !
8 8 ( ( # ( ( "
#(,!,#
( "
,,!# 5,(

( $ ! #(,!,# ! (
,
,!# 5,( . " 7-#
. #- " . ! * ( " 7-#
. !
"
( ( #- !
#
B **!!$ ! 7-# ! 7-#,

B -!*!$ !
9 ! " -# ( ( !

9 ! (
B6 !$#$ ! * ! ( # " . ! #,

7-# ! ! / ( #+
( : :

: ( * ! % ,(' *
!
(
!
. " ! "
( ( ! < #- !
" ( ! B #- 7 . ! #

9
-# # 7-#
BA
$ $

H # " -# .
BA J( :+ # " ! (",
8
! " %6 AA' L G #
# . $& !
,,!# ,(

#(,!,# ! "
,

- %# B' 5,(

#(,!,# ! "
,,!# ! (/

8
! 5,( #(,! # % ( '
&
* "
( : . $
0 #(1 #(,!,# 0J -1
,

-
0&(1 5,( 0&!#1
,,!#
& :+ ! (/ 8
! "
%6 AA' L < = # # . $& !
,

,!#
,

- #(,!,# ! " 5,(
%# ' ! (/ 8
!
,,!# #(,

! # % ( '

:+ ! (/ 8
! "
%6 AA' L 6G A # # . $& !

,,!#
,

- #(,!,# ! "
5,( %# ' ! (/ 8 (
,,!#

#(,!,#
,

- % ( '
( 5,( (/ " ,( !
( :
! . ! " * ( 5,( *
0(1 ( 4 (
* ! .

" 0(1 # 5.
,
,
-
,,!#
# (
0!#(1 -#

&
* "

.
&
* "
.
#,!,# (
/ "

,,!# 5,

( ! . " 7-#
BA # " ! ! (/,
8 ( " * "
( : :

: .
BA6 ! (/ 8 (
,,!#
5,( #(,!,#
,

-
9 #" %6
AA' L A # # . %$&' ! #(,!,

#
,,!# !

,

- 5,( #
! (/ 8

! #(,!,#
,
,!# ! (/ 8
! 5,(
,

-

9 ( " " *

= $$$ ! $$ ( ! . !.,

,

- ( ( ( 4
# ,( *+
,,!# 5,( #,

(,!,# 4 ( -# # !

,

- * !
" ( #-
" # :, " ( ( : -,
# ! #
,
,!# 5,( #(,!,# !

,

-
(/ ( " ( #- "
(
9 #" ! "
9 #" #(,
!,#
" , , !( "
!
,,!#
5,(

,,!# 5,(+
,,!# 5,(

,

- ( #- # " ( $ ! ,
" * ! ## # (
.
,,!#
(
" 9 ( 5,
( ! ( " #- # &
9
/ . ( ( . ! * 5,
( ! ( (
" ( (

(
. ( "
,,!#
5,( ( ( #- # -# "

#(,!,#+ * #
,,!# 5,( !

# # ! 8
,

#" . ( " (
" "
9 ( ! ( ! # !
#
,,!# 5,( 2 ( * !

( 9 #
# ,"
. (/ # ,( * #
( #- # # # # ,
,"
. " #,
E2 (
#,!,# ! #
* (
( ,
* " #(
,,!# 5,( (
,,!#
, 5,(,
,,!# ! (

#,

-# ! =

9 ! -#
"
,,!#, 5,( ( #" " 7-,

# ( " #(,!,# ! #

,

-+ 7-# ! !
"
( ( " ( 2 ! (
,5,
&!# 5,( #(,!,# #,
"
! #" (
,
* * "

/ *
,

- - (
"
- # #"
(

" ( ( (
#- (
,

- * #
(
4 # : #,!,# ! #
/ "

,,
!# 5,( ( ! ! 7-#
( 7-# " . ( 7-#

+ %' 5,( " ( ! * : %'
,,!#
! 5,( ( #- " ( !
( : "
7-# . (
! . (
# " -# (( #(,!,#

/ " #,
" *
# #,

" 8
/ ( . #
8 #
! ,( #
# -# ## E2 ( #(,!,#
8 # ( " #,
##
# # *
## #,

(

" ,
" %' ! . " %
" 5,(' %' ## # # %
#(,!,#' -# ! # # *
#"
,

- ( ( ! * (
" ! ! " (
*

" * ( 5,(
,

-
4
" ! .
##
##
! ( " * " ,( *
# #,
"
#/ &+$
# ##
D,, " &,
/ %H B <' ) " )
#( 7,
7-# & "! 9 " 2 %" ( #
( F#' (" " ! . ## "

" 5
P E " (

9: 1+ ! * + ; * <) = > * ) $ )))* $) ,$
- .//0 +3 +
9: + !
* ?+ * !

*5 ! * * )
,$ - 1220
+,43 +
9: &+ !
)) * + ; @ ) ) ,$ - .//3 + 4
+
9: &+ ?*) A+
B+ . ) + CD)7 &+ *

?+ ?*) * !+ ;

B* * * *75 E) $ )) ) *
* ) )) ,$ .//3 +43, +
94: B+ .+ ? D * B+ @+ FD ) $ ')
$ 5 *

* )) ,%
4 '
+, +4 ,3+
9,: + >+
6 * ) ')
)
)# ,% %5
'
+ , + +, 33+
93: %+ 8* .+ ?
* + ?* #$ 5 F/ ')+ 6
)

) $) ,$ .//0 +, +
9: "+ @7
&+ ?*) 8+ . ) * >+ 8 B) *
)) $ *
) $
)
) ))
,$ - .//6 +4, 4+
9: "+ " + AC7 "+ * "+ B'
$ )* 7 ) ) $ ) )
D ')
*)
)

'
+4 + +,4,4 +
9 : + + ") * + F ) 7 )5 ! D )* ) $
) * )G * )
* 6
) 7) ,7

'
+3 +, ,+
9: 8+ " * -+ ?6 ))) 7 * *
E)
,$ ! 71221 +3 +
9: 8+ " + * -+ ?6 6
$ )
*) ,$
1220 + +

9: -+ # #+ * ?+ - '* &B $ )

) 6 C
)# '
+ +H?I +3 +
9: + ) "+ ")D * 8+ .+ ? * A
)
$) $ ) *

,$ - .//. + +
94: .+
* ?+ )

)
) ) 6 * ,$ ! .//. +
+
9,: + C * + %) ')+
)
) )* *')5 A'* $
*J $ ) $ ) ,$ - .//0 +, 3, +
93: !+ A6* J' D* >)<
D D *)
)7

'
+ + +34 +
9: + & ' + ?
) + - ) * + 7D 8 8 E5

$ *
. ,$ ! 71228 +3 ,+
9: .+ .) A'
*J *)* E)
$ 7 )7
,$ - 7 ) 7% 9 +,3 +
9 : 8+ . ) + ?
) * .+ ?
7) &)) D*) ,$ - .//0 +3 +
9: 8+ . ) * .+ ?
7) ;
*5 >
* C * )
* $ )

,$ ! 7.//6 +4 4+
9: K+ . * + ' )
$ )* ) )5 ! )* $ )*
$ )
)7) 9: $ - '
+3 + +,
+
9: !+ ) * ?+ * @ ) ))5 B) )) * ) D
)
:: 7 '
+ + +4, +
9: ?+ * )) D J
)) ))) 7 '

+ + +
+
94: 8+ -
)
5 4 D ) ) L) ) )M ,$ ))) ;<28 +4,
,+
9,: + ; * .+ ?
7)
')+ * 7
7 ) ,$
! 7.//0 + +
93: K+ ; * K+ . ! J' )
$
$ )
)

'
+ + +
+
c ISSN 1349-4198
A NOVEL MODAL SERIES REPRESENTATION APPROACH TO

SOLVE A CLASS OF NONLINEAR OPTIMAL CONTROL PROBLEMS
Amin Jajarmi1 , Naser Pariz1 , Ali Vahidian Kamyad2 and Sohrab Effati2
1
Advanced Control and Nonlinear Laboratory, Department of Electrical Engineering
2
Department of Applied Mathematics, Faculty of Mathematical Sciences
Ferdowsi University of Mashhad
Azadi Square, Mashhad, Iran
jajarmi@stu-mail.um.ac.ir; n-pariz@um.ac.ir; { avkamyad; eati911 }@yahoo.com
Abstract. This paper presents a new approach to solve a class of nonlinear optimal
control problems which have a quadratic performance index. In this approach, the nonlin-
ear two-point boundary value problem (TPBVP), derived from the Pontryagins maximum
principle, is transformed into a sequence of linear time-invariant TPBVPs. Solving the
proposed linear TPBVP sequence in a recursive manner leads to the optimal control law
and the optimal trajectory in the form of uniformly convergent series. Hence, to obtain
the optimal solution, only the techniques of solving linear ordinary dierential equations
are employed. In order to use the proposed method in practice, a control design algorithm
with low computational complexity and fast convergence rate is presented. Through the
finite iterations of algorithm, a suboptimal control law is obtained for the nonlinear opti-
mal control problem. Finally, numerical examples are included to demonstrate eciency,
simplicity and high accuracy of the proposed method.
Keywords: Nonlinear optimal control problem, Pontryagins maximum principle, Modal
series, Suboptimal control
1. Introduction. One of the most active subjects in control theory is the optimal control
which has a wide range of practical applications not only in all areas of physics but also
in economy, aerospace, chemical engineering, robotic, etc. [1-4]. For linear time-invariant
systems, optimal control theory and its application have been developed perfectly [5,6].
However, optimal control of nonlinear systems is much more challenging which has been
studied extensively for decades [7,8].
One familiar scheme for the optimal control of nonlinear systems is the state-dependent
Riccati equation (SDRE) method [9]. In the SDRE, a direct parameterization is used to
transform the nonlinear system into a linear structure with state-dependent coecients.
Then, a matrix Riccati algebraic equation with state-dependent coecients is solved at
each sample state along the trajectory to obtain a nonlinear state feedback control law.
Although this method has been widely used in various applications, its major limitation
is that it needs solving a sequence of matrix algebraic equations which may take long
computing time and large memory space. Moreover, this method produces a suboptimal
control law even when the analytic solution of SDRE is available.
Another method, called the approximating sequence of Riccati equations (ASRE), has
been introduced in [10]. Although the ASRE is attractive from practical aspects, it suers
from computational complexity due to the need for solving recursively a sequence of linear
quadratic time-varying matrix Riccati dierential equations.
In order to determine the optimal control law, there is another approach using dynamic
programming [11]. This approach leads to the Hamilton-Jacobi-Bellman (HJB) equation
1413
1414 A. JAJARMI, N. PARIZ, A. V. KAMYAD AND S. EFFATI
that is hard to solve in most cases. An excellent literature review on the methods for
solving the HJB equation is provided in [12] where a successive Galerkin approximation
(SGA) method is also considered. In the SGA, a sequence of generalized HJB equations
is solved iteratively to obtain a sequence of approximations approaching the solution of
HJB equation. However, the proposed sequence may converge very slowly or even diverge.
The optimal control law can be derived using the Pontryagins maximum principle
[13]. Unfortunately, for the nonlinear optimal control problem (OCP), this approach
leads to a nonlinear two-point boundary value problem (TPBVP) that in general can
not be solved analytically. Therefore, many researches have been devoted to find an
approximate solution for nonlinear TPBVPs [14]. Over recent years, some better results
have been obtained. For instance, a new successive approximation approach (SAA) has
been proposed in [15] where instead of directly solving the nonlinear TPBVP, a sequence of
nonhomogeneous linear time-varying TPBVPs is solved iteratively to obtain a sequence of
approximations approaching the optimal control law. The sensitivity approach, proposed
in [16], is a similar one to the SAA. It only requires solving iteratively a sequence of
nonhomogeneous linear time-varying TPBVPs to determine the infinite series presenting
the optimal control law. However, solving time-varying equations is much more dicult
than solving time-invariant ones.
The Modal series method, which is recently developed in the nonlinear system analysis
[17-22], was initially introduced in [17]. It provides the solution of autonomous nonlinear
systems in terms of fundamental and interacting modes. This solution, which is called
modal series, yields a good deal of physical insight into the system behavior. In contrast
with the perturbation method [23], the modal series method does not depend on the
small/large physical parameters in system model. In addition, unlike the traditional non-
perturbation methods, such as Lyapunovs artificial small parameter method [24] and
Adomians decomposition method [25], the solution series obtained via the modal series
method converges uniformly to the exact solution.
In this paper, we propose a new approach to solve nonlinear OCPs by extending
the Modal series method. In this approach, the optimal control law and the optimal
trajectory are determined in the form of uniformly convergent series with easy-computable
terms. Moreover, to obtain the optimal solution, only a sequence of linear time-invariant
TPBVPs is solved in a recursive manner.
The paper is organized as follows. Section 2 describes the problem statement. The
optimal control design strategy and convergence analysis are elaborated in Section 3.
Section 4 explains how to use the results of Section 3 in practice. In this section, in
order to obtain a suboptimal control law, an ecient algorithm with low computational
complexity and fast convergence rate is presented. Section 5 introduces two numerical
examples verifying validity of the proposed method. Finally, conclusions and future works
are given in Section 6.
2. Statement of the Problem. Many nonlinear systems of practical importance are

ane in the control; for example, we can mention many fed-batch processes [26], aircraft
systems [2], chemical reactors [27], etc. Besides, optimal control of such systems has
received growing attention due to its tremendous applications including autopilot design
[28], chemical reactor control [3], ducted fan control [29], control of aeroelastic systems
[30], etc. Therefore, consider a nonlinear OCP described by:
t ( )
min J = 12 t0f xT (t)Qx(t) + uT (t)Ru(t) dt
{
x(t)
= F (x(t)) + G(x(t))u(t), t [t0 , tf ] (1)
s.t.
x(t0 ) = x0 , x(tf ) = xf
A NOVEL MODAL SERIES REPRESENTATION 1415
where x Rn and u Rm are respectively the state vector and the control vector,
F : Rn Rn is a nonlinear analytic vector function where F (0) = 0, G : Rn Rnm is
a nonlinear analytic mapping, x0 Rn is the initial state vector and xf Rn is the final
state vector, Q Rnn and R Rmm are positive semi-definite and positive definite
matrices, respectively.
Remark 2.1. Although the above imposed conditions on the OCP (1) may seem to be
conservative from the theoretic aspect, they are satisfied by many practical applications
such as control of a continuous stirred-tank chemical reactor [5], F8 fight aircraft control
[2], optimal maneuvers of a rigid asymmetric spacecraft [31], etc.
According to the Pontryagins maximum principle, the optimality conditions are ob-
tained as the following nonlinear TPBVP:
x(t)
= F (x(t)) G(x(t))R1 GT (x(t))(t), t [t0 , tf ]
T G(x(t)) 1 T
( )T
(t) x1 (t)
R G (x(t))(t)
F (x(t)) ..
(t) = Qx(t) (t) + . , t [t0 , tf ]
x(t)
T (t) xn (t) R1 GT (x(t))(t)

G(x(t))
x(t ) = x , x(t ) = x
0 0 f f
(2)
where R is the co-state vector and xi (t) is the i
n th
element of vector x(t). Also, the
optimal control law is given by:
u (t) = R1 GT (x(t))(t), t [t0 , tf ]. (3)

Let define (x(t), (t)) and (x(t), (t)) as:

1 T
(x(t), (t)) = F (x(t)) G(x(t))R G (x(t))(t)

T G(x(t)) 1 T
( )T (t) x1 (t) R G (x(t))(t)
F (x(t)) ..

(x(t), (t)) = Qx(t) (t) + . .

x(t)
T
(t) G(x(t)) 1 T
R G (x(t))(t)
xn (t)
(4)
Note that : R R R and : R R R are nonlinear analytic vector
n n n n n n
functions since F and G are assumed to be analytic. In the light of (4), problem (2) can
be rewritten in a compact form as follows:

x(t)
= (x(t), (t)), t [t0 , tf ]

(t)
= (x(t), (t)), t [t0 , tf ] (5)

x(t0 ) = x0 , x(tf ) = xf .
3. Optimal Control Design Strategy and Convergence Analysis. Unfortunately,
problem (2) (and equivalently (5)) is a nonlinear TPBVP that in general can not be solved
analytically except for a few simple cases. Here, we extend the Modal series method to
solve this problem. To do so, first we need Taylor series expansion of the nonlinear
non-polynomial terms. Therefore, if and in (5) are not polynomial, they should be
expanded in the Taylor series around the operating point (xo , o ) = (0, 0) which yields:

x(t)
= A10 x(t) + A01 (t)
T T T

1 1 1

x (t)H20 x(t) x (t)H11 (t) (t)H02 (t)

1 + 1

+ ..
. + ..
.
..
. +

2! 2!

xT (t)H20n
x(t) xT (t)H11n
(t) T (t)H02n
(t)

(t) = A10 x(t) + A01 (t)

T 1 x(t) xT (t)H 1 (t) T 1 (t)

x (t)H (t)H

1 20 11
1 02

+ .
.
. + .
.
. + .
.
. +

2! 2!
20
xT (t)H n
x(t) 11
xT (t)H n
(t) 02
T (t)H n
(t)
(6)
where x and belong to the convergence domain of Taylor series in (6) which is denoted
by where Rn and Rn are subsets of the state space and co-state space,
respectively. Also, the coecients in (6) are as follows:

x=0 , 10 =
x=0
A10 = x =0
A x x=0
, A 01 = x=0
=0
, A01 =
,

2 2
=0 =0
j 20
j 2 j
H20 = x2j x=0 , H = x2j x=0 , H11 = xj x=0 , (7)

=0 =0 =0
2
2 2

H 11
j j
= xj x=0 , H02 = 2j x=0 , H 02
j
= 2j x=0
=0 =0 =0
j are the j th components of vector functions and ,

where j and respectively.
Remark 3.1. For many large-scale systems, the analytic partial derivatives as in (7)
are dicult to be obtained symbolically. For these systems, it may be needed to compute
derivatives numerically as in [17]. In addition, in the presence of such systems, complexity
of computations brought by the curse of dimensionality adds another diculty. Future
works are focused on improving the method in order to overcome these diculties.
Theorem 3.1. The solution of nonlinear TPBVP (2) (and equivalently (5)) can be ex-

pressed as x(t) = gi (t) and (t) = hi (t) where the ith order terms gi (t) and hi (t) for
i=1 i=1
i 1 are achieved by solving recursively a sequence of linear time-invariant TPBVPs.
Proof: The solution of problem (2) (and equivalently (5)) for arbitrary boundary
condition xb = (x0 , xf ) and for all t [t0 , tf ] can be expressed as:
{
x(t) = (xb , t)
b , t) (8)
(t) = (x
where : R2n R Rn and : R2n R Rn are analytic vector functions since
and (defined in (4)) are analytic [32]. Since and are analytic vector functions and
xb is assumed to be arbitrary, we can expand (8) as Maclaurin series with respect to xb
which yields:
( )
2 1 (xb ,t)
x T
xb
b x2b

(xb , t)
1
xb =0

x(t) = (xb , t) = x + .
. +
xb xb =0 2! ( . )
b
2 (x ,t)

| {z }
g1 (t)
xT
b
n b
xb2 xb
xb =0
| {z }
g2 (t)
( )
1 (xb ,t)
2
xTb xb
x2b
b , t)
(x 1
xb =0

(t) = (xb , t) = xb + .. +
xb xb =0 2!
( .
)
(9)
| {z } n (xb ,t)
2
h1 (t)
xTb x2b xb
xb =0
| {z }
h2 (t)
where j and j are respectively the j th components of vector functions and , gi (t)
and hi (t) are vector functions which their components are linear combination of all terms
depending on the multiplication of i elements of vector xb . For example, components
of g2 (t) and h2 (t) contain linear combination of all terms of the form xb,k xb,l for k, l
{1, 2, . . . , 2n} where xb,j is the j th element of xb . Moreover, since and
are analytic
vector functions, existence and uniformly convergence of the Maclaurin series in (9) are
guaranteed. Let 1 2 be the convergence domain of Maclaurin series in (9) for all
t [t0 , tf ] where 1 Rn and 2 Rn are subsets of the initial state space and final
state space, respectively. Assume that 1 = 1 and 2 = 2 are non-empty and
let the boundary condition be xb , i.e., x(t0 ) = x0 and x(tf ) = xf where is an arbitrary
scalar parameter such that xb 1 2 . Since 1 and 2 are assumed to be non-empty,
such parameter exists. Moreover, this parameter only simplifies the calculations and its
value does not have any significance. Similarly to (9), we can write:
{
x (t) = (xb , t) = g1 (t) + 2 g2 (t) +
b , t) = h1 (t) + 2 h2 (t) + (10)
(t) = (x
Since xb 1 2 , (10) must satisfy (6). Satisfying (10) in (6) and rearranging terms
with respect to the order of yield:
g 1 (t) + 2 g 2 (t) + = (A10 g1 (t) + A01 h1 (t))
A10 g2 (t) + A01 h2 (t)

g1T (t)H20 1
g1T (t)H11 1
hT1 (t)H02 1

2
g1 (t) h1 (t) h1 (t)
+
+ 1 . . 1 .
+ .. + .. + ..
2! 2!
g1T (t)H20 n
g1 (t) g1T (t)H11 n
h1 (t) hT1 (t)H02 n
h1 (t)
( )
1 (t) + 2 h 2 (t) + = A10 g1 (t) + A01 h1 (t)

h
A10 g2 (t) + A01 h2 (t)

T
g1 (t)H 1 g1 (t) g T (t)H 1 h1 (t) T

h1 (t)H 1 h1 (t)
20 1 11 02
+
2
+ 1 .
. .
. 1 .
.
+ . + . + .
2! n g1 (t) n h1 (t) 2! n h1 (t)
g1T (t)H 20 g1T (t)H 11 hT1 (t)H 02
(11)
Since (11) must hold for any as long as xb 1 2 , terms with the same order of
on each side must be equal. This procedure yields:
{
g 1 (t) = A10 g1 (t) + A01 h1 (t)
: (12a)
h 1 (t) = A10 g1 (t) + A01 h1 (t)

g 2 (t) = A10 g2 (t) + A01 h2 (t)
T T T

1 1 1

g1 (t)H20 g1 (t) g1 (t)H11 h1 (t) h1 (t)H02 h1 (t)

1 + 1

+ ..
. + ..
.
..
.

2! 2!

g1T (t)H20 n
g1 (t) g1T (t)H11 n
h1 (t) hT1 (t)H02 n
h1 (t)
2 :

h 2 (t) = A10 g2 (t) + A01 h2 (t)

T 1 g1 (t) g T (t)H 1 h1 (t) T 1 h1 (t)

g1 (t)H h1 (t)H

1 20 1 11
1 02

+ .
.. + .
.. + .
..

2! 2!
n g1 (t)
g1T (t)H n h1 (t)
g1T (t)H n h1 (t)
hT1 (t)H
20 11 02
(12b)
and so on.
Note that (12a) is a system of homogeneous linear time-invariant ODEs. From (12a),
g1 (t) and h1 (t) can be easily obtained. Assume that g1 (t) and h1 (t) have been obtained
from (12a) in the first step. Then, g2 (t) and h2 (t) can be easily obtained from (12b) in
the second step since (12b) is a system of nonhomogeneous linear time-invariant ODEs.
Notice that, the nonhomogeneous terms of (12b) are calculated using the solution of (12a).
Continuing as above, gi (t) and hi (t) for i 2 can be easily obtained in the ith step only
by solving a system of nonhomogeneous linear time-invariant ODEs. In addition, at each
step nonhomogeneous terms are calculated using the information obtained from previous
steps. Therefore, solving the presented sequence is a recursive process.
To obtain the boundary condition for the above-mentioned sequence, set t = t0 and
t = tf in the first equation of (10) as follows:
{
x0 = x (t0 ) = (xb , t0 ) = g1 (t0 ) + 2 g2 (t0 ) +
(13)
xf = x (tf ) = (xb , tf ) = g1 (tf ) + 2 g2 (tf ) +
Comparing the coecients of the same order terms with respect to yields:
{ {
g1 (t0 ) = x0 gi (t0 ) = 0
and for i 2. (14)
g1 (tf ) = xf gi (tf ) = 0
Thus, the proof is complete.
Corollary 3.1. For the nonlinear OCP (1), the optimal trajectory and the optimal control
law are determined as follows:

x (t) =

gi (t)
i=1 ( )( ) (15)

u
(t) = R 1
G T
gi (t) h i (t)
i=1 i=1
where gi (t) and hi (t) for i 1 are obtained only by solving recursively the presented linear
TPBVP sequence in the proof of Theorem 3.1.
As it can be interpreted from Corollary 3.1, the proposed method only requires solving
a sequence of linear time-invariant TPBVPs; on the other hand, the other approximate
methods such as SDRE, ASRE, SGA, SAA, etc., need solving a sequence of matrix dier-
ential (or algebraic) equations or a sequence of generalized HJB equations or a sequence of
linear time-varying TPBVPs [9,10,12,15,16]. Therefore, the proposed method takes less
computing time and memory space in comparison with the above-mentioned approximate
methods.
The following theorem shows uniformly convergence of the obtained solution series in
(15) to the optimal solution:
{ } { } { }
Theorem 3.2. Define sequences x(k) (t) k=1 , (k) (t) k=1 and u(k) (t) k=1 as follows:

k
x(k) (t) = gi (t)

i=1
(k) k
(16)

(t) = hi (t)

i=1

u(k) (t) =
R1 GT (x(k) (t))(k) (t).
{ } { }
Then, for the nonlinear OCP (1), the sequences x(k) (t) k=1 and u(k) (t) k=1 converge
uniformly to the optimal trajectory and the optimal control law, respectively.

Proof: According to the proof of Theorem 3.1, Maclaurin series gi (t) and hi (t)
i=1 i=1
converge uniformly to the exact solution of nonlinear TPBVP (2), i.e., x(t) and (t),
respectively. That is:

k

x(t) = lim
U
gi (t) x(k) (t) x(t)
k i=1
(17)

k
U
(t) = lim hi (t) (t) (t).
(k)
k i=1
{ }
Besides, the control sequence u(k) (t) k=1 depends on the state vector sequence {x(k) (t)
{ (k) }
}k=1 through an analytic mapping and on the co-state vector sequence (t) k=1
through a linear operator. So, we have:
( )( )
1 T 1 T

u (t) = R G (x(t))(t) = R G gi (t) hi (t)
( ) ( i=1 ) i=1
k k
= R1 GT lim gi (t) lim hi (t)
{ (k
k i=1
) k
( k i=1 )} (18)
1 T

= lim R G gi (t) hi (t)
k
{ i=1 }
i=1
= lim R1 GT (x(k) (t))(k) (t) = lim u(k) (t),
k k
{ (k) }
i.e., the control sequence u (t) k=1 converges uniformly to the optimal control law and
the proof is complete.
Remark 3.2. Let the solution of nonlinear OCP (1) with the boundary condition xb
1 2 be available as in (15). To find the solution for any other boundary condition of
the form xb , as long as xb 1 2 , there is no need to repeat the recursive process.
The optimal solution can be easily obtained as:

x (t) = i gi (t)
i=1 ( )( ) (19)
i i

u (t) = R G 1 T
gi (t) hi (t) .
i=1 i=1
4. Suboptimal Control Design Strategy. In this section, we explain how to use the
results of previous section in practice. In fact, obtaining the optimal control law as in (15)
is almost impossible since (15) contains infinite series. Therefore, in practical applications,
by replacing with a finite positive integer N in (15), an N th order suboptimal control
law is obtained as follows:
( N )( N )

uN (t) = R1 GT gi (t) hi (t) . (20)
i=1 i=1
The integer N in (20) is generally determined according to a concrete control precision.

For example, every time gi (t) and hi (t) are obtained from the presented linear TPBVP
sequence, we let N = i and calculate the N th order suboptimal control law from (20).
Then, the following quadratic performance index (QPI) can be calculated:

(N ) 1 tf ( T )
J = x (t)Qx(t) + uTN (t)RuN (t) dt (21)
2 t0
where uN (t) has been obtained from (20) and x(t) is the corresponding state trajectory
obtained from applying uN (t) to the original nonlinear system in (1) with x(t0 ) = x0 .
The N th order suboptimal control law has desirable accuracy if for two given positive
constants e1 > 0 and e2 > 0, the following conditions hold jointly:
(N )
J J (N 1)
< e1 (22)
J (N )
x(tf ) xf < e2 (23)
where is a suitable norm on R and x(tf ) is the value of corresponding state trajectory
n
at the final time tf . If the tolerance error bounds e1 > 0 and e2 > 0 be chosen small
enough, the N th order suboptimal control law will be very close to the optimal control
law u (t), the value of QPI in (21) will be very close to its optimal value J , and the
boundary condition will be satisfied tightly.
In order to obtain an accurate enough suboptimal control law, we present an iterative
algorithm with low computational complexity as follows:
Algorithm:
Step 1. Let i = 1.
Step 2. Calculate the ith order terms gi (t) and hi (t) from the presented linear TPBVP
sequence in the proof of Theorem 3.1.
Step 3. Let N = i and obtain the N th order suboptimal control law uN (t) from (20),
apply it to the original nonlinear system with x(t0 ) = x0 to obtain the corresponding
state trajectory x(t), and then calculate J (N ) according to (21).
Step 4. If (22) and (23) hold together for the given small enough constants e1 > 0 and
e2 > 0, go to Step 5; else replace i by i + 1 and go to Step 2.
Step 5. Stop the algorithm; uN (t) is the desirable suboptimal control law.
As it can be understood from Theorem 3.2, the above algorithm has a relatively fast
convergence rate. Therefore, only a few iterations of the algorithm are required to get a
desirable accuracy. This fact reduces the size of computations, eectively.
5. Numerical Examples. In this section, two illustrative examples are employed to

show eciency, simplicity and high accuracy of the proposed method.
Example 5.1. Consider the following nonlinear OCP [33]:
1 2
Min J { = 0
u (t)dt
x(t)
= 0.5x2 (t) sin(x(t)) + u(t), t [0, 1] (24)
s.t.
x(0) = 0, x(1) = 0.5.
According to the Pontryagins maximum principle, the following nonlinear TPBVP
should be solved:

x(t)
= 0.5x2 (t) sin(x(t)) 0.5(t)

(t) = (t)x(t) sin(x(t)) 0.5(t)x2 (t) cos(x(t)) (25)

x(0) = 0, x(1) = 0.5
Table 1. Simulation results of Example 5.1 at dierent iteration times

J (i) J (i1)
i (Iteration time) J (i) J (i) x(tf ) xf
1 0.25 1.601 102
2 0.25 0 1.601 102
3 0.2349 6.43 102 4.477 104
4 0.2349 0 4.477 104
5 0.2346 1.3 103 2.8 105
and the optimal control law is given by:

u (t) = 0.5(t). (26)
Based on the proposed method, instead of directly solving (25), a sequence of linear
time-invariant TPBVPs is solved in a recursive manner as follows:

g 1 (t) = 0.5h1 (t) {
g1 (t) = 0.5t
:
h1 (t) = 0 (27a)
h1 (t) = 1
g1 (0) = 0, g1 (1) = 0.5

g 2 (t) = 0.5h2 (t) {
g2 (t) = 0
2
:
h2 (t) = 0 (27b)
h2 (t) = 0
g2 (0) = 0, g2 (1) = 0

g 3 (t) = 0.5h3 (t) + 0.5g13 (t) {
g3 (t) = 0
3
: h (t) = 1.5h1 (t)g1 (t)
2 (27c)
3 h3 (t) = 0.125t3
g3 (0) = 0, g3 (1) = 0
and so on.
In order to obtain an accurate enough suboptimal control law, we applied the algorithm,
proposed in Section 4, with the tolerance error bounds e1 = 210 3 and e2 = 5105 . In
(5) (4)
this case, convergence is achieved after 5 iterations, i.e., J JJ
(5) = 1.3 103 < 2 103
and x(tf ) xf = |x(1) 0.5| = 2.8 105 < 5 105 . In addition, a closed-form
expression for u5 (t) is obtained as follows:

5
895 1 1 5 1 6
u5 (t) = 0.5 hi (t) = t3 + t + t. (28)
i=1
1792 16 384 256
Simulation results including the QPI values, the QPI relative errors and the final state
errors at dierent iteration times are listed in Table 1.
Simulation curves have been obtained by applying (28) to the original nonlinear system
in (24) with x(0) = 0. Also, we obtained the simulation curves by directly solving (25)
using the collocation method [14]. Results of both methods are very close to each other
as shown in Figures 1 and 2. This confirms that the proposed method yields excellent
results. Furthermore, in contrast with the collocation method, computing procedure of
our method is very straightforward that can be done by pencil-and-paper only.
Problem (24) has also been solved by J. E. Rubio in [33] via the measure theory in
which to find an acceptable solution, a linear programming problem with 1000 variables
and 20 constraints should be solved. Therefore, applying Rubios measure theoretical
method is impossible without using computer. Table 2 provides the simulation results of
Example 5.1 obtained using our method and measure theoretical method. Comparing the
results shows the eciency of proposed method for solving nonlinear OCP (24).
Figure 1. Simulation curves of the suboptimal control in Example 5.1
Figure 2. Simulation curves of the state trajectory in Example 5.1
Example 5.2. In this example, a model of F8 fight aircraft is considered as follows [2]:

x 1 = 0.877x1 + x3 + 0.47x21 0.088x1 x3 0.019x22

( )
+ 3.846x31 x21 x3 + 0.215 + 0.28x21 u
(29)

x 2 = x3

x = 4.208x 0.396x 0.47x2 3.564x3 + (20.967 + 6.265x2 ) u

3 1 3 1 1 1
Table 2. Results of the proposed method and measure theoretical method

in Example 5.1
Method Performance index value Final state error

Proposed method with N = 5 0.2346 2.8 105
Measure theoretical method 0.2425 4.3 103
Table 3. Simulation results of Example 5.2 at dierent iteration times

J (i) J (i1)
i (Iteration time) J (i)
J (i) x(tf ) xf
1 0.2497 1.55 102
2 0.2421 3.1 102 1.22 102
3 0.2381 1.6 102 1.57 103
4 0.2371 4.2 103 1.04 103
5 0.2365 2.5 103 1.01 104
where x1 is the deviation of angle attack, x2 is the flight path angle, x3 is the rate of
change in flight path angle, and the control u is the deviation of tail deflection angle. The
QPI to be minimized is given by:
1
( 2 )
J= x1 + x22 + x23 + 10u2 dt. (30)
0
In addition, the following boundary conditions should be satisfied:
{
x1 (0) = x2 (0) = x3 (0) = 0
(31)
x1 (1) = 0.22, x2 (1) = 0.25, x3 (1) = 0.3.
As previous example, in order to obtain an accurate enough suboptimal control law, we
applied the proposed algorithm with the tolerance error bounds e1 = 5 103 and e2 =
2 104 . Simulation results at dierent iteration times are summarized in Table
3. From
(5) (4)
Table 3, it is observed that convergence is achieved after 5 iterations, i.e., J JJ
(5) =
2.5 103 < 5 103 and x(tf ) xf = 1.01 104 < 2 104 . This indicates that
J (5) = 0.2365 is very close to its optimal value J and the boundary conditions are satisfied
tightly.
Simulation curves of the computed 5th order suboptimal control law, i.e., u5 (t), and the
corresponding state trajectories obtained from applying u5 (t) to the nonlinear system (29)
with x1 (0) = x2 (0) = x3 (0) = 0 are shown in Figure 3.
As mentioned before, using the method proposed in [16] needs solving a sequence of linear
time-varying TPBVPs. Applying this method up to 5th iteration results a minimum of
J = 0.2458. On the other hand, our proposed method only requires solving linear time-
invariant TPBVPs and results a minimum of J = 0.2365 in 5 iterations. Therefore, our
method gets together both accuracy and simplicity in comparison with the above-mentioned
method.
6. Conclusions. In this paper, a novel practical approach has been introduced to solve a
class of nonlinear OCPs. In this approach, by introducing a recursive process, the optimal
control law and the optimal trajectory are determined in the form of uniformly conver-
gent series with easy-computable terms. The proposed method avoids directly solving
the nonlinear TPBVP or the HJB equation. In addition, despite of the other approxi-
mate methods such as SDRE, ASRE, SGA, SAA, etc. [9,10,12,15,16], it avoids solving
Figure 3. Simulation curves of the 5th order suboptimal control law and
the state trajectories in Example 5.2
a sequence of matrix dierential (or algebraic) equations or a sequence of generalized

HJB equations or a sequence of linear time-varying TPBVPs. It only requires solving
a sequence of linear time-invariant TPBVPs. Therefore, in view of computational com-
plexity, the proposed method is more practical than the above-mentioned approximate
methods.
Future works are focused on improving the proposed method for the optimal control
of nonlinear interconnected large-scale systems where the complexity of computations
brought by the curse of dimensionality is eectively reduced. Future studies can also be
focused on extending the method for more general nonlinear OCPs than ones which were
considered in this paper.
Acknowledgment. The authors gratefully acknowledge the helpful comments and sug-
gestions of the reviewers, which have improved the manuscript.
REFERENCES
[1] L. Tang, L.-D. Zhao and J. Guo, Research on pricing policies for seasonal goods based on optimal
control theory, ICIC Express Letters, vol.3, no.4(B), pp.1333-1338, 2009.
[2] W. L. Garrard and J. M. Jordan, Design of nonlinear automatic flight control systems, Automatica,
vol.13, no.5, pp.497-505, 1977.
[3] V. Manousiouthakis and D. J. Chmielewski, On constrained infinite-time nonlinear optimal control,
Chem. Eng. Sci., vol.57, no.1, pp.105-114, 2002.
[4] T. Notsu, M. Konishi and J. Imai, Optimal water cooling control for plate rolling, International
[5] D. E. Kirk, Optimal Control Theory: An Introduction, Prentice-Hall, NJ, 1970.
[6] A. E. Bryson, Applied Linear Optimal Control: Examples and Algorithms, Cambridge University
Press, 2002.
[7] N. Boussiala, H. Chaabi and W. Liu, Numerical method for solving constrained non-linear optimal
control using the block pulse functions (BPFs), International Journal of Innovative Computing,
[8] R. Padhi and M. Kothari, Model predictive static programming: A computationally ecient tech-
nique for suboptimal control design, International Journal of Innovative Computing, Information
[9] T. Cimen, State dependent Riccati equation (SDRE) control: A survey, Proc. of the 17th IFAC
World Congress, Seoul, Korea, pp.3761-3775, 2008.
[10] S. P. Banks and K. Dinesh, Approximate optimal control and stability of nonlinear finite- and
infinite-dimensional systems, Ann. Oper. Res., vol.98, no.7, pp.19-44, 2000.
[11] R. Bellman, On the theory of dynamic programming, P. Natl. Acad. Sci., vol.38, no.8, pp.716-719,
1952.
[12] R. W. Beard, G. N. Saridis and J. T. Wen, Galerkin approximations of the generalized Hamilton-
Jacobi-Bellman equation, Automatica, vol.33, no.12, pp.2159-2177, 1997.
[13] L. S. Pontryagin, Optimal control processes, Usp. Mat. Nauk, vol.14, pp.3-20, 1959.
[14] U. M. Ascher, R. M. M. Mattheij and R. D. Russel, Numerical Solution of Boundary Value Problems
for Ordinary Dierential Equations, SIAM, PA, 1995.
[15] G. Y. Tang, Suboptimal control for nonlinear systems: A successive approximation approach, Syst.
Control Lett., vol.54, no.5, pp.429-434, 2005.
[16] G. Y. Tang, H. P. Qu and Y. M. Gao, Sensitivity approach of suboptimal control for a class of
nonlinear systems, J. Ocean Univ. Qingdao, vol.32, no.4, pp.615-620, 2002.
[17] N. Pariz, Analysis of Nonlinear System Behavior: The Case of Stressed Power Systems, Ph.D.
Thesis, Ferdowsi University, Mashhad, Iran, 2001.
[18] N. Pariz, H. M. Shanechi and E. Vaahedi, Explaining and validating stressed power systems behavior
using modal series, IEEE Trans. on Power Syst., vol.18, no.2, pp.778-785, 2003.
[19] H. M. Shanechi, N. Pariz and E. Vahedi, General nonlinear modal representation of large scale power
systems, IEEE Trans. on Power Syst., vol.18, no.3, pp.1103-1109, 2003.
[20] F. X. Wu, H. Wu, Z. X. Han and D. Q. Gan, Validation of power system non-linear modal analysis
methods, Electr. Pow. Syst. Res., vol.77, no.10, pp.1418-1424, 2007.
[21] S. Soltani, N. Pariz and R. Ghazi, Extending the perturbation technique to the modal representation
of nonlinear systems, Electr. Pow. Syst. Res., vol.79, no.8, pp.1209-1215, 2009.
[22] M. Khatibi and H. M. Shanechi, Using modal series to analyze the transient response of oscillators,
Int. J. Circ. Theor. Appl., John Wiley & Sons, 2009.
[23] J. A. Murdock, Perturbations: Theory and Methods (Classics in Applied Mathematics 27), Society
for Industrial Mathematics, 1999.
[24] A. M. Lyapunov, General Problem on Stability of Motion, Taylor & Francis, London, 1992.
[25] G. Adomian and G. E. Adomian, A global method for solution of complex systems, Math. Modelling,
vol.5, no.4, pp.251-263, 1984.
[26] E. Visser, B. Srinivasan, S. Palanki and D. Bonvin, A feedback-based implementation scheme for
batch process optimization, J. Proc. Contr., vol.10, no.5, pp.399-410, 2000.
[27] R. Maya-Yescas and R. Aguilar, Controllability assessment approach for chemical reactors: Nonlinear
control ane systems, Chem. Eng. J., vol.92, no.1-3, pp.69-79, 2003.
[28] P. K. Menon and E. J. Ohlmeyer, Computer-aided synthesis of nonlinear autopilots for missiles,
Nonlinear Studies, vol.11, no.2, pp.173-198, 2004.
[29] J. Yu, A. Jadbabaie, J. Primbs and Y. Huang, Comparison of nonlinear control design techniques
on a model of the Caltech Ducted Fan, Automatica, vol.37, no.12, pp.1971-1978, 2001.
[30] S. N. Singh and W. Yim, State feedback control of an aeroelastic system with structural nonlinearity,
Aerosp. Sci. Technol., vol.7, no.1, pp.23-31, 2003.
[31] H. Jaddu, Direct solution of nonlinear optimal control problems using quasilinearization and Cheby-
shev polynomials, J. Frankl. Inst., vol.339, no.4-5, pp.479-498, 2002.
[32] V. I. Arnold, Ordinary Dierential Equations, Springer-Verlag, New York, 1992.
[33] J. E. Rubio, Control and Optimization, the Linear Treatment of Nonlinear Problems, Manchester
University Press, Manchester, 1986.
c ISSN 1349-4198
ADAPTIVE FRAME LENGTH METHOD FOR HARDWARE

CONTEXT-SWITCHING IN DYNAMIC PARTIAL
SELF-RECONFIGURABLE SYSTEMS
Trong-Yen Lee1 , Che-Cheng Hu1 , Yang-Kun Huang1 and Chia-Chun Tsai2

1
Graduate Institute of Computer and Communication
National Taipei University of Technology
No. 3, Zhongxiao E. Road, Taipei 10643, Taiwan
{ tylee; s4419010; t5418065 }@ntut.edu.tw
2
Nanhua University
No. 55, Sec. 1, Nanhua Road, Zhongkeng, Dalin Township, Chia-Yi 62248, Taiwan
chun@mail.nhu.edu.tw
Abstract. A dynamically and partially reconfigurable system can arbitrarily swap hard-
ware tasks at runtime. However, such FPGA systems require a long time to read and
much memory to store the hardware context, when a hardware task is swapped out. A
solution to this problem is to adjust the amount of read-back data in a hardware task to
reduce the required memory size and reconfiguration time. Therefore, this investigation
develops an adaptive frame length method for hardware context-switching with a context
reading and writing (CRW) algorithm in dynamically partial self-reconfigurable (DPSR)
FPGA systems. The proposed CRW algorithm requires less instruction to execute Read-
back and Writeback procedures. Therefore, the CRW algorithm is one means of reducing
reconfiguration time. Experimental results show that the proposed method reduces hard-
ware reconfiguration time by 7.763%, memory size by 63.79% and resource overhead by
47.33%.
Keywords: Context-switching, Dynamic partial self-reconfigurable system, Readback,
Reconfiguration time, Memory size
1. Introduction. Field programmable gate arrays (FPGAs) that execute at least two
hardware tasks simultaneously are very popular in the design of embedded systems. Dy-
namically partial self-reconfigurable (DPSR) FPGAs allow for the reconfiguring of var-
ious hardware tasks at runtime using early access partial reconfiguration [1] whenever
DPSR FPGA systems have placement and removal requirements. Therefore, the release
of hardware space at runtime is critical. DPSR FPGA systems can thus perform partial
reconfiguration and share FPGA resources among hardware tasks. These useful features
can be applied to save and restore register information in DPSR FPGA systems, which
process is called hardware context-switching.
Hardware context-switching enables hardware resources to be replaced when new high-
priority tasks are to be executed. Hardware tasks can also be reloaded to complete
unfinished work whenever a request is made for swap-in by context-switching. Addition-
ally, a hardware task consists of many configurable logics. For this reason, swap-out by
context-switching requires huge memory to save all of the registers of the configurable
logics that are associated with the hardware task. Therefore, the two main disadvan-
tages of the DPSR FPGA system are the time delay that is associated with the hardware
1427
1428 T.-Y. LEE, C.-C. HU, Y.-K. HUANG AND C.-C. TSAI
configuration [1,2] and the size of the memory that is required to store the hardware con-
text. However, the DPSR FPGA system provides the advantages of both hardware and
software in terms of short execution time, concurrency and flexibility [2,3].
To realize hardware context-switching and improve on the required memory size and
time delay of the DPSR FPGA system, this investigation develops an adaptive frame
length method for hardware context-switching to suspend and restart hardware tasks
[4,5] in DPSR FPGA systems. This investigation also proposes a Context Reading and
Writing (CRW) algorithm to perform Readback and Writeback procedures in hardware
context-switching. Readback and Writeback are applied to save and restore information
in hardware registers when hardware tasks are swapped out and swapped in. The inter-
nal configuration access port (ICAP) interface [6,7] is applied to execute Readback and
Writeback procedures to increase the configuration or reconfiguration speed from ICAP
clock frequency. Additionally, the proposed method uses a frame selector to adjust the
reading frame length to prevent the reading of unnecessary data. The limitation of the
proposed method is that it is suitable only for the Xilinx series of Virtex-II/Pro FPGA.
The rest of this paper is organized as follows. Section 2 discusses related work and
designs for context switching. Section 3 describes the proposed method and architecture
for hardware context-switching. Section 4 presents four example designs to demonstrate
the proposed method, and estimates their performance in terms of reconfiguration time,
required memory size and resource overhead. Conclusions are finally drawn in Section 5,
along with recommendations for future research.
2. Related Work. Dynamic switching or relocation in FPGA design falls into one of
two categories reconfiguration-based (configuration port access) [8-13] and design-based
(task specific access structure) [14-17]. Reconfiguration-based dynamic hardware context-
switching utilizes the Readback function that is based on the FPGA configuration port.
The advantage of the reconfiguration-based technique is that it does not require read/write
interface circuits in the design of the hardware task. Therefore, the design eort and
FPGA resources will be reduced. Another benefit is that the designer does not need
to know the internal behavior or communication infrastructure that is associated with
the task. Furthermore, if the task module is an intellectual property (IP) core, then
the hardware register information can all be saved and restored. However, the greatest
disadvantage of the reconfiguration-based technique is its low data eciency [1,8,17,18]
and huge memory requirement, although these issues have been partially resolved in other
works [8,9]. The proposed method is a reconfiguration-based technique.
Kalte et al. [8] only stored locations of the register that contain information on rows and
columns in FPGA. The information on rows and columns is applied to calculate register
location, and the current register information is extracted by Readback procedures. This
method does not require the reading of the entire CLB (configurable logic block) column
of the hardware, because all of the state registers of a CLB column are grouped into
only two frames (a CLB column has 22 frames) in the Virtex-E/II/Pro series of FPGA.
Kalte et al. [8] used the Virtex-E series with a data width of 19 bits to store content and
the location of a register, and the Virtex-II series with a data width of 17 bits to store
content and the location of a register, because the method of Kalte et al. [8] depends
on the maximum row and column lengths of the CLB in the Virtex series as well as
other important information. Clearly, when reconfiguration occurs more frequently, more
memory is required to store the hardware context, however, a smaller memory is strongly
favored in this investigation.
A design-based technique requires a more complex hardware interface, such as a read/w-
rite interface [15,17] or a software API (application programming interface) [14], than the
ADAPTIVE FRAME LENGTH METHOD FOR HARDWARE CONTEXT-SWITCHING 1429
original hardware circuit. Moreover, the designer must have detailed knowledge of the in-
ternal behavior and communication infrastructure that are associated with the hardware
task, because no task has a standard or general interface. Therefore, the greatest disad-
vantages of the design-based technique are the additional design eort required and the
extra resources consumed. Another drawback is that, if the task module that is used as
an IP core comes from a third party, then no extra read/write interface can be added if no
RTL model of the IP is available. However, the advantage of the design-based technique is
its high data eciency [8,17,18], because the designed hardware, including the read/write
interface, performs all context switching. Moreover, only the required data rather than all
of the frame data are read. Consequently, reconfigurable computing by the design-based
technique is faster than that by the reconfiguration-based technique [8,17].
An adaptable system is more ecient and flexible than an unadaptable system [19-
21]. Therefore, this investigation develops an adaptive frame length method of hard-
ware context-switching to save and restore the significant frame and register data for the
swapping-in and swapping-out of hardware modules in an FPGA. The significant frame
and register data are stored in the same memory space. The memory structure in this
method diers from that in an earlier method [8]. The proposed method uses only a data
width of 11 bits to store the content and location of a register. Only the frame addresses,
bit indexes and some flags are recorded. (See Subsection 3.3 for details.) Consequently,
the proposed method reduces the size of the memory in hardware context-switching. How-
ever, the proposed method can also be applied in a web-based remote platform [22] and
a mechanical controller [23].
Modern electronic products require little design eort and have a small memory, low
power consumption and high resource utilization all of which features ensure short time-
to-market, high performance and low cost. Electronic product designs are moving towards
increased functionality, which requires the implementation of more tasks [24], and makes
these constraints increasingly dicult to satisfy. The problem related to simultaneously
satisfying multiple constraints for a DPSR FPGA system is described as follows.
Short time-to-market: The proposed method does not require to the design of an
extra read/write interface. It can also be applied to the IP core of a third party.
Therefore, the proposed method greatly reduces the design eort of the designer.
High performance: Traditionally, one or more ICs are designed with many functions.
However, electronic products do not perform many functions simultaneously. Con-
sequently, the power consumption is high. If a function has requirements about the
placement or removal of hardware, then the hardware associated with this function
will be loaded to a system or removed from it using the proposed method. There-
fore, the system of electronic products retains only the hardware associated with the
used functions. Consequently, the proposed method reduces power consumption.
Additionally, the proposed method reduces the recoding number of the frame data,
reducing the time delay of system.
Low cost: The proposed method can execute partial reconfiguration and share FPGA
resources among hardware tasks to improve resource utilization. It does not require
an extra hardware interface. Therefore, electronic products can be made smaller.
Additionally, the proposed system is adaptive, and requires a smaller data width
and length to store significantly the hardware context. Therefore, the size of the
memory can be reduced. Restated, the proposed method can store more hardware
context in a limited memory.
3. Proposed Adaptive Frame Length Method for Hardware Context-Switching.

The FPGA structures of current reconfigurable systems fall into three categories [18,25]
Figure 1: Self-reconfigurable FPGA system architecture
single context, multi-context and partially reconfigurable. The structure of Xilinx Virtex-
II/Pro FPGA has been described in detail elsewhere [13,26-33]. To implement run-time
reconfiguration on a single context FPGA, configuration data must be grouped into full
contexts. Complete hardware contexts are swapped in and swapped out whenever the
hardware task is replaced. This swapping is the main cause of delay in hardware configu-
ration. Two structures have been applied to reduce the configuration delay multi-context
and partially reconfigurable. Multiple context configuration memory can be activated at
dierent times. Partial configuration allows selective access to the configuration memory.
The method proposed herein is based on the partially reconfigurable FPGA structure.
This section presents a hardware context-switching method for dynamically partial
self-reconfigurable systems. The proposed method is implemented in a self-reconfigurable
FPGA system, as shown in Figure 1. The proposed method uses the ICAP interface to
execute the Readback and Writeback procedures using a micro processor to control the
FPGA configuration memory data and register information.
The Readback procedure can arbitrarily read frame data in configuration memory, in-
cluding the contexts of registers and RAM. Therefore, the Readback function can read
all or some of the bitstream. In the Readback operation, the state information is filtered
and saved in the output of the Readback stream. During the allocation step, the state
information, such as the flip-flop data or the RAM contents in the FPGA structure, can
be restored into suitable register and RAM locations via the Writeback procedure. The
following subsections will introduce the proposed context-switch flow and method.
3.1. Hardware context-switch flow. Figure 2 shows the hardware context-switch flow.
If a task will be executed in a reconfigurable system, then the system will search for a
free reconfigurable area to perform the task. The reconfigurable area in a reconfigurable
system is in one of two situations, which are 1) free, allowing the task to be loaded
directly into the reconfigurable area, or 2) non-free, meaning that the low-priority working
Figure 2: Hardware context-switch flow
hardware module in the reconfigurable area will be swapped out. Accordingly, the register
information about a swapped-out hardware module must be stored in state memory via
a state filter and a frame selector.
During the Readback operation, the Readback procedure reads the register information
about a hardware module from the module database memory. If the register information
is the frame address, then the Readback procedure uses a frame selector and the adaptive
frame length method to determine the number of the frame. Otherwise, it uses a bit
index calculation unit to filter the location and value of the register within the frame.
The values of these registers are stored in the state memory. Afterward, the Readback
procedure repeats the motions that are described above until the register information
within the hardware has been stored.
3.2. Module database memory and state filter. All significant information associ-
ated with each hardware task is stored to database memory. Figure 3 shows the task
context database generation flow. A logic allocation file (task.ll ) is generated using a
BitGen (bitstream generator) [6], and its content includes the frame address and the bit
index of the used registers. The logic allocation file is parsed by an auxiliary tool to
Figure 3: Task context database generate flow
Table 1: Frame address and bit index of slice in XC2V1000
C1 C2
MJA = 3 MJA = 4
X0 X1 X2 X3
XQ (hex) YQ (hex) XQ (hex) YQ (hex) XQ (hex) YQ (hex) XQ (hex) YQ (hex)
MNA = 1 MNA = 2 MNA = 1 MNA = 2 MNA = 1 MNA = 2 MNA = 1 MNA = 2
00060200 00060400 00060200 00060400 00080200 00080400 00080200 00080400
Y79 (dec) 118 118 116 116 118 118 116 116
R40
Y78 (dec) 158 158 156 156 158 158 156 156
Y77 (dec) 198 198 196 196 198 198 196 196
R39
Y76 (dec) 238 238 236 236 238 238 236 236

Y1 (dec) 3238 3238 3236 3236 3238 3238 3236 3236
R1
Y0 (dec) 3278 3278 3276 3276 3278 3278 3276 3276
extract the important information, and is stored in the database memory. Afterward,
the state filter extracts the locations of the registers within each frame. If the registers
have the same frame address, then only one of them will be saved in database memory,
reducing the computing time and required memory size. The characteristics of the bit
index and the frame address, and the equation of the state filter are as follows.
One CLB consists of four similar slices. One slice has two D-type registers (Flip-Flops)
whose respective outputs are XQ and YQ [34]. Each slice occupies two frame addresses in
a single CLB column. Therefore, one CLB has eight registers [6]. The relationship among
rows, columns, frame addresses and bit index parameters (Y row ) is arranged from the
logic allocation file, as shown in Table 1.
Each column (such as C1), X column, can be separated into an even column (X even)
and an odd column (X odd ) [31]. Each row (such as R40), Y row, can be further separated
into an even row (Y even) and an odd row (Y odd ). For example, the eight registers that
occupy two frame addresses in column C1 are 00060200(hex) and 00060400(hex), as shown
in Table 1. X even and X odd in Table 1 are utilized to derive the MJA [32] of the frame
address of the register, as shown in Equation (1). The MNA [32] depends on XQ and
YQ. If XQ is used in the design, then MNA equals 1. If YQ is used, then MNA equals 2.
M JA = X even/2 + 3 = (X odd 1)/2 + 3 (1)
Table 1 shows the rules for computing the bit indexes of registers in a frame. For
example, eight registers in the C1R40 (column 1, row 40) CLB occupy two frame addresses
in the sub-column (such as X0) and the adjacent sub-column (such as X1). Moreover,
the register data are distributed over bit indexes 118, 158, 116 and 156. Equation (2)
is derived from these rules when the column number is odd. Similarly, Equation (3) is
derived if the column number is even. In Equations (2) and (3), CLB Rows is the number
Figure 4: Storage format in database memory
of CLB rows.
bit idx Xodd = 116 + 40 (2CLB Rows 1 Y row) (2)
bit idx Xeven = 118 + 40 (2CLB Rows 1 Y row) (3)
According to Equations (2) and (3), the bit indexes are related to the number of rows
in CLB. Therefore, the Y row datum is stored in database memory rather than the bit
index. Additionally, when the bit indexes of the registers within a frame are obtained
using Equations (2) and (3), the register data within the bit indexes are stored in the
state memory.
In the Xilinx Virtex-II/Pro FPGA, the maximum numbers in the CLB column are 104
(XC2V8000) and 94 (XC2VP100), and the respective maximum numbers in the MJA are
106 and 96. Each CLB column has a maximum of 112 (XC2V8000) and 120 (XC2VP100)
rows, and the respective maximum Y row values are 223 and 239. Accordingly, the MJA
is stored as 7 bits, and the MNA is stored as 1 bit in database memory. Y row is stored
as 8 bits, and a single bit specifies whether a column is odd or even.
3.3. Frame addresses and addressing bit indexes. The frame addresses and bit
indexes are stored in block memory, as shown in Figure 4. The proposed method initially
stores one frame address of a column (such as C1) to database memory. Then, the k bit
index parameter is stored to the following address in database memory.
The database memory has a width of 11 bits for storing the frame address and bit index,
as shown in Table 2. The two most significant bits (Bits Flag) specify the frame address
or bit index format. In the frame address format, the MJA requires 7 bits because the
maximum value is 106; MNA flag uses 1 bit to represent the MNA attribute; Bits Flag
is permanently 00, and nframe uses 1 bit to specify whether the Readback procedure
will read one frame or two frames. In the bit index format, Y row requires 8 bits because
the maximum value is 239, and X oe uses 1 bit to specify the attribute of the X column.
Bits Flag also specifies the utilization condition of the bit index in the registers. Table
3 presents detailed information on Bits Flag, MNA flag, nframe and X oe. Additionally,
Equations (2) and (3) can be substituted into Equation (4) according to X oe in Table 3.
bit idx = 118 + 40 (2CLB Rows 1 Y row) 2X oe (4)
Table 2: Frame address and bit index parameter for addressing the register
Frame address Bit index parameter

Bits Flag MJA MNA flag nframe Bits Flag X oe Y row
10 9 82 1 0 10 9 8 70
00 XXXXXXX X X XX X XXXXXXXX
3.4. Context reading and writing algorithm. This section presents a context reading
and writing (CRW) algorithm for hardware context-switching in dynamic partially self-
reconfigurable systems. Figure 5 displays this CRW algorithm. Moreover, the CRW
algorithm is a modified version of an algorithm that was developed elsewhere [6,13]. The
CRW algorithm includes the Readback (steps 4 to 9) and Writeback procedures (steps 11
to 16), which are used to implement hardware context-switching: they read data from
the configuration memory and configure data in the configuration memory, respectively.
A microprocessor controls the Readback and Writeback procedures in a reconfigurable
FPGA system using the ICAP interface.
Now, the CRW algorithm is discussed in detail. In step 1, the CRW algorithm receives
a hardware context-switching requirement from the system input. Next, it sends the
synchronization word to the ICAP interface to synchronize the configuration logic. Before
the configuration logic is synchronized, all frame data on the configuration interface are
neglected. In step 3, if other high-priority tasks must be executed in the self-reconfigurable
FPGA system, then the Readback procedure (steps 4 to 9) is executed. In step 10, if the
other unfinished tasks must be reloaded in the self-reconfigurable FPGA system, then
the Writeback procedure (steps 11 to 16) is executed. After a read-back or a write-back
operation has been completed, in step 17, the configuration logic is desynchronized before
any further operations can be executed in the ICAP interface. The final step in the CRW
algorithm is to flush the 64-bit buer.
The details of the Readback procedure in the CRW algorithm are as follows. The
purpose of the Readback procedure is to capture and save the states of the current reg-
ister before a task is swapped out. In steps 4 and 5, the Readback procedure sends the
GCAPTURE command to the CMD register, to capture the states of all slices to the
configuration memory. It also sends the RCFG command to the CMD register, to cause
the ICAP interface to prepare to read the frame data from the FDRO register. In steps 6
to 9, the address of the starting frame and its reading length are sent to the FAR register
and the FLR register, respectively. Afterward, the frame data are read from the FDRO
register via the ICAP interface.
The details of the Writeback procedure in the CRW algorithm are as follows. The
purpose of the Writeback procedure is to reload and complete an unfinished task if the
hardware task has been swapped out via the Readback procedure. In step 11, to ensure
that the CRC calculation is correct, the CRC register must be reset. In step 12, an
IDCODE must write successfully into the device ID register before the frame data are
read, because each type of FPGA has a device-specific identification code. In step 13, the
Writeback procedure sends the WCFG command to the CMD register, causing the ICAP
interface to prepare to write frame data to the FDRI register. In steps 14 and 15, the
address of the starting frame and its writing length are sent to the FAR register and the
FLR register, respectively. Afterward, the frame data are written into the FDRI register
via the ICAP interface. In step 16, a CRC value is always created to enable the frame
data to be checked during a write-back operation.
Table 3: Compatibility of bit index and frame address in database memory
Bits Flag Description

00 The frame address
01 The bit index occupied on the first frame address
10 The bit index occupied on the second frame address
11 The bit index occupied on the both frame addresses
X oe Description
0/1 X Column is in the even/odd
MNA flag Description
0/1 MNA = 2/1
nframe Description
0/1 Read two frames/read one frame
Figure 5: Context reading and writing (CRW) algorithm
3.5. Performance analysis. Kalte et al. [8] used Virtex-II series of FPGA with a data
width of 17 bits to store the frame address and bit index information of a register. This
Figure 6: Comparison of Readback procedure commands
information is used to calculate the frame addresses and bit indexes in the Readback
operation. Koester et al. [9] used the number of bytes to carry significant information
about a frame, however, this method depends on the size of the hardware task and the
frame length of the target device. Therefore, a data width of 16 bits was used to store
significant information about the hardware because, in the Virtex-II/Pro series of FPGA,
all of the state registers of a CLB column are located in only two frames. Jovanovic et
al. [35] used a data width of 16 bits to store significant information about a register. The
method of Jovanovic et al. [35] depends on the number of slices associated with hardware
task. Clearly, it requires more memory than the other methods to store the hardware
context. The method proposed herein uses a data width of 11 bits to store the frame
address and bit index information (as detailed in Subsection 3.3). Clearly, the proposed
method requires less memory to store the hardware context.
A series of commands must be sent before the Readback procedure can be executed.
The proposed Readback procedure uses half as many commands as the Shutdown Readback
procedure [8,9,13], as shown in Figure 6. Consequently, the proposed method reduces the
time to send the Readback commands and increases the speed of the Readback operation.
Figure 7 compares the procedures for reading the CLB column materials that are associ-
ated with the methods of Kalte et al. [8], Koester et al. [9] and Lee et al. [13] with those of
proposed in this investigation. In the other methods [8,9], a single frame address requires
one pad frame to flush the frame buer. Therefore, when the Readback procedure reads
two frame addresses, four cycles of reading time per frame are used, as shown in Figure
7(a). In another method [13], a single frame address is saved for each CLB column, and
the three frames are read. The three frames are two frames of data and one pad frame,
as shown in Figure 7(b). However, the proposed method reads out two or three frames,
which include one or two frames data and one pad frame, as shown in Figure 7(c). As
mentioned above, the proposed method uses fewer commands and less memory than the
methods proposed elsewhere [8,9,13].
The reconfiguration time in an FPGA depends on the working frequency in ICAP
and the Readback procedure. Equation (5) yields the setting time of overall command
Figure 7: Comparison of procedures for reading CLB column materials
register for the Readback procedure; the number of commands is the sum of Synch cm,
NCap Rcf g cm , NCap F ar cm Nclb and NDesynch cm . Nclb denotes the number of used CLB
columns. NCap Rcf g cm and NCap F ar cm are the number of commands that must be sent to
read the material in a CLB column. Since one command has four bytes, the numerator in
Equation (5) must be multiplied by four. Tread f rame data is the sum of the reading times of
the data in all frames and Tinstruction . In the expression for the reading time for data in all
frames, the term 3Nboth reg represents three cycles of reading time per frame, when a CLB
uses both XQ and YQ, and the term 2Nsingle reg represents two cycles of reading time per
frame, when a CLB uses either XQ or YQ. The term NByte/F rame denotes the frame length
of FPGA. Therefore, Equation (6) yields the total time required for reading the frame
data. A CLB of Xilinx Virtex-II/Pro has 22 frames (22Nclb ). Therefore, Equation (7)
yields an estimate of the configuration time, Tconf ig . The reconfiguration time (Treconf ig )
in a task is the sum of Tread f rame data and Tconf ig , which is given by Equation (8).
4 [Synch cm + NCap Rcf g cm + (NCap F ar cm Nclb ) + NDesynch cm ]
Tinstruction. = (5)
fICAP
(3Nboth reg + 2Nsingle reg ) NByte/F rame
Tread f rame data = + Tinstruction ,
fICAP (6)
where 3Nboth reg : both XQ and YQ are used in a CLB,
where 2Nsingle reg : XQ or YQ is used in a CLB
22Nclb NByte/F rame
Tconf ig = (7)
fICAP
Table 4: Comparison of memory sizes
Kalte [8] Koester [9] Jovanovic [35] Our Reduced

32-bit Divider (73) 1241 Bits 1168 Bits 2275 Bits 484 Bits 66.09%
GCD32 (83) 1411 Bits 1328 Bits 4130 Bits 715 Bits 59.39%
DES (84) 1428 Bits 1344 Bits 7901 Bits 704 Bits 63.13%
DCT (1421) 24157 Bits 22736 Bits 29609 Bits 8415 Bits 66.57%
Average 63.79%
Reduced: % of overheads of memory size in the proposed method compared to [8,9,35]. For example
(DES), {[(1428 704)/1428] + [(1344 704)/1344] + [(7901 704)/7901]} /3 = 63.13%
25Nclb NByte/F rame

Treconf ig = Tread f rame data + Tconf ig = + Tinstruction (8)
fSelectM AP
4. Experimental Results. In this section, four example designs of a 32-bit divider,

a 32-bit greatest common divisor (GCD32 ), a data encryption standard (DES) and a data
transformation component (DCT) are used to demonstrate the benefits of the proposed
method. In practice, all design examples must complete the generation of a task con-
text database, as shown in Figure 3, before hardware context-switching can be executed.
Figure 4, Tables 2 and 3 present the format for storing a task context database. When
DPSR FPGA systems have swap-in or swap-out requirements, significant information
about these design examples will be saved or restored by CRW algorithm, according to
the task context database.
An experiment is performed on Xilinx XC2V1000 and XC2VP20 FPGA platforms.
The XC2V1000 FPGA has a frame length of 424 bytes (106 words 4 byte/word,
NByte/F rame ), and the XC2VP20 FPGA has a frame length of 584 bytes (146 words
4 byte/word, NByte/F rame ). The working frequency in ICAP is 50 MHz. The reconfigura-
tion time and memory size determine the hardware context-switching performance in the
reconfiguration-based technique.
Table 4 compares the memory sizes in the methods of Kalte et al. [8], Koester et
al. [9], Jovanovic et al. [35] and this investigation. The memory structure of the proposed
method diers from that in the others [8,9,35]. For example, that in the method of Kalte
et al. [8] depends on the maximum row and column length of the CLB in the Virtex series.
However, the proposed method records only frame address values and bit index values.
Therefore, Kalte et al. [8], Koester et al. [9] and Jovanovic et al. [35] used 17 bits, 16
bits and 16 bits registers, respectively, to store the parameters of the frame addresses, bit
indexes and significant information. The method proposed herein uses an 11 bits register
to store the frame addresses, bit indexes and some flags. The experimental results show
that the proposed method requires on average 63.79% less memory than the other three
methods. Therefore, the proposed method can store more hardware context in a memory
of limited size. The values in the Reduced column are explained further under Table 4.
Table 5 compares the reconfiguration times in the methods of Kalte et al. [8], Koester
et al. [9], Lee et al. [13] and this investigation. As shown in the rows Reading Frame
and Reading Frame Data in Table 5, Kalte et al. [8] and Koester et al. [9] used four
cycles of reading time per frame (4Nclb ), as shown in Figure 7(a), while Lee et al. [13]
used three (3Nclb ), as shown in Figure 7(b). The proposed method can flexibly choose the
number of frames, requiring only two or three cycles of reading time per frame (3Nboth reg +
2Nsingle reg ), as shown in Figure 7(c).
The rows Number of Commands and Time of Comm. Reg. Setting in Table 5 show
that the methods of Kalte et al. [8] and Koester et al. [9] used ten Readback operations,
while that of Lee et al. [13] used five Readback operations in the Shutdown Readback
procedure, as shown in Figure 6(b). However, the proposed method uses only five such
operations in the Readback procedure, as shown in Figure 6(a). Clearly, the proposed
method requires less processing of the Readback operation than the other methods [8,9,13]
in all design examples. Equation (5) yields the results of rows, Number of Commands
and Time of Comm. Reg. Setting. The proposed method requires on average 44.26%
less command setting time than the other three methods.
The row Reading Time of Frame in Table 5 presents the total time to read all frames,
which is given by Equation (6). The proposed method requires on average 33.6% less
reading time of frame than the other three methods. The row Reconfig. Time in Table
5 denotes reconfiguration time, which is given by Equations (7) and (8). The proposed
method requires on average 7.763% less hardware reconfiguration time than the other
three methods. The values in the rows Reduced are explained further under Table 5.
Section 2 pointed out that design-based reconfigurable computing is faster than the
reconfiguration-based technique. Restated, the reconfiguration-based technique has a
longer reconfiguration time than the design-based. However, the design-based technique
requires an extra read/write interface to execute hardware context-switching. Therefore,
the reconfiguration-based technique has a lower resource overhead than the design-based.
Table 6 compares the resource overheads of the reconfigurable-based and the design-based
techniques. The values in the row Flip-Flops in Table 6 for the method of Huang et
al. [17] range from 169 to 2094 in all design examples, while those for the proposed
method range only from 83 to 1421 in all design examples. Clearly, the proposed method
consumes fewer resources than the method of Huang et al. [17] in all design examples.
Consequently, it has on average a 47.33% lower hardware resource overhead. The values
in the row Reduced rate of resource are explained further under Table 6.
Above experimental results demonstrate that the reconfiguration time and required
memory size of the proposed hardware context-switching methodology are lower than
those of other reconfiguration-based techniques. Additionally, the proposed reconfigurati-
on-based technique uses fewer resources than the design-based techniques.
5. Conclusions. This work developed an adaptive frame length method for hardware
context-switching to reduce the required memory size and reconfiguration time in dy-
namically partially self-reconfigurable systems. A context reading and writing (CRW)
algorithm for hardware context-switching to handle Readback and Writeback procedures
was also proposed. The CRW algorithm reduced the read-back time and the complexity
of the command registers setting. The method of hardware context-switching was suit-
able for all FPGAs of the Virtex-II/Pro series. The experimental results showed that
the proposed method reduced the size of the required memory by 63.79% for storing the
hardware context, and had on average a 7.763% lower hardware reconfiguration time than
previous reconfiguration-based techniques. The proposed method has on average a 47.33%
lower resource overhead than design-based techniques. This method was applicable only
to the Virtex-II/Pro series of FPGA. Therefore, the proposed method cannot be applied
directly to other series of FPGA, such as Virtex-4 or Virtex-5 series, because dierent
series of FPGA have dierent structures. Future work should aim to determine the com-
patibility of the structures of various series of FPGA. Eorts are underway to develop a
hardware context-switching method that is appropriate for various Virtex series of FPGA
by utilizing the mutual compatibility of FPGA structures.
1440
Table 5: Comparison of reconfiguration times for saving and restoring context
Design
32-bit Divider GCD3 2 DES DCT
Examples
Items Kalte Koester Lee Kalte Koester Lee Kalte Koester Lee Kalte Koester Lee
XXX
XXX Our Our Our Our
Method XX [8] [9] [13] [8] [9] [13] [8] [9] [13] [8] [9] [13]
Reading Frame 20 20 15 13 16 16 12 11 24 24 18 18 36 36 27 27
Reading Frame
17520 8480 6360 5512 19136 9344 7008 6424 28704 14016 10512 10512 43056 21024 15768 15768
Data (Byte)
Number of
284 284 184 132 244 244 164 112 324 324 204 152 444 444 264 212
Commands (Byte)
Times of Comm.
Reg. Setting 5.68 5.68 3.68 2.64 4.88 4.88 3.28 2.24 6.48 6.48 4.08 3.04 8.88 8.88 5.28 4.24
(s) (Tinstruction )
Reduced 45.10% 46.63% 43.89% 41.40%
Average 44.26%
Reading Time of
Frame (s) 356.08 175.28 130.88 112.88 387.6 191.76 143.44 130.72 580.56 286.8 214.32 213.28 870 429.36 320.64 319.6
(Tread f rame data )
Reduced 39.22% 35.66% 29.79% 29.72%
Average 33.60%
Reconfig. Time
1288.88 1108.08 1063.7 1045.68 1415.4 1219.6 1171.3 1158.6 2122.3 1828.6 1756.1 1755 3182.64 2742 2633.3 2632.24
(s) (Treconf ig )
Reduced 8.73% 8.08% 7.13% 7.11%
T.-Y. LEE, C.-C. HU, Y.-K. HUANG AND C.-C. TSAI
Average 7.763%
Reduced: % of time overheads in the proposed method compared to [8,9,13]. For example (GCD32 ), {[(387.6 130.72)/387.6] + [(191.76
130.72)/191.76] + [(143.4 130.72)/143.4]}/3 = 35.66%
Table 6: Comparison of resource overheads
Design Examples GCD32 DES DCT

PP
PPMethod
PP Huang [17] Our Huang [17] Our Huang [17] Our
Item PP P
Flip-Flops 169 83 207 84 2094 1421
Reduced rate
51% 59% 32%
of resource
Average 47.33%
Reduced rate of resource: % of resource overheads in the proposed method compared
to [17]. For example (GCD32 ), [(169 83)/169] = 51%
Acknowledgment. The authors would like to thank the National Science Council, Tai-
wan, for partially supporting this research under Contract No. NSC-98-2221-E-027-071.
The authors also gratefully acknowledge the helpful comments and suggestions of the
reviewers, which have improved this work.
REFERENCES
[1] J. Phillips, A. Sudarsanam, H. Samala, R. Kallam, J. Carver and A. Dasu, Methodology to derive
context adaptable architectures for FPGAs, IET Computers & Digital Techniques, vol.3, pp.124-141,
2009.
[2] F. Berthelot, F. Nouvel and D. Houzet, A flexible system level design methodology targeting run-time
reconfigurable FPGAs, EURASIP J. on Embedded Systems, vol.2008, pp.1-18, 2008.
[3] S. M. Scalera and J. R. Vazquez, The design and implementation of a context switching FPGA,
Proc. of the IEEE Symposium FPGAs for Custom Computing Machines, pp.78-85, 1998.
[4] C.-H. Huang, S.-S. Chang and P.-A. Hsiung, Generic wrapper design for dynamic swappable hard-
ware IP in partially reconfigurable systems, Int. J. of Electrical Engineering, pp.229-238, 2007.
[5] C.-H. Huang, K.-J. Shih, C.-S. Lin, S.-S. Chang and P.-A. Hsiung, Dynamically swappable hard-
ware design in partially reconfigurable systems, Proc. of the IEEE Int. Symposium on Circuits and
Systems, pp.2742-2745, 2007.
[6] Xilinx UG002, Virtex-II Platform FPGA User Guide, 2007.
[7] Xilinx DS208, OPB HWICAP (v1.00.b), 2006.
[8] H. Kalte and M. Porrmann, Context saving and restoring for multitasking in reconfigurable systems,
Proc. of Int. Conf. on Field Programmable Logic and Applications, pp.223-228, 2005.
[9] M. Koester, M. Porrmann and H. Kalte, Relocation and defragmentation for heterogeneous re-
configurable system, Proc. of Int. Conf. on Engineering of Reconfigurable Systems and Algorithms,
pp.70-76, 2006.
[10] H. Simmler, L. Levinson and R. Manner, Multitasking on FPGA coprocessors, Proc. of the 10th Int.
Conf. on Field-Programmable Logic and Applications, pp.121-130, 2000.
[11] S. A. Guccione, D. Levi and P. Sundararajan, JBits: A java-based interface for reconfigurable
computing, Proc. of the 2nd Annual Military and Aerospace Applications of Programmable Devices
and Technologies Conf., 1999.
[12] H. Simmler, L. Levinson and R. Manner, Multitasking on FPGA coprocessors, Proc. of the 10th Int.
Workshop on Field Programmable Gate Arrays, pp.121-130, 2000.
[13] T.-Y. Lee, C.-C. Hu, L.-W. Lai, C.-C. Tsai and R.-S. Hsiao, Hardware context switching methodology
for dynamically partially reconfigurable systems, Proc. of the National Computer Symposium, pp.300-
310, 2007.
[14] K. Puttegowda, D. I. Lehn, J. H. Park, P. Athanas and M. Jones, Context switching in a run-time
reconfigurable system, The J. of Supercomputing, vol.26, pp.239-257, 2003.
[15] M. Ullmann, M. Huebner, B. Grimm and J. Becker, An FPGA run-time system for dynamical
on-demand reconfiguration, Proc. of the 18th Int. Parallel and Distributed Processing Symposium,
pp.135-142, 2004.
[16] J.-Y. Mignolet, V. Nollet, P. Coene, D. Verkest, S. Vernalde and R. Lauwereins, Infrastructure for
design and management of relocatable tasks in a heterogeneous reconfigurable system-on-chip, Proc.
of the Design, Automation and Test in Europe Conf. and Exhibition, pp.986-991, 2003.
[17] C. H. Huang and P. A. Hsiung, Software-controlled dynamically swappable hardware design in
partially reconfigurable systems, EURASIP J. on Embedded Systems, vol.2008, pp.231940, 2008.
[18] K. Compton and S. Hauck, Reconfigurable computing: A survey of systems and software, ACM
Computing Surveys, vol.34, pp.171-210, 2002.
[19] Q. Wang, M. Oya, T. Kobayashi and K. Okumura, Adaptive steering controller to improve handling
stability of combined vehicles, International Journal of Innovative Computing, Information and
Control, vol.4, no.12, pp.3273-3283, 2008.
[20] H. C. Cho, M. S. Fadali and K. S. Lee, Intelligent control of mechanical systems with random friction,
2009.
[21] X. Zhao and J. Han, Yaw control of helicopter: An adaptive guaranteed cost control approach,
2009.
[22] W.-J. Shyr, K.-C. Yao, D.-F. Chen and C.-Y. Lu, Web-based remote monitoring and control system
for mechatronics module, ICIC Express Letters, vol.3, no.3(A), pp.367-371, 2009.
[23] Y. Konishi, N. Araki, Y. Iwai and H. Ishigaki, Optimal integrated design for mechanical structure
and controller using bargaining game theory, ICIC Express Letters, vol.3, no.1, pp.41-46, 2009.
[24] T.-Y. Lee, Y.-H. Fan, Y.-M. Cheng and C.-C. Tsai, Hardware-software partitioning for embedded
multiprocessor FPGA systems, International Journal of Innovative Computing, Information and
[25] Y. Birk and E. Fiksman, Dynamic reconfiguration architectures for multi-context FPGAs, Computers
and Electrical Engineering, vol.35, no.6, pp.878-903, 2009.
[26] Xilinx UG012, Virtex-II Pro and Virtex-II Pro X FPGA User Guide, 2007.
[27] H. Kalte and M. Porrmann, REPLICA2Pro: Task relocation by bitstream manipulation in virtex-
II/pro FPGAs, Proc. of the 3rd Conf. on Computing Frontiers, pp.403-412, 2006.
[28] C. Steiger, H. Walder and M. Platzner, Operating systems for reconfigurable embedded platforms:
Online scheduling of real-time tasks, IEEE Trans. on Computers, vol.53, pp.1393-1407, 2004.
[29] M. Hubner, C. Schuck and J. Becker, Elementary block based 2-dimensional dynamic and partial
reconfiguration for virtex-II FPGAs, Proc. of the 20th Int. Parallel and Distributed Processing Sym-
posium, pp.8, 2006.
[30] H. Kalte, G. Lee, M. Porrmann and U. R uckert, REPLICA: A bitstream manipulation filter for mod-
ule relocation in partial reconfigurable systems, Proc. of the 19th IEEE Int. Parallel and Distributed
Processing Symposium, pp.151b, 2005.
[31] Y. E. Krasteva, E. de la Torre, T. Riesgo and D. Joly, Virtex II FPGA bitstream manipulation:
Application to reconfiguration control systems, Proc. of the Int. Conf. on Field Programmable Logic
and Applications, pp.1-4, 2006.
[32] Xilinx XAPP151, Virtex Series Configuration Architecture User Guide, 2004.
[33] Xilinx XAPP138, Virtex FPGA Series Configuration and Readback, 2005.
[34] B. Blodget, C. Bobda, M. Huebne and A. Niyonkuru, Partial and dynamically reconfiguration of
Xilinx virtex-II FPGAs, Proc. of the Int. Conf. on Field-Programmable Logic and its Applications,
pp.801-810, 2004.
[35] S. Jovanovic, C. Tanougast and S. Weber, A hardware preemptive multitasking mechanism based on
scan-path register structure for FPGA-based reconfigurable systems, Proc. of the NASA/ESA Conf.
on Adaptive Hardware and Systems, pp.358-364, 2007.
c ISSN 1349-4198
AN ID-BASED ACCESS CONTROL IN A HIERARCHICAL KEY

MANAGEMENT FOR MOBILE AGENT
Chia-Hui Liu1 , Yu-Fang Chung2 , Tzer-Shyong Chen3 and Sheng-De Wang1

1
Electrical Engineering Department
National Taiwan University
No. 1, Sec. 4, Roosevelt Road, Taipei 10617, Taiwan
{ d96921027; sdwang }@ntu.edu.tw
2
Electrical Engineering Department
3
Information Management Department
Tunghai University
No. 181, Sec. 3, Zhonggang Road, Xitun District, Taichung 407, Taiwan
{ yfchung; arden }@thu.edu.tw
Abstract. The related techniques and applications on e-commerce have been concerned
by a great number of researchers. The use of a mobile agent, in particular, is an important
breakthrough on the e-commerce applications. With the abilities of higher autonomy and
mobility, a mobile agent can move freely among dierent execution environments, can
automatically detect its resided environment, and can react itself accordingly. Besides, a
mobile agent itself can complete the tasks assigned by the users. Because of these charac-
teristics, a mobile agent becomes the most suitable application for e-commerce. However,
it is always a risk to transfer confidential information over an open Internet environ-
ment. When a mobile agent roams itself among the servers over Internet or the mobile
agents exchange information with each other, the users would concern whether the mobile
agent was attacked by some manipulated servers or the carried confidential information
is stolen or tampered by the others. All these worries make the safety of a mobile agent
on the Internet be an important issue. Thus, this paper will propose a suitable and se-
cure scheme for the mobile agent. The scheme, based on the bilinear pairing over elliptic
curves, takes the concept of identity-based access control on a hierarchical key manage-
ment. This paper also aims to increase improvements on the scheme presented by Volker
and Mehrdad to resolve the problem of storage waste in their scheme because of storing
the overlapping decryption keys of a mobile agent. From the results of the security and
performance analysis in this paper, the proposed scheme is proven to protect the mobile
agent in an ecient and secure way.
Keywords: Mobile agent, Access control, Key management, Information security, Bi-
linear pairings
1. Introduction. With the speedy development of the Internet, there is a growing de-
mand on information sharing, information management, improvement of network e-
ciency and secure transmission. Therefore, more complex technologies for distribution
systems are required because bulk information and data are transmitted forward and
backward among numerous servers, which were spread out in an open network. Due to
the great deal of data transferring in such an open distributed system, the problem of
overloading transmission occurs. However, the present information management on facing
toward a large and distributed network framework meets the problems such as depend-
ability, expandability, interactivity and inelasticity. The development of mobile agents is
the answer to these problems.
1443
1444 C.-H. LIU, Y.-F. CHUNG, T.-S. CHEN AND S.-D. WANG
Figure 1. Mobile agent system architecture
A mobile agent is a software program, migrating between servers and performing dis-
tributed computation independently [1,2]. Mobile agents can migrate themselves among
dierence servers and can interact with each other. They also can perform assigned tasks
and return the results to the agent owner autonomously. Therefore, a mobile agent is
autonomous as well as mobile. Figure 1 is a simple illustration of the structure of the
mobile agent system.
The owner of a mobile agent, in Figure 1, makes a request to a mobile agent. The
request could be the assignment of some tasks. According to the assignment, a mobile
agent will migrate itself among servers to deal with the tasks and return the final result
to the owner of the mobile agent.
Mobile agents have various benefits. They reduce network loads and transmission de-
lays, and are applicable in dierent environment platforms. They are also fault-tolerant
and highly adaptable. Thus, mobile agents can be applied in widespread of application do-
mains, including distribution information retrieval, remote processing, workload control,
network management, online bidding, E-commerce, mobile computing, etc. [21].
A mobile agent which is migrating itself from server to server executes the tasks on
a heterogeneous environment; it comes into contact with the other servers or agents for
interchanging information. Thus, it confronts with some security problems. The security
problems occur for a mobile agent generally during its insecure contact. Security threats
on the mobile agent can be classified into the subsequent four types: integrity attacks,
denial of service, confidentiality attacks and authentication risks.
1. Integrity attacks: A malicious server can modify a mobile agent by adding or deleting
the mobile agents information, execution code, status, etc., or the malicious server
AN ID-BASED ACCESS CONTROL 1445
can interrupt the execution of the mobile-agent program to threaten the integrity of
the mobile agents information and execution.
2. Denial of service: The requests issued by a mobile agent are denied such that the
mobile agent is unavailable to its intended servers. For example, a mean server
refuses its data being accessed by a mobile agent; it delays services required by the
mobile agent, or it refuses to send the mobile agent to the next server.
3. Confidentiality attacks: A malevolent server monitors or analyzes the mobile agents
data, execution code or status to obtain its confidential data during the execution
of the mobile agent.
4. Authentication risks: A malicious server masquerades as the next server on-going
visited by the mobile agent to cause the mobile agent executing its codes on its plat-
form so as to clone the mobile agent; thus, the mobile agent cannot be authenticated
by the other servers.
The above security threats would arise due to the characteristics of mobile agents. They
switch themselves among remote servers, exchange confidential information with each
other and send confidential data to the servers asking for processing. Therefore, a suitable
access control protocol is a must to ensure the resources of the mobile agent is reasonably
distributed and the mobile agent can receive eective protection. A number of related
studies have been proposed in recent years. Corradi et al. presented an SOMA-structure
for the mobile agent [3]. Their structure includes mobile agents, a mobile-agent server,
a management system and a security protocol. Later, Karnik and Tripathi proposed an
Ajanta-structure for the mobile agent [4]. Their structure has the same functions as in
SOMA.
Volker and Mehrdad [5] proposed a tree-based key management and access control struc-
ture. Though their structure can protect the mobile agent from malicious attacks when
roaming around the Internet, their scheme repeats the designation of keys to make the
size of the mobile agent increase dramatically and to let the calculations of key derivation
be extensive, which causes the lower eciency of the mobile agent executing the assigned
tasks. Therefore, this study aims to examine Volker and Mehrdas scheme and further to
correct their problems by proposing an ID-based hierarchical key management and access
control structure for the mobile agent with bilinear pairing [14]. In the proposed scheme,
the identities of servers are taken as the parameters to generate public keys. Thus, the
management of public keys is easier, the convenience for users is enhanced and the se-
curity of the mobile agent is improved as well. Based on the hierarchical structure, the
proposed scheme can prevent the problems of the designation of keys repeatedly and the
increasing size of the mobile agent dramatically and further can have the consumed band-
width wasted less and the network overloads reduced more. Thus, the proposed scheme
will be suitable and practical for the more and more popular ubiquitous networks. To
avoid the problem of the extensive calculations of key derivations, we adopt the technique
of the bilinear pairings. Because of the bilinear pairing, the mobile agent is easier to
carry out its assigned tasks on those portable devices such as i-phone or PDA with fewer
storage and lower computational power. Thus, mobile agents can be applied suitably to
the future applications on e-commerce.
2. Related Work. In this section, a brief introduction of the mobile agent system is
made, and the tree-based mobile agent structure of Volker and Mehrdads scheme is
explained.
2.1. Brief introduction of a mobile agent. A mobile agent is a software agent that
is autonomous, social and able to learn itself. Most importantly, it has the ability of
mobility. It can migrate freely among heterogeneous network systems, autonomously

execute assigned tasks, and interact with other mobile agents, achieving the aim of division
of labor and collaboration, bringing further benefits and improving convenience.
Mobile agents have the following abilities: (1) Observation: Mobile agents can be sent to
some remote server to monitor and report the occurrence of a certain event. (2) Searching:
Mobile agents can migrate to a remote server to search for and access information. (3)
Orchestrating: A mobile agent owner can divide a job into several tasks and assign one of
them to another mobile agent so that all the mobile agents collaborate to finish the job
altogether. (4) Migrating: A mobile agent can autonomously migrate itself to a remote
server when it is necessary to complete its assigned task. (5) Meeting: Mobile agents
interact with each other to achieve the target of collaboration [6].
Besides its execution code, the mobile agent also carries information referred during
the execution of tasks. The information consists of the following components.
1. Agent Owner: The owner of the mobile agent, and generally including the owners
digital signature, server name and IP address, etc.
2. Identifier: The identification information of the mobile agent. A mobile agent with-
out an identifier is regarded as an illegal agent by the other agent servers and the
services it requires will be denied.
3. Goal/Result: This component consists of information regarding to the assigned task
of the mobile agent and the possible outcome after the execution.
4. Life Time: A mobile agent is given a life time. The life time of a mobile agent varies
from task to task. The owner of the mobile agent will add a time stamp to the
mobile agent at the onset, so as to inform other agent servers of the life time of the
mobile agent.
5. States: A mobile agent records the execution status before migrating from one agent
platform to the next. The internal state of the agent includes read-write file records,
resource consumption time and the last action taken, etc.
2.2. Advantages of a mobile agent. Mobile agents can simultaneously migrate among
a great number of computers. They can also autonomously distribute computational
process. The mobile agents have the following advantages:
1. Oine operations: In the mobile computational environments, a user will set up a
connection with a remote server. However, the connection can be stopped [24] after
the launch of the mobile agent. In the following, the mobile agent, according to the
assigned task, will migrate among various nodes over the network to complete its
task and to report the final-gotten result to the user after the completion of its task.
Therefore, the user does not have to remain online for the task to be completed.
2. Network trac reduction: In most network communication protocols [22], security
considerations in them focus on the interaction among computers, bringing network
delays and network overloads. So, after a sender sends the mobile agent to the
receiver, the connection between the sender and the receiver does not keep main-
taining, and the mobile agent can autonomously execute tasks at the receiver end.
Thus, network trac can be reduced.
3. Roaming ability in heterogeneous environments: Mobile agents are originally de-
signed for use in heterogeneous environments. Therefore, they are created to be
a kind of software program, which is platform independent. Their operations and
transportations are through the Internet, and their runtime environments are the
only concern. Therefore, mobile agents are able to roam in a heterogeneous environ-
ment.
4. Supports electronic commerce: Since mobile agents can work oine and roam in het-
erogeneous environments, e-commerce system with the help of mobile agents is able
to collect customer habits, to manage customer information, to run distributed data
mining and further to improve e-commerce system eciently and system functions
eectively [7].
5. Ease of development: Since mobile agents can encapsulate communication protocols,
software developers are able to focus entirely on agent applications and not on con-
struction of communication protocols. Thus, the diculty of development is reduced
to focus on system development only.
6. Real time applications: Since mobile agents are able to work oine, they can execute
their tasks on the receiver end and thus reduce network delays. Obviously, mobile
agents are suitable for real time applications.
2.3. Volker and Mehrdads scheme. When the mobile agent is roaming around the
Internet, it is possible to expose the code, information and data that it carries with.
Therefore, to protect them and to fight against some security threats and malicious at-
tacks, Volker and Mehrdad proposed a tree-based mobile agent structure [18] in 1998 (as
shown in Figure 2) where the code and data carried by the mobile agent are arranged in
a tree structure, and the protection and access control of confidential data are achieved
by the process of key management [23]. The entire mobile agent structure is like a tree-
structured file system, and in this structure, each node represents either a folder or a file.
The main nodes of the tree structure are defined below:
1. Agent: Agent is the root node. All data of the mobile agent are stored in this folder.
Under this folder are two sub-folders called Static and Mutable respectively.
2. Static: Contents inside the Static folder are unchanged data during the lifetime of
the mobile agent.
3. Mutable: This folder stores all changed data when the mobile agent executes its
assigned task. When the agent completes its task on a certain server, the Mutable
folder of the mobile agent changes its status and the data stored in this folder may
be changed by the server.
4. Classes: The class code is required by a mobile agent during its execution of tasks.
5. Sctx: The security context node represents all of the security information relating
to the mobile agent. This folder contains access control keys (acl), certificate (cert)
and security policies, etc.
6. Heap: This folder is a place used by the mobile agent to store data during its
execution of tasks primarily.
Figure 2 shows the tree structure for a mobile agent. It is obvious that there are two
main sub-trees, Static and Mutable and some main nodes set up under either of them.
The tree-structured access control and key management scheme proposed by Volker
and Mehrdad can prevent mobile agents from being eavesdropped and malicious attacks.
Remote access control to a mobile agent can be done by any server to achieve the assigned
task securely. A brief description of their scheme is as follows. Before the owner of a mobile
agent sends it onto the Internet for carrying out its assigned task, he/she must designate
a traveling route first and set up access policies at the same time. He/she has to construct
a corresponding sub-folder (as shown in Figure 3: static/sctx/acl/S1 , static/sctx/acl/S2 ,
, static/sctx/acl/S6 ) for each legitimate remote server in the static/sctx/acl folder.
Next, the owner applies some symmetrical encryption-decryption algorithm, like AES, to
encrypt the decryption key of the particular confidential file, and to store the decryption
key in the pertaining folder according to the access policies. Then, the corresponding
public key of the server is used to encrypt its exclusive folder. So, when the mobile agent
Figure 2. Tree structure of the mobile agent
Figure 3. Example of Volker and Mehrdads method
reaches a certain server, the said server can use its private key to decrypt the pertaining
folder and obtain the decryption key which is used to decrypt the confidential files it
has been given access to. Therefore, each server can use its corresponding private key to
decrypt and access the decryptions keys in the corresponding folder only.
Figure 3 illustrates the tree-structured access control and key management scheme. In
the folder Classes, there are four confidential files, A.zip, B.zip, C.zip and D.zip encrypted
by four decryption keys, DK 1 , DK2 , DK3 and DK4 , respectively. In the folder acl, an
exclusive one is constructed for each server on the traveling route, and the folders are
encrypted by the corresponding private keys of the servers; the exclusive folder stores the
decryption keys accessible to the particular server. In Figure 3, S1 , S2 , S3 , S4 , S5 and S6
represents six dierent servers, ntu.ee.edu.tw, ntu.im.edu.tw, thu.ee.edu.tw, thu.im.edu.tw,
ncyu.ee.edu.tw and ncyu.cs.edu.tw, respectively. Therefore, the owner of the mobile
agent constructs six folders, static/sctx/acl/S 1 , static/sctx/acl/S 2 , static/sctx/acl/S 3 ,
static/sctx/acl/S 4 , static/sctx/acl/S 5 and static/sctx/acl/S 6 . Since server S1 is autho-
rized to access four confidential files, A.zip, B.zip, C.zip and D.zip, therefore, the owner
of the mobile agent stores four decryption keys, DK 1 , DK2 , DK3 and DK4 under
static/sctx/acl/S 1 . Server S2 is authorized to access three confidential files, A.zip, B.zip
and C.zip, therefore, the owner stores three decryption keys, DK1 , DK2 and DK3 under
static/sctx/acl/S 2 . Likewise, the owner of the mobile agent stores the corresponding de-
cryption keys of confidential files in the respective folders of S3 , S4 , S5 and S6 . Through
the above tree-structured access control and key management method, confidential files
can be protected.
However, Volker and Mehrdads scheme did not take the eciency of key management
into considerations. Their scheme has the following faults:
1. Increasing the size of a mobile agent: Since each server on the traveling route of
the mobile agent is given access to quite a couple of similar files, the decryption
keys in their scheme are repeatedly stored under dierent folders. Take Figure 3 as
an example, decryption key DK 1 is stored in six folders, belonging to S1 , S2 and
S4 , respectively. Decryption keys DK 2 , DK 3 and DK 4 are also repeatedly stored.
Their scheme obviously wastes a lot of storage space in storing decryption keys,
causing space wastage of a mobile agent and increases the size of the agent; this in
turn causes the mobile agent to consume excessive Internet bandwidth when it is
roaming around the Internet.
2. Excessive computations for public keys: Because decryption keys are repeatedly
stored under the folder static/sctx/acl/, the owner of a mobile agent needs to perform
more public key computations for the mobile agent to protect the security of the
related folders.
For better eciency in transmission of a mobile agent over the Internet, a mobile agent
should be designed to a form of the smaller size and fewer key computations. In the next
section, we will propose an improvement to amend the faults in Volker and Mehrdads
security scheme by using the bilinear method based on the elliptic curve. The improvement
is a hierarchy-based access control and key management scheme for the mobile agent [8-
13]. Along with the proposed scheme, we also explain the concepts cited for the proposed
scheme, its workings and an example.
3. The Proposed Scheme. To improve the drawbacks on the Volker and Mehrdads
scheme, this paper proposes an ID-based access control and key management scheme for
the mobile agent based on bilinear pairings over elliptic curves [14]. We shall explain the
fundamental concepts of bilinear pairing over elliptic curve first before we introduce our
proposed scheme.
3.1. Basic concepts of bilinear pairing on elliptic curves. The usage of bilinear
pairing over elliptic curves assures the high security and eciency of our proposed method.
Bilinear pairings over elliptic curves have been used to construct many Identity-based
Cryptographic Schemes [16]. The two most common bilinear pairings are Weil pairing
and Tate pairing [15-17], where the bilinear map between the two cyclic groups is defined.
The details are as follows.
Let G1 be an additive group of an elliptic curve E, and let its order be a large prime
number q. Let G2 be a multiplicative group with the same order as G1 . A bilinear map
e : G1 G1 G2 between these two groups satisfies the following properties.
1. Bilinear: Let point P , Q, R G1 , thus,
e(P + Q, R) = e(P, R) e(Q, R)
e(P, Q + R) = e(P, Q) e(P, R)
e(aP , bQ) = e(bP , aQ) = e(P, Q)ab ,
where a and b Zq .
2. Non-degeneracy: P G1 exists such that e(P, P ) = 1.
Figure 4. Example of an ID-based hierarchical access control and key

management scheme for mobile agents
3. Computable: If P , Q G1 , then e(P, Q) can be computed within a polynomial time.
3.2. An ID-based key management and access control scheme. This paper pro-
poses an ID-based hierarchical access control and key management for mobile agents (as
in Figure 4).
First, the owner of the mobile agent builds a hierarchical structure for controlling
accesses according to access control policies. In the hierarchical structure, the leaf node
represents the decryption key used to decrypt confidential files, while the internal node
stands for the folders corresponding to the server. The relationship Sj Si means that if
server Si has access rights to some confidential data, Si can open the corresponding folder
succeeding the leaf node of internal nodes that correspond to the server. Based on this
description, it charts a hierarchical access policy.
Before the owner of the mobile agent assigns a mobile agent to execute tasks on the
Internet, he/she sets up the following job first, deciding the target servers, building a
predefined travel route for the agent, defining the data which visited servers can access,
and determining the access policy of the agent. Also, the owner has to assign a secret key
to the internal node corresponding to each server, and to announce some of the public
parameters of the mobile agent. In the proposed scheme, there are two major phases,
namely the key generation phase and the key derivation phase, described in the following.
1. Key Generation Phase
The owner of the mobile agent executes the following algorithm step by step.
Step 1: Select non-repeated integers DK t for each confidential file randomly, and
choose a bilinear map e : G1 G1 G2 , along with a public parameter
P0 G1 .
Step 2: Generate two one-way hash functions:
H1 : {0, 1} G1 H2 : G2 {0, 1}
Step 3: Select a secret key Zq randomly and calculate a public parameter Ppub =
P0 .
Step 4: For all servers Si :
To calculate a public key Qsi = H1 (IDsi )

To calculate a private key Dsi = (Qsi )
End for
Step 5: Choose a random value r and calculate U = r P0 .

Step 6: For all decryption key (DK t ).
To construct the public function FDKt (x) given below:

FDKt (x) = DKt [x H2 (gSr j )] , where gSj = e(QSj , Ppub ).

Sj Si
To announce (U, FDKt (x))

End for
2. Key Derivation Phase
Step 1: When a mobile agent arrives at server Si , Si can derive the decryption key
DK t for accessing the specific file.
Step 2: Si takes the corresponding private key DSi and the public parameter U as
input to the function FDKt (x) for obtaining DK t , where x = H2 (
e(DSi , U )).
Example 3.1. As shown in Figure 4, a mobile agent has four decryption keys, denoted
as {DK1 , DK2 , DK3 , DK4 }. The owner of the mobile agent will construct the relating
four public functions: FDK1 (x), FDK2 (x), FDK3 (x) and FDK4 (x) as follows:
FDK1 (x) = DK1 [x H2 (gS1

r
)] [x H2 (gS2
r
)] [x H2 (gS4
r
)]
r
)] [x H2 (gS2
r
)] [x H2 (gS3
r
)]
[x H2 (gS4
r
)] [x H2 (gS5
r
)]
r
)] [x H2 (gS2
r
)] [x H2 (gS3
r
)]
[x H2 (gS5
r
)] [x H2 (gS6
r
)]
r
)] [x H2 (gS3
r
)] [x H2 (gS6
r
)]
If server S2 wants to obtain decryption key DK1 to decrypt the file, B.zip, it can take
the corresponding private key DS2 , the public parameter U = rP0 and x = H2 (3(DS2 , U ))
as input to the equations below to obtain DK1 .

r
)] [x H2 (gS2
r
)] [x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] [H2 (
e(DS2 , U )) H2 (
e(QS2 , Ppub )r )] [x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] [H2 (
e(DS2 , r P0 )) H2 (
e(QS2 , Ppub )r )]
[x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] [H2 (
e( QS2 , r P0 )) H2 (
e(QS2 , Ppub )r )]
[x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] [H2 (
e(QS2 , P0 )r ) H2 (
e(QS2 , Ppub )r )]
[x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] [H2 (
e(QS2 , P0 )r ) H2 (
e(QS2 , Ppub )r )]
[x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] [H2 (
e(QS2 , Ppub )r ) H2 (
e(QS2 , Ppub )r )]
[x H2 (gS4
r
)]
= DK1 [x H2 (gS1
r
)] 0 [x H2 (gS4
r
)]
= DK1
4. Security and Performance Analysis. The proposed scheme will be discussed and
analyzed in this section according to the possible attacks it could face. Through the
security analysis, we will prove that the proposed scheme is feasible as well as secure in
implementations.
4.1. Reverse attacks. When the relationship Sj Si exists between server Sj and
server Si , Sj cannot use its corresponding secret parameter DSj along with the public
parameter U to derive the secret parameter DSi of Si . Thus, the attacker, Sj , cannot
launch a reverse attack on server Si eectively.
The related parameters in the proposed scheme are calculated through one-way hash
functions and bilinear pairings over elliptic curves. Since the one-way hash function is
irreversible, an attacker should solve the Elliptic Curve Discrete Logarithm Problem first
for obtaining the exact elliptic curve base point. Therefore, the attacker, Sj , cannot derive
the secret parameter DSi of server Si and neither can it destroy or access data that is
accessible by server Si . Hence, due to the lack of information, it is extremely dicult for
the attacker to obtain secret parameters through a reverse attack on server Si .
4.2. Cooperative attacks. Another is the cooperative attack. Suppose server Si can
use its secret parameter DSi to access certain private files. By collecting all related public
system parameters, some unauthorized servers attempt to derive DSi through cooperative
attacks. For example, suppose server S4 and server S5 in Figure 4 try to derive the secret
parameter D1 of server S1 through cooperative attacks.
However, with the function FDKt (x) as the guardian, the proposed scheme can avoid this
kind of attack. Obviously, even if several attackers cooperate to collect all the relating
public parameters and all of their own secret parameters, they still cannot derive the
secret parameter DSi of server Si via the function FDKt (x). This is because within the
function, FDKt (x) included the public parameter QSi of the relevant server, but not the
secret parameter DSi . Therefore, cooperative attacks will not work.
4.3. External collective attacks. As to the external collective attack, an intruder

wants to obtain a desired file by attacking the mobile agent with all the public parameters
he has. In our proposed scheme, the intruder has to reverse the given polynomials to their
original form. However, it is the one way hash function, irreversible and distinct outputs
when given any two inputs. That means it is impossible to derive the same outputs. In
addition, the elliptic curve encryption technology based on the bilinear map e makes it
extremely dicult for the intruder to derive the desired secret parameter. Therefore, it
is impossible for the intruder to succeed within the validity period.
4.4. Equation hacking attacks. Finally, it is the equation hacking attack. Let servers
Si , Sj and Sk be able to access the decryption key DK t . Can server Si hack the polynomial
of FDkl (x) to obtain the secret parameters of servers Sj and Sk ?
Take Figure 4 as an example. Suppose servers S5 and S6 can legally obtain DK3 by
the function FDk3 (x). However, S5 wants to hack the equation of FDk3 (x) to determine
the secret parameter of S6 . So, S5 will convert FDk3 (x) to the following form:
FDK3 (x) = DK3 [xH2 (gS1
r r
)][xH2 (gS2 r
)][xH2 (gS3 r
)][xH2 (gS5 r
)][xH2 (gS6 )]
r r r r r
FDK3 (x)DK3 = [xH2 (gS1 )][xH2 (gS2 )][xH2 (gS3 )][xH2 (gS5 )][xH2 (gS6 )]
FDK3 (x) DK3
[x H2 (gS5
r
)] [x H2 (gS6
r
)] =
[x H2 (gS1 )] [x H2 (gS2
r r
)] [x H2 (gS3r
)]
Server S5 uses its own secret parameter to substitute x = H2 ( e(DS5 , U )) for the above
equation, where U = r P0 :
r r FDK3 (x) DK3
[H2 (
e(DS5 , U ))H2 (gS5 )][xH2 (gS6 )] =
[x H2 (gS1 )] [x H2 (gS2
r r
)] [x H2 (gS3
r
)]
DK3 DK3
0 [x H2 (gS6
r
)] =
[x H2 (gS1 )] [x H2 (gS2
r r
)] [x H2 (gS3r
)]
According to the above calculations, server S5 cannot determine the secret parameter
of server S6 by FDk3 (x) even if S5 has its own secret parameter and decryption key DK3 .
Therefore, our proposed scheme can eectively guard against this type of attack.
4.5. Performance analysis. In the following, we would like to compare our proposed
method with Mehrdads scheme [5] and Chang and Lins scheme [18] in terms of the
consumed storage and computational cost.
First, let us suppose a mobile agent will visit server v and carry m number of confidential
data {F1 , F2 , . . . , Fm }. The number of confidential data that each visited server can access
is {W1 , W2 , . . . , Wv }; that is to say, the number of decryption keys server Si can access is
Wi . Assume the length of the decryption keys is equal or less than L. Hence, the total
v
number of DK stored by a mobile agent is Wi in Volker and Mehrdads scheme, which
i=1

v
means that the largest storage space required in their scheme is L Wi . In Chang
i=1
and Lins scheme, the total number of DK will be m number of decryption keys and v
number of private key, making the largest storage space required to be L (m + v). In
comparison, our proposed method would require only m number of decryption keys and
v number of private keys. Assume the longest length of the private keys to be L . The
largest storage space required in our scheme will be L (m + v). If length L is 1024bits,
to achieve the similar level of security, L requires only 160bits. Compared with Chang
and Lins scheme, our scheme requires much less storage space.
In computational cost, the folder in each server must keep confidential to prevent illegal
access. In Volker and Mehrdads scheme, each folder is encrypted by the accessed servers
public key, and decrypted by the accessed servers private key. Hence, their scheme

v
requires 2 Wi exponential computations. In Chang and Lins scheme, the owner of
i=1
the mobile agent requires individually v number of modular exponential computations
to generate the secret key, v number of modular exponential computations to obtain the
DK, v number of modular exponential computations to encrypt DK, and also v number of
v
modular exponential computations to decrypt DK. Since their scheme will require Wi
i=1
number of modular exponential computations to compute DK, it will need a total of

v
4v + Wi modular exponential computations. In addition, Wi is the number of files that
i=1

v
server Si can access; therefore, Wi will be much greater than v as shown in Table 1.
i=1
Also, it is worth mentioning that modular exponential computations being the primary
cost in both of their schemes are much more complicated than the multiplication and
addition computations, listed in Table 1.
In our method, the owner of the mobile agent requires 4v number of pairing computation
to generate the secret key, to obtain, to encrypt and to decrypt the decryption key. It

v
will also require Wi number of pairing computation to compute the decryption key,
i=1

v
requiring a total number of 4v + Wi pairing computations.
i=1
From the above analysis, the decryption keys in Volker and Mehrdads scheme are
repeatedly stored in the tree-structured mobile agent; therefore, when larger number of
confidential files required, equally larger storage space is needed to increase the size of the
mobile agent dramatically. This causes the increases of the communication cost because
the wider bandwidth is required when the mobile agent is roaming around the Internet.
To improve these faults, Chang and Lin decreased the size of the mobile agent by using
RSA. However, their scheme will increase the computational cost because the modular
exponential computations are required to compute the secret key in their scheme. In
comparison, our proposed scheme uses bilinear pairings over the elliptic curves, and has
the benefits that the required calculated pairing for computing private key Dsi is only
about 1/3 [19,20] of modular exponential computations.
Table 1 shows the comparisons of our proposed scheme with the other two, Volker
and Mehrdads scheme and Chang and Lins scheme, in terms of the total storage space
required for the decryption keys, the modular exponential computations required and
the pairing computations required. From this table, the storage space required in our
proposed scheme is the least, while the requirement in Volker and Mehrdads scheme
is the most. As to the computational cost, we require the lower pairing computations
instead of the modular exponential computations used in both of the other two schemes.
As mentioned above, the calculated pairing computation for private keys is only about
1/3 of the modular exponential computation. Therefore, the proposed scheme is obviously
more significant than the other two in the storage space required and the computational
cost.
Due to the advanced technology in the wireless networking, the mobility of e-commerce
is required more eectively and more eciently than before. The concept of mobile agents
shows the suitability to the new application environments. For example, a user carrying
a mobile device would like to purchase some product via the network. With the help of
a mobile agent, he can set up the item he prefers, his budget, and brands he likes to the
agent. Then searching on the authorized servers, the mobile agent matches the possible
Table 1. The comparisons of three schemes
Total storage space of Required modular Required pairing

Schemes
the decryption keys exponential computations computations
Volker and
v v
L Wi 2 Wi
Mehrdads scheme i=1 i=1
Chang and v
L (m + v) 4v + Wi
Lins scheme i=1

v
Our scheme L (m + v) 4v + Wi
i=1
v: the number of servers visited by the mobile agent; m: the number of confidential files; Wi : the
number of files that the visited servers are allowed to access.
items that the users want and get more detailed information about the wholesaler back
to the user. Since the e-commerce world is restricted in a limited broad bandwidth and
the mobile devices carrying by the users are generally of fewer storage and lower compu-
tational power, the hierarchical structure proposed in this paper shows the reliability and
availability of the mobile agents to give a hand to the users.
5. Conclusions. This paper proposed a hierarchical-structured key management and
access control scheme for mobile agents. The proposed scheme was based on bilinear
pairings over elliptic curves, and made use of the one way hash functions.
The advantages of the scheme are shown in the following. A hierarchy structure makes
it easier to construct the mobile agent. Compared with Volker and Mehrdads scheme, the
storage space for duplicated keys is eectively reduced. The derivation of decryption keys
is relatively simple. Eective protection against malicious attacks can be done. There-
fore, our proposed scheme can eectively provide mobile agents with a secure runtime
environment.
The hierarchical structure for the mobile agent proposed in this paper can have the
consumed bandwidth wasted less and the network overloads reduced more to show the
abilities of reducing the storage required for the decryption keys and the extensive cal-
culations of key derivations. However, if the mobile agent should carry a great number
of confidential files in the future, it will be unavoidable to increase the size of the mobile
agent though the number of duplicated keys can be under control. Fortunately, a way
to classify the confidential files under the hierarchical structure could shorten the size of
the mobile agent. Otherwise, using the concept of the group-key management is another
solution to deal with the problem of decryption-key increase when the mobile agent has
to carry plenty of confidential files.
Acknowledgement. This work was supported partially by National Science Council of
Taiwan under Grants NSC 99-2221-E-029-023.
REFERENCES
[1] A. Karmouch, Mobile software agents for telecommunications, IEEE Communications Magazine,
vol.36, no.7, pp.24-25, 1998.
[2] D. Chess, B. Grosof, C. Harrison, D. Levine, C. Parris and G. Tsudik, Itinerant agents for mobile
computing, IEEE Personal Communications, vol.2, no.5, pp.34-49, 1995.
[3] A. Corradi, R. Montanari and C. Stefanelli, Security issues in mobile agent technology, Proc. of the
7th IEEE Workshop on Future Trends of Distributed Computing Systems, Cape Town, South Africa,
pp.3-8, 1999.
[4] N. M. Karnik and A. R. Tripathi, A security architecture for mobile agents in ajanta, Proc. of the
the 20th International Conference on Distributed Computing Systems, Taipei, pp.402-409, 2000.
[5] R. Volker and J. S. Mehrdad, Access control and key management for mobile agents, Computers and
Graphics, vol.22, no.4, pp.457-461, 1998.
[6] N. Borselius, Mobile agent security, Electronics and Communication Engineering Journal, vol.14,
no.5, pp.211-218, 2002.
[7] P. Maes, R. H. Guttman and A. G. Moukas, Agents that buy and sell, Communications of the ACM,
vol.42, no.3, pp.81-91, 1999.
[8] M. S. Hwang, Extension of CHW cryptographic key assignment scheme in a hierarchy, IEE Proc. of
Computers and Digital Techniques, vol.146, no.4, pp.219, 1999.
[9] S. G. Akl and P. D. Taylor, Cryptographic solution to a problem of access control in a hierarchy,
ACM Trans. on Computer Systems, vol.1, no.3, pp.239-248, 1983.
[10] M. S. Hwang, An asymmetric cryptographic scheme for a totally-ordered hierarchy, International
Journal of Computer Mathematics, vol.73, pp.463-468, 2000.
[11] M. S. Hwang, An improvement of novel cryptographic key assignment scheme for dynamic access
control in a hierarchy, IEICE Trans. on Fundamentals of Electronics, Communications and Computer
Science, vol.E82-A, no.3, pp.548-550, 1999.
[12] M. S. Hwang, An improvement of a dynamic cryptographic key assignment schemes in a tree hier-
archy, Computers and Mathematics with Applications, vol.37, no.3, pp.19-22, 1999.
[13] M. S. Hwang, A dynamic key generation scheme for access control in a hierarchy, Nordic Journal of
Computing, vol.6, no.4, pp.363-371, 1999.
[14] Y. J. Choie, E. K. Jeong and E. J. Lee, Ecient identity-based authenticated key agreement protocol
from pairings, Applied Mathematics and Computation, vol.162, no.1, pp.179-188, 2005.
[15] C. H. Lin, K. J. Huang and H. H. Lin, Improving Shims tripartite authenticated key agreement
protocol based on Weil pairing, Proc. of the 14th Information Security Conference, pp.250-255,
2004.
[16] D. Boneh and M. Franklin, Identity-based encryption from the Weil pairing, CRYPTO 2001, LNCS,
vol.2139, pp.213-229, 2001.
[17] N. P. Smart, Identity-based authenticated key agreement protocol based on Weil pairing, Electron.
Lett., vol.38, no.13, pp.630-632, 2002.
[18] C. C. Chang and I. C. Lin, A new solution for assigning cryptographic keys to control access in mobile
agent environments, Wireless Communications and Mobile Computing, vol.6, no.1, pp.137-146, 2006.
[19] P. Barreto, H. Y. Kim, B. Lynn and M. Scott, Ecient algorithms for pairing-based cryptosystems,
CRYPTO 2002, LNCS, vol.2442, pp.354-368, 2002.
[20] S. Lee, Y. Kim, K. Kim and D. H. Ryu, An ecient tree-based group key agreement using bilinear
map, Applied Cryptography and Network Security, vol.2846, pp.357-371, 2003.
[21] Y. Huo, W. Li and D. Yang, On dynamic mechanisms of mobile services adoption in social network,
[22] K. Matsumoto, A. Utani and H. Yamamoto, Adaptive and ecient routing algorithm for mobile
ad-hoc sensor networks, ICIC Express Letters, vol.3, no.3(B), pp.825-832, 2009.
[23] M.-S. Hwang, C.-C. Lee, S.-K. Chong and J.-W. Lo, A key management for wireless communications,
2008.
[24] C.-H. Hsu, Downlink MIMO-SDMA optimization of smart antennas by phase-amplitude perturba-

!" # $%!&
' ' (") * + #! ,
+- .''+',
/"" +

)" $%!&
' 0, * + , " 1 ,
-!''+',
" &!!

" #!" $%!&
' 2 ( 3" * + " ,
! +" 4 &3-'+',

" " $%!&
' 5 $%!& * + !" ,! & & 5 ,

5-!''+',

!"!#$

% ! & '
'

" +" !
& "& !
! !"
!
! "
# $%& !
!
$'!&! $(! !!)&!
* "
+ ! , - .
/ $0& 1234 !
!
"

5 23
- ! " 12//4 $0&
! " 2// ! 6#
7
!
+ ! #

2// ! -
#
- ! -
*
# ,
#!
5
* - !
2// $0!%!(& ,!
6 !

/ 2//
* -
23 8 $'& , -
6-
8 $%!(&
-
,
# 3 $&

#6 2// 6* -
39 $&!
#6
6
#
#
,6*
# : 9
$)&! 6 , #
6 + ! ; *
-
* -

,
5

! 3 $& 6
1<=2//4 5
#
6*
# * - 5
,! ,* ,* !
<=2//
* -
2// !
5 $& 2//
+ !
* - !

5
! 1! 4 6

* - 1</<=4
5 </<= ;
* - !

</<= ! !
5 "

5 </<=
5 6
5
#
-
#$
$# 7 6 6
! -! </<= !
1! 4 2// ! 6
! # # # ! ! #
# !
5 </<= 5

5
/ !

/#
3 - !

! 7

# ! > !

5 ! !

=
6
7 6

#
#
- -

#! # # 5 </<=

5 </<= ,* $& 6

# ,*
,!
# -
- =

- 6 !
- 6
#
,* ,
% ! & '"

5
6
6 ()* ! 6

#!

# #
#

#
# -

#

#
# - 5 </<=
!

# - 5

#

#!

#
#!

#
# 5

# -

#

# -

#

5 </<=

#
& ' 5

#

#
&
' & & ( 5!

#

& ' & & ( 5 "

,! </<=
#!

#!
#
5 !
# ! </<=

# !
# !

# +! ! ! !
! ! ! = -

!

5

? ! '
( ?

$$ +, $-./
7 6
# !
@ @ -

# -! %

3 -

#

-

$$ +,' $-./
7 6
# !
&

!

! %

!

3 -

!
#
-

$$ +,( $-./
3 & '
-

#

7 6
# !
& &'& !

%

) 3 - & & '&

# -
/
$$ +, $-.

#
& & '

7 # !

/
#-0 +,' $-.

# &
5 !

7 !

'

/
#-0 +,( $-.
#
!

A # !

'

!
#

# / !
! !

# !
" 7 1 4!

# 1 4 !
" 5 !

-

# ! ;
# !
#

5# 1 1 )'(! 1 1 2! 1 2! 1 2 - 5

! "

! "

) #$ 34$5$-+ $6+6 5

</<= B
2// $'&! <=2//$&
59 $& 5 7*8 )'( - 14 C 14!

7*8 )'( - 14 C 14!

289 ()* - 14
5 " - * ;

# * ,
-! ;
! 1 1 )

</<= 1 2 2! 9 9
) ) 5 - -
6
5 ;

+ !
! -
!

/
#
# ;
#

5

) !
; -, -
! ;

5 - - D ,
1 2 '
- - !
!
, ,

! ,
,
"

5 !

</<= ;6

"
,
,!
;
5 - </<=
B
2// $'&! <=2// $& 59 $& %

B
2//
5 6#! , 6
(

<=2// 6,*
1 2 2! 9 9 ) )! 5 <=2// ;
</<= + !
6# 0 6

59 *
1 ( ( * 5 6

5 *
- 59 !

/*
- ! B
2//
! <=2//
, ! 6
# 59
6 + !
- * 5 </<= ; ,
* - < ! </<=

#
5

</<= ;
! E
9

' +66 5 1! 4

* - 1</<=4 5 </<=
! *
- ! 5 -
# " !
# # !
</<= ! E ,
, #!
- 1! 4 </<= ! " </<=

5 !

1! 4 B
2//
5

6

1! 4 <=2//

67 ' 8' " ( ! !!

$%!&
, '
67 8' )' 8' (' ' ' #" ' #' ! + 8' 9' :

%
%!
!
!" ! , " !. ;! " %
'11 ' '551
1'
67 ' ( !""

! !
!!"
"! & <%

&!! ) *** +,,- ) :
=: '2
25 5'

5 6

59
67 '' > 8'#' + '' " >?!+ "& + ++" ! +"
"!
%
'2 '@?A
' '
627 ' ' > 8' ( + ' 8 : " ,3" ! , + + %&

%
' '5@:A '1
1 1'
657 #' > )'B' 8 '' + C'' > #+"

,3! !! %!
& "&

%
'2 '1
'12 '
617 ' + :' !
& "& . / *012 #3"
%
'2 ' 2'
67 ' D' & ' 8' #" 8' (' > *' C' 0" + (' "

!! %!
& "& ) 4 %

' ' '552 1'
67 0'>' + '' " +" !+ E !!+ "! !" " +! +
+
!"
%
'2 '1
' 2 '
6 7 ' ' + D' ' > " !" 5 %
'5 '2 '15211
'
67 ' ' ! ' + =' # " F "" "
!! %!
! !" !
"
%
'25 ' '22 '
67 ' 0" >' C" ' + B' > , ! !" !! !+ ?
!
) 4 %
' ' '11512 1'
67 *' C' 0" + ' #' " !" , !
!+ , "! ) 4

# %
'1 '5 '22222 5'
67 ' ' 8" , %!
! !" !! !"
! + ) 4
# %
'2 ' ' '
627 ' ' 8" + ' ' :! % %!
! !" !! , ;
;! ) 4 # %

'5 ' ' 5 2'
657 ' ' 8" + ' ' *+ !+ , !. ! % %!
! !"
!! !" !<
3,! ) 4 %
' '1 ' 5'
c ISSN 1349-4198
SERVICE-ORIENTED ROUTING AND CLUSTERING STRATEGIES

FOR VEHICLE INFOTAINMENT DISSEMINATION
Chenn-Jung Huang1, , Chin-Fa Lin1 , Ching-Yu Li2 , Che-Yu Lee2

Heng-Ming Chen2 , Hung-Yen Shen1 , You-Jia Chen3 and I-Fan Chen4
1
2
Institute of Electronics Engineering
National Dong Hwa University
No. 1, Sec. 2, Da Hsueh Rd., Shoufeng, Hualien 97401, Taiwan

Corresponding author: cjhuang@mail.ndhu.edu.tw
3
Institute of Computer Science and Information Engineering
National Taiwan University
No. 1, Sec. 4, Roosevelt Road, Taipei 10617, Taiwan
4
National Tsing Hua University
No. 101, Sec. 2, Kuang-Fu Road, Hsinchu 30013, Taiwan
Abstract. The establishment of vehicle infotainment systems in heterogeneous vehic-

ular networks needs to pay special attention to the distinctive network characteristic in
wired/wireless networks, such as dynamic topology and less available bandwidth and the
exploitation of the underutilized peers resources to support large-scale services. This work
proposes a service-oriented vehicle infotainment dissemination management scheme to
ensure the satisfaction of varied requests service requirements in heterogeneous vehicular
communication. A novel group mobility-based clustering algorithm and a service-oriented
routing mechanism that fits the distinct characteristics of heterogeneous vehicular net-
works are proposed to enhance the quality of communication. A series of experiments
was conducted and the experimental results exhibit the eectiveness and practicability of
the proposed work.
Keywords: Vehicular infotainment, Clustering, Group mobility, Heterogeneous vehicu-
lar network, Service-oriented routing
1. Introduction. Recent advances in wireless inter-vehicle communication systems have

enabled the establishment of vehicular ad-hoc networks (VANET) and have created sig-
nificant opportunities for the deployment of a wide variety of applications and services
for vehicles. The large-scale media streaming is either collected or disseminated from two
ways, that is, vehicle to road base station (V2R) or vehicle to vehicle (V2V). Routing
algorithms have been a challenge issue in a VANET due to the quick change of network
topology. There are several representative routing protocols for VANET in the litera-
ture. The algorithm proposed in [1] improved the packet delivery rate by grouping the
vehicles based on the distance between neighboring vehicles. However, merely taking the
distance between vehicles into consideration might be a mistaken judgment. For instance,
two vehicles will be assigned into one group while the two vehicles move in the opposite
direction on the road, and get close to each other within a short period of time. The
regrouping inevitably occurs frequently when two vehicles move away from each other.
Vehicles are grouped according to their moving directions in [2]. The authors claimed
that this kind of grouping ensures that vehicles in the same group are more likely to
1467
1468 C.-J. HUANG, C.-F. LIN, C.-Y. LI ET AL.
establish reliable transmission paths in VANETs. Nevertheless, frequent regrouping still

occurs even though the members of a group move at the same direction owing to the
speed discrepancy.
Recently, the service oriented architecture (SOA) [3-5] has been widely adopted to dis-
seminate infotainment data in the heterogeneous wireless networks in the literature. The
requirement of agile adaptation to varying resource constraints in mobile systems has mo-
tivated the use of SOA. The elements of a SOA include service provider, service registry,
service broker and service customer. A service provider publishes a service description at
a service registry. The service description is a machine-readable document that provides
information for the accessing interfaces and functions of the service. A service customer
first looks up the dynamic service location by querying the service registry for a service
that matches its requirements. If such a service exists, the registry provides the customer
with the location and interface information for that service. The service customer then
invokes the service by following the conventional request/reply mechanism. A service
broker is adopted to maintain the service registry and process service discovery requests
for service customers. It was reported that the SOA possesses several essential benefits,
such as better reuse of existing resources, the ability to adapt to changing requirements in
an agile manner, significant potential cost savings and independence from the front-end
applications that invoke the service. We thus employ the SOA to alleviate the unsta-
ble transmission situation caused by the selection of inappropriate paths in the HVN
environment and achieve the acceptable and predictable performance.
It was reported in the literature [6] that hierarchical clustering architecture can ef-
fectively improve scalability of the routing mechanism for the wireless ad hoc network.
Clustering divides a network into several interconnected clusters, and each cluster has a
cluster head to coordinate with the member nodes. A cluster head not only records the
topology information, but also can act as a temporary base station and communicate with
other cluster heads. Nevertheless, the assumption of mobile nodes with low-mobility in
most clustering methods fails to fit the distinct characteristic of VANET, in which the
vehicles might move at high-speed.
There are several major contributions in this research. Firstly, a service-oriented vehicle
infotainment dissemination management scheme for a HVN is presented in this work to
fully adapt to the high volatility of the vehicular networks and ensure the reliable dissemi-
nation of infotainment data. A group mobility-based clustering algorithm for a VANET is
presented in this work to enhance the quality of communication and service management
in the HVN. Secondly, the deployment of V2R or V2V is used in the traditional vehicular
networks environment. This work integrates wireless cellular networks with traditional
vehicular networks and a network selection mechanism is adopted to establish a reliable
service delivery so that the guaranteed service quality is maintained during dissemination
of the vehicle infotainment. Thirdly, cache relay nodes are deployed and intelligently se-
lected by a fuzzy logic inference system to forward the data streams in a HVN based on
the users hardware capability and network resource.
The cluster structure is determined by the geographic position of nodes and the pri-
orities associated with the vehicle trac information in the proposed algorithm. Each
cluster elects one node as its cluster head. The cluster size is controlled by a predefined
maximum distance between a cluster head and its members. Clusters are independently
controlled and dynamically reconfigured as nodes move. The establishment of clusters
can achieve steady service quality of peer-to-peer (P2P) transmission in the routing path.
The velocity and acceleration of vehicles are taken as the parameters to form the vehicular
routing clusters in our work, which can improve scalability of VANET with high-mobility
characteristics and achieve eectiveness in region with either dense node or scattered node
SERVICE-ORIENTED ROUTING AND CLUSTERING STRATEGIES 1469
distributions. In addition, compared with two representative clustering techniques in the

literature, position-based prioritized clustering (PPC) [7] and Lowest-ID [8-10], the result
demonstrates that the proposed algorithm in this work outperforms them in several di-
mensions such as the number of re-clustering, average lifetime of the cluster head, cluster
members, average delay, packet loss rate and throughput.
The remainder of this paper is organized as follows. Section 2 presents the service-
oriented vehicle infotainment dissemination management scheme for the HVN. The sim-
ulation results and analysis are given in Section 3. The conclusion is in Section 4.
2. Problem Statement and Preliminaries. Figure 1 shows the architecture of the

proposed service-oriented vehicle infotainment dissemination management scheme. Adopt-
ed from the concept of service-oriented architecture, the vehicles which ask for service, are
regarded as the service requestors; while the data content servers or the vehicles that act
as the cache relay nodes for forwarding the data stream are regarded as the cache relay
nodes; and the ones that search for services and release the provider lists to the clients as
a service broker.
Figure 1. Architecture of service-oriented routing and clustering mechanisms
2.1. Group mobility-based clustering algorithm. The mobility group-based clus-

tering module is used to group vehicles in VANET. A cluster head communicates with a
base station via the Vehicle-to-Roadside (V2R) protocol, and liaises with and its member
vehicles or other cluster heads via IEEE 802.11p vehicle-to-vehicle (V2V) protocol. By
using the hierarchical communication approach, clustering can eectively alleviate the
redundancy problems caused by wireless peer-to-peer (P2P) transmission, make services
much easier to be integrated and reduce the consumption of network resources.
It was reported that vehicles tend to move in procession most of the time, and the
drivers behavior significantly influences the moving pattern of vehicles such that the
formation of vehicle procession is indirectly aected. We thus attempt to detect the
formation of vehicle procession based on the drivers behavior and other parameters, such
as direction and speed of the vehicles. A measurement of so-called grouping indicator (GI)
is used to represent the possibility of grouping for vehicles that are located within each
others transmission range. The GI is helpful in the determination of the establishment
of clusters. The maintenance of clustering topology is also an important issue because
the steady service quality of the networks can be continuously provided if the formation
of the clusters is kept intact.
2.1.1. Grouping indicator. Suppose that there are k vehicles within the transmission range
of a vehicle x. The grouping indicator, GI, for the vehicle x can be expressed by,
k ( )( )
|V V | |Ax Aj |
1 xM j 1 Max(A x ,Aj )
j=1
GIx = (1)
k
where the index j = 1, 2, , k. M represents the upper limit of the vehicle speed.
Vx and Vj are the average velocity of vehicles x and j, and Ax and Aj represent the
acceleration of the two vehicles, respectively. It can be observed from Equation (1) that
GI is comparatively smaller when the gaps between the velocity and the acceleration of
the vehicle x and its neighboring vehicles are larger, and vice versa.
2.1.2. Establishment of clusters. The establishment of each cluster can be summarized as

below.
Step 1: Each vehicle is treated as an isolated node at the initial state.
Step 2: Each isolated node first broadcasts a message with its radius being equal to half
of the transmission range to see if there is any opportunity for joining in a cluster.
After each neighboring vehicle that moves in the same direction with the isolated
node receives the broadcast message, the GI value for each neighboring vehicle is
then computed by the neighboring vehicle and replied back to the isolated node.
Step 3: The GI value for each neighboring vehicle is compared to that of the isolated
node.
Step 4: The newly elected cluster head will broadcast to all neighboring vehicles moving
in the same direction and invite them to join in its cluster.
Notably, the cluster which is constructed by the above algorithm has a feature, whereby
the cluster head is not easily changed, and the computation overhead for frequent change
of cluster heads as observed in the literature is thus avoided.
2.1.3. Maintenance of clusters. There are three cases considered in the maintenance mech-
anism.
Case 1: When the vehicle is a cluster member and it separates from other members of the
cluster it belongs to owing to the speed discrepancy and will attempt to reconnect
to a nearby cluster moving in the same direction.
Case 2: When the vehicle itself is a cluster head, the distance to the heads of its neigh-
boring clusters is kept under surveillance. Two clusters will be merged if their
cluster heads are getting close and have similar speeds.
Case 3: When the node is an isolated node and it moves close to a nearby cluster that is
moving in the same direction, it will attempt to reconnect to the cluster.
Case 4: When two or more isolated nodes are close to each other and have similar speeds,
a new cluster will be formed and the new head will be elected from these isolated
nodes depending on which one has the largest GI value.
Notably, the above-mentioned cluster maintenance mechanism avoids several needless

merging situations that might occur. Take Figure 2 as an example. The two clusters
marked in blue and yellow are moving in the opposite direction to each other. As shown in
Figure 2(a), the two clusters have to be split even they were merged into one, when the two
clusters meet each other. The proposed clustering algorithm takes the moving direction
of the clusters into consideration and thus avoids unnecessary merging as illustrated in
Figure 2(b).
(a) (b)
Figure 2. (a) Merging without considering the moving direction of two

clusters; (b) Avoidance of unnecessary merging after considering the moving
direction of two clusters
The second example that illustrates how the clustering algorithm considers the moving
speeds of two clusters to avoid needless merging is given in Figure 3. The two clusters
are merged without considering the moving speeds of the two clusters and are then split
very shortly after they are merged, as given in Figure 3(a). On the contrary, the proposed
clustering algorithm takes the gap between the moving speeds of the two clusters into
consideration and thus keeps the two clusters separated even though they are close to
each other as illustrated in Figure 3(b).
(a) (b)
Figure 3. (a) Merging without considering the gap of the moving speeds
of two clusters; (b) Consideration of the gap of the moving speeds between
two clusters
2.2. Service-oriented routing mechanism.

2.2.1. Cache relay node selection. The purpose of the proposed cache relay node selection
mechanism is to find out the cache relay node that can satisfy the requirement of the
requestor. When the requestor requests a service, the management servers first provide
the list of the candidate cache relay nodes, and the request node sends the query messages
to all the cache relay nodes on the list. If no appropriate cache relay node is found, the
request node will directly receive the data from the service content server in the wired
environment instead. In case all channels of the content server are all occupied, the
request will wait for service until its preset timeout expires.
2.2.1.1 Determination of locations of the cache relay nodes.
In this work, the locations of the cache relay nodes in the HVN need to be determined
for ease of further processing. Three cases are considered as follows:
Case 1: When the cache relay nodes and the requestor are located in the same
cluster, which is a simple intra-cluster V2V network, the requestor asks the cluster-
head to provide a candidate cache relay node list, which contains the cache relay
nodes ID.
Case 2: When the cache relay nodes and the requestor are separated in the dierent
V2V network clusters the requestor asks the local cache relay node management
server (LCNMS), which acts as a service broker, to record the history of the services
and the Internet protocol address, to provide a cache relay node list. After the
cache relay nodes are determined, the infotainment data stream will be sent to the
requestor via a roadside base station.
Case 3: When the cache relay nodes and the requestor are distributed in a het-
erogeneous wide-area network as illustrated that the requestor will ask the global
cache relay node management server (GCNMS), which serves as a service broker to
record the service information received from LCNMS, to provide a cache relay node
list. After the cache relay nodes are determined, the infotainment data stream will
be sent to the requestor via roadside base stations, such as BS1 and BS2 and the
Internet. Notably, BS1 and BS2 in this example are in dierent network protocols.
2.2.1.2 Fuzzy logic-based cache relay node selection.
In order to reduce the control overhead, a fuzzy cache relay node determination algo-
rithm is employed to find the most appropriate relay nodes for forwarding the infotainment
data stream as illustrated in the preceding subsection. The fuzzy logic techniques have
been used to solve several resource assignment problems eciently in the ATM and wire-
less networks in the literature [12]. The inputs of the fuzzy logic inference system for
the selection mechanism include received signal strength, available bandwidth and delay
observed at the cache relay node. The output of the fuzzy logic system is the estimated
appropriateness level for the candidate cache relay node.
Stability of signal strength
It was reported that the received signal strength (RSS) can be adopted as one essen-
tial parameter to reduce the unnecessary vertical hando and accordingly minimize the
dropping probability [13]. The hando process can be performed before the RSS becomes
weak and thus obtains better QoS and higher data rate. We thus adopt the received signal
strength at the cache relay node as the first input parameters for the fuzzy logic system.
The received signal strength that can be predicted in the next period is expressed by

S(t + 1) = Spre (n + 1) S (n) (2)
where S (n) is the summation of previous n data, Spre

= Ppre (n + 1) is a polynomial
function with M + 1 unknown coecients linguistic, 0 i M , and
Ppre (t) = a0 + a1 t + a2 t2 + + aM tM (3)

where ai is determined by the Native Gauss Elimination method [13].
The stability of the signal strength in the next period T as employed in the fuzzy logic
system is then expressed by
T
1
(S(i + 1) S(i))2
i=1
Ssta = (4)

T
S(i)
i=1
Available bandwidth
The second parameter that considers the available bandwidth of node x is expressed
by, {
Bx f ree Bx need
Bx
, if Bx f ree Bx need 0
Bx = f ree , (5)
0, if Bx f ree Bx need < 0
where Bx f ree denotes the free bandwidth at node x and Bx need is the amount of the
required bandwidth for the request. It can be seen from the equation that it is more
possible for node x to be selected as the cache relay node if it can provide more bandwidth.
The fuzzy linguistic variables used for this parameter are high, medium and low.
Delay
The third input parameter used for the fuzzy logic system, the delay, is defined as
the delay for the intermediate nodes along the routing path in a HVN. The linguistic
variables of the delay are long, medium and short. The non-fuzzy output of the
defuzzifier can then be expressed as the weighted average of each rules output after the
Tsukamoto defuzzification method is applied. For example, if the received signal strength
of the candidate cache relay node is strong, and its available bandwidth is medium,
and the delay between the candidate cache relay node and the request node is short,
then the estimated appropriateness level of the candidate cache relay node is high.
2.2.2. Network selection mechanism. Besides the two parameters, the trac load and the
received signal strength, we further consider the average delay and packet loss rate for
satisfying the QoS requirement of the real time services in selecting the suitable wireless
access technologies,
( ) ( ) ( )
Bia log(ri ) Di Pis
Qi = wB + wS 1 + wD 1 + wP , (6)
Bi log(Ri ) Dmax Pis + Pil
where Bia is the number of available bandwidth units in network Ni , and Bi is the total
number of bandwidth units in network Ni . Bia /Bi represents the complementing of the
normalized utilization of network Ni . Ri is the transmission range of network Ni and
ri denotes the distance between the mobile client and the BS/AP of network Ni . Thus,
1 log(ri )/log(Ri ) stands for the relative received signal strength from network Ni . Di is
the summation of its forward and backward delays with respect to the cache relay nodes,
and Dmax is the maximum delay that the client can tolerate. 1 Di /Dmax thus stands
for the delay degree whether it satisfies the users requirement. Pil denote the number
of dropped packets on the transmission path during a fixed period of time and Pis is the
number of all / successfully transmitted packets through the same link over the same period
s s l
of time. Pi Pi + Pi is thus viewed as the end-to-end loss rate. Besides, the constraint
for the four predetermined weights as given in Equation (3) is wB + wS + wD + wP = 1.
3. Simulation Results. A discrete-event simulator using the JAVA language was adopt-
ed to evaluate the performance of the proposed work. Extensive simulations were per-
formed to study the characteristics of the proposed algorithm using the microscopic vehicle
trac simulator. The simulation scenario developed by the Federal Highway Authority
[15] was applied in this work because it provides vehicle trac simulation data that is
very close to real trac data. The network used in the simulations consists of 100 vehi-
cles moving in a highway with 10 km length. The packet length including the position
information is 75 bytes. The first part of the simulation compares the proposed group
mobility-based clustering algorithm (GMC) with two representative clustering techniques
in the literature, including Position-based prioritized clustering (PPC) and Lowest-ID
clustering methods, and the second part of the simulation is to examine the performance
of the proposed service-oriented vehicle infotainment dissemination management scheme,
which consists of the above-mentioned GMC algorithm and the service-oriented routing
(SOR) mechanism, in terms of metrics average delay, packet loss rate and the throughput.
The comparison target is a state-of-the-art routing algorithm that suits heterogeneous
wireless networks, a named policy-based QoS routing (PBQR) algorithm [16].
The clustering in PPC is determined by the geographic position of nodes and the
priorities associated with the vehicle trac information, such as travel time and velocity.
Each cluster elects one node as its cluster head. The cluster size is controlled by a
predefined maximum distance between a cluster head and its members. Similar to the
proposed work, clusters are independently controlled and dynamically reconfigured as
nodes move. In the Lowest-ID algorithm, the central controller selects arbitrarily one
node to be a cluster head, draws a circle of preset radius around that node and declares
all nodes captured in that circle to belong to the first cluster. From among the nodes
not in the first cluster, the central controller chooses arbitrarily the node with the lowest
ID number to be the next cluster head. The controller draws a second circle of preset
radius around the new cluster head and declares all nodes within this second circle to be
members of the second cluster. The process repeats until every node belongs to at least
one cluster and has its own cluster head.
The first part of our simulation indicates the comparison of the simulation outcomes
for the three compared clustering schemes under dierent transmission distances in terms
of performance metrics, the number of re-clustering, average lifetime of the cluster head
and that of cluster members.
Figure 4 demonstrates the number of re-clustering that the three algorithms need on
the condition of dierent transmission range. Figure 4 indicates that the number of re-
clustering significantly decreases when the transmission range increases.
The parameters of velocity, distance and direction are used in the proposed GMC
algorithm to appropriately select the cluster head. Apparently, the cluster head selected
in the proposed GMC algorithm is the most stable when compared with the other two
cluster mechanisms, including PPC which considers the factor of location and Lowest-ID,
which uses the comparison of groups weight to select the cluster head. With the stability
of the cluster head, we can greatly decrease the number of re-clustering and avoid the
waste of system resources and computation time. Meanwhile, the iterations of executing
the Lowest-ID algorithm are significantly increased when the communication range is
shorter. PPC considers the stability of cluster head, and re-clustering occurs when the
cluster head leaves. Thus, the number of re-clustering in PPC is less than that in Lowest-
ID. Although PPC sets a limit to determine whether re-clustering should be performed
owing to the leaving of the cluster head, the setting lacks for flexibility of adapting to the
volatile change of the HVN.
Figure 4. Comparison of number of re-clustering for transmission range
Figure 5 illustrates the comparison for the average lifetime of cluster heads on the
conditions of dierent transmission range. Figure 5 shows that the average lifetime of the
cluster head is prolonged by the increase of the transmission range.
Figure 5. Comparison of average lifetime of cluster head for transmission range
The proposed GMC algorithm achieves the best performance among the three. The
grouping indicator, GI, as given in Equation (1), plays an important role in the GMC
algorithm. It not only helps the determination of the establishment of clusters but also
prolongs the average lifetime of the cluster head. The average lifetime of the cluster head
of GMC grows up significantly while the transmission range of GMC is greater than 600.
The selection of the cluster head without careful consideration in Lowest-ID and the usage
of the traveling time as the parameter for choosing the cluster head in PPC deteriorate
the performance of the two schemes.
Figure 6 demonstrates the average lifetime of cluster members compared with the other
two algorithms on the condition of dierent transmission range. As shown in Figure 6,
we can observe that the average lifetime of cluster members increases when the transmis-
sion range increases. We present a mechanism for maintenance of the cluster members
to strengthen the stability of the cluster head, whereby the cluster members will not be
split because of the factors of distance or velocity. The clustering mechanism algorithm
mentioned in the literature often makes the cluster members unstable without considering
the factor of velocity and thus significantly decreases the lifetime of the cluster members.
Therefore, the average lifetime of the cluster members can be much higher than that in
other clustering mechanism algorithms in the literature. Meanwhile, it can be observed
that it is infrequent for a cluster member to switch to another cluster for the three schemes
when the communication range becomes larger. The estimation of the node stability in
PPC results in its better performance than Lowest-ID. The robust maintenance mecha-
nism employed in the proposed GMC algorithm can eectively prolong the average time
of a cluster member staying at the current cluster and thus achieve the best performance.
Figure 6. Comparison of average lifetime of cluster member for transmis-

sion range
The second part of our simulation is to examine the performance of the proposed service-
oriented vehicle infotainment dissemination management scheme, which consists of the
above-mentioned GMC algorithm and the service-oriented routing (SOR) mechanism, in
terms of metrics average delay, packet loss rate and the throughput.
Figure 7 shows the average delay of each algorithm on the condition of dierent num-
ber of vehicles respectively. Figure 7 shows that if the number of vehicles increases, the
average delay will also increase. All services are provided by the global cache relay node
management server in the other two compared algorithms. On the other hand, the pro-
posed SOR scheme can eectively choose the nearest server to obtain the information the
clients need and significantly decrease the delay.
Figure 8 demonstrates the percentage of the packet loss for each algorithm on the
conditions of dierent number of vehicles. The compared clustering mechanisms make
the cluster members unstable without considering the vehicle velocity, and the packet
losses for these two mechanisms are obviously increased. On the contrary, the stability of
the cluster head is strengthened in our work, and thus the cluster members will not re-
cluster because of the velocity variation. This is why the proposed clustering mechanism
combined with SOR scheme can significantly decrease the packet loss. Besides, the load-
balancing approach to forward the data streams adopted in this work also contributes to
decreasing the packet loss.
Figure 9 illustrates the throughput compared with each algorithm on the condition of
dierent number of vehicles. Figure 9 shows that the system throughput will gradually in-
crease when the number of vehicles becomes larger. It can be observed from Figure 9 that
Figure 7. Comparison of average delay for number of vehicles
Figure 8. Comparison of packet loss for number of vehicles
there is a significant gap in the throughput performance between GMC and PPC cluster-
ing algorithms. Again, the appropriate selection of cache relay nodes and load-balancing
approach adopted in the proposed SOR scheme significantly decreases the service time
and improves the throughput. The instability of cluster formation results in the frequent
service interruption and thus lowers the transmission throughput in PPC algorithm.
4. Conclusion. A service-oriented infotainment dissemination management scheme is

proposed in this work to provide a scalable framework for delivering real-time and best
eort services over heterogeneous vehicular networks and ensure interoperability, roaming
and end-to-end session management. The clustering algorithm takes the speed and accel-
eration of vehicles as the parameters to elect cluster heads and form vehicle clusters. The
emerging and splitting of the clusters and the concept of self-healing are considered in the
algorithm to strengthen the maintenance of cluster topology. Another focus of this work
is to implement a service-oriented routing mechanism, in which relay nodes that forward
the data streams and the wired/wireless access technologies for the data transmission
between the relay node and the requestor can be intelligently determined. A series of
Figure 9. Comparison of throughput for number of vehicle
simulations were run to evaluate the performance of the proposed infotainment dissemi-
nation management schemes. The experimental results demonstrate the eectiveness and
practicability of the proposed work.
Acknowledgment. The authors would like to thank the National Science Council, Tai-
wan for financially supporting this research under contract numbers NSC 96-2628-E-259-
022-MY3 and NSC 98-2220-E-259-003.
REFERENCES
[1] H. Y. Huang, P. E. Luo, M. Li, D. Li, X. Li, W. Shu and M. Y. Wu, Performance evaluation of
SUVnet with real-time trac data, IEEE Trans. on Vehicular Technology, vol.56, no.6, pp.3381-3396,
2007.
[2] T. Taleb, E. Sakhaee, A. Jamalipour, K. Hashimoto, N. Kato and Y. Nemoto, A stable routing
protocol to support ITS services in VANET networks, IEEE Trans. on Vehicular Technology, vol.56,
no.6, pp.3337-3347, 2007.
[3] E. Park and H. Shin, Reconfigurable service composition and categorization for power-aware mobile
computing, IEEE Trans. on Parallel and Distributed Systems, vol.19, no.11, pp.1553-1564, 2008.
[4] S. Oton, A. Ortiz, J. R. Hilera, R. Barchino, J. M. Gutierrez, J. J. Martnez, J. A. Gutierrez, L.
de Marcos and M. L. Jimenez, Service oriented architecture for the implementation of distributed
repositories of learning objects, International Journal of Innovative Computing, Information and
[5] P.-C. Tseng, C.-Y. Chen, W.-S. Hwang, J.-S. Pan and B.-Y. Liao, Qos-aware residential gateway
supporting Zigbee-related services based on a service oriented architecture, International Journal of
[6] H. Nakano, A. Utani, A. Miyauchi and H. Yamamoto, Grouping of mobile nodes in MANET based
on location and mobility information using an ART network, International Journal of Innovative
Computing, Information and Control, vol.5, no.11(B), pp.4357-4365, 2009.
[7] Z. Wang, L. Liu, M. C. Zhou and N. Ansari, A position-based clustering technique for ad hoc
intervehicle communication, IEEE Trans. on Systems, Man, and Cybernetics, Part C: Applications
and Reviews, vol.38, no.2, pp.201-208, 2008.
[8] A. Amis, R. Prakash, T. Vuong and D. T. Huynh, Maxmin d-cluster formation in wireless ad hoc
networks, Proc. of the IEEE Conf. Comput. Commun., Tel Aviv, Israel, vol.1, pp.32-41, 1999.
[9] C. Alwan, R. Bagrodia, N. Bambos, M. Gerla, L. Kleinrock, J. Short and J. Villasenor, Adaptive
mobile multimedia networks, IEEE Pers. Commun., vol.3, no.2, pp.34-51, 1996.
[10] D. J. Baker, A. Ephremides and J. A. Flynn, The design and simulation of a mobile radio network
with distributed control, IEEE J. Sel. Areas Commun., vol.SAC-2, no.1, pp.226-237, 1984.
[11] Y.-F. Huang, Y.-F. Chen, T.-H. Tan and H.-L. Hung, Applications of fuzzy logic for adaptive in-
terference canceller in CDMA wireless communication systems, International Journal of Innovative
[12] K. Hirota, Industrial Applications of Fuzzy Technology, Springer-Verlag, 1993.
[13] B. J. Chang and J. F. Chen, Cross-layer-based adaptive vertical hando with predictive RSS in
heterogeneous wireless networks, IEEE Trans. on Vehicular Technology, vol.57, no.6, pp.3679-3692,
2008.
[14] E. Stevens-Navarro, Y. Lin and V. W. S. Wong, An MDP-based vertical hando decision algorithm
for heterogeneous wireless networks, IEEE Trans. on Vehicular Technology, vol.57, no.2, pp.1243-
1254, 2008.
[15] B. van Arem, C. J. G. van Driel and R. Visser, Impact of cooperative adaptive cruise control on
trac-flow characteristics, IEEE Trans. on Intelligent Transportation Systems, vol.7, no.4, pp.429-
436, 2006.
[16] K. Yang, Y. Wu and H. H. Chen, QoS-aware routing in emerging heterogeneous wireless networks,
IEEE Communications Magazine, vol.45, no.2, pp.74-80, 2007.
c ISSN 1349-4198
AN OPTIMAL FUZZY CONTROLLER FOR A HIGH-PERFORMANCE

DRILLING PROCESS IMPLEMENTED OVER AN
INDUSTRIAL NETWORK
Rodolfo E. Haber1,4 , Agustn Gajate1 , Steven Y. Liang2

Rodolfo Haber-Haber3 and Rau l M. del Toro1
1
Center of Automation and Robotics (CSIC-UPM)
Spanish Council for Scientific Research (CSIC)
Ctra. Campo Real Km. 0.200, 28500 Arganda del Rey (Madrid), Spain
rhaber@iai.csic.es
2
George W. Woodru School of Mechanical Engineering
Georgia Institute of Technology
801 Ferst Drive, Atlanta, GA 30332-0405
steven.liang@me.gatech.edu
3
Escuela Tecnica Superior de Ingenieros Industriales
Universidad Politecnica de Madrid
c/Jose Gutierrez Abascal, 2, 28007 Madrid, Spain
4
Escuela Politecnica Superior - UAM
C/Francisco Tomas y Valiente 11, 28049, Madrid, Spain
Abstract. Network-based applications can provide intelligence to high-performance dril-

ling (HPD) processes through networked control systems (NCS), with many advantages
such as flexibility, and reduced maintenance time and cost. In order to improve drilling
eciency while preserving tool life, this study focuses on the design and implementation
of an optimal fuzzy-control system for drilling force. This 49-rule controller is networked
and operates on a computer numerical control (CNC) machine tool. It is optimally tuned
using a known maximum allowable delay to deal with uncertainties in the drilling pro-
cess and delays in the network-based application. Experimental results demonstrated that
the proposed control strategy provides a good transient response (without overshoot) and
better error-based performance indices than the CNC working alone (uncontrolled), two
versions of a PID controller and a 9-rule fuzzy controller.
Keywords: High-performance drilling process, Fuzzy control, Force
1. Introduction. Recent advances in communications, artificial intelligence based tech-

niques and computing technologies have favoured an increasing trend in manufacturing
towards the distribution of function and resources. Network-based control applications
have been extensively applied to complex and large-scale processes. The principal char-
acteristics of the next generation of manufacturing processes will be distinguished by four
adjectives: adaptive, digital, networked and knowledge-based [1]. The link joining these
characteristics is the network, which allows dynamic and value-added cooperation in a
global production environment. Trends in this area have led to the use of networked
intelligent control to increase productivity and performance for example, in the high-
performance machining processes which have become an integral part of manufacturing
over the past decade [2].
In particular, high-performance drilling (HPD) processes have had a major impact on
production technology in many areas, including the automotive, die/mould and aerospace
1481
1482 R. E. HABER, A. GAJATE, S. Y. LIANG, R. HABER-HABER AND R. M. DEL TORO
industries. In a networked enterprise, the machine tool is a peripheral device in a computer

network. The network architecture is therefore an essential element for intelligence to
be eectively provided to the machining platform through a networked control system
(NCS), yielding many advantages such as flexibility, and reduced maintenance time and
cost. The literature has many examples of work dealing with new network protocols and
improvements to existing protocols for networked model-based machining force control
approaches [3].
Drilling is one of the commonest operations in manufacturing processes, comprising up
to 50% of all machining operations. The major aims of control in any drilling process
are to increase drilling rates and improve drilling accuracy [4]. One way to compensate
for the influence of drilling depth on useful tool life while enhancing productivity is to
introduce real-time control of the cutting force [5]. Research into drilling-process control
has intensified in recent years. Furness et al. [6], in a seminal paper, describe the dynamic
modeling and real-time control of a drilling process. A third-order linear model with time
delay is identified and used in conjunction with an input/output pole-placement controller
with gain scheduling for the real-time process control. Oh, Kim and Chu [7] utilize a PID
controller designed by means of the Ziegler-Nichols tuning procedure. The controller gain
is re-tuned using the root-locus method in order to eliminate response overshoot, since
overshoot typically provokes instability.
Most of the work on machining control has been devoted to the issue of classical and
adaptive techniques [8]. But on the one hand, PID control and gain scheduling strate-
gies have generally not yielded the expected improvement in drilling processes, and on
the other, adaptive controllers are extremely time intensive since the parameters are es-
timated on-line and controller gains are adjusted accordingly. Furthermore, adaptive
control systems must be carefully tuned, and sometimes exhibit complex and undesirable
behaviour.
Artificial-intelligence-based control techniques have aroused considerable interest in the
scientific community regarding the design of networked control systems (NCS) [9]. How-
ever, neural-network and evolutionary-algorithm-based strategies involve long computa-
tion times (e.g., on-line learning algorithms), which limits their performance in network-
based applications [10].
The main focus of the present work on designing and implementing a networked fuzzy
system was to study, analyze and verify the behaviour of a 49-rule, two-input/one-output,
fuzzy controller in the control of a complex process. The theoretical foundation of the
work was twofold. The first issue was to eectively combine a simple and computationally
ecient numerical method for minimizing an objective function in a multi-dimensional
space with a performance index that penalizes long-duration transients and generally
produces smaller overshoots and oscillations [11-13]. Those selected and applied were
the Nelder-Mead method and the integral of the time weighted absolute error (ITAE )
performance index. The second issue was to use the theoretical foundation provided
by fuzzy logic and the MacVicar-Whelan template approach for rule base creation to
design an accurate and simple control law [14-16]. A 49-rule, two-input/one-output, fuzzy
controller was then designed for the present case study. On the basis of this theoretical
foundation, a method is proposed for controlling a complex process in a network-based
application with a known bound on the maximum allowable delay.
We also investigate specific issues such as the Nelder-Mead approach for the optimal
setting of fuzzy control systems, and the computational eciency and viability of applying
an algorithm which implements 49-rule fuzzy control systems in real-time. The advantages
and drawbacks of using this 49-rule fuzzy controller in the field of control are compared
with a PID controller [17] with a weighting setpoint, and with a 9-rule fuzzy controller
AN OPTIMAL FUZZY CONTROLLER FOR A HIGH-PERFORMANCE DRILLING PROCESS 1483
[18]. This comparative study involved both simulation and a real-world application of the
entire system. The main technical motivations are the need to improve the eciency of
complex processes which have a major economic impact (drilling), and to optimize the
drilling process eectively by taking the useful tool life into account.
It is beyond the scope of the present study to compare the Nelder-Mead method with
other local or global optimization techniques such as genetic algorithms, simulated an-
nealing and cross entropy [19-21]. The main point is that the Nelder-Mead approach is
very easy to program (with fewer parameters than many other optimization strategies),
and it yields consistently accurate results. Also beyond the scope of this study is a com-
parison with long-computation-time strategies (i.e., neuro-fuzzy and adaptive strategies)
which require training stages and on-line controller gain estimation [22].
The paper is organized as follows. Section 2 presents a brief study of the complex
process of high-performance drilling; Section 3 describes the design of a fuzzy controller
to control HPD; Section 4 addresses the implementation of this fuzzy controller and its
connection to the computer numerical control (CNC) machine tool through a field bus;
Section 5 presents the experimental results and a comparison of the 49-rule fuzzy controller
with two versions of a linear controller (PID) and a 9-rule fuzzy controller; and Section 6
presents the conclusions.
2. Modeling of the Drilling Operation. Metal cutting operations, of which drilling
is a subset, have been modeled in a variety of manners. Many approaches explicitly model
the cutting force through geometric and process variables, e.g. spindle speed, feed rate,
depth of cut, etc as described in the overview by Landers et al. [23]. For the drilling
process in particular, Kim et al. analyzed the thrust force (F ) and feed rate (f ) of the
drilling operation on a CNC machine tool, creating a linear, first-order system as shown
in (1), where the system gain and time constant are dependent on machining conditions
and the materials properties [5].
F (s) K
G(s) = = (1)
f (s) Ts + 1
The entire system of feed drive, cutting process and dynamometer platform was modeled
as a third-order system [17], with the following transfer function:
F (s) 1958
GP (s) = = 3 (2)
f (s) s + 17.89 s + 103.3 s + 190.8
2
where s is the Laplace operator, f is the command feed and F is the cutting force.
The above model has certain limitations in representing the complexity and uncertainty
of the drilling process. However, it does provide a rough description of the processs
behaviour that is useful for the optimal tuning of a network-based control system.
3. A Network-Based Fuzzy-Control System for Drilling Force. In order to avoid
rapid tool wear or catastrophic tool breakage, conservative and constant drilling param-
eters are set by human operators on the basis of past experience. However, a fuzzy con-
troller can adjust the feed rate to regulate the drilling force, optimizing the production
rate without there being any need for concern about drill damage due to over-force/torque
acting on the tool. In particular, the controller can adjust this variable to respond to in-
creased friction between the drill and workpiece or changes in work-material hardness.
3.1. Design of the HPD fuzzy controller. A typical fuzzy controller describes the
relationship between the change of the control signal (e.g., the feed-rate increment f )
and the error (e.g., force error eF and its variation eF ) [15]. This type of fuzzy controller
was suggested as early as the mid 1970s, and remains a standard in the field of fuzzy control
[24]. We followed the well established steps for defining input and output membership
functions and for constructing fuzzy-control rules in designing a two-input/one-output
fuzzy-control system. The fuzzy controllers core performs on-line actions to modify the
feed rate. The manipulated (action) variable is the feed-rate increment (f as a fraction
of the initial value programmed into the CNC).
The reference force value (Fr ) is derived from a learning phase of uncontrolled drilling
tests. For each sampling period k, the cutting-force error and the variation in cutting-force
error are calculated as:
eF (k) = Fr F (k) (3)
eF (k) = eF (k) eF (k 1) (4)
where eF and eF are calculated at each instant k.
The output (u) and the error vectors are, respectively:
[ ]
[e ec] = Ke eF Kce eF ; u = f (5)
where e and ec are the discrete input variables, eF is the cutting-force error (in newtons),
eF is the variation in cutting-force error (in newtons) and Ke , Kce are scaling factors
for the inputs (error and variation in error, respectively).
How the knowledge base of any fuzzy controller is created strongly influences the con-
trollers overall dynamic behaviour. The universe of discourse of input variables is par-
titioned into seven fuzzy sets characterized by triangular-shaped membership functions
in the range [30, 30]. Seven partitions with triangular membership functions are also
defined for the output (action) variable. The output variables membership functions are
equally spaced, and set according to the maximum modification of the nominal feed rate
under nominal cutting conditions. Figure 1 shows the resulting fuzzy partition for the
output u and the error vectors e and ec. The seven fuzzy sets used for the inputs and the
output are NB, negative large; NM, negative medium; NS, negative small; ZE, zero; PS,
positive small; PM, positive medium; and PB, positive large.
Figure 1. Fuzzy partitions and membership functions for (a) e, ec; (b) u
The selection of the shape of the membership functions is the result of combining a trial-
and-error process with empirical design guidelines established on the basis of a knowledge-
engineering standpoint and simulation studies. Triangular-shaped membership functions
are the most frequently used in fuzzy controllers because of their computational eciency
and ease of implementation in real-time applications.
One method of constructing the rule base is to use a template (standard). The template
rule-base can be regarded as unifying engineering common-sense with experience in fuzzy
control. We chose the MacVicar-Whelan template approach for rule-base creation [14],
after appropriate modification of the linguistic labels of the output variable. The 49
control rules are summarized in Table 1. A similar rule base has been applied to the
networked fuzzy control of an electromechanical system [25].
Table 1. Rule base to handle the feed rate
e/ec NB NM NS ZE PS PM PB
NB NB NB NB NB NM NS ZE
NM NB NB NB NM NS ZE PS
NS NB NB NM NS ZE PS PM
ZE NB NM NS ZE PS PM PB
PS NM NS ZE PS PM PB PB
PM NS ZE PS PM PB PB PB
PB ZE PS PM PB PB PB PB
The sup-product composition operator is selected for the compositional rule of inference.
An in-depth study of the behaviour of numerous fuzzy operators in dierent fuzzy control
roles can be found in [26]. The controller obtained via sup-product composition will work
well for many types of application with reasonable accuracy.
Taking the maximum number of rules activated at any given sampling instant to be
four, and using the algebraic product operation, fuzzy implication and the maximum
(union) operation, one has:
4
R (e, ec, u) = max {prod {ei (e), eci (ec), ui (u)}} (6)
i=1
The crisp controller output, which changes the feed rate, is obtained by defuzzification.
There are three basic traditional defuzzification methods: the maximum, mean-of-maxima
(MOM) and centre-of-area (COA). A survey of defuzzification strategies can be found in
[27].
We chose the COA strategy for defuzzification because of its appropriate steady state
performance, and its use as a standard defuzzification method in experimental and indus-
trial fuzzy controllers. It is expressed as:

R (ui ) ui
u= i
(7)
R (ui )
i
where u is the crisp value of ui for a given crisp input (e, ec).
The output-scaling factor (GC ) multiplied by the crisp-control action (generated at
each sampling instant) provides the final actions that will be applied to the CNC machine
tools cutting parameters. Finally, the feed rate f of the CNC drilling machine can be
written as:
f (k) = f (k 1) + GC u(k) (8)
3.2. Network-based fuzzy control using a field bus. PROFIBUS is an extensively
used field bus that operates via a master-slave relationship among devices connected to
the network. The multipoint interface (MPI) is a programming interface for the Simatic
S7 series that closely follows the PROFIBUS protocol [28]. The MPIs physical interface
is identical to the PROFIBUS RS485 standard. The transmission rate can be increased
up to 12Mbits/s with the use of an MPI. The control-system architecture for a machine
tool using an MPI network is shown in Figure 2.
Figure 2. Network-based fuzzy-control system architecture for a high-

performance drilling process
Since the control signal (command feed rate) is transmitted through the MPI, some
amount of network-induced delay is unavoidable. A comprehensive treatment of network-
induced delay can be found in [29].
Figure 3. Histogram of the transmission delay time over an MPI network
Figure 3 is a histogram of 100 experimental transmission samples between a personal

computer and an open CNC connected through the network. In order to ensure the
suitability of this experimental test, every sample is written in a single parameter to a
register on the CNC PLC through a numerical control-dynamic data exchange interface.
The CNC PLC then modifies another register, which depends on the previously written
parameter, and requests this value. The cycle is completed when the second register
has the initial value of the first register, with the total time delay being recorded for
this sample. The maximum (41 ms) and minimum values (16 ms) of the total elapsed
time required to write a single variable (i.e., feed rate) to the CNC machine tool are
then estimated. However, network-induced delays are usually unknown a priori and are
dicult to measure and estimate on-line.
The maximum delay estimated from open-loop experiments is 0.4s, including both
dead-time process and network-induced delay [30]. The whole model is represented as:
F (s) 1958
GP (s) = = 3 e0.4s (9)
f (s) s + 17.89 s + 103.3 s + 190.8
2
The results given in this section serve to provide a rough estimate of the maximum
delay, which is essential for the design and implementation of the fuzzy control system to
be presented in the following sections. Two remarks are important regarding Equation
(9): it is only an approximate description of the dynamic behaviour of the drilling process,
and it has been verified experimentally under just certain cutting conditions. Therefore,
it does not describe the whole operating region and there will be a certain limitation on
the control system design based on this model (the desired closed-loop behaviour can not
be guaranteed). Indeed, the main bottleneck encountered in designing a control system
for drilling is that the dynamics cannot be accurately modeled mathematically. Model
parameters such as gains and time constants depend on the workpiece material, cutting
conditions and tool wear.
4. Implementation of the Fuzzy Control System. This section briefly explains the
tuning and implementation of the fuzzy controller (whose design was described in Section
3) that communicates with the open-architecture CNC through the MPI network.
The data-acquisition system consists of a dynamometer, a charge amplifier and hard-
ware and software modules, as depicted in Figure 4. The cutting forces are measured with
a Kistler 9257B piezoelectric dynamometer mounted between the workpiece and the ma-
chining table. The electric charge is then transmitted to the Kistler 5070A four-channel
charge amplifier. The interface hardware module consists of a connecting block and an
AT-MIO-16E-1 16-channel A/D acquisition card. The A/D device transforms the ana-
logue signal into a digital signal such that the Simulink program reads the three axial
force components simultaneously, and processes and displays these data. The Real-Time
Windows Target (RTWT) allows for real-time execution of Simulink models [31].
Figure 4. Scheme of the experimental set-up and fuzzy-control system

The output of the fuzzy controller is connected to the process through a multipoint
interface (MPI) with a default transmission rate of 187.5 kbits/s. A CP5611 card connects
the PC that implements the fuzzy-control system to this network. The MPI messaging
interface is a master-client relationship (active station), and manages exchanges with the
Siemens S7 PLC. The system has two masters, the man-machine control (MMC 103) with
a numerical-control kernel and a numerical-control unit (NCU 573.3) [32].
The application software developed runs under MATLAB c
. The toolboxes used in
conjunction with MATLAB are Simulink , Real-Time Workshop
c c
(RTW) and Windows
Target c
(WT). The Simulink Toolbox is an interactive graphic environment for modeling
and simulating dynamic systems. Real-Time Workshop gives Simulink the capacity to
interface in real-time to real world devices (or, in the RTW vernacular, to targets). Real-
time targets are (I/O) devices utilizing their own processors, running real-time tasks and
communicating with the PC/RTW.
These hardware and software tools are used to create the integrated environment. Two
simulation schemes in the same environment (i.e, two Simulink-based models) are used.
One is used for networked control system simulation (i.e., linear and fuzzy controllers)
and the other for DAQ and controller implementation. This setup helps the real system
to be designed, implemented and verified as rapidly and easily as in the simulation.
5. Simulation and Experimental Validation. The control literature of the past few
decades abounds in various performance criteria, with the most significant being functions
of time and error [33]. The optimal tuning of the present fuzzy controller was based on
the integral of the time weighted absolute error (ITAE ) performance index in order
to obtain a short settling time and overshoot-free time response. Other performance
indices such as the integral of the time weighted squared error (ITSE ), integral of
the absolute error (IAE ) and the integral of the time-squared weighted squared error
(IT 2 SE ) were also evaluated. Specifically, the ITAE penalizes long-duration transients
and is very sensitive to parameter variations in comparison with other criteria. Moreover,
the ITAE criterion generally produces smaller overshoots and oscillations than the IAE
and ITSE criteria.
Before experimental testing of the control scheme, the performance of the fuzzy-control
system was examined by simulation using Simulink as depicted in Figure 5. The ITAE
index describes how well the system responds to an outside disturbance. In this study, a
step in the force reference signal is considered a disturbance, and one attempts to assess
how well the system follows setpoint changes using the ITAE criterion.
Figure 5. Overall model of the fuzzy-control system
The goal is to obtain the optimal parameters for the input scaling factors [Ke , Kce ]OP T
such that the ITAE performance index has its minimum:
( T )
KOP T = [Ke , Kce ]OP T = arg min |eF (t)| tdt (10)
0
Ovt = (max(eF ) Fr )/Fr 100 (11)

where T is the overall drilling time (i.e., time elapsed between tool entrance and exit) and
Ovt is the overshoot of the closed-loop dynamic response.
The optimal tuning of the fuzzy controller is carried out using the input scaling factors
[Ke , Kce ]OP T . While high values of Ke yield good responsiveness of the system (low
steady-state error and settling time), they also lead to poor stability (large overshoot).
The fastest convergence is obtained with low value of Kce . There are many methods for
tuning fuzzy controllers, and an in-depth study of them is beyond the scope of the present
work.
The optimal setting of the input scaling factors [Ke , Kce ]OP T was determined using
the Nelder-Mead search algorithm [11]. This is a numerical method for minimizing an
objective function in a multi-dimensional space. It finds a locally optimal solution to a
problem when the objective function varies smoothly. The fundamental step is to replace
the worst point with a point reflected through the centroid of remaining data, increasing
the size of the jump in this direction in the case of successfully proceeding downhill.
The resulting variable-size downhill simplex is corrected with a constant small-size uphill
simplex that climbs local gradients to local maxima to avoid entrapment in local minima.
The cost function was evaluated through simulation based on Equation (9). Initially
[Ke , Kce ]IN IT = [1.0, 1.0] and GC = 0.500, corresponding to an ITAE = 0.709 and an
overshoot of 4.090%. The output scaling factor GC was kept constant during the tuning
procedure. The minimum ITAE = 0.469 was reached at [Ke , Kce ]OP T = [0.0125, 0.0374].
The ITSE criterion was 0.292, the IT 2 SE criterion 0.164, and the overshoot Ovt 0.15%.
The simulation results are depicted in Figure 6. The systems step response with
a setpoint of 1000 N is depicted in Figure 6(a). As shown in the figure, the cut-
ting force is well regulated with respect to the reference value. Figure 6(b) shows the
control signal (feed rate modification). The transient response is clearly improved af-
ter optimization [Ke , Kce ]OP T = [0.0125, 0.0374] relative to the initial scaling factors
[Ke , Kce ]IN IT = [1.0, 1.0].
(a)
(b)
Figure 6. Simulation results: (a) step response of the cutting force; (b)
feed rate
There have been studies of constraints in communications from the control system
viewpoint. Some have analyzed the delays in a switched network control system with a
defined latency time using the TCP/IP transmission protocol. In general, these delays are
modeled using such statistical techniques as negative exponential distribution functions
and Poisson and Markov chains [34].
These models have a limited validity, however, due to the non-deterministic behaviour
of some networks. To study the influence of delay on controller performance, a simple
delay generator was included in the simulation stage. The maximum delay was assumed
to be known, and, to keep the computational cost down, a uniform distribution function
was selected given that the approximation with this delay generator is close enough for the
current study. The time-varying delay, L, in the network-based application was simulated
by assuming a uniformly-distributed delay between 0.138s (i.e., twice the sampling time)
and 0.8s (i.e., twice the maximum delay), and 100 simulation tests were run for the
optimal fuzzy control system. The maximum, minimum, and mean values of the ITAE
performance index and the overshoot were calculated from these simulations.
For a randomly generated minimum delay of 0.1469s, the overshoot was 0.044% and
the ITAE performance index was 0.288. The worst case occurred for the maximum delay,
0.7962s, where the overshoot was 1.701% and the ITAE was 1.125. Figure 7 shows the
simulation results corresponding to the systems step response for all the simulated delays.
Figure 7(a) represents the cutting-force behaviour and Figure 7(b) the feed-rate variations
in each case simulated. One infers from the simulation results that the optimal tuning
achieved in the simulation ensures an overshoot-free transient response with a settling
time of less than 2 seconds.
Figure 8 shows the dependence of the performance indices on delay time. The dete-
rioration of the ITAE (solid line) and IT 2 SE (dashed line) indices increases nonlinearly
with the delay. While this deterioration in the two indices is as was to be expected, it
demonstrates the robustness of the 49-rule, two-input/one-output, fuzzy controller in the
presence of dead-time plus network-induced delay.
In addition, the initial condition (i.e., the nominal feed rate set by the CAD/CAM or
the operator) adds a further diculty against the control systems satisfying the tran-
sient response and settling time requirements. In order to form a basis for performance
(a)
(b)
Figure 7. (a) Cutting-force behaviour; (b) feed-rate variations for the

minimum, mean and maximum delays
Figure 8. Behaviour of the ITAE (solid line) and IT 2 SE (dashed line)

performance indices in the presence of delays
comparison, a classic PID controller was designed [25]:

f (s) I
GP ID (s) = =P + +Ds (12)
E (s) s
where GP ID (s) is the transfer function of the controller, and P = Kp , I = Kp/Ti , D =
Kp Td are the proportional, integral, and derivative gains, respectively. The discrete
implementation chosen for Equation (12) was:
f [k] = fp [k] + fi [k] + fd [k]
fp [k] = Kp e [k]
( )
fi [k] = fi [k 1] + Kp h/Ti e [k] (13)
Td Kp T d N
fd [k] = fd [k 1] + (e [k] e [k 1])
Td + N h Td + N h
where N = 10 is the filter coecient for the discrete-time derivative and h = 0.069s is
the sampling time.
The model of Equation (9) was used to design this PID controller based on the Ziegler-
Nichols (Z-N) tuning method for compensation of a time delay. In order to accommodate
the delay (network-induced delay plus process dead-time), it was treated as the plants de-
lay time. In addition, a setpoint weighting algorithm was applied to improve the transient
response [35].
It is important to remark that the Z-N method and related approaches are not ideally
suited to the drilling control problem. Classic tuning procedures are typically designed
for an optimal transient response with an overshoot that is inappropriate (too large) for
the present cases technological constraints (preserving useful tool life, avoiding the risk
of tool breakage, etc). The Z-N method was therefore only used as a starting point, with
subsequent adjustments being necessary.
Considering a maximum delay of L = 0.4s and applying the ZieglerNichols method, we
obtained the following controller parameters PZN = 0.1086, IZN = 0.1483, DZN = 0.0219.
These parameters needed then to be adjusted to improve the transient response. In
particular, we reduced the controller gain tenfold (0.1), giving the new settings PZN =
0.01086, IZN = 0.01483, DZN = 0.00219. In order to avoid large overshoots, the setpoint
was weighted using a low-pass filter Gf (z) = (1 )/(z ) with = 0.95, as detailed
in previous work [17].
Experimental drilling tests were carried out on an HS1000 Kondia milling machine
equipped with a Sinumerik 840D open-CNC controller. A 10-mm diameter Sandvik cut-
ting tool was used for the drilling operations. The workpiece material was GGG40 with
a Brinell hardness of 233 HB. The nominal feed rate and spindle speed were f0 = 100
mm/min and n0 = 870 rpm, respectively, with a maximum drill depth of 14 mm. The
setpoint was set at 1000 N. Photographs of the experimental setup are shown in Figures
2 and 5.
A comparative study showed that the main dierence between the 9-rule and 49-rule
fuzzy controllers was of course their rule base. That earlier study [18] presented the 9-rule
fuzzy controller designed by combining trial-and-error processes with simulation studies,
while the present work describes the design of a 49-rule fuzzy controller inspired by the
MacVicar-Whelan template approach for rule-base creation. Moreover, triangular mem-
bership functions are used in the present work, whereas trapezoidal-shaped membership
functions were employed in [18]. The universe of discourse of input and output variables
is also dierent in the two approaches. The static characteristics (control surfaces) of the
two controllers before and after the optimal tuning procedure are shown in Figure 9.
(a)
(b)
Figure 9. Lateral view of the static characteristics (control surface) of the

9-rule and 49-rule fuzzy controllers: (a) initial design; (b) after optimal
tuning
(a)
(b)
Figure 10. (a) Cutting force; (b) feed rate variations in uncontrolled
drilling and drilling with a linear controller (PID)
(a)
(b)
Figure 11. (a) Cutting-force; (b) feed-rate variations in drilling controlled

by a PID with a weighting setpoint and by fuzzy controllers
A few dozen trials were conducted in the design stage. The results shown are taken
from the best experimental results with both the linear and the fuzzy approaches. Figure
10 shows those corresponding to drilling the GGG40 workpiece. Figure 10(a) shows the
behaviour of the uncontrolled force (solid line) and the PID controller with the Z-N tuned
parameters reduced by a factor of ten (0.1) (dash-dotted line). The setpoint is shown
by a dashed line. Figure 10(b) shows the corresponding control signal.
Figure 11 shows the experimental results under the same cutting conditions as ana-
lyzed previously, plotting the controlled force behaviour (Figure 11(a)) and the feed-rate
variations (Figure 11(b)) for the three cases analyzed (a PID controller with a weighting
setpoint [17], the 49-rule fuzzy controller and a 9-rule fuzzy controller [18]). Techni-
cally, an improved transient response diminishes the risk of catastrophic drill failure and
maximizes the useful tool life.
It is beyond the scope of the present work to compare the Nelder-Mead method with
other local or global optimization techniques [20]. The main point is that the Nelder-Mead
approach is very easy to program (with fewer parameters than many other optimization
strategies), and it yielded consistently accurate results for the optimal tuning of this
49-rule fuzzy controller.
A stability analysis of networked control systems [36] is also beyond the scope of the
present work. Indeed, in recent years several fuzzy control systems have been designed
taking advantage of the knowledge of expert operators, and, after their implementation,
a satisfactory dynamic performance exhibited during real tests has frequently been con-
sidered sucient to ensure stability. Nevertheless, stability is a basic condition that every
control system must satisfy, and its analysis requires a plant model. If a linear approxima-
tion of the plant is available that characterizes its main dynamics at least in the vicinity
of an operating point then a test for local asymptotic stability becomes feasible. The
feedback interconnection of the plants linear transfer function and the nonlinear fuzzy
controller leads to a nonlinear standard control loop that can be analyzed in the absolute
stability theoretical framework with a finite domain, e.g., with the circle criterion [37].
Table 2 lists the performance indices (10-11), (14-16) calculated using actual data (i.e.,
drilling force, time) recorded in real-time from experimental tests in an industrial envi-
ronment. At first sight, the behaviour of the 49-rule fuzzy controller and the others might
seem to be very similar. However, the 49-rule fuzzy controller has generally better per-
formance indices and good dynamic responses. The experimental results confirmed that
increases and fluctuations in the drilling force can be suppressed, with the improvement
being most marked in the 49-rule fuzzy control system. Overshoots and oscillations of
the drilling force are prejudicial for useful tool life (as well as for the drilling operations
quality) and need to be avoided.
T
IAE = |eF (t)|dt (14)
0
T
IT SE = te2F (t)dt (15)
0
T
2
IT SE = t2 e2F (t)dt (16)
0
6. Conclusions. We have presented a 49-rule, two-input/one-output, fuzzy controller

to regulate the force in drilling processes while maximizing useful tool life. The fuzzy-
control system adjusts the feed as needed to regulate the drilling force using the CNCs
own resources to modify the feed rate and a dynamometer to measure the drilling force.
Table 2. Comparative study of the control strategies applied
Controller Parameters ITAE ITSE IT 2 SE IAE Ovt(%)

CNC alone f0 = 100 mm/min
4.319 0.755 2.959 1.651 14.275
(without control) n0 = 870 rpm
PZN = 0.01086
Modified
IZN = 0.01483 1.259 0.448 0.376 1.229 17.732
PID Z-N
DZN = 0.00219
PID Z-N PZN = 0.01086

modified with IZN = 0.01483
2.174 0.412 0.569 1.347 0.0
setpoint DZN = 0.00219
weighting [17] = 0.95
Ke = 0.0559
9-rule fuzzy
Kce = 0.1156 1.990 0.297 0.660 1.089 6.575
control [18]
GC = 1.0
Ke = 0.0125
49-rule fuzzy
Kce = 0.0374 0.967 0.346 0.242 1.087 1.470
control
GC = 0.5
The main novelty of the approach lies in the design of an optimal 49-rule fuzzy con-
troller using an approximate linear model and a known maximum allowable delay to deal
with uncertainties and nonlinearities in the drilling process and delays in the network-
based application. The optimal setting of the input scaling factors of the fuzzy controller
is done using the ITAE criterion. The result is that the 49-rule fuzzy controller can deal
with nonlinearities due to tool degradation, machine tool vibration and workpiece prop-
erties (anisotropy) better than the other controllers with which it was compared in this
study. In particular, performance indices demonstrated that the 49-rule fuzzy controller
outperforms both a PID with a weighting factor and a 9-rule fuzzy controller, providing
a good transient response, small settling time, and minimal steady-state error.
The proposed controller yields a non-oscillating system regardless of disturbances and
nonlinearities, making it the appropriate choice for coping with such disturbances as
increases in drilling depth. This approach therefore represents an interesting alternative
to other control strategies for drilling process control.
Acknowledgment. This work was conducted under project DPI2008-01978 Networked

cognitive control system for high-performance machining processes (COGNETCON) fi-
nanced by the Spanish Ministry of Sciences and Innovation.
REFERENCES
[1] J. Wang, X. Xu, J. Sun, R. T. Li and W. S. Wang, Development of an NC controller for next
generation CNCS, International Journal of Innovative Computing, Information and Control, vol.4,
no.3, pp.593-604, 2008.
[2] S. Dongfeng and N. N. Gindy, Industrial applications of online machining process monitoring system,
IEEE/ASME Trans. on Mechatronics, vol.12, no.5, pp.561-564, 2007.
[3] Y. Tipsuwan and M. Y. Chow, Gain scheduler middleware: A methodology to enable existing con-
trollers for networked control and teleoperation Part I: Networked control, IEEE Trans. on Indus-
trial Electronics, vol.51, no.6, pp.1218-1227, 2004.
[4] Y. Sheng and M. Tomizuka, Intelligent modeling of thrust force in drilling process, Journal of
Dynamic Systems Measurement and Control-Transactions of the ASME, vol.128, no.4, pp.846-855,
2006.
[5] J. B. Kim, S. J. Lee and Y. P. Park, Stable and ecient drilling process by active control of the
thrust force, Mechanical Systems and Signal Processing, vol.8, no.5, pp.585-595, 1994.
[6] R. J. Furness, T. C. Tsao, J. S. Rankin, M. J. Muth and K. W. Manes, Torque control for a form
tool drilling operation, IEEE Trans. on Control Systems Technology, vol.7, no.1, pp.22-30, 1999.
[7] Y. T. Oh, G. D. Kim and C. N. Chu, Design of a drilling torque controller for a machining center,
International Journal of Advanced Manufacturing Technology, vol.22, no.5-6, pp.329-335, 2003.
[8] S. Y. Liang, R. L. Hecker and R. G. Landers, Machining process monitoring and control: The state-
of-the-art, Journal of Manufacturing Science and Engineering, Transactions of the ASME, vol.126,
no.2, pp.297-310, 2004.
[9] P. Mendez-Monroy and H. Benitez-Perez, Supervisory fuzzy control for networked control systems,
[10] X. M. Zhu, C. C. Hua and S. Wang, State feedback controller design of networked control sys-
tems with time delay in the plant, International Journal of Innovative Computing, Information and
Control, vol.4, no.2, pp.283-290, 2008.
[11] J. C. Lagarias, J. A. Reeds, M. H. Wright and P. E. Wright, Convergence properties of the Nelder-
Mead simplex method in low dimensions, SIAM Journal of Optimization, vol.9, no.1, pp.112-147,
1998.
[12] F. L. Lewis and V. L. Syrmos, Optimal Control, Wiley-Interscience, New York, 1995.
[13] R. Middleton and G. Goodwin, Digital Control and Estimation: A Unified Approach, Prentice Hall,
New York, 1990.
[14] P. J. MacVicar-Whelan, Fuzzy sets for a man-machine interactions, International Journal of Man-
Machine Studies, vol.8, no.6, pp.687-697, 1977.
[15] R. R. Yager and D. Filev, Essentials of Fuzzy Modeling and Control, John Wiley&Sons Inc., 1994.
[16] M. Zhang and S. Hu, Fuzzy control of nonlinear systems with input saturation using multiple model
structure, ICIC Express Letters, vol.2, no.2, pp.131-136, 2008.
[17] R. Haber-Haber, R. E. Haber-Guerra, M. C. Schmittdiel and R. M. del Toro, A classic solution
for the control of a high-performance drilling process, International Journal of Machine Tools and
Manufacture, vol.47, no.15, pp.2290-2297, 2007.
[18] R. E. Haber-Guerra, R. Haber-Haber, D. M. Andres and A. A. Palomar, Networked fuzzy control
system for a high-performance drilling process, Journal of Manufacturing Science and Engineering,
Transactions of the ASME, vol.130, no.3, pp.0310091-0310096, 2008.
[19] R. E. Haber, A. Gajate and R. del Toro, Optimal fuzzy control system using the cross-entropy
method. A case study of a drilling process, Information Sciences, vol.180, no.14, pp.2777-2792, 2010.
[20] R. E. Haber, R. Haber-Haber, A. Jimenez and R. Galan, An optimal fuzzy control system in a
network environment based on simulated annealing. An application to a drilling process, Applied
Soft Computing, vol.9, no.3, pp.889-895, 2009.
[21] D. Martin, R. del Toro, R. Haber and J. Dorronsoro, Optimal tuning of a networked linear con-
troller using a multi-objective genetic algorithm and its application to one complex electromechanical
process, International Journal of Innovative Computing, Information and Control, vol.5, no.10(B),
pp.3405-3414, 2009.
[22] A. Gajate, R. E. Haber and R. M. del Toro, Neurofuzzy drilling force-based control in an ethernet-
based application, International Journal of Innovative Computing, Information and Control, vol.6,
no.1, pp.373-386, 2010.
[23] R. G. Landers, A. G. Ulsoy and Y. H. Ma, A comparison of model-based machining force control
approaches, International Journal of Machine Tools and Manufacture, vol.44, no.7-8, pp.733-748,
2004.
[24] E. H. Mamdani and S. Assilian, An experiment in linguistic synthesis with a fuzzy controller, Inter-
national Journal of Man-Machine Studies, vol.7, pp.1-13, 1975.
[25] K. C. Lee, S. Lee and M. H. Lee, Remote fuzzy logic control of networked control system via
Profibus-DP, IEEE Trans. on Industrial Electronics, vol.50, no.4, pp.784-792, 2003.
[26] O. Cordon, F. Herrera and A. Peregrn, Applicability of the fuzzy operators in the design of fuzzy
logic controllers, Fuzzy Sets and Systems, vol.86, no.1, pp.15-41, 1997.
[27] S. Roychowdhury and W. Pedrycz, A survey of defuzzification strategies, Int. Journal of Intelligent
Systems, vol.16, no.6, pp.679-695, 2001.
[28] H. Kleines, J. Sarkadi, F. Suxdorf and K. Zwoll, Measurement of real-time aspects of Simatic PLC
operation in the context of physics experiments, IEEE Trans. on Nuclear Science, vol.51, no.3,
pp.489-494, 2004.
[29] S. Vitturi, Stochastic model of the PROFIBUS-DP cycle time, IEE Proc. of Conf. on Science Mea-
surement Technology, vol.151, no.5, pp.335-342, 2004.
[30] R. del Toro, M. Schmittdiel, R. Haber-Guerra and R. Haber-Haber, System identification of the
high performance drilling process for network-based control, ASME IDETC 2007: Proc. of the 21st
Biennial Conference on Mechanical Vibration and Noise, Las Vegas, NV, 2007.
[31] MathWorks: Real-Time Windows Target, MathWorks Inc., Natick, 2006.
[32] Siemens AG., Sinumerik 840d, OEM-package NCK. Software release 4, Users Manual, Erlangen,
Germany, 1999.
[33] R. C. Dorf and R. H. Bishop, Modern Control Systems, 9 Edition, Prentice-Hall, Upper Saddle River,
2001.
[34] T. C. Yang, Networked control system: A brief survey, IEE Proc. of Control Theory and Applications,
vol.153, no.4, pp.403-412, 2006.
[35] K. J.
Astrom and T. Haglund, PID Controllers: Theory, Design and Tuning, 3rd Edition, Interna-
tional Society for Measurement and Control, 1998.
[36] H. B. Li, Z. Q. Sun, M. Y. Chow, H. P. Liu and B. D. Chen, Stabilization of networked control
systems with time delay and packet dropout Part II, 2007 IEEE International Conference on
Automation and Logistics, vol.1-6, pp.3012-3017, 2007.
[37] R. E. Haber, G. Schmitt-Braess, R. H. Haber, A. Alique and J. R. Alique, Using circle criteria for
verifying asymptotic stability in Pl-like fuzzy control systems: Application to the milling process,
IEE Proc. of Control Theory and Applications, vol.150, no.6, pp.619-627, 2003.
c ISSN 1349-4198
NEW SECRET KEY TRAITOR TRACING SCHEME WITH DISPUTE

SETTLEMENT FROM BILINEAR MAPS
Kuo-Yu Tsai1 , Tzong-Chen Wu1,2 and Chien-Lung Hsu1,3

1
Taiwan Information Security Center
2
No. 43, Sec. 4, Keelung Rd., Taipei 106, Taiwan
nicklas@twisc.org
3
Chang Gung University
No. 259, Wen-Hwa 1st Rd., Kwei-Shan, Tao-Yuan 333, Taiwan
Abstract. We propose a new secret key traitor tracing scheme using bilinear maps, in
which the size of the enabling-block in a broadcast message is independent of the number of
subscribers. The proposed traitor tracing scheme can identify malicious subscribers (also
called traitors) from a pirate decoder, which is collusively made by k or fewer traitors.
Further, any malicious subscriber cannot falsify an unintended content for framing the
broadcast center and cause all other personal keys to be exposed. We also give formal
analyses to discuss the security of the proposed scheme in the random oracle model.
Keywords: Traitor tracing, Dispute settlement, Bilinear map
1. Introduction. Broadcast encryption [1] is one of the mechanisms to securely tackle

the distribution of digital content to a specific set of authorized subscribers [2, 3, 4, 5, 6, 7].
In various well-known practical and useful applications of broadcast encryption, such as
pay-per-view or subscription television broadcasts, online database publicly accessible on
the Internet, or distribution of multimedia (e.g., files in MP3 or JPEG), a subscriber is in
possession of a decoder that allows them to access the broadcast. The roles of a broadcast
encryption mechanism can be categorized as a broadcast center and a set of subscribers
each equipped with a decoder. Initially, the broadcast center assigns an authorized key,
stored in a decoder, for each subscriber. Thereafter, the broadcast center encrypts an
intended content with an encryption key and broadcasts the encrypted content. Only
the authorized subscribers can reconstruct the encryption key with their authorized keys,
and further, they recover the intended content with their encryption keys. Basically, a
broadcast message is divided into two parts: an enabling-block and a cipher-block. The
cipher-block is the ciphertext form of an intended content, which is encrypted by using
the encryption key. The enabling-block contains the public information that is used by
a decoder to reconstruct the encryption key for obtaining the intended content. Upon
receiving the broadcast message, each authorized subscriber can use his/her authorized
key to reconstruct the encryption key from the enabling-block and then use it to decrypt
the cipher-block to obtain the intended content.
Consider such scenario in which some malicious subscribers (so-called traitors) may
collude to create a pirate decoder with their authorized keys or copy intended content
to non-subscribers in the above applications. Traitor tracing [8] is one of the important
mechanisms to resist the piracy. Chor et al. [8] introduced the concept of a traitor tracing
1499
1500 K.-Y. TSAI, T.-C. WU AND C.-L. HSU
scheme to discourage malicious subscribers from creating a pirate decoder. In Chor

et al.s scheme, each subscriber is assigned a distinct personal key for decrypting the
broadcast messages. Some traitors may try to mix parts of their personal keys to create
a pirate decoder. In case of detecting and confiscating any pirate decoder, one or more
personal keys will be traced. According to traced personal keys, at least one traitor will be
identified. A traitor tracing scheme is called k-resilient traceability if at least one traitor
can be correctly identified when up to k traitors collude to make a pirate decoder. Since
then, several traitor tracing schemes have been developed [9, 10, 11, 12, 13, 14, 15, 16,
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]. Among these schemes, some of them are secret
key traitor tracing schemes, e.g., in [14, 19, 20, 23, 24, 26] and some of them are public
key traitor tracing schemes, e.g., in [9, 12, 17, 18, 21, 25]. The former schemes allow
the broadcaster to use a secret key to broadcast messages and hence are applicable to
the case that only the broadcast center can broadcast messages to the set of subscribers.
The latter schemes allow the broadcaster to use a public key to broadcast messages and
hence are applicable to the case that anyone could serve as the broadcaster. In general,
public key traitor tracing schemes are more costly than the secret key ones in storage
requirements and bandwidth. However, to earn cost eectiveness and to achieve security
robustness are two major design principles regarding to the key management issues for
traitor tracing. To design a traitor tracing scheme, a flexible and ecient approach to
key management is one of the major factors. Besides, it should provide a robust traitor
detection mechanism against piracy without harming the innocent subscribers.
According to Boneh et al.s discussions [11], traitor tracing can be classified into two
categories: trace and trace-and-revoke. The former focuses on the capabilities of tracing
a traitor. The latter provides not only tracing capabilities but also revoking capabilities.
As compared with these two categories, the former requires shorter key length and lower
message transmission. The former does not provide a technical solution for revoking but
it can solve the revoking problem by some non-technical methods, such as seizing the
decoder, punishment, etc.
Following up the above design principles, Mitsunari et al. [19] proposed a secret key
traitor tracing scheme using bilinear maps, in which the size of the enabling-block in a
broadcast message is independent of the number of subscribers. However, To et al. [25]
have shown that Mitsunari et al.s scheme cannot withstand the so-called linear attack.
That is, any k out of n subscribers acting as colluding traitors can always find a linear
combination of their personal keys to generate a pirate personal key. The pirate personal
key is dierent from their pre-assigned personal keys on avoiding traitor tracing. To et al.
also proposed three public key traitor tracing schemes and claimed that their proposed
schemes are resistant to the linear attack. However, Chabanne et al. [12] pointed out
that To et al.s schemes [25] are still insecure. Chabanne et al. further introduced the
property of public traceability, which means that the tracing capability can be delegated
to a dishonest party and publicly computable by the delegated party.
In this paper, inspired from the original system model proposed by Mitsunari et al. [19]
and the definition of collusion-resistance presented by Boneh et al. [11], we shall present
a new secret key traitor tracing scheme with dispute settlement using bilinear maps, in
which the size of the enabling-block in a broadcast message is independent of the number
of subscribers. In the proposed scheme, it is able to identify malicious traitors from a
pirate decoder which is collusively made by k or fewer traitors, and no innocent subscriber
is convicted, where k is the number of traitors. Furthermore, any malicious subscriber
neither has the ability to falsify an unintended content for framing the broadcast center
nor causes all other personal keys to be exposed. We also prove the security of the
proposed scheme in the random oracle model [28].
NEW SECRET KEY TRAITOR TRACING SCHEME 1501
The paper is organized as follows: In Section 2, we give some definitions of the com-
putational problems and a secret key traitor tracing scheme with dispute settlement. We
also discuss a formal security model in this section. Our scheme is given in Section 3. In
Section 4, we discuss the security and eciency analyses. Finally, conclusions are given
in Section 5.
2. Preliminaries. In this section, we first introduce the properties of a bilinear map,

and then give definitions of the one-way hash function (OHF) [29] assumption, the weak
Die-Hellman problem (WDHP) [30], and the collusion attack algorithm with k-traitors
(k-CAA) [26, 31]. Finally, we give a formal definition and a security model for a secret
key traitor tracing scheme with dispute settlement.
2.1. Bilinear maps. Let G1 and G2 be two cyclic groups of prime order q, where G1 is an
additive group and G2 is a multiplicative group. We use a bilinear map e: G1 G1 G2
in our system, having the following properties:
1. For any P, Q, R G1 and a, b Zq , we have e(aP + bQ, R) = e(P, R)a e(Q, R)b and
e(R, aP + bQ) = e(R, P )a e(R, Q)b .
2. Non-degenerate: There exists a point P G1 such that e(P, P ) = 1.
3. Computable: There is an ecient algorithm to compute e(P, Q) for any P, Q G1 .
2.2. Computational assumptions.
Definition 2.1 (OHF assumption). A secure one-way hash function H operates on an
arbitrary length input m and outputs the fixed length y = H(m) such that (i) Given m, it
is easy to compute y = H(m). (ii) Given y, it is computationally infeasible to derive m
such that y = H(m). (iii) Given m, it is computationally infeasible to find m (m = m)
such that H(m ) = H(m). (iv) It is computationally infeasible to find m and m such that
H(m) = H(m ).
Definition 2.2 (WDHP). Let G1 be an additive group of the prime order q. Given
(P, aP, W ) for P, W G1 and a Zq , it is computationally infeasible to compute aW .
Definition 2.3 (k-CAA). Let G1 be an additive group of the prime order q. Given
(P, yP, h11+y P, h21+y P, , hk1+y P ) for P G1 , y Zq , and h1 , h2 , , hk Zq , it is
computationally infeasible to derive h+y 1
P for some h h1 , h2 , , hk .
2.3. Formal definitions and security model. A secret key traitor tracing scheme with
dispute settlement is comprised of the following algorithms:
Setup: takes a security parameter and returns system parameters and a secret key
SK.
Extract: takes system parameters, a secret key SK and an arbitrary identity. This
algorithm is used to generate a personal key Ki that is given to subscriber i associated
with the identity ui .
Encrypt: takes a secret key SK, a random encryption key s and a message M . This
algorithm is used to encrypt M to generate a broadcast message, consisting of an
enabling-block EBK and a cipher-block CBK. Note that there exists coherent rela-
tionship between M and its corresponding EBK.
Decrypt: takes a broadcast message (EBK, CBK ) and a personal key Ki for ui . If the
coherent relationship between M and the corresponding EBK is correct, it outputs
a random encryption key s and a message M ; otherwise .
Trace: takes a secret key SK and a pirate personal key K and outputs a set of malicious
subscribers (also called traitors).
Settlement: takes a secret key SK and a broadcast message (EBK, CBK ) and out-
puts True or False.
Security requirements of traitor tracing schemes with dispute settlement can be defined
as collusion-resistance, falsification-free and dispute-settlement, and exposure-resistance
below.
1. Collusion-resistance: Let n and k denote respectively as the total number of
subscribers and the number of malicious subscribers, where k n. According to the
property of collusion-resistance defined by Boneh et al. [11], a traitor tracing scheme
is said to be k-collusion resistant if the tracing algorithm is able to identify malicious
traitors from a pirate decoder which is collusively made by k or fewer traitors. The
notion of collusion-resistance against k traitors is defined by Game 1 played between
an adversary A and a challenger C.
Game 1
Start: For a given security parameter , the challenger C runs Setup to generate
system parameters and a secret key SK. Then, C gives the adversary A the
system parameters and keeps SK secret.
Query: A issues a number of personal key extract queries with respect to identities
{u1 , u2 , , uk } to C, where k < n. The challenger C responds personal keys Ki s
(1 i k) to A by running Extract.
Collusion: With the knowledge of {K1 , K2 , , Kk }, A decides a set of the col-
luding personal keys to form a pirate personal key K.
Trace: The challenger C runs Trace to obtain a set S.
We say that A wins the game if both of the following conditions hold.
C1: The pirate personal key is said to be valid if it can decrypt all broadcast
message with overwhelming probability , where = 1/f () for some polynomial
f.
C2: The detected set S is either empty, or is not the subset of {K1 , K2 , , Kk }.
2. Falsification-free and dispute-settlement: This game ensures that no one can
falsify an illegal broadcast message to set a trap for a broadcast center. We define
falsification-free and dispute-settlement by Game 2 played between A and C below.
Game 2
Phase 1: A issues a number of queries to C. Three types of queries are as follows:
-Extract query: A submits an identity to C. Then, C responds a personal
key by running Extract.
-Encrypt query: A submits a message M to C. Then, C responds a broadcast
message (EBK, CBK ) by running Encrypt.
-Decrypt query: A submits a broadcast message (EBK, CBK ) to C. Then,
C responds a random encryption key and a message M by running Decrypt.
Challenge: Once Phase 1 is finished, A decides a target message M dierent
from the messages M s as in Phase 1.
Phase 2: A issues more queries as in Phase 1. It is not allowed to submit an
Encrypt/Decrypt query for the target message M .
Output: Finally, A outputs a broadcast message (EBK , CBK ) for the target
message M . If the coherent relationship between M and the corresponding
EBK is correct, A wins the game. That is, A successfully forges an illegal
broadcast message for M . Otherwise, falsification-free requirement is achieved.
A secret key traitor tracing scheme with dispute settlement is falsification-free

and dispute-settlement if the successful probability of any polynomially bounded
adversary in the above game is negligible. That is, a broadcast message for the
target message cannot be forged by any polynomially bounded adversary. In case of
a dispute, the broadcast message can be used to settle the dispute.
3. Exposure-resistance: According to evil manufactures problems [15], a malicious
subscriber may abuse the key assigned to him and cause all other personal keys to
be exposed. We define exposure-resistance by Game 3 played between A and C.
Game 3
Phase 1: A issues a sequence of Extract, Encrypt and Decrypt queries to C.
These are answered by C as in Game 2.
Challenge: Once Phase 1 is finished, A decides a target identity dierent from
the identities as in Phase 1.
Phase 2: A issues more queries as in Phase 1. It is not allowed to submit an
Extract query for the target identity.
Output: Finally, A outputs a personal key for the target identity. If the personal
key is valid, A wins the game. That is, A successfully causes the other personal
key to be exposed. Otherwise, exposure-resistance requirement is achieved.
3. The Proposed Scheme. For simplicity, the following symbols are defined in advance:
p, q: are prime numbers and q < p.
G1 , G2 : are two cyclic groups of order q, where G1 is an additive group and G2 is a
multiplicative group.
e : is a bilinear map defined as e : G1 G1 G2 .
H0 , H1 : are secure one-way hash functions, where H0 : G1 Zq and H1 : {0, 1}
G1 .
Elaborating on Mitsunari et al.s traitor tracing scheme [19] and Boneh et al.s short
signature scheme [32], we apply short signature to provide evidence for dispute settlement
dispute settlement in a secret key traitor tracing scheme. The proposed scheme consists of
five phases: Initialization, Distributing, Decoding, Traitor Detection and Dispute
Settlement. The tasks performed in Initialization, Distributing, Decoding and Traitor
Detection are like those in the Mitsunari et al.s scheme [19]. Additionally, in Initialization
and Dispute Settlement phases, an adjudicator or a trusted third party (or TTP for short)
is required in the case that a malicious subscriber has falsified an unintended content for
framing the broadcast center.
Denote BC as the identity of broadcast center and U as the set of subscribers in the
broadcast system. Without loss of generality, let ui Zq be the identity associated to
the i-th subscriber in U . Each subscriber possesses a decoder designated in hardware
box which stores the owners identity, personal key, and public functions. The decoder is
equipped with the computation ability to reconstruct an encryption key from an enabling-
block (or EBK for short) and then to obtain an intended content from the cipher-block (or
CBK for short). We describe the Initialization, Distributing, Decoding, Traitor Detection
and Dispute Settlement phases in detail as follows.
Initialization Phase First, TTP runs Setup to generate system parameters (p,
q, G1 , G2 , H0 , H1 ) and to choose a secret key SK = (a, P ) for BC, where a is a ran-
dom number in Zq and P is an element in G1 . For the purpose of dispute settlement,
TTP computes authentic information for a broadcast center by computing Authcode =
H1 (BC||a||H0 (P )), and stores Authcode in the directory. Then, TTP runs Extract to
compute a personal key Ki for a subscriber i with an identity ui , where
Qi = H1 (ui ) (1)
and
Ki = (ui + H0 (Qi ) + a)1 P. (2)
After that, TTP stores (ui , Ki , H0 , H1 ) in a decoder and delivers the decoder to the
subscriber i. Note that TTP serves as a key generation center in this phase and as an
adjudicator in the subsequent Dispute Settlement phase.
Distributing Phase Let M be an intended content to be broadcasted by BC at
a timestamp t. To do that, the broadcast center first chooses a random encryption key
s G2 and then runs Encrypt to construct an enabling-block EBK = [V, R, P1 , t] and
cipher-block CBK = [Es (M )], where
P1 = H1 (M ||t), (3)
V = s e(tP, P1 ), (4)
and
R = aP1 . (5)
Afterwards, BC broadcasts (EBK, CBK ) to all subscribers.
Decoding Phase Upon receiving (EBK, CBK), a decoder of an authorized sub-
scriber i triggers Decrypt to perform the following tasks:
1. Reconstruct the decryption key s from EBK with the personal key Ki by computing:
s = V e(tKi , (ui + H0 (Qi ))P1 + R)1 . (6)
2. Use s to obtain M = Ds (CBK) by decrypting CBK.
3. Verify if P1 = H1 (M t). If the equality holds, then outputs s and M ; otherwise .
Traitor Detection Phase In traitor tracing schemes, the trace algorithm works
based on the assumption that the suspected pirate decoders are confiscated before per-
forming the trace algorithm [8]. Such an assumption is feasible in real world. In practice,
it can be done by some non-technical methods, such as some intellectual property pro-
tection mechanisms. Anyone can report a pirate of music, software or movies to the
authorities and the authorities can seize the pirate objects. Our proposed scheme focuses
on the tracing technique based on the same assumption. Suppose that a confiscated pirate
personal key from a pirate decoder is collusively made by k traitors. BC first retrieves a
set of suspected personal keys from the pirate decoder, BC runs Trace to identify possible
traitor(s) and then verifies Equations (1) and (2) by the secret key SK = (a, P ). If the
two equations hold, then the suspected personal key(s) used to form the pirate personal
key is identified. The subscriber(s) with the identified personal key(s) are traitor(s).
Dispute Settlement Phase Suppose that a malicious subscriber has falsified an
unintended content M at a timestamp t for framing BC. For settling such a dispute,
TTP requires BC to present his/her secret key (a, P ), a suspected broadcast message
consisting of EBK = [V, R, P1 , t ] and CBK = [Es (M )], and the intended personal key via
a secure channel. TTP first verifies the correctness of (a, P ) using Authcode stored in the
directory. Then, TTP runs Settlement to compute P1 = H1 (M t ) and verify Equation
(5). If Equation (5) holds, then TTP concludes that BC cannot deny broadcasting M .
Otherwise, M is not broadcasted by BC.
4. Security and Eciency Analysis. In Section 4.1, we will show that the pro-
posed scheme achieves collusion-resistance, falsification-free and dispute-settlement, and
exposure-resistance. We brief the eciency of the proposed scheme in Section 4.2.
4.1. Security proof. The security of the proposed scheme is based on the intractability
of the one-way hash function (OHF) assumption [29], the weak Die-Hellman problem
(WDHP) [30] and the collusion attack algorithm with k-traitor (k-CAA) [26, 31].
Correctness. The proposed secret key traitor tracing with dispute settlement is correct
if a decoder for an authorized subscriber i can correctly reconstruct an encryption key s
from EBK by Equation (6), which is identical to that chosen by BC.
Proof: From Equations (2), (4) and (5) we have
s= V e(tKi , (ui + H0 (Qi ))P1 + R)1
= s e(tP, P1 ) e(tKi , (ui + H0 (Qi ))P1 + R)1
= s e(tP, P1 ) e(t(ui + H0 (Qi ) + a)1 P, (ui + H0 (Qi ))P1 + aP1 )1
= s e(tP, P1 ) e(t(ui + H0 (Qi ) + a)1 P, (ui + H0 (Qi ) + a)P1 )1
= s e(tP, P1 ) e(tP, P1 )1 .
Theorem 4.1. The proposed secret key traitor tracing scheme with dispute settlement is
said to achieve collusion-resistance if there exists no polynomial-time algorithm that solves
the diculty of the OHF assumption with non-negligible probability.
Proof: We build a polynomial-time adversary A that wins the collusion-resistance
game with non-negligible probability , where = 1/f () for some polynomial f . Consider
the following conditions.
C1: After queries of personal key extract, the adversary can obtain k personal keys
Ki s for 1 i k. By the same approach demonstrated by [25], A produces a pirate
personal key that can be used to decrypt all broadcast messages. With the knowledge
of K1 , K2 , , Kk , A can produce the following pirate personal keyK
= (K0p , K1p ),
where
K0p = 1 K1 + + k Kk (7)
K1p = 1 (u1 + H0 (Q1 ))K1 + + k (uk + H0 (Qk ))Kk (8)
C2: A necessary condition for a trace algorithm is to satisfy properties of non-ambiguity
of collusion and key-user correspondence [16]. First of all, we consider non-ambiguity
of collusion. From K0p , the trace algorithm may detect one or more sets that satisfy
Equation (7) with at least probability q1n .
Secondly, colluding personal keys and their corresponding identities must be ex-
actly traced to satisfy key-user correspondence. From K1p , the trace algorithm checks
the identities and the corresponding hashed values that satisfy Equation (8) with
probability 1 1 , where 1 is the probability that successfully solving the OHF
assumption.
Let Pr[A wins Game 1] be the probability that the adversary A wins the collusion-
resistance game. ( )
1 1
Pr[A wins Game 1] = 1 n + n 1
q q
Since happens with non-negligible probability, we can see that the probability 1 is also
with non-negligible advantage. This contradicts the fact that solves the OHF assumption
is computationally infeasible.
Theorem 4.2. The proposed scheme is falsification-free if there is no polynomial-time
algorithm that solves the WDHP with non-negligible probability.
Proof: Let A be a polynomial-time algorithm that outputs an existential forgery with
probability . Let qH0 , qH1 , qext , qenc and qdec denote the number of queries to H0 , H1 ,
Extract, Encrypt and Decrypt oracles, respectively. In the following, we demonstrate

how to use the algorithm A to construct a polynomial-time algorithm 1 that outputs the
solution of the WDHP with probability 2 .
Start: For a given security parameter , the algorithm 1 runs Setup to generate sys-
tem parameters (p, q, G1 , G2 , H0 , H1 ) and a secret key SK = (a, P ). Let (P1 , aP1 , W )
be an instance of the WDHP, where P1 , W G1 . Then, 1 gives the algorithm A
the system parameters and keeps (a, P ) secret. Here, H0 and H1 are random oracles
controlled by 1 .
Phase 1: A issues a number of queries to 1 . Types of queries are described as follows:
Q1: H0 -query: A can query a random oracle H0 at any time. 1 maintains a list
list0 to respond to As queries. Upon receiving a H0 -query for some Qi , the
algorithm 1 performs the following:
1. Check list0 . If there exists (Qi , ri ) in list0 , it returns ri and terminates the
simulation.
2. Choose a random number ri Zq . If there exists ri in list0 , repeats Step 2.
3. Put (Qi , ri ) in list0 .
Q2: H1 -query: A can also query a random oracle H1 at any time. 1 maintains
another list list1 to respond to As queries. Upon receiving a H1 -query for some
message Mi and ti , the algorithm 1 performs the following:
1. Check list1 . If there exists (Mi , ti , H1 (Mi ti ), ci , bi ) in list1 , it returns
H1 (Mi ti ) and terminates.
2. Generate a random number ci Zq .
3. Flip a random coin bi with probability P r[bi = 1] = .
If bi = 0
Compute and return H1 (Mi ti ) = ci P1 + W
else
Compute and return H1 (Mi ti ) = ci P1
4. Put (Mi , ti , H1 (Mi ti ), ci , bi ) in list1
Q3: Extract query: A submits an identity ui to 1 . The algorithm 1 responds
the personal key Ki = (ui + H0 (Qi ) + a)1 P for ui by running Extract.
Q4: Encrypt query: A submits a message Mi and ti to 1 . Then, the algorithm
1 performs the following:
1. Run the above algorithm for simulating H1 -oracle to get (Mi , ti , H1 (Mi ti ), ci ,
bi ).
2. If bi = 0
Abort the simulation
else
Compute Ri = ci aP
3. Choose an encryption key si and compute Vi = si e(tP, ci P1 )
4. Return (si e(tP, ci P1 ), ci aP1 , ci P1 , ti ) to A.
(si e(tP, ci P1 ), ci aP1 , ciP1 , ti ) is a valid enabling block since any authorized sub-
scriber i can recover the decryption key s by using his personal key Ki , where
s = Vi e(tKi , (ui + H0 (Qi ))ci P1 + ci aP1 )1 . Hence, 1 returns a valid broadcast
message for Mi to A with successful probability .
Q5: Decrypt query: A submits a broadcast message (EBKi , CBKi ) to 1 , where
EBKi = (si e(tP, ci P1 ), ci aP1 , ci P1 , ti ) and CBKi = Es (Mi ). The algorithm 1
recovers a random encryption key si and a message Mi by running Decrypt. If
ci P1 = H1 (Mi ti ), 1 returns si and Mi to A. Otherwise, it outputs .
Challenge: A decides a target message M which is dierent from the messages Mi s
as in Phase 1.
Phase 2: A issues more queries as in Phase 1. It is not allowed to submit an En-
crypt/Decrypt query for the target message M .
Output: Suppose that A successfully outputs a broadcast message (EBK , CBK )
for M with probability . If (EBK , CBK ) is a valid broadcast message, there
exists coherent relationship between M and the corresponding EBK such that
H1 (M t ) = c P1 + W and R = aH1 (M t ). Substituting H1 (M t ) into R , we
can have R = ac P1 + aW , which implies aW = ac P1 R . Hence, 1 can carefully
derive the solution aW of the WDHP.
During the simulation, 1 does not abort the game with the probability qenc . At the
end of simulation, A successfully outputs the broadcast message for the target message
M with the probability . Finally, 1 can transform As forgery into the solution of the
WDHP with the probability at least (1 ).
We can conclude that 1 can solve the WDHP with the probability 2 qenc (1 ).
Theorem 4.3. The proposed scheme is exposure-resistant if there is no polynomial-time
algorithm that solves the k-CAA with non-negligible probability.
Proof: We suppose that H0 and H0 are random oracles, and there exists a polynomial-
time algorithm A that wins the exposure-resistance game with non-negligible probabil-
ity . Let qH0 , qH1 , qext , qenc and qdec denote the numbers of queries to H0 , H1 , Ex-
tract, Encrypt and Decrypt oracles, respectively. In the following, we will demonstrate
how to use A to construct a polynomial-time algorithm 2 that solves the k-CAA. Let
(P, yP, h11+y P, h21+y P, , hk1+y P ) be a random instance of the k-CAA. The goal of 2 is
to derive h1+y P for some h h1 , h2 , , hk .
Start: For a given security parameter , the algorithm 2 runs Setup to generate
system parameters (p, q, G1 , G2 , H0 , H1 ) and a secret key SK = (a, P ). Then, 2
gives the algorithm A the system parameters and keeps (a, P ) secret.
Phase 1: A issues a number of queries as in Theorem 4.2.
Challenge: A decides a target identity u dierent from the issued identities in Phase
1.
Phase 2: A issues more queries as in Phase 1. It is not allowed to submit an Extract
query for the target identity u .
Output: After Phase 2, A can get k personal keys Ki = (ui +H0 (Qi )+a)1 P , where k
is the number of Extract queries and 1 i k. Suppose that A successfully outputs
a personal K = (u + H0 (Q ) + a)1 P for u with probability . Let ui + H(Qi ) and
a be hi and y, respectively. The personal keys Ki s can be rewritten as hi1+y P for
1 i k. Hence, K = (u + H0 (Q ) + a)1 P = h1+y P is a solution to the k-CAA.
After the simulation, 2 does not abort the game with the probability qenc as in Theo-
rem 4.2. At the end of simulation, A successfully outputs the personal key for the target
identity u with the probability . Finally, 2 can obtain the solution to the k-CAA with
at least (1 ).
We can conclude that 2 can solve the k-CAA with the probability 3 qenc (1 ).
4.2. Eciency analysis. A comparison of the proposed scheme with previous works
is shown in Table 1. Suppose that the size of the plaintext is , where is the secu-
rity parameter. Let t denote the number of traitors and n denote the total number of
subscribers. The rates of the ciphertext and the key are defined as ciphertext/plaintext
and key/plaintext, respectively. According to Table 1, the sizes of the ciphertext, the
Table 1. Comparison of the proposed scheme with previous work
Ciphertext Personal Key Encryption Max Traceable

Rate Rate Key Rate Collusion
Chor et al.s scheme [8] O(t2 log n) O(t2 log n) O(t2 log n) t
Mitsunari et al.s scheme [19] O(1) O(1) O(1) 1

Boneh et al.s scheme [11] O( n) O(1) O( n) n
Boneh and Naors scheme [10] O(1) O(t log n) O(t log n)
2 2
t
The proposed scheme O(1) O(1) O(1) t
personal key, and the encryption key are independent of the number of traitors and sub-
scribers in the proposed scheme and Mitsunari et al.s scheme [19]. However, Mitsunari
et al.s scheme cannot resist a collusion attack. An important advantage of the proposed
scheme is that the lengths of the broadcast message and the personal/encryption key
are independent of the number of subscribers. That is, the broadcast message and the
personal/encryption key sizes in the proposed scheme are shorter than those in previous
works. Furthermore, the proposed scheme will work as long as the pirate decoder has
fewer than t personal keys.
5. Conclusions. This paper has presented the first secret key traitor tracing scheme
with dispute settlement that achieves the security requirements of collusion-resistance,
falsification-free and exposure-resistance. Note that in most of the previously proposed
traitor tracing schemes, the enabling-block is independent of the intended content. This
may cause a dispute in the case that any malicious subscriber may have the ability to
falsify an unintended content for framing the broadcast center. In the proposed scheme, a
coherent relationship between the intended content and the corresponding enabling-block
is made up, and this provides robust evidence for dispute settlement by the adjudica-
tor in the system. However, making up such coherent relationship between the intended
content and the corresponding enabling-block just takes a few computational and com-
municational overheads by introducing P1 in EBK.
Acknowledgement. This work was supported in part by Taiwan Information Security

Center (TWISC), National Science Council under the grants NSC 97-2219-E-001-001,
NSC 97-2219-E-011-006, NSC 96-2221-E-011-148-MY2 and NSC 97-2221-E-182-032.
REFERENCES
[1] A. Fiat and M. Naor, Broadcast encryption, Advances in Cryptology CRYPTO93, LNCS, vol.773,
pp.480-491, 1993.
[2] C. L. Chen, An all-in-one mobile DRM system design, International Journal of Innovative Com-
puting, Information and Control, vol.6, no.3(A), pp.897-911, 2010.
[3] T.-H. Chen, T.-H. Hung, G. Horng and C.-M. Chang, Multiple watermarking based on visual
secret sharing, International Journal of Innovative Computing, Information and Control, vol.4,
no.11, pp.3005-3026, 2008.
[4] M. Gulbis, E. M uller and M. Steinebach, Content-based audio authentication watermarking, In-
ternational Journal of Innovative Computing, Information and Control, vol.5, no.7, pp.1883-1892,
2009.
[5] H. Luo, F.-X. Yu, S.-C. Chu and Z.-M. Lu, Hiding multiple watermarks in transparencies of visual
cryptography, International Journal of Innovative Computing, Information and Control, vol.5, no.7,
pp.1875-1881, 2009.
[6] S. Weng, Y. Zhao and J.-S. Pan, A novel reversible data hiding scheme, International Journal of
[7] F.-X. Yu and Z.-M. Lu, A DCT-VQ based multipurpose image hashing scheme for copyright pro-
tection and content authentication, International Journal of Innovative Computing, Information
[8] B. Chor, A. Fiat and M. Naor, Tracing traitor, Advances in Cryptology CRYPTO94, pp.257-270,
1994.
[9] D. Boneh and M. Franklin, An ecient public key traitor tracing scheme, Advances in Cryptology
CRYPTO99, pp.338-353, 1999.
[10] D. Boneh and M. Naor, Traitor tracing with constant size ciphertext, Proc. of the 15th ACM
Conference on Computer and Communications Security, pp.501-510, 2008.
[11] D. Boneh, A. Sahai and B. Waters, Fully collusion resistant traitor tracing with short ciphertexts
and private keys, Advances in Cryptology EUROCRYPT06, pp.573-592, 2006.
[12] H. Chabanne, D. H. Phan and D. Pointcheval, Public traceability in traitor tracing schemes, Ad-
vances Cryptology EUROCRYPT05, pp.542-558, 2005.
[13] M. Dobashi, H. Nakayama, N. Kato, Y. Nemoto and A. Jamalipour, Traitor tracing technology
of streaming contents delivery using trac pattern in wired/wireless environments, Proc. of IEEE
Global Telecommunications Conference, pp.1-5, 2006.
[14] A. Fiat and T. Tassa, Dynamic traitor tracing, Journal of Cryptology, vol.14, no.3, pp.211-223,
2001.
[15] H. Jin and J. Lotspiech, Hybrid traitor tracing, Proc. of IEEE International Conference on Multi-
media and Expo, pp.1329-1332, 2006.
[16] A. Kiayias and M. Yung, Self protecting pirates and black-box traitor tracing, Advances in Cryp-
tology CRYPTO01, pp.63-79, 2001.
[17] A. Kiayias and M. Yung, Traitor tracing with constant transmission rate, Advances in Cryptology
EUROCRYPT02, pp.450-465, 2002.
[18] K. Kurosawa and Y. Desmedt, Optimum traitor tracing and asymmetric schemes with arbiter,
Advances in Cryptology EUROCRYPT98, pp.145-157, 1998.
[19] S. Mitsunari, R. Sakai and M. Kasahara, A new traitor tracing, IEICE Trans. on Fundamentals,
vol.E85-A, no.2, pp.481-484, 2002.
[20] M. Naor and B. Pinkas, Threshold traitor tracing, Advances in Cryptology CRYPTO98, pp.502-
517, 1998.
[21] K. Ogawa, G. Hanaoka and H. Imai, Traitor tracing scheme secure against adaptive key exposure
and its application to anywhere TV service, IEICE Trans. on Fundamentals, vol.E90-A, no.5,
pp.1000-1011, 2007.
[22] B. Pfitzmann, Trials of traced traitors, Proc. of the 1st Information Hiding Workshop, pp.49-64,
1996.
[23] R. Safavi-Naini and Y. Wang, Sequential traitor tracing, IEEE Trans. on Information Theory,
vol.49, no.5, pp.1319-1326, 2003.
[24] D. Stinson and R. Wei, Combinatorial properties and constructions of traceability schemes and
frameproof codes, SIAM Journal on Discrete Mathematics, vol.11, no.1, pp.41-53, 1998.
[25] V. D. To, R. Safavi-Naini and F. Zhang, New traitor tracing schemes using bilinear map, Proc. of
the 3rd Digital Rights Management Workshop, pp.67-76, 2003.
[26] C. Yang, W. Ma and X. Wang, New traitor tracing scheme against anonymous attack, Proc. of the
1st International Conference on Innovative Computing, Information and Control, Beijing, China,
pp.389-392, 2006.
[27] B. Zhu, M. Feng, F. Liu and L. Hu, Analysis on AACS traitor tracing against mix-and-match
attacks, Proc. of the 5th IEEE Consumer Communications and Networking Conference, pp.1097-
1101, 2008.
[28] M. Bellare and P. Rogaway, Random oracles are practical: A paradigm for designing ecient
protocols, Proc. of the 1st ACM Conference on Computer and Communications Security, pp.62-73,
1993.
[29] A. J. Menezes, P. C. van Oorschot and S. A. Vanstone, Handbook of Applied Cryptography, CRC
Press, Boca Raton, 1996.
[30] M. Zhang, G. Chen and J. Li, Ecient ID-based proxy chameleon signature from bilinear pairings,
Proc. of the 1st International Multi-Symposiums on Computer and Computational Sciences, vol.2,
pp.135-141, 2006.
[31] R. Dutta, R. Barua and P. Sarkar, Pairing-Based Cryptographic Protocols: A Survey, http://eprint.
iacr.org/2004/064.pdf, 2004.
[32] D. Boneh, B. Lynn and H. Shacham, Short signatures from the Weil pairing, Advances in Cryptology
ASIACRYPT01, pp.514-532, 2001.
Observer-Based Iterative Learning Control with Evolutionary Programming Algorithm

for MIMO Nonlinear Systems 1357
Yan-Yi Du, Jason Sheng-Hong Tsai, Shu-Mei Guo, Te-Jen Su and Chia-Wei Chen
Reversible Data Hiding for VQ Indices Using Prediction Errors 1375
Zhi-hui Wang, Chin-Chen Chang, Huynh Ngoc Tu and Ming-Chu Li
A Novel Technique for Designing Decentralized Stabilizers for Robust Control
in Power Systems Using an H Criterion 1387
Alexandre C. Castro, Jose M. Araujo, Eduardo T. F. Santos, Fabricio G. S. Silva
and Clivaldo S. de Araujo
A Study of Three Novel Line-Based Techniques for Multi-Target Selection 1397
Jibin Yin and Xiangshi Ren
A Novel Modal Series Representation Approach to Solve a Class of Nonlinear
Optimal Control Problems 1413
Amin Jajarmi, Naser Pariz, Ali Vahidian Kamyad and Sohrab Effati
Adaptive Frame Length Method for Hardware Context-Switching in Dynamic Partial
Self-Reconfigurable Systems 1427
Trong-Yen Lee, Che-Cheng Hu, Yang-Kun Huang and Chia-Chun Tsai
An ID-Based Access Control in a Hierarchical Key Management for Mobile Agent 1443
Chia-Hui Liu, Yu-Fang Chung, Tzer-Shyong Chen and Sheng-De Wang
Meaningful Shadow Based Multiple Gray Level Visual Cryptography without Size
Expansion 1457
Pei-Yan Pai, Chin-Chen Chang, Yung-Kuan Chan and Chi-Cheng Liao
Service-Oriented Routing and Clustering Strategies for Vehicle Infotainment
Dissemination 1467
Chenn-Jung Huang, Chin-Fa Lin, Ching-Yu Li, Che-Yu Lee, Heng-Ming Chen
Hung-Yen Shen, You-Jia Chen and I-Fan Chen
An Optimal Fuzzy Controller for a High-Performance Drilling Process Implemented
Over an Industrial Network 1481
Rodolfo E. Haber, Agustin Gajate, Steven Y. Liang, Rodolfo Haber-Haber
and Raul M. del Toro
New Secret Key Traitor Tracing Scheme with Dispute Settlement from Bilinear Maps 1499
Kuo-Yu Tsai, Tzong-Chen Wu and Chien-Lung Hsu

An Efficient Simulated Annealing with a Valid Solution Mechanism for TDMA

Broadcast Scheduling Problem 1181
Gon Kim, Sung Soo Kim, Il-Hwan Kim, Dong Hoi Kim, V. Mani and Jae-Ki Moon
Scalable and Systolic Dual Basis Multiplier over GF(2m) 1193
Liang-Hwa Chen, Po-Lun Chang, Chiou-Yng Lee and Ying-Kuei Yang
A Fuzzy-Based Rough Sets Classifier for Forecasting Quarterly PGR in the Stock
Market (Part II) 1209
Ching-Hsue Cheng and Yoush-Yang Chen
Re-Evaluation of Adaptive X Control Charts: A Cost-Effectiveness Perspective 1229

Yan-Kwang Chen, Hung-Chang Liao and Hsu-Hwa Chang
Inventory Management in a (Q,r) Inventory Model with Two Demand Classes
and Flexible Delivery 1243
A Novel Text Watermarking Algorithm Using Image Watermark 1255
An Efficient and Secure Three-Pass Authenticated Key Agreement Elliptic Curve
Based Protocol 1273
Zeyad Mohammad, Chien-Lung Hsu, Yaw-Chung Chen and Chi-Chun Lo
Indicator of the Quality of the Sensor Set Up: A Study Using Surface EMG
on Sub-Band ICA 1285
Ganesh R Naik and Dinesh K Kumar
H Filtering for Markovian Jump Systems with Time-Varying Delays 1299
Jinliang Liu, Zhou Gu and Songlin Hu
Surface Texture Characterization of Fibers Using Fractional Brownian Motion Model 1311
Jiun-Jian Liaw, Chuan-Pin Lu and Lin-Huang Chang
Power Grid Node and Line Delta Centrality Measures for Selection of Critical Lines
in Terms of Blackouts with Cascading Failures 1321
Hwachang Song, Rodel D. Dosano and Byongjun Lee
Optimizing MIMO-SDMA Smart Antennas by Using Linear Array Phase Perturbations
Based on Particle Swarm Optimization 1331
Chao-Hsing Hsu
An Environmental Visual Features Based Navigation Method for Autonomous
Mobile Robots 1341
Fairul Azni Jafar, Yasunori Suzuki, Yuki Tateno, Kazutaka Yokota
and Takeshi Matsuoka
(Continued)

INFORMATION FOR AUTHORS
1. Submissions. The International Journal of Innovative Computing, Information and Control
(IJICIC) is welcoming contributions related with the journal's scope. In general, decisions will
be made after peer-review, within 3-5 months of receipt. Accepted articles will generally be
published within 18 months thereafter.
Manuscripts (in a single PDF format only, no more than 3MB size) should be submitted to
the IJICIC Editorial Office by email: office@ijicic.org.
2. Manuscript Layout. Manuscripts should be typewritten in English on one side only with
wide margins on high quality paper using double spacing throughout. Each page of the
manuscript should be numbered. The front page should contain the article title, author's name
and affiliation, abstract and keywords. The abstract should not exceed 200 words and should
be intelligible to general readers without recourse to the main text. A list of notation should be
avoided, but new notation and departures from standard use should be explained in text.
Articles should, in general, begin with an introduction which adequately covers the
literature and scope of the paper. Each section in the paper should be numbered. Each article
should have, in general, a conclusion or discussion section. Any appendices should follow the
Reference section.
3. Illustrations. Proper figure preparation is essential for optimum reproduction of line
drawing and photograph. Figures and Tables should be placed as part of the text and should
be numbered. Reprints, if needed, should be ordered as soon as the article is accepted for
publication.
4. References. References should be listed at the end of the main text. All in-text citations
(references) should be listed in order, that is, [1], [3,4], [6-9], etc. In the References, each
citation should be written in the order of appearance in the text. The normal form of listed
references is author's initials, surname; article title; journal name (abbreviated in accordance
with the World List of Scientific Periodicals, 4th edn); volume number; issue number; page
numbers; year.
5. Copyright/Offprint. Manuscript submitted to this journal signifies that it has not been
published previously, has not been copyrighted, is not been submitted elsewhere and the
author(s) agrees that the copyright for article will be transferred to the publisher upon
acceptance of the article.
After a manuscript has been accepted for publication, the author's institution or company
will be approached with a kind request to pay a reasonable charge to cover part of the cost of
publication. The author will receive a free copy of the journal with a CD containing his/her
paper after the page charge is honored. Detailed instructions for the payments will accompany
the proofs. Further offprint can be ordered at extra cost at the proof stage.
Acceptance of LaTex by the IJICIC. If accepted papers written by LaTex, the authors will
be asked to provide the LaTex source files of the paper. In preparing the paper, please use the
attached IJICIC style files for submission and ensure that all the macros used to write your
paper accompany the manuscript. If you have any questions about the use of this file in
preparing your paper, please contact the editorial office. Please follow all the usual
instructions to authors as detailed above. It is very important to follow the correct referencing
system.

Ijicic Complte Book

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ijicic Complte Book

Uploaded by

Copyright:

Available Formats

International Journal of

Innovative Computing, Information and Control

2011 ICIC INTERNATIONAL ISSN 1349-4198 PRINTED IN JAPAN

IJICIC Volume 7, Number 3, March 2011

International Journal of Innovative

Sponsored by Science Communication Institute, China

Published by ICIC International

The Improvement of all-Digital Amplitude-Locked Loop Separation Analysis

2011 ICIC INTERNATIONAL ISSN 1349-4198 PRINTED IN JAPAN

An Efficient Simulated Annealing with a Valid Solution Mechanism for TDMA

2011 ICIC INTERNATIONAL ISSN 1349-4198 PRINTED IN JAPAN

Observer-Based Iterative Learning Control with Evolutionary Programming Algorithm

2011 ICIC INTERNATIONAL ISSN 1349-4198 PRINTED IN JAPAN

AIMS AND SCOPE OF IJICIC

1. Intelligent Systems and Innovative Computing

2. Information and Control

IJICIC Editorial Office

2011 ICIC INTERNATIONAL ISSN 1349-4198 PRINTED IN JAPAN

THE IMPROVEMENT OF ALL-DIGITAL AMPLITUDE-LOCKED

Gwo-Jia Jong and Gwo-Jiun Horng

Received July 2009; revised June 2010

A. Basic Model of All-Digital Phase-Locked Loop. A block diagram of ADPLL

Figure 1. Block diagram of ADPLL

Figure 2. Phase model of ADPLL

Vo (k) = Vo cos [0 T k + o (k)] , k = 0, 1, 2, . . . (2)

3. All-Digital Amplitude-Locked Loop of CCI.

A. Frequency Modulation (FM). A narrowband digital FM with frequency detection

B. Co-channel Interference under AWGN Channel. Typical methods for reducing

Figure 3. Phase diagram of phase modulation with noise

b(t) mAc sin d t m sin d t

Figure 4. ADPLL FM model with noise

Figure 5. System-level depiction of ALL application

d(t) rn (t) sin n t

C. ADALL Analysis in the Discrete-Time. The conventional amplitude-locked loop

Figure 6. Frequency-domain analysis model of ADALL

It is demodulated by an ideal ADPLL demodulator for the output signal fADP LL ( ).

The subdominant signal is derivative as follows:

4. Multi-Channel FM Signals with Multi-Input Multi-Output (MIMO) Model

given in accordance with as:

5. Simulated Performance. The CCI problem could be become a serious impairment

Figure 8. The separation dominant and subdominant signals for adaptive

Figure 9. The performance of MMSE via SNR ratio

(a) Proposed method for ADPLL and ADALL

(b) Traditional method for PLL and ALL

wd log (eiwd + m) log e d 4

VALUE BASED INTELLIGENT REQUIREMENT PRIORITIZATION

Muhammad Ramzan, M. Arfan Jaffar and Arshad Ali Shahid

1. Introduction. Software Engineering is one of the youngest engineering domains which

Requirement prioritization enables us to understand the significance of requirements

The technique generated a lot of conflicts at the end of requirement prioritization

1.1. Major contributions. Following are the major contributions:

3, we have presented an elaborate overview of existing requirement prioritization tech-

3. Requirement Prioritization Techniques: An Overview. As mentioned in the

Figure 1. Sample cost-value diagram for AHP

4. Value Based Intelligent Requirement Prioritization (VIRP): The Proposed

Figure 2. Flow graph of fuzzy based intelligent requirement prioritization process

4.1. Requirement elicitation and stakeholder level prioritization. Requirement

(a) Success in prioritization achieved for dierent process models

Figure 3. Experimental evaluation results of various prioritization techniques

Table 1. Project specific requirement specification factors (pRCF)

Table 2. Requirement specific requirement specification factors (rRCF)

5. Theoretical and Experimental Evaluation of Requirement Prioritization

5.1. Analysis of existing requirement prioritization techniques. Literature oers

5.2. Experimental results for requirement prioritization techniques. The theo-

!""

!

$! +!

. & #'% ( % ) @2

- 0 #/ - 0 * $ + $ * !

7

E8 -0

F ./ $

J $! -0 +

+

/ =+

$ #" # %## $ !"!& !#!! #!#! #&!"

0 3

H :0 J/ H > !:::! FJ H .;; G

H 90 J/ H .;! ! FJ H .9.;

* +! $ $ $ 9 - H 90

$ % ) $

- 0 $

%

! $ !

.

) +

23 1' ' D" ' ' D" *%