Ine eqs ON SYSTEMS MAK AND CYREANEIES NOU 2D.S0, 2 MCHA 1 Fuzzy Logic in Control Systems: Fuzzy Logic Controller, Part II CHUEN CHIEN LEE, srupewr MEER, IEEE Abas — During the pst several years zy cone has emerged a6 nea he ost tives ral areas for rosarch i the applications [St toy set theory, espe nthe ele af dst processes which ‘dnt lend themaees to antral by eomentonal methods eae of & tack of quantiatie data regarding the input-outpt ration. Pur2y ‘condi based on tz lale—a hal stem tht much loner Soin inking and natural language than radon legis {Seem The fry lege contoler(FLCY hase om fa ope provides 4 Imean of converting nguitic conta raey bated om expert kao ge ita am automate contr strategy. A survey ofthe FLC i pre enti «general ethology for contracting bevfermance deere and problems tht eed farther research ate Punt et. In partcalar, tbe expontionincades 8 discussion of Feesifeation and deorscaon state, he derivation of the database Sn fray control rules, the dtiniion of farsy implication a Stays of fas retoning mechanisms 1. Decisionwaxine Locic 'S. WAS noted in Part I of this paper [150}, an FLC may be regarded as a means of emulating a skilled human operator. More generally, the use of an FLC may be viewed as stil another step in the direction of model ing human decisionmaking within the conceptual frame work of fuzzy logic and approximate reasoning. In this context, the forward data-driven inference (generalized ‘modus ponens) plays an especially important role. In what follows, we shall investigate fuzzy implication functions, the sentence connectives and and also, compositional ‘operators, inference mechanisms, and other concepts that ‘are closely related to the decisionmaking logic of an FLC. A. Fuzzy Implication Functions In general, a fuzzy conteol rule isa fuzzy relation whieh js expressed ss a fuzzy implication. In fuzzy logic, there fare many ways in which a fuzzy implication may be defined. The definition of fuzzy implication may be expressed as a fuzzy implication function. The choice of a fuzzy implication function reflects not only the intuitive {or implication but also the effect of connective Mnesrpt reed May 27. 18: revied July 11980, This work support im prt ty NAGA under grunt NCCITS and nya By DPD SI unde gra 3st, “he suthor th the Electonics Research Laborlon, Departmen ‘of Ec! Engineering and Computer Sencen, Unbertyof Calor te erickey Ca S470 i tag Nomes HONS 1) Basie Properies of a Fuzzy Implication Function: The choice of a fuzzy implication function involves number of criteria, which are discussed in (3}, (241, (2) [71], (18), (52), (19), (116), (85), (72), and (96). In particular, Baldwin and Pilsworth [3] considered the following basic characteristics of a fuzzy implication function: fundamen- tal property, smoothness property, unrestricted inference, symmetty of generalized modus ponens and generalized ‘modus tollens, and a measure of propagation of fuzziness. All of these properties are justified on purely intuitive ‘grounds, We prefer to say’ that the inference (conse- uence) should be as close as possible to the input truth function value, rather than be equal to it, This gives us a more flexible criterion for choosing @ fuzzy implication function, Furthermore, in a chain of implications, itis necessary to consider the “fuzzy syllogism” [147] associ- ‘ted with each fuzzy implication function before we can talk about the propagation of fuzziness. Fukami, Mizumoto, and Tanaka [24] have proposed a set of intuitive criteria for choosing a fuzzy implication function that constrains the relations between the an- tecedents and consequents of a conditional proposition, with the latter playing the role of a premise in approxi- mate reasoning. As is well known, there are two impor- tant fuzzy implication inference rules in approximate rea- soning. They are the generalized modus ponens (GMP) land the generalized modus tollens (GMT). Specifically, premise I: x is A premise 2:if xis A (Gop) consequence: y is BY premise I: y is BY premise 2:if x is A then y is B (om) ‘consequence: xis A’ in which A, a’, B, and B' are fuzzy predicates. The propositions” above the line are the premises: and the proposition below the line is the consequence. ‘The pro- posed criteria are summarized in Tables I and I. We fe that if a causal relation between “x is A” and “y is is not strong in a fuzzy implication, the satisfaction of criterion 2-2 and criterion 3.2 is allowed. Criterion 4-2 is incerpreted as: if x is A then y is B, else y is not B. (0018-9472 /90,/0300-0819501.00 ©1990 IEEEIyrurrve Crrena Rotrine Pat a> Cons Fox Given Pat GM TABLE IL Inve Caren Reais PAE Asi Coma eo Given Pa OMT re athe Te RCo Tem) Team Gieiet eA vee Gaenons yaa emt 4 Criterion 24x every 4 yer Gitenon? —}itaotmafeorles Bhat mee or les A Criterion 2-2 x isyery A yi eee ibs eee Citesoas1 —rmeorien A yemacorten s SHER me Criterion 3-2 x more or less A yisB = Hate Ghemeet Fond yi Although this relation is not valid in formal logic, we often make such an interpretation in everyday reasoning. The same applies to criterion 8 2) Families of Fuzzy Implication Functions: Following Zadeh’s (1465) introduction of the compositional rule of inference in approximate reasoning, a number of re- searchers have proposed various implication functions in which the antecedents and consequents contain. fuzzy variables. Indeed, nearly 40 distinet fuzzy implication functions have been described in the literature. In gen eral, they can be classified into three main categories: the fizz) conjunction, the fuzzy disjunction, and the fuzzy implication. Tre former two bear a close relation to @ fuzzy Cartesian product. The latter is a generalization of implication in multiple-valued logic and relates to the extension of material implication, implication in proposi tional calculus, modus ponens, and modus tollens [18]. In what follows, after a short review of triangular norms and triangular co-norms, we shall give the definitions of fuzzy conjunction, fuzzy disjunction, and fuzzy implication. Some fuzzy implication functions, which are often em: ployed in an FLC and are commonly found in the litera- ture, will be derived. Definition I: Triangular Norms: The triangular norm is a two-place function from [0,1]%(0,1] to [0,1] ie. # [0,11%(0,1}+0,1}, which includes intersection, algebraic product, bounded product, and drastic product. The {greatest triangular norm is the intersection and the least fone is the drastic product. The operations associated with triangular norms are defined for all x,y £10, 1 intersection Ayo min(x,y} wyeay Oy = maxl0.2+ xayaly 0 Definition 2: Triangular Co:Nonms: ‘The triangular co: ‘norms + isa two-place function fron (0, 1]%I0, 1] t0 [0 1, ile. +: (011101) to [0,1], which includes union, alge braic sum, bounded sum, drastic sum, and disjoint sum. ‘The operations associated with triangular co-norms are algebraie product bounded product -1) drastic product sel xyel defined forall x,y €(0, 1 union xvy= marty) algebraic sum xtyertyoy bounded sum x@y=min(I,x +9) . yao drastic sum xuyely 2=0 T xy>o disjoint sum say = max(min(a,1~ y), min(= x, ‘The triangular norms are employed for defining conjune- tions in approximate reasoning, while the tiangulat co- rhorms serve the same role for disjunctions. A fuzzy con- trol rule, “if x is A then y is B.” is represented by 2 fuzzy implication function and is denoted by A> B, where A and B are fuzzy sets in universes U and V with ‘membership functions 4 and 4», respectively Definition 3: Fuzzy Conjunction’ The fuzzy conjunction is defined for all weU and vEV by A>B=AXB fi gandhi) where + is an operator representing a triangular norm, Definition 4: Fuzzy Disjunction: The fuzzy disjunction is defined for all weU and v eV by A> B=AXB = fh Helin + al) where + is an operator representing triangular co-norm. Definition 5: Fuzzy Implication: The tuzzy implication is associated with five families of fuzzy implication functions in use. As before, * denotes a triangular norm and ++ is a triangular co-norm. 4.1) Material implication: A> B=(n0t A)+B 42) Propositional calculus: A> B= (not A)+ (Ae B) 43) Extended propositional calculus A> Be (not AXnot B+ B 44) Generalization of modus ponens: A> B=sup(ce [0,1], Aree B)45) Generalization of modus tollens: A> Bm int(ee [0,1], B+rs A) Based on these definitions, many fuzzy implication func- tions may be generated by employing the triangular norms and co-norms. For example, by using the definition ofthe fuzzy conjunction, Mamdani’s mini-fuzzy implication, R, is obtained if the intersection operator is used. Larsen’s product fuzzy implication, R, is obtained if the algebraic product is used, Furthermore, Ry, and Ry, are obtained iF the bounded product and the drastic product are used, respectively. The following fuzzy implications, which are ‘often adopted in an FLC, will be discussed in more detail ata later point Mini-operation rule of fuzzy implication [Mamdani) R.=AXB J Pawn plo) at. Product operation rule of fuzzy implication {Larsen Ryn AX dug) a) ‘Aithmetic rule of fuzzy implication (Zadeh (not AxV) OW xB) TAC = ww) + ago Au, Maxmin rule of fuzzy implication (Zadeh) Ry =(AX BUtnot AXxV) =f, aed ugh fed) Standard sequence fuzzy implication: Ry= AXV >UXB = fp kd mao) where (; salt) < teal) 0 du) > sal. Boolean fuzzy implication: Ry (not Ax VIUW x B) [ft mwdrvaglod9/0u 0). Goguen’s fuzzy implication: Ry=AXVUXB = fede wad where fod > mal 1 du) 0) Ho) We note that Zadeh’s arithmetic rule follows from Definition 5.1 by using the bounded. sum operator Zadeh’s maxmin rule follows from Definition 8.2 by using the intersection and union operators, the standard se- {quence implication follows from Definition 5.4 by using the bounded product; Boolean fuzzy implication follows from Definition 5.1 by using the union; and Goguen’s fuzzy implication follows from Definition 5.4 by using the algebraic product. 3) Choice of a Fuzzy Implication Function: Firs, we investigate the consequences resulting from applying the preceding forms of fuzzy implication in fuzzy inference And, in particular, the GMP and GMT. The inference is based on the sup-min compositional rule of inference. In the GMP, we examine the consequence of the following compositional equation: 8 R where fuzzy implication (elation, ® sup-min compositional operator, A fuzzy set which has the form Am fopese)/u very A= A?= fypiu)/u more oF less A= a= ful) /u not A= ful ps) Similarly, in the GMT, we examine the consequence of ‘he folowing equation ARB where R_ fuzzy implication (relation) BY a fuzzy set that has the form: not B= fyl~ sigue not very B= fl — ns6v/¥ not more or less B= Jul whiv)/v B fy go)/» The Case of R,: Larson's Product Rule: A method for ‘computing the generalized modus ponens and the gener alized modus tollens laws of inference is described in (3. ‘The graphs corresponding to Larsen’s fuzzy implication R,, are given in Fig. 1. The graph with parameter pi used for the GMP, and the graph with .y is used for the GMT. LLarsen’s Product Rule in GMP: Suppose that A'~ A* (ce> 0); then the consequence Bis inferred as follows Bye ATR, sm fpednr/ue fwd malo) C0) ‘The membership function jeg, of the fuzzy set By is pointwise defined forall v= by sup min (45(u), wal )sea( ¥)} sup S,(1~ 24) rnc where SCC) # min (5) Cada) (A' Ak The values of $,(y4,(u)) with a parameter 460), S49 #g(v)= 0.3 and 08, are indicated in Fig. 2by a broken line and dotted line, respectively. The member2 {E25 Yea oe SIREN MAN A CORFARITIOS os os ° age » Fig 1. Diagrams for caclation of membership fenton. (2) jy Nees na nth the parameter ky tb) ay erty wth pram Spi) ue” ae Fe. 2. Approximite_resoning: peneraized mess ponees_ with ares product operton rl ship function jp, is obtained by as(2) = sup min (1.4). s (e909) maw) = 1 Hal), la'~ 4°} The values of $, (aw) with a parameter al) say wg(0) = 03 and 08, re indicated broken line and dotted line, respectively. ship function jp, may be expressed as, Fig. 3bya 16 member aso) = sup min (uu), edial 9) = ual) (A'= AP% The values of $480) with a parame- ter walv), say wy(v) = 0.3 and 0.8, are indicated in Fig. 4 2m. 2aecn/aran 980 Su) 10 ue 10 os os os y "aye Apri ong geet mae Wh se ago os ° + ye Fig 4. Approximate reasoieg: generalized mas ponens with arcs produc operation rs Sit) ° + yw Fig. Approsimate reasoning: generalized mats poncns. with nc’ product operation rule by a broken line and dotted lin, respectively. The mem: bership function jp, is given by dias) ~ sup min 0) ede 9) a) (A'= not AP The valves of Sy(— yeu) with a pa rameter p4(0), Say se,(0)= 03 and 08, are indicated in Fig. 5 by a broken line and dotted line, respectively. The membership function jy, is given by oC sup min (1 sal u) malta) _ #9) Teal)Sift) 10 ten os ° + ade Fir & Approuimate reasoning! generalized modus tllens_ with Laren product operation ule Syrup te os ° + ade Fig 7 Approimate_restonings generized modus sollens wi ‘nse product operation rae Larsen's Product Rule in GMT: Suppose that B’= not B® (a>); then the consequence A), is inferred as fol- lows: = Ry o(not B°) =f, pdeduat) [aed 0-H60)7 ‘The membership function p, of the fuzzy set Ay is pointwise defined for all wel by aa) sup in (eu) ng). 1= wD) = sup $,(1= m0) where SU u9(0)) 2 min (1, 4(1) 490) = WHO) (B'= not B) The values of $,(1~ gC) with a pa- rameter j.(u), Say uu) 03 and 0.8, ae indicated in Fig. 6 by a broken line and dotted lin, respectively. The membership function sis given by Haga) = sup mine Wg) 1~ Hal) al) eal) {8'= not B27} The values of S,0~p4)) with a parameter 1 fu), say fu) = 03 and 08, ar indicated in Fig Approximate reasoning: oneaized meds tllens wits’ product operation rile Sy) ro) Fie. 9. Approximate reasoning: generalized modes tllens with TABLE ML Seuyany oF erenence Rests Fo Vea Fig. 7 by a broken line and dotted line, respectively. The ‘membership function p14, is given by dau) ~ sup min (ged). — a) _ ale) Y wu) +4 = ale) Bec neeeruascaaeeinrea (8'= not B% The values of S,1~ Cv) with & parameter pu), say .4u) = 0:3 and 0.8, are indicated in Fig. 8 by 2 broken line and dotted line, respectively. The‘TABU Not Very 5 wv e TABLE v oo x 6 6 & 6 6 x Giese xx Gems ox x Sims xX Gamer x kG Giemsa 6 6 ‘membership fonction jis given by Ha) = sup min (a dCudsea( v1 wa (0) 2a glu) +1 Vig rh Balu (= BE The values of S(sp(v)) with a parameter aud, say wu) = 03 and O8, are indicated in Fig 9 by a broken line ‘and dotted line, respectively. The member- ship function 4, is given by (A) = su (AWD al Do dal OD) ase) ‘The remaining consequences [24] inferred by Ry, R., Roy RyyRyoRy can be obtained by the same method as juSt described: The results are summarized in Tables IIL and IV. By employing the intuitive criteria in Tables Land 11 in ‘Tables III and IV, we can determine how well a fuzzy implication function satisfies them. This information is summarized in Table V. In FLC applications, a control action is determined by the observed inputs and the control rules, without the XxKKK KOKKERK KOKKX KOKKXEX consequent of one rule serving as the antecedent of another. In effect, the FLC functions as « one-level fo ward data-driven inference (GMP). Thus the backward sgoal-driven inference (GMT), chaining inference mechs- nisms (sjllogisms), and contraposition do not play a role in the FLC, since there is no need to infer a fuzzy control action through the use of these inference mechanisms, Although R, and R, do not have a welldefined logical structure, the results fabulated in Table V indicate that they are well suited for approximate reasoning, especially for the generalized modus ponens. R,, has logical structure which is similar to Ry. R, is ‘based on the implication rue in Lukasiewice's logic Lea: However, R, and R, are not well suited for approximate reasoning since the inferred consequences do not always fit our intuition. Furthermore, for multiple-valued logical systems, R, and Ry have significant shortcomings. Over- all, R, yields reasonable results and thus constitutes an appropriate choice for use in approximate reasoning. B. Inverpretation of Sentence Connectives “and, also” In most of the existing FLC’s, the sentence connective and" is usually implemented as a fuzzy conjunction in @ Cartesian product space in which the underlying variablestake values in different universes of discourse, As an illustration, in “if (A and B) then C," the antecedent is interpreted as a fuzzy set in the product space U x V, ‘with the membership function given by peanol 0) = min (1 4(00) si )) Hal) = He) Bgl) where U and V are the universes of discourse associated with A and B, respectively. ‘When a fuzzy system is characterized by a set of fuzzy control rules, the ordering of the rules is immaterial. This necessitates that the sentence connective “also” should hhave the properties of commutativity and associativity (see sections III-A and ILC in Part land Part D in this section). In this connection, it should be noted that the ‘operators in triangular norms and co-norms possess these properties and thus qualify as the candidates for the interpretation ofthe connective “also.” In general, we use the triangular co-norms in assoctation with fuzzy conjune- tion and disjunction, and the triangular norms in associa tion with fuzzy implication, The experimental results {52]-{54), {96} 73] and the theoretical studies [18], [85], [116}, (19) relate to this issue. Kiseka er al. [52] described a preliminary investigation of the fuzzy implication funetions and the sentence con- nective “also” in the context of the fuzzy model of a de series motor. In later work, they presented additional results for fuzzy implication functions and the connective “also” in terms of the union and intersection operators (53), [54 ‘Our investigation leads to some preliminary conelu- sions. First, the connective “also” has a substantial influ- tence on the quality of 2 fuzzy model, as we might expect. Fuzzy implication functions such as R,, Ry, and R, with the connective “also” defined as the union operator, and R., Ry, Roy, and Ry, defined as the intersection, yield satisfactory results. These fuzzy implication functions dif- fer in the number of mathematical operations which ate needed for computer implementation Recently, Stachowicz and Kochanska [96] studied the characteristics of 38 types of fuzzy implication along with nine diferent interpretations (in terms of triangular norms and co-norms) of the connective “also,” based on various forms of the operational curve of a series motor. Based oon their results, we tabulate in Table VI a summary of the ‘most appropriate pairs for the FILC of the fuzzy implica- tion function and the connective “also.” ‘Adltional results relating to the interpretation of the connective “also” as the union and the intersection are reported in [73]. The investigation in question is based on a plant model with fistorder delay. It is established that the fuzzy implication functions R,, Ry, Rypy Rup with the connective “also” as the union operator yield the best control results. Furthermore, the fuzzy implications R, and R, are not well suited for control applications even 5 as ‘TABLE VI Suerane Pains or Pury Tate cio Posen "ano Conte le Tepheation Roe Conese Ae ene Aon ® . RRR, inant 1 depends onthe shape of reproduced curve which forme the et of fuzzy conta aes though they yield reasonably good results in approximate reasoning. From a practical point of view, the computational as pects of un FLC require a simplification of the fuzzy control algorithm. In this perspective, Mamdanis R, and Larsen's R,, with the connective “also” as the union ‘operator appear to be better suited for constructing fuzzy siodels than the other methods in FLC applications. We will have more to say about these methods at 2 later point C. Compositional Operators In a general form, a compositional operator may be expressed as the sup-star composition, where “star” de- notes an operator—eg., min, product, ete.—which is chosen to fita specific application. In the literature, four kinds of compositional operators can be used in the ‘compositional rule of inference, namely sup-min operation (Zadeh, 1973), sup-product operation [Kaufmann, 1975), sup-bounded-product operation [Mizumoto, 1981), sup-drastic product operation (Mizumoto, 1981]. In FLC applications, the sup-min and sup-product ‘compositional operators are the most frequently used, “The reason is abvious, when the computational aspects of an FLC are considered. However, interesting results can be obtained if we apply the sup-product, sup-bounded- product, and sup-drastic-product operations with differ ent fuzzy implication functions in approximate reasoning {70}, (721. The inferred results employing these composi- tional operators are better than those employing. the sup-min operator, Further investigation of these issues in the context of the accuracy of fuzzy models may provide interesting results. D. Inference Mechanisms ‘The inference mechanisms employed in an FLC are generally much simpler than those used in a typical expert system, since in an FLC the consequent of a rule is not applied to the antecedent of another. In other words, in FLC we do not employ the chaining inference mecha- nism, since the control actions are based on one-level orward data-driven inference (GMP) ‘The rule base of an FLC is usually derived from expert kknowiedge. Typically, the rule base has the form of @‘MIMO system = (Rhawos Rianos"**+Riawo) where Rignyo represents the rule: if (x is A, and--+, and y is B) then (2, is C,:"»,2, is D,). The antecedent of Ryayao forms a fuzzy set 4) x'++ +> B, in the product space UX + XV. The consequent is the union of q independent control actions. Thus the ith rule Ryaaso may be represented as a fuzzy implication Rhayyo! (Ai 00 XB) (2,4 + from which it follows thatthe rule base R may be repre- sented as the union 42) XB) (2,4 +l} = {Ua xB)>2)), lax xa) lax xa) +29} ={U Ota xa)s91} = {RB has» RBiaso." “+ RBSuso}: In effect, the rule base R of an FLC is composed of a set of sub-rule-bases RBlyso, with each sub-rule-base RBjyso) consisting of m fuzzy control rules with multiple process state variables and a single control variable. The general rule structure of a MIMO fuzzy system can there fore be represented as a collection of MISO fuzzy sj5- tems: R= [RBhuso, RBiasor “+ RBXaso} where RBtyso represents the rule: if (x is 4, and---, and y is B,) then (2, is Dp, = 1,2.---sm. Let us consider the following general form of MISO fuzzy control rules in the ease of two-input/single-output xis A’and y is BY Ry: if-xis A, and y is By then 2 is C, if is A, and y is By then z is C, iflx is A, and y is B, thon 2 is C, 2c where x, y, and 2 are linguistic variables representing the VERE TRANSACHONS UM SYREN MAK AN CYBERNETICS 0120.80, 2, SERA 190 process state variables and the control variable, respec: lively; A,, B,, and C, are linguistic values of the linguistic variables x,y, and = in the universes of discourse U, V, and W, respectively, with f= 1,2,-*+57 ‘The fuzzy control rule “if(x is A, and y is B,) then (2 is C,)" is implemented as a fuzzy implication (relation) R, and is defined as Fim Bessa esto) = foea(u) and pa()] > wel") where “A, and B)" is @ fuzzy set A, xB, in UXVs R,2(A, and B,)-°C, is 3 fuzzy implication (relation) in UXV XW; and —+ denotes a fuzzy implication function, ‘The consequence C’ is deduced from the sup-star compositional rule of inference employing the definitions of a fuzzy implication function and the connectives “and” and “also.” In what follows, we shall consider some useful proper- ties of the FLC inference mechanism, First, we would like to show that the sup-min operator denoted by * and the connective “also” as the union operator are commutative. “Thus the fuzzy control action inferred from the complete set of fuzzy control rules is equivalent to the aggregated result derived from individual control rules. Furthermore, as will be shown later, the same properties are possessed by the sup-product operator. However, the conclusion in question does not apply when the fuzzy implication is used in its traditional logical sense (18), (19]. More specifi- cally, we have Lemma I: (4,B)> Proof = UL MABIER, 4.8)° UR, =(4,B)° U (A, and B, +0) The membersipfnction wf the faz set Cs pointwise defined for all we by c(W)= (y(t) ssp e))> max (np (0, in (a¥aM essen (8) ~ supmin (a (1) ag») may (1 (uo¥o¥)s Hg (a¥sW oo stg (sv) = sup max (min f(a eC) tel) maleoron)]> im (aye) Har) teg(tes¥ | = max ([(ae(u) nel) malar], [46 ae) aa (aren )])Therefore C= [A.B RUA BY Ra] Ur UAB Rul = Urs) R, U(4.896¢A,and 846) Lemma 2: For the fuzzy conjunctions Roy Rys Ryps and Rips We have (4,8) (A,and B, +6) AA, C)] [BBC] pagan Ha, A, (A.B (A, and B, +6) = [4A CIB (BC) Maen = aH, Proof: (4,8) (A, and BC) Hep eon) Cia, Be) = (Hertie) (min (a4 sttn) He) = (este) min (Hs, > He,)s(sea, > He,)) = supmin ([(4440) min [( 4, He.) (Ha, He] = supmin (min [1 y.(4, + He,)]y min [este 7 He) = min {a2 4, > He,)] [Ha Cee, He.) Hence we obtain C= [A9(4, 40) 1B (B,+C))- OED. Let us consider two special cases that follow from the preceding lemma and that play an important role in FLC pplication. Lemma 3: If the inputs are fuzzy singletons, namely, Araug, B=, then the results dervied by employing ‘Mamdani’s minimum operation rule R, and Larsen’s product operation rule R,, respectively, may be expressed simply as en where ay = tg)’ maf) and; By) myo Proof » Cj [AAC] 08°, we; = min {[ a9 °(u4(u) > ucw))], [vo o(eale) > ac(w))]} = min [1 4(uo) + eh) o[otal Yo) > He") 2) CLA AC) [BBC] ey = [moe (wate) 4c ))] [vara Y) Hc #))] = [eal ta) ed] [ant %0) te] As will be seen in following seton, the last lemma not only simplifies the process of computation but also pro- vides a graphic interpretation of the fuzzy inference ‘mechanism in the FLC. Turning to the sup-product oper- ator, which is denoted us +, we have the following Lemma 1: (AB UR,= QUA, BYR, Lemma 2: For the fuzzy conjunctions R,s Ryy Ros and Ry, We have (A,B)-(4, and B, +6) [44,6] 01848, +69] if bape, = Ha, Oe, (4,8)(A, and 8, +6,) [4446 [BO] sean Bah Lemma 3: Ifthe inputs are fuzzy singletons, namely, Armity By, then the results derived by employing Mamdani’s minimum operation rule R, and Larsen's product operation rule K,,fespectively, may be expressed Simply as Re a Atel) R Ry a ael) Rye aiA wel) ai el) where «= se4(yA yl) AME = ply) He) Therefore we can assert that Re we= Ua Mite, arm, where the weighting factor (firing strength) a, is a mea sure of the contribution of the ith rule to the fuzzy control action. The weighting factor in question may be determined by two methods, The first uses the minimum ‘operation in the Cartesian product, which is widely usedon He, ae * ‘ a 4, | 4 3 2 (of Lemma Pander and R %, a 4, Fe 1 in FLC applications. The second employs the algebraic product in the Cartesian product, thus preserving the ‘contribution of each input variable rather than the domi- nant one only. In this respect, it appears to be a reason- able choice in many FLC applications. For simplicity, assume that we have two fuzzy control rules, as follows: Ry: if-xis Ay and y is By then zis Cy, Ry: ix is Ay and y is B, then 2 is Cy Fig. 10 ilustrates a graphic interpretation of Lemma 2 under R, and a,*. Fig. 1 shows a graphic interpretation of Lemma 2 under Ry and a,’ In on-line processes, the states of a control system play an essential role in contiolaetions. The inputs are usually ‘measured by sensors and are crisp. In some cases it may be expedient to convert the input data into fuzzy sets, In general, however, a crisp value may be treated as a fuzzy singleton. Then the firing strengths «7, and a ofthe first ‘Graphical interpretation of Lemme? under «and Ry and second rules may be expresed a8 ,= Hao) A ea 90) 5~ hal) Han) where 424(r») and 419y) play the role of the degrees of patil match between the useraupplied data and the data in the rule base. These relations pay a central role in the four spes of fuzzy reasoning curtently employed in FLC application, and are described in the following 1) Fuzzy Reasoning ofthe Fast Type — Mamdanis Min- imum Operation Rule as a Fuzzy Implication Function Fuzzy reasoning ofthe firs type i associated with the use of Mamdan’s minimum operation rule Ras & fuzzy ‘implication function. In this mode of reasoning, the 1th rule leads to the contol decision weil) =a, A Helo) which implies that the membership function 4c of the inferred consequence C is pointwise given by Helo) = Me: te for me(w)] V[as A we]Ys Fig. 13. Disgrammatc representation of zy remoing 2 To obtain a deterministic control action, a defuzzifica- tion strategy is required, as will be discussed at a later point. The fuzzy reasoning proces is illustrated in Fig, 12, which shows a graphic interpretation of Lemma 3 in terms of Mamdani’s method R.. 2) Fuzzy Reasoning of the Second Type—Larsen’s Product Operation Rule as a Fuzzy Implication Function: Fuzzy reasoning of the second type is based on the use of Larsen’s product operation rule Ry, as a fuzzy implication function. In this case, the ith rule leads to the control decision es) 4H) Consequently, the membership function yee of the i {erred consequence C is pointwise given by Helo) = he; ¥ es a-ak] [aswell]. From C, a crisp control action ean be deduced through the use of a defuzzifiction operator. The fuzzy reasoning process is illustrated in Fig, 13, which shows a graphic interpretation of Lemma 3 in terms of Larsen’s meth- od R,. 3)" Fuzzy Reasoning of the Third Type—Tsukamoto’s Method with Linguistic Terms as Monotonic Membership Functions: This method was proposed by Tsukamoto (117) It is a simplified method based on the fuzzy reasoning of the first type in which the membership functions of fuzzy sets 4, B, and C, are monotonic. However, in our derivation, , and B, ate not required to be monotonic but Cis. In ‘Tsukamoto's method, the result inferred from the first rule is @, such that a, = C\(y,). The result inferred from the second rule is a such that a, = C,(y,). Corre- sponding, a crisp control action may be expressed as the Weighted combination (Fig. 14) spn et aay 4) Fuzzy Reasoning of the Fourth Type—The Conse quence of a Rule isa Function of Input Linguistic Variables: Fuzzy reasoning of the fourth type employs a modifieda rary oe Os, Fg. 4 Digrammatic representation of xy reasoning version of state evaluation function. In this mode of reasoning, the ith fuzzy control rule is of the form Ryo (x I Aj and y is B) then 2=fre--9) where x,°+*, y, and 2 are linguistic variables represent- ing process State variables and the control variable, re- spectively; 4,,-*+,B, are linguistic values ofthe linguistic variables x,-">,y' in the universes of discourse U.=:,, respectively, with /=1,2,"":,n and f, isa function of the process state variables x,:-',y defined in the input sub- spaces. For simplicity, assume that we have two fuzzy control rules a$ follows Ry: if is A, and yis By then == fxyy) Ry: ifxis Az and y is By then 2~ f(x,y). ‘The inferred value of the control action from the frst rule is aj fxq, Jo). The inferred value of the control action from the second rule is a f(y.) Correspondingly, a crisp control action is given by afl FoI) + afl Koro) ata This method was proposed by Takagi and Sugeno [103] and has been applied to guide a model ear smoothly along a crankshaped track [98] and to park a car in a garage (971, 991. I. Deruzziricanion Sreatraues Basically, defuzification is @ mapping from a space of fuzzy control actions defined over an output universe of discourse into a space of nonfuzzy (crisp) control actions. 11 is employed because in many practical applications & crisp control action is required, A defuzzifcation strategy is aimed at producing a non- fuzzy control action that best represents the possibilty distribution of an inferred fuzzy control action. Unfortu- nately, there is no systematic procedure for choosing a \efuzzification strategy. Zadeh [142] first pointed out this probiem and made tentative suggestions for dealing with it. AL present, the commonly used strategies may be described as the max criterion, the mean of maximum, and the center of area. A. The max erterion method The max criterion produces the point at which the possibility distribution of the control action reaches a ‘maximum value, B. The Mean of Maximum Method (MOM) ‘The MOM strategy generates a control action which represents the mean value of all local control actions whose membership functions reach the maximum, More specifically, in the ease of a discrete universe, the control Action may be expressed as = where w, is the support value at which the membership function reaches the maximum value y,(w,), and I is the ‘number of such support values. mie ©. The Center of Area Methoil (COA) ‘The widely used COA strategy generates the center of gravity of the possibilty distribution of a control action, In the case of a discrete universe, this method yields E ntodom Law) where is the number of quantization levels of the ‘output. Fig. 15 shows a graphical interpretation of various defuzzification strategies. Braae and Rutherford [5] pre- sented a detailed analysis of various defuzzification strategies (COA, MOM) and concluded that the COA Strategy yields superior results (also see (58), However, the MOM strategy yields a better transient performance while the COA strategy yields a beter steady-state per- formance [94] It should be noted that when the MOM strategy is used, the performance of an FLC is similar 10 that of a multilevel relay system [48], while the COA strategy yields results which are similar to those obtain- able with a conventional PI controller (46]. An FLC based fon the COA generally yields a lower mean square error than that based on the MOM [111]. Furthermore, the MOM strategy yields a better performance than the Max criterion strategy (S21. TIL Apruicaions ano Recent DeveLorments A. Applications During the past several years, fuzzy logic has found. ‘numerous applications in fields fanging from finance to earthquake engineering [62]. In particular, fuzzy controlWe Fig 15, Diagrammatic representation varius dluziation state s has emerged as one of the most active and fruitful areas for research in the application of fuzzy set theory. In ‘many applications, the FLC-basedsjstems have proved to bbe superior in performance to conventional systems. ‘Notable applications of FLC include the heat exchange {80}, warm water process (47) activated sludge process (013), 35), wafic junction (82) cement kiln (59, 118, aircraft flight control (58), turning process (92), robot conto! {119}, (94, (106), [8 [34], model-car parking and turing (97]-(99), automobile speed conto (74), [75], war ter purification process [127], elevator control (23), auto- ‘mobile transmission control (40), power sjstems and nu Clear reactor contol (4 [S1, uzzy memory devices (107 {ios}, 120}, 1128, (125), 133} and the fuzzy computer [132] tn this connection, it should be noted thatthe first successful industrial application of the FLLC was the c2- ment kiln control system developed by the Danish cement plant manufacturer FL. Smidth in 1979. An ingenious pplication is Sugeno’ fuzzy car, which has the capability of learning fom examples, More recently, predictive fuzzy control systems have been proposed and succesfully ap plied to automatic tain operation systems and automatic Container crane operation systems [135]-(139. In parallel with these developments, & great deal of progress has been made in the design of fuzzy hardware and its use in so-called fuzzy computers {132} B. Recent Developments 1). Sugena's Fuzzy Car: One of the most interesting applications of the FLC is the fuzzy car designed by Sugeno. Sugeno's car has successfully followed a crank: shaped track and parked itself in a garage (98}-[99) The contol policy incorporated in Sugeno’s car is ep: resented by a Set of fuzzy control rules which have the form: Ry: ifr is Ajo and y is B, then reay tales os taly where xy: and y are linguistic variables representing the distances and orientation in relation to the bound- aries of the track; 4,,°*-, and B, are linguistic values of avo, and yi 2 i the value of the control variable of the ith control rule; and a},*°, and a are the parameters ‘entering in the identification algorithm [103], (99) The inference mechanism of Sugeno’s fuzzy car is based ‘on fuzzy reasoning of the fourth type, with the parameters TABLE Vit Fuzzy Corn. Roars rom Ive dhs, and a, identified by taining, The training process involves a skilled operator who guides the fuzzy model ear under different conditions. In this way, Sugeno's car has the capability of learning from examples, 2). FLC Hardware Sysiems: A higher-speed FLC hard ‘ware system employing fuzzy reasoning of the first type has been proposed by Yamakawa (130), [131]. It is com- posed of 15 control rule boards and an action interface (ie. defuzzifier based on the COA). It can handle fuzzy linguistic rules labeled as NL,.NM,NS,ZR, PS, PM, PL. ‘The operational speed is approximately 10 mega fuzzy logical inferences per second (FLIPS). "The FLC hardware system has been tested by an appli- cation to the stabilization of inverted pendulums mounted fon a vehicle, Two pendulums with different parameters were controlled by the same set of fuzzy control rules (Table VID). It is worthy of note that only seven fuzzy control rules achieve this result. Each control rule board and aetion interface has been integrated to a 40-pin chip. 3) Fuzzy Automatic Train Operation (ATO) Systems: Hitachi Ltd. has developed a fuzzy automatic train opera- tion system (ATO) which has been in use in the Sendai- City subway system in Japan since July 1987. In this ‘stem, an object evaluation fuzzy controller predicts the performance of cach candidate control command and Selects the most likely control command based on a skilled hhuman operator's experience More specifically, fuzzy ATO comprises two rule bases which evaluate two major functions of a skilled operator based on the criteria of safety, riding comfort, stop-gap accuracy, traceability of target velocity, energy consump- tion, and running time. One is constant-speed control (CSO), which starts a train and maintains a prescribed speed. The other is the train automatic stop control (TASC), which regulates a train speed in order to stop at the target position at a station. Each rule base consists of twelve object-valuation fuzzy control rules. The an- tecedent of every control rule performs the evaluation of train operation based on safety, riding comfort, stop-gap accuracy, ete, The consequent determines the control action {© be taken based on the degree of satisfaction of each criterion, The control action isthe value of the train control notch, which is evaluated every 100 ms from the ‘maximal evaluation of each candidate control action, and takes asa value a diserete number; postive value means ‘power notch,” negative value means “break notch.a ‘The Sendai-City subway system has been demonstrated to be superior in performance to the conventional PID ATO in riding comfort, stop gap accuracy, energy con- ‘sumption, running time, and robustness [135], [136] 139). 4). Fussy Automatic Container Crane Operation (ACO) ‘Systems: In the application of FLC to the automatic oper- ‘ation of container-ship loading cranes, the principal per- formance criteria are safety, stop-gap accuracy, container sway, and carrying time, Fuzzy ACO involves two major operations: the trolley ‘operation and the wie rope operation. Each operation comprises two function levels: a decision level and an activation level. Field tests of fuzzy ACO systems with real container cranes have been performed at the port of Kitakyusyu in Japan. The experimental results show that ccargo handling ability of Fuzzy ACO by an unskilled ‘operator is more than 30 containers per hour, which is ‘comparable 10 the performance of a veteran operator. ‘The tests have established that the fuzzy ACO controller has the capability of operating a crane as safely, accu- rately, and skilfully as a highly experienced human opera- tor [1371-[139} 3) Fuzzy Logic Chips and Fuzzy Computers: The first fuzzy logic chip was designed by Togai and Watanabe at ATAT Bell Laboratories in 1985 [107]. The fuzzy infer- fence chip, which can process 16 rules in parallel, consists fof four major parts: a rule-set memory, an inf processing unit, a controller, and an input-output cir- cuitry. Recently, the rule-set memory has been imple- mented by a static random access memory (SRAM) to realize a capability for dynamic changes in the rule set ‘The inference-processing unit is based on the sup-min compositional rule of inference. Preliminary timing tests indicate that the chip can perform approximately 250000 FLIPS at 16-MHz clock. A fuzzy logic accelerator (FLA) based on this chip is currently under development (108), 120}, Furthermore, in March 1989 the Microelectronics Center of North Carolina successfully completed the fab- rication ofthe world’s fastest fuzzy logic chip, designed by Watanabe. The full-custom chip comprises 688000 tran- sistors and is capable of making $80000 FLIPS. In Japan, Yamakawa and Miki realized nine basic fuzzy logic functions by the standard CMOS process in ccurrent-mode circuit systems (128), Later, a rudimentary concept of a fuzzy computer was proposed by Yamakawa and built by OMRON Tateishi Electric Co, Ltd [132]. The Yamakawa-OMRON computer comprises a fuzzy mem ‘ory, a set of inference engines, a MAX block, a defuzzi- fier, and a control unit. The fuzzy memory stores lin- uistic fuzzy information in the form of membership functions. 1thas a binary RAM, a register, and a member ship function generator [128] A. membership function generator (MFG) consists of a PROM, a pass transistor array, and a decoder. Every term in a term set is repre- sented by a binary code and stored in a binary RAM. The corresponding membership functions are_generatcd by the MFG via these binary codes. The inference engine employs MAX and MIN operations, which are imple imented by the emitter coupled fuzzy logic gates (ECF gates) in voltage-mode circuit systems. The linguistic in- puts, which are represented by analog voltages distributed fon data buses, are fed into each inference engine in parallel. The results inferred from the rules are aggre gated by a MAX block, which implements the function of the connective “also” as a union operation, yielding a consequence which is a set of analog vollages distributed fon output lines. In the FLC applications, a crisp control command necessitates an auxiliary defuzzitier. In this implementation, a fuzzy computer is eapable of process- ing fuzzy information at the very high speed of approx: mately 10 mega-FLIPS, Iti indeed an important step not nly in industrial applications but also in common-sense Knowledge processing IV. Furure Srupirs ano Pronuents In many of its applications, FLC is either designed by domain experts of in close collaboration with domain experts. Knowledge acquisition in FLC applications plays fan important role in determining the level of performance of a fuzzy control system. However, domain experts and skilled operators do not strueture their decisionmaking in any formal way. As a result, the process of transferring expert knowledge into a usable knowledge base of an FLC is time-consuming and nontrivial. Although fuzzy logie provides an effective tool for linguistic knowledge representation and Zadeh’s compositional rule of infer tence serves a a useful guideline, we are still in need of more efficient and more systematic methods for knowk edge acquisition, ‘An FLC based on the fuzzy model of a process is needed when higher accuracy and reliability ae required, However, the fuzzy modeling of a process is sill not well understood due to difficulties in modeling the linguistic structure of a process and obtaining operating data in industrial process control (13), (84), (111), (128), (104), tou, Classical control theory has been well developed and provides an effective tool for mathematical system analy: sis and design when @ precise model of a system is available. In a complementary way, FLC has found many practical applications as a means of replacing a skilled hhuman operator. For further advances, what is needed at this juncture are well-founded procedures for system de- sign. In response to this need, many researchers. are {engaged in the development of a theory of fuzzy dynamic systems which extends the fundamental notions of state (6), controllability 31], and stability [77], [4], [891 [5]. Another direction of recent exploration is the concep- tion and design of fuzzy systems that have the capability to learn from experience. In this area, a combination of techniques drawn from both fuzzy logic and neural net- work theory may provide a powerful tool forthe design of| systems which can emulate the remarkable human ability to learn and adapt to changes in environment.Acknowtcnamenr 1am greatly indebted to Professor Lotfi A. Zadeh of the University of California-Berkeley for his encourage- ment of this research. The assistance of Professor Zadeh is gratefully acknowledged. The author would like to thank Professor M., Tomizuka of the University of Califor nia-Berkeley and the reviewers for their helpful com- ments and suggestions Reverences LI). Boaklni apd EM. Madani, “Preserve, methods for crtlng nel poy na ay togic coil.” Elecon, Let So Tu pp. 65-500, 1975. toy patie and We Guid, “Moding comoters using 827 felon” Ayre, a9; pp. 223-22, 18 ta) Tr Boldwin and BW, Plot “Atomatic apycach to ‘eplation for epprinimateTeavning wh foxy Ios” Pus) Set rs ol 3 pp 98-219, 1980, tat JAD Bernard“ of rule-based tem for proces conta TEEE Con Sy Beg so 8.3 pp. 3-18 18 {5) Me Bic ond B. A Keron “Fuzzy flame ina conta Ging Kierstesp7 -3.pp [88-185 97K toy SMP" Aortic tnd impute aspects of the furry ls Gaile ama, vl 18.408 pp. 38-57, 178 ee Son of parameters fora fy loge como.” Fue Fae Sons so no pp 18S 198, 19. (a) Rick, 1Fey Use), and Ck "Praia aspects of the Kole teed como!» mobile robot moti in Foe Si Mies Sm on Cian and ystems Syacse. NY. 187 19) E'Ceoply, Pub Sein Decson Maing and Con ‘Cologne: Verlag TUY Rhian 98 Copan, “Prbabii faz) controler asa generation of fhe comeep a feny sont” Fay See Spt. ol 2. PP 215-23 ss Cropland K. Hits, Poa Sep: Faz and Stchate chin Deon Cit and.Reoion Pros loge! veras TUY Rein, AE 'chtgaa and W. Feder “Some problems concerning the onsrston of slot Df desi making In fuzzy sens net Maw, Mock Ser 0115, 9. 2012221, 19681 sm AG ibenticaton i zy Stems ad Hs appcations in po SS pA com tel pric nf stems” Py Ses to Foe. pp 287274 ne Bsa regeneration fr farey control” Chem. St SATS opp 8-298 1 TE Copal an, Rie "Modeling of» fers controler with ppheston fo te cond of bagel process. Fey Sets ‘Suv 3, pp 3-22, 1988 Binarah! Best aad Ve VS, Sarma, “Estimation of uz rept oe rf: pp 8 De abois and H, Prade “Fane lie and the generaied ‘modu pens rested Cem sow 15, pp. 3-4 1984 ress Ete ecneralard mode ponens nde spin eompose ea tncoreeal study” in Appa Resong ov Esper ‘Suns: MM. Gupta A. Kandel W. Bande, and J. KRONA, Amira North otand, 188, pp 217-232. tao) ena sone ad nocessy meastes: Toward» os ‘Sie iterpretaton of stgram” say Ss Sr v0 pp Sea - 11. sii Theory An Apprase to Computed Poesing FDieemany.” Now York. Pram Pres. 188 1 Efsetou,"Rolebated process cond using fuzz lng” 0 rit Reagent Shem, Deion and Con ere Sanchen and L.A. Zadeh, Ede New York Pergamon Tete pp. 5-158. Fr Fujiee “FLEX:A00 series elvtor group contol sem, Fujtee 3 Lid, Oso, Japan, 198 9) uu 12) bs) us us) i} un 08) tus) en wa bs) ba best bal bn ea bo bo) on wa os) ba) bs) be br es be eo. wn ta to oy Ws) les wn ts wo (sa. sn 9 5. Fahaml, M. Micumoto, and K. Tana, “Some consideration St any condom ference Pacsy Set Sty vl fr. MSC BIR Gaines an LJ. Kobo. The fuzzy decade: A big phy ol fons aps ape osc felted tops.” ft. J. Man Mach states vo, pp 881077 Goutal snd W, Peiner, “Problems of the design of fury Somers: pian Resmi Ext Sens, MM Gopi Kendal W Bander aed’ B. Keka, ES.” AMS ‘Som: Norio 98, pp 38-88, BBP Graham an Ru, Newel "Fuzzy identification and ont igi eve ig Facy Sor Sv 3 pp. 255-278 von IOS funy adaptive contol of Sena App 7-45, 1988 T"Grahan and PC Done, Expr Sons Knowle, Uncertain Gnd Dre New York: Chama and Hal 18, Mr Mt Gupia and Ys Toshamot, "Fuzz lope conoler— IX perpecines Prot dow Automate Coral Cor San Franeasa, Avg. 980. pp. FAIO- MUM Cope. Me Tipian and J.B. Kis, “Control of hex control stems ZEEE. Tras, Sa Man Chem, a SMe. nt pp. 36-52. ima YA and 8: Hashing, “Moving mark recognition find moving sbject manipulation in faz conte ob.” Con. Theory Aut Tel wl 20 pp. 39048, 186 order process” Fuzy Rfvota and We edger, "Anabsis and sithess of wee uci ty te ue of fry lek” Pay Sete Sat 8a 1, m0 pple 1. Ei ideniation and uzzy rae based com! of a sie feoot motion iw Pa. IEEE Dt Simpson Inlien Conta, Pep, PAL 198. Oiitontk Gone Nakayama, and S, Takamizama, “Applic ona zy contol ty acvated sod prowess In Pcs 2nd TES4 Congo Togo Japa, Sly. 198, pp. 282-28, A domes A. Keston, and H1Zimmermann, Py Se Tico and Apes, Boson: B. Reidel Passing Com aes 1586 JF Racpegk and SA Orlow, Opuinion Models ang Fy Sasand Posty Phe” Boston: D. Reel Publishing Con any, 198 Bicone a M, Fc, Combining Puzsy Imreion with Praatine ‘Uncenainy ot Devon Making New "York Springer Ver. 188 Radel Fat Tocges ix Pater Recon. New York ‘leynterctence. 82 YR and. Morinots, “Electronically cooled conti hsb sartle waren in Pro. Conges Promo fo Elecomc, Desborn. M1988 [RT Keutmenny Inaction 10 Theory of Faz Suber ane Academie 975 2A Kuan and MM. Gupta, draduton Puy Avie ew York Van Nowrand, RS "hy Morhemateal Models Engen and Monae ‘a Scenes" Amend! Noth tala. 18h IM'Ketemen, Fueines and sacral stabi,” Kyra, 0 10, pp eta WEFT Kickers “Towards am apabss of igus moding Pray Se Sts vl 3p 298-307. 197, Scr anaj and appli of zy li onal Rep F/R "5 Queen Mary Colege, London, 195. ier and HR, Van Nauta Lemke. Appeation of Tet conta o's warm water plat” Auama vol 12.004, fps toe, SEIN Ricker snd EH. Mardi Anas of zy ie Unive Pas’ Ser Sv eo. I pp 29-44, 1978 Pra Ring an EW. Mandan, The appeaton of fz contol Ippo dasa prcese,"aIRAC World Congres, MI. Boston, 1978 Pes. he appli of fzry conta tems 1 instr roesese saturn. 3 pp 28-282 1977 BeRinoshias 7 Fauak 1 Satoh sad ML Miyake, “AD Aromat operation mete! for control ras in BWR plans.” Pron Speci” Meng In-CoreInuretatin and Recor (Coe Aeron, Catach, France. 18H Se Rina E Kocharha, and DS, Shins, “The Taunt some parameter om the acuracy of 7 mde,” Newis) 9 us) (9) 7 (58) 9) (oo) tou ay sl (6) cor (4) to ww ro) ru vray vs pal vs. va 7 vos) om tree rane ion OM SYA Ma AND EYRERNETI WOH 30, N0. 2, ARCH Am 1D nds Aptcations of Fry Con M Sugeno, 8. Amster ‘Sim: Nowland 108 pp dBA “Te nace of ime azz impiation aerator the aay ofa fey meat” Psy Sa Sr 915. Hie 98S see ntuence of some faze imple operators on the aray ofa fan mosel-Part I” Pay Se Sr 1 SSS na Ma 17 Kola, MM. Gupta and P. 8. Nikoru, “Energetic Sabi of zy dynamic stems” EEE. rms St Man {ohms SMC. m0. 5 pp 792 1S, Goa" Kiatnd TA. lger Fay So Unerain an Infr ton. "Enpewood Cis Ni Prentice ial, 1, SU". Kamion. S.'PMakoo, and IF. Shaken, “Optimal Sontol of afte svlomation wth fz constrain ey target" Cytem, so emo fp. ash 1 Tarkan "A faze lope conte fo aera Right conto ts Iti ptf ay Conn Sige 8.” A Br Larsen “Indus ppiatins a zy logs onto” 4 Mom Mock, Stade, vo 0 pp 3-1 E, Temes “Dynami Tearing chaos of 2 rule-based ‘citorranaing contre.” Pai. thse, Quen May Coleg YTE Ui and CC Lay, “Development of fez algorithms for Ser systems.” IEEE Comp Sst Mags 9. pp 8-2 oes 4. Maes and. 5, Sherif, Applications of fizzy st theo, TERE Trane Sp, Man Cem ak SMCS. 90 1 pp. 17S toss Evi Maman, “Appisitine of faze algorithms for simple ‘ynamis plan Poe PEE vl, 12, a 12 Pp 1586 1S TP ETH Maman’ and 8. Asian, “An experiment ia lings Sates wth ony lg controle". Mn Math Sal few 7 1 ppd 1975 EH Mamas ’Agances fs the tngutic synthe of for fonts nd. Man Mach: Sta ol 8: rp 09-07%. 1908 “Application of fry lie 10 approximate resoning ‘Sing ngutte omens" IEEE Tron. Compute, Ch 1B. 182-1101, 177 “Applations of fy vet theory to conta stems: & ney” in aay Automate and econ Poss, M, Wh pts GOR sure ad BR. Goines Eds "Amica North otand 1977 pp. 77-8 EVE. Mamdatt and BR. Gaines, Facy Rewoning and ds Applicaton Lond: Acaseme, 138) Nahin a tga Haman oper fr mae I Conf on Faz Ifomation Presi. 1384. NM. Miamot, “Noe onthe athmelc fle Bt Zadeh fr tuxy mdiuonat inference" Cem. St vl 12, pp. 24-306, 181 MMaumoto sna. Zmmcrman, “Comparon of fy fs Soaing methods” Racy Ser end Sons ol % pp. 25888 io M Mizumon Comparison of ato tury reasoning metho in Proc: Dud PSA Cones. Toyo. Japan ly 987 pp 2-7 Fuzzy conto under various spss fessoning ‘matbosin Pace. 2nd IFS4 Conger, Tey, Jap, Jy, 1 pr iseise SP urakami, “Application of far cone to avons Speed conta ssem,” in Prac of the TEAC. Symp om Pay Infomation. Koowiedge Representation and Deion Ana, Marcie, France 98, pp aa S'Marakam and Nt Maets,"Applcation of fz conte 0 Sutomodie speed conta Stemi Padua ppcesons of uz Comal M. Sigena, Ed Amster Nowland, 13K pp. 105-128 Yo Miryam and T. Terapo, “Optimizing consol of die npn” in Indu Apeanone of Pay Con M: Super FE Areserdam, Noth old 195, pp. 83-72 {CV Negota, "On te say of fy stems in Pre. IEEE lin Conf on Cooma and Sony, TTB pp P89 wept Stems and Fay Stina. Mela Park: Be {jin Cummings Pushing Company, RS TDI Nola and AG. 8) Vente The Mathematics of Fics ‘Shuens” Gesamterstlline Verig TUV Rian 1988 (80) 4.4, Oserpaad, “Fuzzy oie cont of heat change proces” {nf Autor and Besson Poses, MM. Capa, GW. Sud. and BGues. Eas." Amsterdam’ NortHelland, 1907 pp 285-308 SK Thi and DK. Dutta Mander, Fsy Mathomael 4p rach i Patem Recaption. New Delhi Wiey Eastern Lim CF Paps and, 1. Mama “A fuzzy ope conto for a ‘efi jmction FEEE Trans. Sy, Man Cyber. ol. SMC, no. Tipp nm, 197, 1, Planes "An spprosch othe ani offs tems” It J Coo sol 30 pp A, 3 Sar 8p. 155-16. 184 “appteaions of zy relational equations for methods of ‘easing m presence of zy data" Fc Ss SW 16 PP H. Prade and CV. Noite, Fay Lage in Kendee Engine Ing. Gosamtberstellas: Veg TUV Rhian. 88. TS), ook and Es He Maman, "A gus sel-ngaiing Proves comer” omar sl 15,0. pp. 15-30 9 GP Rao and D. A" Ruheiord, "Approximate reconstruction ‘of mapping fantom frm inputicusnrptn in problem zy te applied to atem ton,” Cyemet, wo12 0.3, pp 228 26, A'S. Ray andD.D, Majumder, “Application of ie cites for ‘abilty anise of tncar SISO ad MIMO stems sesciated ‘ty zy Tg concer" IEEE Tron. Sp. Man Cer Vl SMC!TE na pp 345-59, 188 D'A Retherioet and GC. Biaore, “The implementation of fea lgotms for antral Proc. IEEE, wat 690. Pp sista te DA Rauteriod and G, A. Carte, “A Heute adaptive trier for sinter pnt” In Pc. 2nd TEAC. Symp on Rtomatin Ann Nina and Metal Presng,Heannee- ‘ig RSA, Sop, 1b Y'Eitat A fyconrlle ia tring process atemation in Inara Apeatons of Fay Corr M Supers, Ea Ate” ‘Sm: NorthHolland, 1985, pp. 89-132, EE Sancher and LA Zadeh Approx: Reson in Ilene Som Deron and Con! New York Pergamnon, 1988 E'M. Soho and NJ Mnde. "The appiation Of zy In Instat Aptian of Fay Conve! MSagenn, ES. Am ‘erdam: Nort Hold, 1S. pp. 41-62 5S, "Fuze selfoveanizhg contol apd is epplcaton for Sma proces” Puc Ss Sst v.28 pp 11-16, 188 (M'S"Stlcnmicr and ME” Kochanat,“Fuzey modling othe process Proc. 2nd TSA Congas, Tokyo, Sapam uy 1987, pp a9. At Sogeno and K. Murakami, “Furry paring control of adel fii tn 23 IEEE Conon Deion and Cone Las Vers, ik, IM Sugeno and M. Nida, "Fuzzy control of model ca,” Fay ‘Sen Ste pp. TOSI, 8S 1M Sogesa and urahanh “An experimental study om fury Pang contol sing mosel ct” it Indl Apeatons of Fiz Conia Ni"Sugeno, Ea.” Amsterdam NewHolland, 188 pp. 125-138, IM. Sepeno, "An introductory sey of fuze contol” dom, IM. Sogeno dt GT” Kang, “Stace iertieation of furry Pol Fez Sn Su. 28. pp Is" 388. I'Sigyama, "Anahis and sjtbei ofthe ule sed sl ‘xpansing contol,” PRD thes Queen Mary College, Un ‘ot Landon, 1 ‘T. Tabag and M. Sugeno, “Derivation of fuze conuol rules from han operators control acon” In. Pro. ofthe IFAC ‘Syms an Fes Inemation conan ad on inahat tare, France, Jy TAR pp. 35-60 Pan senticton of stems ans applications 10 ‘Rateling and control” IEEE. Trot Sht- Man Cer ol SMCS; 001. pp. 116-132, 1985, KUL, Tang and R 1. Mulholand, *Compsring zy lope with nse contoler desis. IEEE Trans. Man Cyn. vl. BR Tanscheit aad Sohal, "Experiments wih the ase of i) 2) 3) ws) ws) assy See (sa) ws (os) two) bo, wou, ba (3) a 3) 9) wn 8) oy tio tion ay (003) 04) 110ton to) Lion i) un a) us) na) i) 116) um 9) MU tia) an aa) tas) 4) ws) 9) wn 1128 tule-based selLorganising conto for robotics applications ‘sg So Stn Oh pp S218, AS 1M Toga and, Wataae, "Expert stem om 2 chip. An engine {or teatime spproimateremoning” IEEE" Expo Sor Meg IM Tog and 8. Ch, “A fuzy accleratar an» programing ooment for sel-tie Tuzzy contol" iPr! FSA Gong Toyo, pana 1987, pp 1-151 RON Tonge "Ano fory Contol algrithms using the Felton matin” Ia) Ma Mack. Studs vol. . #P roan, 19 SPSS coil engineering review of Farry ems” Audra SOT, po. pp 388-808 197. soe Syne of fey modes for instil process: Gensel. pp- Hed, 178 2 Ste ehalution of fz models derived (com experimen {Sr ina" Pus Ses St v4 pp 1-12 180 RM. Tong, MLB Beek, and AU Taten “zy conto of the ‘Seed sige wastewater treme proces suet Tot 6p. 995-71, 18. RUM. Tong, "A retroiperine view of fuzzy contol ssems” Pray Se Si tl 19-210, 188 PSS San ante Wlogrephy of Tz) conto” in nds ‘at Applesion of Pass Conta, M Sogeno, Ea Antena Norihoiand 1985, pp. 28-208 Erle ant Vaherde, “Oe made and inpicton in approx Irate reasoning” in Appagmae Reason i Eipe Se MINA Gupea, A, Kandel, We Banter, and J.B. Rika Eas ‘Asters North Hollnd 1588, pp 157166. YT Teaiamot "An approsch Yo any reasoning math” dsnees In Faz Set Theo and depicts, MM. Gop ROK Ragade, and RR Yager Eas Amsterdam Noth 1G Unters and. J King, “An anasis of human-ecison raking im cement sn control und the mplistioe for sors tiny J. Man Mach Suds, vl 121, pp 1-23 80. mi, Mi Minamoto, and Tana “Pussy robo Sonsrole® Coke vol 8 pp. Heat 78. Watanabe and W. Det, "Reconfiguae faz loge po fessor A fl custom digi VLSI ia fa. Wonshoy on Pacey Se Arpicaon, aa, Japa, Aus 198, pp. 39-0 Wines, “Optial cont of faery tem” in Pro. fn ng pd Ss Recah and Cet, Aca. “Decca mang ln porate dacs rank IEEE Conon Deco and Con, San Dio, Dee Be D: Willys and N. Maat, Use of feray mode! for proces font In Prac TEBE In Conf om Cohemets and Sec 1978 pp, 92-5, J Contain of te funy sts theory © man-mschine ‘Siem Adee Pe St Thay and Apcatns. MM. Supa, RA Ragade, and RoR Vege Bde Amer: ‘Nowh Hota 199, 9p. 48-50, CW. Xuand ¥. Zao, "Furry model entiation and sl Herne amc adem TLEE am once RR Yager, §Ovchinnion, Ru M. Tong, snd HT, Neen ‘iat Set and Applicaton Sond Paper by. i Bah Now ork: Wie trerstence 987 Sapna 0. Hoh and Mt. Sugeno, “Application of fury ‘Rasning tte water purieaton process” ih aduarl Ap. ator of Puzty Cool. M. Sugeno. Ed Anserdam: North Holland 1988 p. 128 ‘Te Vanaknva nd T. Miki “The cartent mode fey lapc Integrated crest abit bythe standard CMOS poses TREE Trae Computer sl. C3, ta 2p. 151-167. 1986 hm [121 7. Yamaha und K. Saiki “Fuzry memory dee.” in Pr Ind FH Cons, Tokyo apan ay THR, pp 81-585, Vlas igh sped ay corer hardware tema Pro Bnd Facey Seo Simp. Sapa, 198, pp 1213. (aan) oe sey cmt hardware stem oe Pr 2d TFS ara Toyo pan ly 8 (as) eile fay computer hardware system employing ‘inn man eration “A chulenge tosh generation om pate In Proc. Dud TESA Congress, Toby, Japan Tul 198. JP amakans Furry meroprssiar Rae chip aed dela fie chp inf Warkshoy Fy Stem Aptis zh Sapam Rug pp ste lus) Vina and M Sugeno, °A, microprocessor based tury otro forint porposes* Indust! Ape of ‘ing’ Ct Me Senor Ed” Amsterdam Neth Hold TORS pp. 31-200 SYatenoba, 5: Miymoto and HL tara, “Fuzzy contol for Susomate tsi operation yam,” in P. 4 TEAC TEP THORS ne Congets on Conte n Tromponanon Syms. Bae Boden, Ap ORS, 'S"Yaeundb an 8. Miyamoto, “Astomati tain operation by Proviivefzy coro im Ina. of Fas Co fot Mt Sagena. Ea Amsterdam NothsHoland, 1985, pp ."Yasunobu and. Hasegaa, “Evaluation of an automatic aie crane operation stem Based on pred fz eon Atal Cal They at, Technol, va 2.'m. Sy ppe 419-13, 2 predictive fu conto andi appieation for atomic oaaier crane option stem ih Po. nd TFA Cones “Tolgo.fapan. Jy 987 pp. 349-152 5" Yesunabie S.'Sekino, end T- Hisepana, “Automatic tin Speraion an automa cane operation stem fased om pe ‘Sane tery contol in Pros Su TESA Cong, Toy, Japan, M'Yonokis. “The aplcstion of faze ws they tthe tem perature of tox anne Trnaces using imultontechiques” Fr Pre Sh TEAC Wir Congress. Tokyo, 18 A°Dadeh "Fuzzy 5." Informa Cont 8 pp 338-383, rons tis ss) us 39) 19) an be) zealot norma. Conta, vl 12, pp 84-102, 1043) roman tony of zy astm” ia Age of Network and Spt They. RE. Kalman an Deh Ea. New ‘Yor Hot Rinhart snd Winton. 1871, pp 29-190. ‘Simany relatos ind faze orderings” norma, Se Sor pp 7 30,17 Se Mena Control 3. pp eens Ouiine of 2 new apiouc 0 the_anlss complex ‘Sree and decson proces” IEEE Tram Sr Men tem S01 SMCS, pp 2848 1973 “The Concept os Haga arise ud ite aplication vo hfomats Seis wl 8 PO. tna tus) tu) bey ig 1O5°SSte pp 3a 397 9p 9-80, TS Zimmermann, Pass) Set Theory and 1s Applaions. Bostn, Kluwer Aceme Publisher KS Bs" Zimmermann, Fuss. So Drcion Mating end Fagor ‘Shem "sto: Kluwer Academie Pushers 187. °C" Les, "Pury og in contol pte: Fury lage conor Pant 1 IEEE Trane St Ba Cher. S30, Wo. 2. PD Sat Mey Ape 9 tue iu) uso) ‘Chuen Chien Le (588) for biography and photograph, pease see page #18 af this Taassacrows