Professional Documents
Culture Documents
tional skills. To solve its ongoing problem Tennessee, Heery tried to remove some of
of assigning managers to projects, Heery the subjectivity and second guessing and
uses an optimization model that we at the same time tried to simplify the pro-
developed. cess of assigning managers to projects. In-
Previous Solution Methods stead of estimating the hours for each of
There are different ways to assign man- the many tasks, it used a monthly inten-
agers to building projects. One is the stan- sity, which is a dimensionless number for
dard resource-loading process of listing in each project, to quantify the management
detail all the tasks required to manage all effort. The intensity function assigns rela-
projects. Each task would have an esti- tive values for construction projects based
mated duration, and managers would be on the projects’ dollar values. Towle [1990]
assigned to multiple projects to work on
concurrent tasks. The number of hours per One questionable assumption
individual per week would be tallied, and is that the first project was
the workload shifted appropriately among well managed.
individuals as conditions change. For large
projects, listing each task in such detail is discusses the use of these intensity func-
very laborious. tions for predicting future management ef-
Because of its simplicity, companies fort. He notes that project dollar value im-
more often use the method of historical plies management workload because it
comparison to staffing levels and tasks in implies pay requests, change orders, de-
previous similar projects. However, in do- sign reviews, value engineering studies,
ing this they make several questionable as- scheduling, construction coordination, and
sumptions. One assumption is that the so forth.
first project was well managed. Another is This intensity function is traditional in
that neither project has any unique aspects managing construction for the Tennessee
(for example, architectural intricacies or government. It has a logarithmic shape
unusual construction materials) that will similar to the shape of curves that have
cause extra or lesser workloads. Another long been in general use to determine ar-
assumption is that the contracting envi- chitects’ fees. The logarithmic shape is ap-
ronments are similar. Staffing levels for a propriate because as projects get larger,
project in Nashville where little labor strife there is less difference between the intensi-
occurs and a spirit of cooperation exists in ties necessary to manage them. The func-
the construction community would not be tion is strictly empirically based; Towle
the same as those for a project in New [1990] notes that he tried various curves to
York City, where union issues and adver- find the best fit to management efforts for
sarial relations prevail. Thus, this method different size projects before selecting one
of determining staffing requires a great (Figure 1). Ever since then, the State of
deal of judgment to adjust for differences Tennessee has successfully used it to de-
in locations. termine total staffing for managing
Quantifying the Assignment Process capital-improvement projects.
During earlier projects for the State of Prior to our modeling efforts, Heery de-
INTERFACES 30:6 96
OPTIMIZATION MODEL
Figure 1: The intensity curve for various projects is given by project intensity ⴔ 6 • log (project
cost in $000,000) ⴐ 1. There is a large difference between intensities for a $2 million and a $3
million project but a much smaller difference between those for $28 million and $29 million
projects.
Figure 2: This shows a comparison of intensities for three construction representatives. Obvi-
ously, transferring some projects that are ongoing in the last few months from rep 1 to rep 3
would result in a more balanced workload. Thus, intensity curves facilitate a comparison and
adjustment of anticipated future workloads.
November–December 2000 97
LEBLANC, RANDELS, SWANN
INTERFACES 30:6 98
OPTIMIZATION MODEL
An important consideration in calculating the seven managers and some of the 114
intensities was the driving time for the projects to be assigned in a recent problem
managers to the various projects, so we re- (Table 1). Changing cells equal one if the
defined intensities as follows: manager is assigned to the project and
zero if not. This method of using ones to
project intensity ⳱ [1 Ⳮ driving time]
indicate manager-project assignments is
• [6 • log (project cost) Ⳮ 1] .
consistent with algebraic models’ use of
Thus, if driving time is one hour, then the zero-one decision variables defined for
intensity is half that of a project with iden- each manager-project combination as
tical cost but with three-hour driving time.
冦
1 if the manager is
We could calibrate this intensity by scaling
XManager,Project ⳱ assigned to the project,
(multiplying) it by a small constant such
0 if not.
as 0.01. This is not necessary, however, be-
cause the solution to any optimization For example, XDenise,3 equals 1 if Denise is
model is unchanged regardless of any assigned to manage project 3 and equals 0
scaling factor. These intensities were calcu- if she is not. These variables are often re-
lated in a separate Excel file and then in- ferred to as go/no-go variables.
put as data cells in the optimization In many cases, certain managers are as-
model. signed a priori to certain projects. For ex-
We let CManager,Project denote these inten- ample, the most experienced manager
sity values (C denotes cost). With this defi- might be assigned to very important pro-
nition of manager-project intensity, Heery jects, while recent hires are assigned to
can include in the model the fact that a smaller ones. Also, maintaining continuity
new hire is not able to manage as effi- of previous assignments to ongoing pro-
ciently as more experienced managers. jects is important. The user specifies a
However, Heery has no new managers, so priori assignments by manually entering
in the current model the assignment inten- 1s and 0s in the changing cells’ range of
sities do not reflect this variation. the spreadsheet (Table 2A) and reducing
We will explain the changing cells in the changing cells’ range so that the opti-
our Excel optimization model in terms of mization model excludes these assign-
November–December 2000 99
LEBLANC, RANDELS, SWANN
Table 2A: The Excel model shows input data in cells with double borders and changing assign-
ment cells (decision variables) in single heavy orders (actually red borders). Rows 4 through 10
give the intensity data cells for each manager-project combination. Rows 14 through 20 show
the changing cells using 1s and 0s, sometimes input by the user, and sometimes chosen by the
model, to specify manager-project assignments. Each column must have a single 1, since each
project must have only one manager assigned to it. This shows that Denise is assigned to the
first project, Tim is assigned to the next three, Chuck to projects five and six, and so forth. The
intensity of the assignments shown is [4.8] ⴐ [8.7 ⴐ 0.3 ⴐ 6.6] ⴐ [0.5 ⴐ 2.6] ⴐ [0.4] ⴐ [2.8 ⴐ
0.5] ⴐ [10.4]. . .ⴐ [68.4] ⴔ 106.0 plus the intensity of assigning projects 11 through 113, which
are not shown.
projects to minimize the sum of the inten- B of Table 2B shows that this intensity will
sities corresponding to the chosen assign- last during months 4, 5, and 6.) Finally,
ments. The model chooses assignments constraints ensure that project manage-
subject to constraints that balance the ment is not split among different manag-
workload given to each manager in each ers—each construction project must have
month. The first set of constraints is that exactly one manager assigned to it. In a
each manager’s total intensity cannot ex- problem instance with 114 projects, with
ceed a specified amount in each of the 54 projects assigned a priori (Table 1),
next 12 months. In a problem instance there are 60 such constraints.
with seven managers (Table 1), there are The key data that our model requires
7 • 12 ⳱ 84 such constraints, which ensure are the intensity values for manager-
that the chosen assignments are not unfair project assignments. We did not have data
to any individual. The next constraint set on the maximum or minimum intensities
also has 84 constraints to ensure overall allowed for each manager. (These were
fairness—each manager’s total intensity available for the earlier single intensity
must be at least a specified amount in each values for projects, but not for the
month. (Column B of Table 2A shows that manager-project intensities that we used.)
if Mike is assigned to manage project 1, he Therefore, we used the following proce-
will be busy at intensity level 2.8. Column dure. We first solved the model with zero
Table 2B: Rows 24 through 35 of the Excel model contain data cells specifying the months in
which each project is active, using ones to indicate active months. Rows 39 onward contain
coefficients for the minimum and maximum intensity constraints for each manager in each
month. Row 39 shows the two constraints for Denise in month 1. These constraints sum the
intensities for projects 2, 3, 4, 6, 8, . . . , 114. Projects 1, 5, 7, 9, and 10 are ignored in these two
constraints because they are not active in month 1.
minimum and very large maximum allow- The difficulty with this approach is that
able intensities. Next, we observed the projects at a site may not all have the
minimum and maximum values that the same schedule, with some active all year
model assigned to any manager. We then and others only in some months. With
tightened these values by one unit and re- each project no longer described by its
solved the model. We continued this pro- own spreadsheet column specifying its in-
cedure until further tightening resulted in tensities and active months, the partially
no solution to the model. We implemented active nature of the combined project com-
the solution to the model with the tightest plicates the maximum and minimum in-
minimum and maximum allowable inten- tensity constraints. To cope with this diffi-
sities for which a solution exists. This took culty, we used another approach that
only a few iterations to complete. keeps the larger number of variables
The Revised Manager-Project (separate project-manager combinations)
Assignment Model and adds additional constraints (Table 4).
After solving the assignment- These constraints ensure that a single
optimization model, we realized that occa- manager handles all projects at one site.
sionally the solution did not have one im- Experience with the Model
portant characteristic—all construction The natural progression of construction
projects in certain outlying locations must projects as they begin, end, and occasion-
have the same manager. (We had assumed ally stall dictates a monthly update of proj-
that the solution to the assignment model ect data and model solution at Heery’s
would always have this property.) For ex- Nashville office. Since the model’s incep-
ample, three projects (projects 4 through 6 tion, Heery’s program administrator
in Table 1—MTSU Stadium, MTSU Book- (Swann) has used it in each of the nine
store, and MTSU Library) should all be successive months. During this time, one
managed by one person. We refer to this manager resigned, and many projects
as the single-site policy. Since all projects were added and removed (completed). It
in some locations must have only one is a tribute to the model’s simplicity that
manager, we could reduce the number of since the initial run, the program adminis-
decision variables by defining some vari- trator has needed no expert assistance to
ables to refer to a certain location (Tables modify the model to reflect the new pro-
3A and B). jects and managers even though he had no
prior exposure to optimization modeling —“Using the model greatly simplifies the
or to the Solver software. However, the process of assigning managers to
single-site-policy aspect of the model has projects.”
proven more difficult. Because this prob- —“Assignments are now much less con-
lem occurs for very few projects and be- troversial. In the past there were time-
cause the model’s solution often satisfies consuming and sometimes heated discus-
these constraints even when they are not sions about assignments.”
formally included, the program adminis- Solution time using the Premium Solver
trator has ignored the constraints and Plus add-in for Microsoft Excel 97 for the
manually overridden the model’s solution model with 420 go/no-go variables varied
when a difficulty occured. For example, in considerably, depending primarily on the
a recent month the model’s solution as- minimum and maximum intensities al-
signed separate projects at Tennessee lowed. Many problems required only sec-
Technological University to both Mike and onds or a few minutes to solve for the ex-
Chuck, but the program administrator act optimal solution, but some took a few
overrode this, assigning them all to Mike, hours (on a Pentium 133 Mhz laptop with
after verifying that the total intensity as- 32 meg running Windows 95). A very few
signed to each manager was still within did not solve even after running over-
the allowable range. Except for this type of night. There may be no feasible solution to
change, the program administrator made these problems, meaning that the mini-
actual assignments according to the mum and maximum total allowable inten-
model’s solution in each of the past nine sities for each manager in each month
months. were inconsistent with the intensities for
Some managers reacted to the use of the managing the projects. From a practical
model as follows: point of view, there is no difference be-
—“I initially thought that Julie should tween a model with no feasible solution
manage the TTC McKinsey project, but the and one with a feasible solution that the
optimal solution called for Dean to man- available software cannot find in accept-
age it. After thinking it over, it does make able time. Solver quickly identified those
more sense for Dean to manage this problems that were obviously infeasible.
project.” Generally, for problems that had no feasi-
ble solution but were almost feasible, four hours during the last week of each
Solver took a long time to conclude that month updating and running the model
there was no feasible solution. and then notifies managers of their assign-
In a real sense, there is no cost of using ments. They have made essentially no
the model. The Premium Solver Plus up- complaints and only occasionally ques-
grade (which solves larger models more tioned their assigned tasks. This modeling
efficiently than the basic Microsoft Excel effort succeeded because all the affected
Solver) did cost $800, but it has paid for
itself just by reducing the time it takes to The affected managers see the
assign managers to projects. model’s assignments as fair.
Heery has saved money by using the
intensity function and our assignment managers wanted a better method of as-
model. Heery officially reported to the signment, and they see the model’s assign-
State of Tennessee that it has been saving ments as fair. We anticipate continued
the state $1 million per year in construc- positive results, with the model eventually
tion management costs. Most of this was being offered to all departments in the
due to the earlier use of the intensity worldwide Heery organization.
formula rather than our assignment- We combined model development and
optimization model. However, shortly af- user training during a six-month period.
ter Heery began using the assignment The program administrator and the first
model, one manager transferred out and author (consultant) met for a few hours
has not been replaced. Heery attributes its one or two days each month. We could
successful project management with the have completed model development and
reduced staff entirely to the use of our op- training in one month, but Heery gave pri-
timization model. It has also saved travel ority to obtaining new projects, so we con-
costs because the model’s solution assigns ducted training sessions intermittently.
managers to projects close to their homes It may be necessary in the future to pre-
(subject to allowable total intensities) more vent certain managers from being as-
often than the previous method. This is re- signed to certain projects, such as a new
flected in the fact that Heery has not in- hire to a multimillion dollar project. We
creased its travel budget even though the can do that by setting the intensity of that
state has imposed additional travel assignment to a huge number so that the
requirements. solution will not include that assignment.
Nonfinancial benefits include improved We think that the alternative of deleting
morale, because managers now perceive the corresponding XManager,Project variable
workloads as fair. Managers are given a would be more difficult for the users.
clear picture of their job responsibilities To test the robustness of our model for
and schedules well into the future, allow- larger problems, we also solved several in-
ing them to keep the big picture in view stances in which all 114 projects had to be
and plan for maximum efficiency. The assigned to seven managers. This model
program administrator now spends about has 7 • 114 ⳱ 798 go/no-go decision vari-
0 100 5 0:0:10
3 79 5 0:1:30
10 55 5 0:2:03
12 40 10 2:33:12
12 39 10 1:16:55
12 38 10 0:26:31
12 37 10 No solution
Table 5: The solution time on a Pentium 200 PC varies for different model instances. Solution
times generally (but not always) increased for tighter constraints. Solution times for exact-
optimal solutions would have been substantially higher. No solution means that no feasible
solution was found in eight hours.
ables (changing cells). Since the Premium problem.
Solver Plus software that we used for so- APPENDIX
lution is limited to 800 changing cells, any Notation for the algebraic optimization
model for assigning managers to projects
additional managers or projects would re-
is the following:
sult in a model too large for it to solve. Parameters:
To reduce the solution times for these
larger problems, we set the Solver toler- Cij ⳱ intensity of assigning manager i to
project j.
ance equal to five percent or 10 percent for
these problems. As a result, the solution
found was known to be accurate to within
Ajt ⳱ 冦10 if project j is active in period t
if not.
five percent (or 10 percent) of optimality
mit ⳱ minimum intensity allowed for man-
but was not necessarily exactly optimal ager i in period t.
(Table 5).
Mit ⳱ maximum intensity allowed for man-
Many other organizations could use a
ager i in period t.
similar approach to assigning managers. Decision variables:
Examples include assigning bus drivers to
routes, sales representatives and stock bro-
kers to customers, programmers to mod-
Xij ⳱ 冦10 ifif manager
not.
i is assigned to project j
mit ⱕ 兺
projects
[Ajt Cij]Xij ⱕ Mit Butterbaugh, K.; and Atherton, S. 1988,
“Maximizing profits for North American Van
j
Lines’ Truckload Division: A new framework
for each manager i and each period t. for pricing and operations,” Interfaces, Vol.
18, No. 1, pp. 21–41.
Because of the coefficients Ajt, this is the
Savage, S. 1997, “Weighing the pros and cons
sum of active projects in each period. Be-
of decision technology in spreadsheets,”
cause of the variables Xij, this is the sum OR/MS Today, Vol. 24, No. 1, pp. 42–45.
of projects assigned to manager i that are Straver, E. 1998, private communication, techni-
active in period t. cal support manager, Frontline Systems Inc.,
Additional constraints are that each pro- PO Box 4288, Incline Village, Nevada 89450
ject must have exactly one manager as- (http://www.frontsys.com).
signed to it: Thompson, G. 1997, “Assigning telephone op-
erators to shifts at New Brunswick Tele-
兺
managers
Xij ⳱ 1 for each project j. phone Company,” Interfaces, Vol. 27, No. 4,
i pp. 1–11.
Towle, T. 1990, “Determining staffing require-
References ments for managing a large capital improve-
Abara, J. 1989, “Applying integer linear pro- ment program,” Cost Engineering, Vol. 32,
gramming to the fleet assignment problem,” No. 10, pp. 15–18.
Interfaces, Vol. 19, No. 4, pp. 20–28. Willemain, T.; Jordan, E.; Lasdon, L.; Lenard,
Awad, R. and Chinneck, J. 1998, “Proctor as- M.; Moore, J.; and Powell, S. 1997, “OR/MS
signment at Carleton University,” Interfaces, and MBAs: Mediating the mismatches. Re-
Vol. 28, No. 2, pp. 58–71. port of the operating subcommittee of the
Conway, D. and Ragsdale, C. 1997, “Modeling INFORMS Business School Education Task
optimization problems in the unstructured Force,” OR/MS Today, Vol. 24, No. 1, pp. 36–
world of spreadsheets,” Omega International 41.
Journal of Management Science, Vol. 25, No. 3,
pp. 313–332.
Fylstra, D.; Lasdon, L.; Watson, J.; and Waren, Emory F. Redden, Vice President, Herry
A. 1998, “Design and use of the Microsoft International, Inc., 999 Peachtree Street
Excel Solver,” Interfaces, Vol. 28, No. 5, pp. NE, Atlanta, Georgia 30367-5401, writes:
29–55.
Grandine, T. 1998, “Assigning season tickets “The optimization model described . . . has
fairly,” Interfaces, Vol. 28, No. 4, pp. 15–20. been very helpful for assigning managers
Grandzol, J. and Traaen, T. 1995, “Using math- to projects . . .. We have been satisfied
ematical programming to help supervisors
with the assignments chosen at the Nash-
balance workloads,” Interfaces, Vol. 25, No. 4,
pp. 92–103. ville office . . .. We look forward to using
Grossman, T. 1997, “End-user modeling,” the model in our Atlanta office and else-
OR/MS Today, Vol. 24, No. 5, p. 10. where in the Heery organization.”
Hillier, F. and Lieberman, G. 1986, Introduction
to Operations Research, fourth edition,
Holden-Day, Oakland, California.
Jarrah, A. and Diamond, J. 1997, “The problem
of generating crew bidlines,” Interfaces, Vol.
27, No. 4, pp. 49–64.
Powell, S. 1997, “The teacher’s forum: From in-
telligent consumer to active modeler. Two
MBA success stories,” Interfaces, Vol. 27, No.
3, pp. 88–98.
Powell, W.; Sheffi, Y.; Nickerson, K.;