A Modified Game Theory Approach

to Multiobjective Optimization
S. S. Rao Many mechanical and structural design problems encountered in practice require
solutions which balance several conflicting objectives. The vector, scalarization, and
trade-off-curve methods have been developed to achieve multiobjective solutions.
T. I. Freiheit One of the best known methods for generating a compromise solution, based on
the concept of Pareto minimum solution, is the cooperative game theory method
School of Mechanical Engineering, since it uses a scalarized approach and has a numerical measure of compromise.
Purdue University, However, game theory is hard to automate due to a two step optimization process
West Lafayette, IN 47907 involved. Hence, in this work, a modification to the game theory is introduced in
which the two optimization steps are combined and an algorithm for its implemen-
tation is developed. The algorithm is tested on two numerical examples, including
one dealing with the probabilistic design of an eighteen speed machine tool gear
train. The probabilistic theory necessary for the design of the gear train is also
introduced. The examples validate the modified game theory.

Introduction
In today's competitive economy, it is no longer possible for inferior solution in which the user indicates which objective is
the designer to find solutions to conflicting objective problems to be favored over another (Mukai [3]) and multicriteria dy-
which are merely adequate. The designer must select the design namic programming (Rosenman and Gero [4]).
variables such that there is a superior compromise between the A method related to the vector methods is the e-constraint
objectives. To this end, a large number of methods have been method. This method optimizes one of the objectives subject
developed for the solution of multiobjective optimization to the other objectives being constrained to the neighborhood
problems. The earliest reported in-depth work on the for- of their respective goals. It is mathematically formulated as:
mulation of the multiobjective problem is that of Kuhn and Find the design vector X e S such that it:
Tucker [1]. The methods for optimizing multiple objectives
minimizes f(X) (2)
may be categorized into two types, leaving the objectives in
vector form and scalarizing the objectives into one equation. subject to
In the vector form, the multiobjective (or multicriteria) design //X)<e,;y=l,2,...,
problem may be mathematically defined as: Find the design
vector X = (xl,x2,...,xk) that will where n is the number of objectives, j^i, and e may be varied
T
to generate the entire Pareto optimal set. This method has a
minimize f(X) = (/,(X), / 2 (X), , /(X)) (1) major drawback in that the value of e in defining the constraints
subject to: is not known.
The technique for the optimization of single objective prob-
X e S = {Xl^(X)<0,y=l,2,...,m) lems are well established and have been proven effective and
where S is the feasible design space. The best known method efficient. It is desirable, therefore, to scalarize the multiob-
of optimizing several objectives while leaving them independ- jective problem into a single objective such that a noninferior
ent of one another is the mini-max method (Osyczka [2]). This compromise solution may be found by any standard single
entails minimizing the maximum deviation from the objective objective optimization program. The most common scalari-
goals (desired values) with the constraint that the solution must zation method is the goal programming method. Goal pro-
be Pareto optimal. A Pareto optimal (noninferior) solution is gramming is the minimization of the aggregate deviation of
a solution to a multicriteria problem in which any decrease in the objectives from predetermined goals. Mathematically it
one objective results in a simultaneous increase in one or more can be stated as follows:
of the other objectives. Mathematically, a solution X* e S is
Pareto optimal if and only i f X e S and f(X)<f{X*) for
/' =1,2,..., k imply that fi(X)=f(X*) for i=\,2,...,k. Two other minimize /(X) = ) E ( d + ; + <*-/("'. ^ 1 (3)
vector methods are an algorithm which will generate a non- S/=i
subject to
Contributed by the Design Automation Committee for publication in the
XeS
JOURNAL OF MECHANICAL DESIGN. Manuscript received April 1990. f(xy-d; + dj = bj

286 / Vol. 113, SEPTEMBER 1991 Transactions of the ASME

Copyright 1991 by ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27589/ on 07/31/2017 Terms of Use: http://www.asme.org/
 where bj is the goal and df, dj are the respective over- or This paper presents an algorithm utilizing a modification of
under-achievement of the goal for the j t h objective function, cooperative game theory to generate compromise solutions to
/}(X). Several systematic methodologies have been developed multiobjective design problems. The importance of this al-
which utilize goal programming for the solution of multicriteria gorithm is that it creates a systematic method of solving con-
problems (Dlesk and Liebman [5]). These methods also include flicting multiobjective problems with no need for interaction
slight variations on goal programming by placing weights on with the designer during optimization. Furthermore, the de-
the goals. Another common scalarization method is the utility signer is given a measure of how good a compromise solution
function method. This method formulates the objectives into is by the formulation of supercriterion within the problem.
one expression using a product or summation. The most widely The algorithm is tested on two problems including the prob-
used expression is the weighted sum, where the weights indicate abilistic design of a machine tool gear train.
the relative importance of each objective:
Game TheoryCooperative and Noncooperative
minimize U(X) = cJ^X) (4) A multiobjective design problem may be envisioned as a
game in which each objective is a player competing to optimize
his standing in a system subject to limited resources. Operations
subject to X ( S, J ] Cj = 1, and c, is the weight on the ;'th researchers and economists, in studying competitive systems,
/=i
have developed theories for games which are readily applicable
objective. The weighted sum method generates a Pareto op- to engineering design. Two theories have been used to describe
timal solution but is deficient in that it is difficult to determine the interaction of players; the noncooperative theory, based
which weights will produce the best compromise solution. Fur- on the concept of Nash equilibrium, and cooperative game
thermore, a systematic redistribution of the weights on the theory, based on the concept of Pareto minimum solution.
objectives does not produce a continuous Pareto optimal so- The noncooperative theory of games assumes that each player
lution set, making it difficult to compare different compromise is looking out for his own interests. Each player selects his
solutions found using this technique. share of the resources only with the view of optimizing his
Other methods of scalarization include the global criterion own objective and does not care how his choice will affect the
method, the ratios method, and the game theory method. The objectives of the other players. The players then bargain with
global criterion method (Rao [6]) finds a compromise solution each other, exchanging resources, until an equilibrium is
by minimizing the sum of the squares (or a power greater than reached. The resultant solution, referred to as a Nash equilib-
one) of the relative deviation of the objectives from their fea- rium, is a solution where no player may improve his objective
sible ideal points (from the Pareto optimal curve): by attaining some different amount of the resources as long
r/xx*,) - f&) as the other players maintain their resource choices. This so-
minimize F(X) = V* (5) lution can be dependent on the order in which the resources
7XX*,-) are allocated and the order in which the players choose the
subject to the constraints X e S, and p > 1. The ratios method resources when more than one equilibrium point exists. It
(Metwalli et al. [7]) is used to optimize multicriteria problems should be noted that solutions exist in which both players have
which have some objectives that need to be minimized and better function values than those found at the Nash equilibrium
other that need to be maximized. The method is preformed points.
by nondimensionalizing the objectives and minimizing the ratio Cooperative game theory assumes that each player is a mem-
of the minimization objectives to some power over the max- ber of a team willing to compromise his own objective to
imization objectives to some power. The problem with this improve the solution as a whole. In the cooperative solution,
method is that the powers on the objectives must be varied to the team would want to allocate the resources with the intent
generate enough solutions to make a choice on the compromise that all players should be as optimal as possiblein other
solution; this gets computationally very expensive. The game words, a Pareto optimal solution. The team then must decide
theory method treats each objective as a player in a game and how to distribute the resources such that a gain for one player
formulates a supercriterion to model the interaction of the does not result in an unacceptable loss for another player. One
players. Rao [6] compares nine different multicriteria methods method is to distribute the resources such that all players are
through the design of a vibration isolation system. He studies as far from their worst cases as possible. To find a solution
the efficiency of the different methods and compares the final which is as far as possible from the worst cases, it is necessary
designs. to formulate a supercriterion as a measure of compromise.

Nomenclature

A = center distance
d, = pitch circle diameter of / th
gear / = speed ratio ^(^iv) = bending (surface wear)
E = Young's modulus M, = torque on, gear stress
fj = / h objective function nw = speed of wheel T = number of teeth on wheel
/, = / h normalized objective P = power transmitted by gear U = face width of ith gear
function Rs = reliability of system X = vector of design variables
fx(x) probability density func- R
cij = reliability of ith component Xf = optimum solution of f,{X)
tion of X in / h load condition Xi = ith design variable
Fiu = worst value of /th objective = feasible design space; su- z = standard normal variate
function
s percriterion = mean (standard deviation)
< ).*()
gj = / h inequality constraint Sw) = bending (surface wear) of()
function strength a = pressure angle

Journal of Mechanical Design SEPTEMBER 1991, Vol. 113 / 287

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27589/ on 07/31/2017 Terms of Use: http://www.asme.org/

 This supercriterion, a function of the objectives, is essentially (3) Formulate the supercriterion S:
a penalty functionpenalizing solutions with objectives that
are too close to their predefined worst cases. A formulation
for a supercriterion which has been used effectively (Rao and * = iii - /,(X) (9)
Hati [8]) is:
Note that due to the normalization, S will always have a value
between zero and oneof the same magnitude as the nor-
F,u - F,(X%) (6) malized objectives.
(4) Formulate a Pareto optimal objective FC using the nor-
where Fiu is the worst value, X*c is a Pareto optimal solution, malized objectives. Suggested methods are:
and a better compromise solution occurs when S is maximized. (a) Weighted sum:
Implementing cooperative game theory requires an algo-
rithm which will perform two steps. First, the set of Pareto FC
optimal solutions are to be generated. Then the Pareto optimal
= c,/(x) (10)
solution which has the most optimal supercriterion is to be
selected as the best compromise design. Any method which
subject to J ^ c, = 1.
can generate a Pareto optimal solution is suitable for the first
;'=1
step. Methods which generate Pareto optimal solutions include
weighted sum, e-constraint, and goal programming. The (b) Goal programming:
weighted sum method is particularly suitable to cooperative \/P
game theory by eliminating the need to generate the entire = (L/-AX)) (11)
Pareto optimal solution set [8]. This is done by an iterative FC =
process involving two optimizations: First minimize the where 2 </?<<.
weighted sum of the objectives to produce a Pareto optimal (c) e-constraint variant:
solution and calculate its supercriterion. Then maximize the
supercriterion by modifying the weights and iterate until an FC = /,(X) (12)
acceptable solution is achieved. Although this process is dif- subject t o / / X ) - e / < 0 , i^j, and e determined by
ficult to implement in an automated routine because of the
double optimization, it has been applied to the optimal design tj=MX*j) + Cj Fju-fnj(X*j)
of vibration isolation systems (Rao and Hati [8]) and structures
(Rao [9]). A general discussion of the concepts of game theory and 0 < c y < l .
in the design process may be found in Vincent [10], (5) Since FC has to be minimized and S has to be maximized,
a new objective is constructed as (for minimization):
OBJ=FC-S (13)
Modified Game Theory
subject to:
To date, the procedure for finding an optimal cooperative
game theory solution has been inefficient and required much XeS
manual effort for the designer. It is proposed then to increase and
the efficiency by modifying the game theory method such that
the Pareto optimal solution generation and the maximization fni(X*i)^fni(X)^Fiu; i=l,2,...,m
of the supercriterion may be performed simultaneously. This with the design vector, X, now including the weights on the
may be done by combining the Pareto optimal solution gen- objectives (weighted sum) or the constraints (e-constraint). It
eration and the supercriterion into one objective by subtracting is important to include the constraints on the normalized func-
the supercriterion from the Pareto optimal objective. This new tions to insure that S remains between zero and one.
objective may be minimized by a standard, single objective The utilization of goal programming in the Pareto optimal
optimization program. This formulation, however, does change objective is effective because it already generates fairly good
the characteristic of the optimization problem such that it is compromise solutions. However, the weighted sum method
neither truly minimizing the Pareto objective nor is it truly and the e-constraint variant have some special characteristics.
maximizing the supercriterion. The solution of this new ob- In the weighted sum method, the addition of the weights to
jective, as demonstrated by the numerical examples, is near- the design vector allows the algorithm not to get "stuck" on
optimal and, from an engineering point of view, the compu- local maximum solutions of the supercriterion, a problem that
tational savings warrant the acceptance of a solution which is frequently encountered when the supercriterion is maximized
deviates slightly from a true game theory optimum. An al- alone. In the e-constraint variant, the final solution favors the
gorithm for implementing the modified game theory follows. objective which is used as the Pareto optimal objective. This
(1) From a constant initial design vector, minimize each of is due to the minimization/maximization that is occurring in
the n objectives separately, recording the values of the other the modified game theory objective; the e-constraint objective
objectives at each optimal design vector. is being minimized while the supercriterion is being maximized.
minimizefj(X), i = 1,2,...,. (7)
subject to
Probabilistic Design of Gear Trains
XeS
The multiobjective probabilistic design of an eighteen speed
(2) Normalize the objectives so that no objective due to its gear train is considered as an example to demonstrate the
magnitude will be favored. The following normalization pro- computational efficiency of the modified game theory tech-
cedure gives zero as the optimum value and one as the worst nique. The reliability of a multicomponent system can be found
value of the ;'th objective function: using a series model as (Rao and Das [11]):

/m(X)
fm-ffx-T) (8)
Fiu-fi&t)
R,
II IK (14)
where Fiu is the worst value, and/(X*) is the optimum value / = i y'=i
of the ;' h objective. where Rc,, is the reliability of the; component in they' loading

288 / Vol. 113, SEPTEMBER 1991 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27589/ on 07/31/2017 Terms of Use: http://www.asme.org/

 condition with u and v denoting the number of components 0M,
and number of load conditions, respectively. The reliability of sb (22)
A ' t
any component is the integral of the product of the probability
distribution function of the load on the component and the where /S is a constant given by
probability density function of the strength of the component: = kc kd (/ + 1)
(23)
= (' m y cos a
*' f FL{r).fs(r)-dr (15)
and y is the Lewis form factor given by

"H1 + f)
The commonly used probability density function, namely,
the normal distribution, can be expressed for the variable L y (24)

The wear (contact) stress on the gear tooth is given by:

(16)
-.Fhw
2-n- a, V la
L / 7
aL sw

where y and ay are given by

A
(?) (25)

The product of two normal probability density functions can

be simplified by defining a dimensionless standard normal Ekckd{i+\)
variate z as: T - a(<I) sin 2a
(26)

izl (17)
<7 7 = (27)
2 2 1/2
IE
where ,=S-L and = S - L and oj = (os + aL) . This The mean and standard deviations of sb and sw can be computed
formulation assumes that the two variables are independent as outlined in Rao and Das [11]. The equations for sb and
of each other. Using the standard normal variate, the integral as (sw and os ) can be used to find the reliability of one gear
for reliability now becomes:
pair in bending (surface wear). The overall reliability of the
R ex 2 dz (18) gear train for failure due to bending stresses and failure due
'= i 1 p(-^ ) to wear stresses can be calculated using the weakest link chain
where the lower limit of integration Z\ is the standard normal model. Depending on the output speed of the gear box, each
variate with the random variable set to zero. The evaluation gear pair will rotate at one of several speeds or will be dis-
of this integral is readily found from standard normal tables engaged. For a worst case reliability analysis of any gear pair,
or approximation equations. The random variable is set to the lowest reliability calculated from the several speeds will be
zero because it is desired to find the probability that the strength selected as the reliability of that particular gear pair. Mathe-
is greater than the load, i.e., = S - L > 0 . matically, the reliability in bending of the gear train will be:
In the case of gear design, it is desired to find the reliability
R =
of the gear teeth under the influence of bending and contact
stresses. Therefore, the standard normal variate formulation % n ^v (28)

of the lower bound for calculating the reliability in bending

where u is the total number of gear pairs in the gear train and
will be:
min(.#j tj) is the minimum reliability in bending of the i gear
sb J
Z\b 2.1/2 (19) when it turns at the j speed. The equation for reliability of
K a, ) the entire gear train for failure in surface wear can be found
and the reliability in wear can be found by replacing b with w by replacing in equation (28) w for b.
in equation (19).
Assuming that the kinematic design of the gear train is
known, the strength design involves finding the face width of Numerical Examples
each gear pair such that the stresses in the gear teeth do not Example 1-Optimization of Three Algebraic Functions
exceed the material strength and no extra material is wasted.
The possible objectives are to minimize the total weight of the The ability of the modified game theory algorithm to achieve
gear train and to maximize the reliability of the gear teeth. a solution close to the maximum supercriterion is tested by
The total mass of a gear train, Wt, may be calculated by considering the minimization of the functions:
summing the mass of each gear in the train: 1 2

k ,
*=2X
Wtit) (20) yi (pc-iy+i (29)
2
y3= (x-3)
where p is the density, d, is the pitch circle diameter of the /
gear, t, is the face width of the /' gear and k is the number of yx, y2, and y}, have minima of 0, 1, and 0, respectively. The
gears in the gear train. worst values for yu y2, and y3 are 9/2, 5, and 9 and occur for
The mean and standard deviation of the stresses in bending the functions when x is at 3, 3, and 0, respectively. The func-
and wear on the gear teeth for any loading condition can be tions are normalized as:
calculated as follows (Rao and Das [11]). A gear pair trans- 2
mitting powr p(kW) at speed nw (rpm) will experience a torque t - i
M, (kgf-cm) given by:
Jn\ - g*
M, = 97500 (21) (30)

The stress in bending may be calculated, assuming the model f2=\(X-D2

of the tooth as a cantilever beam with an end load, as:
fm= (*-l) 2
Journal of Mechanical Design SEPTEMBER 1991, Vol. 113 / 289

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27589/ on 07/31/2017 Terms of Use: http://www.asme.org/

 Table 1 Compromise solutions for Example 1 Table 2 Initial design vector for eighteen speed gear train
yi y3
fi f2 f3 t > (0.57,0.94,0.81,0.69.1.97.1.25.0.81.6.64.2.22)
Modified Game Theory
Wti Rb; Rw, S
122.426 -.685128 -.932433 0.387938
Weighted Sum 0.728940 1.04303 3.21332 0.533017
normalized 0.323656
cl=0.0,c2=1.0 0.161987 0.01755 0.35703
Func.
0.988965 1.16515 . 2.53959 0.536943 Eval.
Goal Programming
0.21977 0.041288 0.282177 Min Wt 101.826 -.0151408 -.791253
Max Rb 233.153 -.999994 -.999982
E-constraint var. 0.548960 1.00229 3.81102 0.505903 Max Rw 220.982 -.999994 -.999993
FCy, 0.000572 0.423446
Note - Face widths have upper and lower bounds of 10.0 and 0.5 cm
0.729322 1.04316 3.21218 0.533050 except the eighth pair, which has an upper bound of 11 cm.
FC=y 2
0.162072 0.010789 0.356909

1.43615 1.48273 1.70359 0.485365

FC=y 3
0.319144 0.120682 0.189287

the worst case is 2.25 percent, well within an acceptable

Olhcr Multiobjeclive Optimization Methods
engineering tolerance. It should be noted that the relative
Sid Game Theory 1.1387 0.648336 1.01924 3.46438 weights calculated in the modified game theory using the
d=.000I,c2=.9999 0.144075 0.004810 0.384931 weighted sum as the Pareto optimal objective are very close
1.107019 1.21437 2.36236
to the best relative weights found using standard game theory.
Sid Goal Programming
0.237819 0.053592 0.262485

Example 2 - Probabilistic Design of an Eighteen Speed Gear

input: Shaft I
Train
output: Shaft 5
The design of the eighteen-speed machine tool gear train
f shown in Fig. 1 is considered with three objectives: (1) Min-
Group A
imization of weight, (2) Maximization of reliability in bending,
and (3) Maximization of reliability in wear. All parameters of
fc3VA /~\ Group 8
the gear train are specified with the design vector representing
the face widths of nine gear pairs. The reliability in bending
^ 7 2*^_^ Group C
and in wear of each gear pair is constrained not to fall below
50 percent. The number of teeth on the (pinion, wheel) for
Group 0
each of gear pairs 1,2,...,9 are (26,40), (33,66), (38,61), (44,55),
(24,96), (40,80), (60,60), (24,96), (40,80), respectively. The
mean center distances are (13.2,19.8, 24.0, 24.0) cm, the bend-
ing strengths are (2000, 2200, 2700, 3000) kfg/cm 2 , the surface
Fig. 1 Schematic of the eighteen speed gear train
wear strengths are (16800, 16800, 17050, 19530) kfg/cm 2 and
the coefficients of variation of the shaft speeds are (0.10, 0.07,
0.05, 0.03) for the gear groups (A,B,C,D). Other data are:
The supercriterion to be maximized then becomes: gear module = 4 mm, pressure angle = 20 deg, face width
c.v. (coefficient of variation) = 0.05, center distance c.v. =
1 0.05, stress and strength c.v. = 0.10, mean power transmitted
S = [x6-8x5 + 90x3-81x2- 162.x] (31)
324 = 7.35 kw, power transmitted c.v. = 0.10, Young's modulus
The maximum value of the supercriterion is found to be = 2.1 x 10 kgf/cm 2 , density = 0.00775 kg/cm , kc = 1.5
0.539387 which occurs at the design parameter x = 1.33017. and kd = 1.1.
The objective function values corresponding to this maximum The first step in the modified game theory algorithm is to
supercriterion are: minimize each of the objectives separately. Table 2 shows the
initial design vector, the initial values of the objectives and the
>>, = 0.884682 /, = 0.196596 values of all of the objectives corresponding to the minimum
y2 = 1.109015 f2 = 0.027253 of each individual function. The minimization of the weight
j>3 = 2.788317 / 3 = 0.309813 has produced the worst values for the reliabilities and the
Note that y2 is the closest and y3 is the farthest from their minimization of the negative of reliability in bending has pro-
respective optima. duced the highest weight. When the weight was minimized,
Table 1 presents the solutions of the three modified game four reliability constraints, all in bending, became active or
theory formulations presented earlier, namely, weighted sum, were very close to their minimum allowable value. The face
goal programming, and e-constraint variant. For comparison, widths of the first two gear pairs had also reached their min-
Table 1 also presents the solutions generated by the standard imum value of 0.5 cm at the minimum weight. The face width
game theory and goal programming methods. Of the five mod- for the eighth gear pair was active at its maximum of 11 cm
ified game theory formulations listed in Table 1, the best com- when both the reliability in bending and in wear were maxi-
promise solution was produced when goal programming was mized.
used in generating the Pareto optimal objective. Its supercri- In Table 3, five different optimal compromise solutions for
terion deviates from the maximum value by only - 0.4 percent, the face widths of the gear train (given by modified game
better than the standard goal programming result whose su- theory) are presented along with the corresponding design ob-
percriterion errs by - 1 . 4 percent. The worst modified game jectives. Note that three of the five solutions correspond to
theory solution deviated from the optimum supercriterion by the e-constraint variant, one for each objective used as the
- 1 0 percent. This was produced by the e-constraint variant Pareto optimal objective. As indicated earlier, the best com-
which used FC=y3. However, this large error is to be expected. promise solutions are those solutions whose supercriterion is
As noted earlier, the e-constraint variant generates solutions maximized. All of the modified game theory solutions have
which favor the Pareto optimal objective, which in this case supercriterion values of almost twice the supercriterion cor-
was also the objective farthest from its optimal solution at the responding to the intial design vector. Furthermore, the su-
maximum supercriterion. The average error of the supercri- percriteria at the optimal solutions differ by no more than 4.1
terion for the modified game theory technique not including percent. From an engineering standpoint, any of these solu-

290 / Vol. 113, SEPTEMBER 1991 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27589/ on 07/31/2017 Terms of Use: http://www.asme.org/

 Table 3 Eighteen speed gear train Modified game theory results For highly complex design spaces such as in this problem, this
Objectives
S Furtc. sensitivity contributes to the large number of solutions con-
Wl j Rb | Rw Eval.
verged upon in the neighborhood of the maximum supercri-
Weighted Sum 128.369 -.908074 -.988511 0.683623 terion.
cl=0.0,c2=.0116: 0.202115 0.0933341 0.0550054
Table 4 gives four solutions for the face widths of the gear
t*=(0.75034,0.97238,0.83857,0.71323,2.0996.1.2953,0.83693,6.9496,2.2961) train using standard game theory and other multicriteria op-
,,,. 128.092 -.906165 ' -.995118' 0.706873
timization techniques. The best solution found using standard
Goal Programming ^ ^ ^ . . ^ ^ ^ game theory (which required 2790 function evaluations) had
Tl=(0.94591,0.96361.0.82903,0.70420,2.1636,1.2776,0.82358,6.8674,2.2632) relative weights whose values were very close to the values
E-constraint var. FC=Wl 123.578 -.846875 ..993448 0.682552
obtained with the modified game theory using the weighted
Cl=.66073,c2..66976 0.165631 0.155474 0.0313551 sum Pareto objective.
t=(0.94595,0.93604,0.80586,0.68521,2.1092,1.2389,0.80110,6.5727.2.1897)

FC=Rb 136.108 -.963867 -.996497 0.699931 Conclusions

Cl=.29855,c2=.99851 0.26104 0.0366821 0.0167475
r=(090724,1.0l77.0.87649.0.74492,2.2512,1.3569,0.87606,7.3775,2.4000)
An algorithm for the solution of multicriteria design problem
has been presented. It utilizes a modification to cooperative
FC=Rw
cl-22216,c2=.1000
130.450
0.217963
-.926698
0.0744235
-.996223
0.0180627
0.710761 game theory which automates the design process. The modi-
t*=(0.99745,0.97887,0.84239,0.71693,2.2157,1.2984,0.84031,6.9840,2.2924)
fication is accomplished by combining the generation of the
Pareto optimal solution and the maximization of the super-
criterion. The algorithm is numerically tested on two mul-
Table 4 Eighteen speed gear train results given by other multicriteria tiobjective optimization problems. One of the examples
optimization methods represents the probabilistic design of a machine tool gear box.
Objectives Thus use of the modified game theory algorithm is quite general
Method Wl Rb Rw
and it may be applied to any multicriteria optimization prob-
fl f2 f3
lem. Of the three Pareto optimal objective formulations sug-
Standard Game Theory
c2=.011,c3=.988
125.236
0.178257
-.873447
0.128493
-.992256
0.0370654
0.6 gested in this paper, the goal programming Pareto objective
produced the best results. If the designer is interested in giving
"t^=(0-88813,0.96406,0.83252,0.71062,2.0455,1.2843,0.83797,6.6883,2.2438)
one objective more weight than the other objectives, it is sug-
126.181 -.886944 -.993259 gested that the e-constraint variant be employed.
Goal Programming
0.185454 0.114787 0.0322592
l
t =(0.87426,0.95487,0.82403,0.70212,2.1069,1.2683.0.82009,6.7649,2.2521)
References
123.596 -.852516 -.990332 0.676479 1 Kuhn, H. W., and Tucker, A, W., "Nonlinear Programming," in Pro-
Global Criterion
0.165769 0.149746 0.046284
ceedings of the Second Berkeley Symposium on Mathematical Statistics and
"^=(0.81831,0.94086,0.81101,0.68951,2.0451,1.2487,0.80702,6.6339.2.2172) Probability, J. Neyman, ed., University of California, Berkeley, CA, 1951, pp.
481-491.
Dynamic Programming [12]
135.637 -.878272 -.996946 0.641270 2 Osyczka, A., Multicriterion Optimization in Engineering with FORTRAN
0.257457 0.123594 0.14597 Programs, Halsted Press, a division of John Wiley and Sons, New York, 1984.
f>=(1.1746,1.1682.0.99509.1.7085,2.2928.1.3298,0.83116,6.8622,1.9906) 3 Mukai, H., "Algorithms for Multicriterion Optimization," IEEE Trans-
actions on Automatic Control, Vol. AC-25, No. 2, April 1980, pp. 177-186.
4 Rosenman, M. A., and Gero, J. S., "Pareto Optimal Serial Dynamic
Programming," Engineering Optimization, Vol. 6, 1983, pp. 177-183.
5 Dlesk, D. C , and Liebman, J. S., "Multiple Objective Engineering De-
tions would be acceptable as they are all in the neighborhood sign," Engineering Optimization, Vol. 6, 1983, pp. 161-175.
of the best compromise solution. Of the different modified 6 Rao, S. S., "Design of Vibration Isolation Systems Using Multiobjective
game theory solutions, the largest supercriterion is the one Optimization Techniques," American Society of Mechanical Engineers Paper,
No. 84-DET-60, 1984.
found using the e-constraint variant which had reliability in 7 Metwalli, S. M., et al., "Multiple Design Objectives in Hydrodynamic
wear as its Pareto optimal objective, followed closely by the Journal Bearing Optimization," ASME JOURNAL OF MECHANISMS, TRANSMIS-
result found using goal programming as the Pareto optimal SION, AND AUTOMATION IN DESIGN, Vol. 106, No. 1, March 1984, pp. 54-60.
objective. The lowest value of the supercriterion is the one 8 Rao, S. S., and Hati, S. K., "Optimum Design of Shock and Vibration
Isolation Systems Using Game Theory," Engineering Optimization, Vol. 4, No.
found using the e-constraint variant with weight as the Pareto 4, 1980, pp. 215-226.
optimal objective. This is due to the combination that the 9 Rao, S. S., "Game Theory Approach for Multiobjective Structural Op-
weight is the most dynamic objective and the e-constraint var- timization," Computers and Structures, Vol. 25, No. 1, 1987, pp. 119-127.
iant favors the Pareto optimal objective. 10 Vincent, T. L,, "Game Theory as a Design Tool," ASME JOURNAL OF
MECHANISMS, TRANSMISSIONS, AND AUTOMATION IN DESIGN, Vol. 105, June 1983,
A comparison of the weighted sum and the goal program- pp. 165-170.
ming Pareto optimal objectives shows how sensitive the su- 11 Rao, S. S., and Das, G., "Reliability Based Optimal Design of Gear
percriterion is to changes in the design vector. Increases of 2.5 Trains," ASME JOURNAL OF MECHANISMS, TRANSMISSIONS, AND AUTOMATION
and 1 mm in the first and fifth gear pair face widths with IN DESIGN, Vol. 106, No. I, March 1984, pp. 17-22.
12 Dhande, S. G., "Reliability Based Design of Gear Trains A Dynamic
minor adjustments to the other face widths produced a sig- Programming Approach," Proceedings 1st Design Engineering Technical Con-
nificant increase in the supercriterion between the two designs. ference, New York, NY, Oct 5-9, 1974, pp. 413-422.

Journal of Mechanical Design SEPTEMBER 1991, Vol. 113 / 291

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jmdedb/27589/ on 07/31/2017 Terms of Use: http://www.asme.org/