ON FUZZY GOAL PROGRAMMING*

Edward L. Hannan,** State of New York Department of Health

ABSTRACT
This paper illustrates how the goal programming problem with fuzzy goals having
linear membership functions may be formulated as a single goal programming problem.
Also, a previously defined method for dealing with fuzzy weights for each of the goals is
reexamined.
Subject Areas: FUZZJ Sets, Goal Programming, and Decision Processes.

INTRODUCTION

In a recent paper in Decision Sciences, Narasimhan [ 3 ] discusses the use of

fuzzy set theory in goal programming. Fuzzy goals and constraints are defined
and are incorporated within the traditional goal programming formulation.
Under the assumption of linear membership functions, the goal programming
(GP) problem with fuzzy goals is formulated as a set of 2K linear programming
(LP) problems, each containing 3 K constraints, where K is the number of goals in
the original problem. The methodology is then extended to accommodate fuzzy
priorities corresponding to the different goals in the problem.
The purposes of this paper are (1) to indicate how the fuzzy G P problem
with linear membership functions may be formulated as a single LP problem with
2K goal-related constraints and ( 2 ) to illustrate a conceptual problem that arises
when defining fuzzy priorities as in [ 3 ] and indicate an alternative way of defining
fuzzy priorities.

FUZZY GOAL PROGRAMMING WITH

LINEAR MEMBERSHIP FUNCTIONS

In the presence of a fuzzy decision environment, the goals in a GP problem

cannot be articulated precisely. Instead, fuzzy goals such as a profit of “consid-
erably more” than $lO,OOO, a “sufficiently high” rate of return, and a sales
revenue of “close to” $4O,OOO are postulated by the decision maker.
In j3, p. 3271, the fuzzy G P problem is represented as the determination of
the optimal decision D

XrO

*The author would like to acknowledge the valuable suggestions of two anonymous reviewers.
**The author is on leave of absence from Florida International University.

5 22
19811 Hannan 523

-
where “ ” is a “fuzzifier” representing the imprecision in the stated goals, X is
an n x 1 column vector, b is a K x 1 column vector, and A is a K x n matrix with
typical components Xi,b , and aij, respectively.
Also, membership functions are defined [3, p. 3271 as

p J A X ) = 1, if (AX),=bi

= f l ( A x ) ,b,), if ( A m i # bi

where (AX)iis the ith equation of AX.

The membership function of the decision set, po(x), is given by (3, p. 3271

and the maximizing decision is 13, p. 3271

When the membership functions are linear [3, p. 3281,

/ 0 if ( A m i sbi- Ai

\ 0 if (AX),? bi+ Ai
where the Ais are subjectively chosen constants.
This approach is applied by Narasimhan to the solution of a problem involv-
ing K = 3 goals, each with a linear membership function, and 2K= 8 subproblems,
each with 3K = 9 constraints. The j t h subproblem involves the maximization of
the quantity Aj subject to some constraints of the type [3, pp. 328-3291

bi + A , - (AX)i
Aj I
Ai
and
5 24 Decision Sciences [Vol. 12

for some goals and some of the type

and

for other goals.

I f there are K goals, there are 2K different combinations of constraints such
that for a given goal i, either type (2) constraints or type (3) constraints is in-
cluded.
The following theorem establishes the fact that the best of the 2K solutions to
the subproblem can be obtained by formulating a single equivalent GP problem
with only 2K constraints (see [2]or ( 1 ) for GP formulations).
Theorem: Let A * = Max Aj where Aj is the optimal solution to the
j =1 , . . ., 2 K
subproblemj mentioned above (A* is the optimal solution of a fuzzy GP problem
with 2K subproblems). Let A " be the optimal solution of the following problem:

Problem I Max A

A +d,:+dTs 1 i= 1 , . . ., K (5)

Xj' 0 j=l, . . ., n
A, d,:, d: 1 0 i=l, ..., K.
Then A* =A".
Proo$ Suppose A* =Aj; that is, suppose the j t h subproblem produces the op-
timal solution. Also suppose, without loss of generality, that after renumbering
the constraints of t h e j t h subproblem they are of the form

bj+ Aj-(AX)j
Aj s i= 1 , . . ., i, (6)
Ai
b i S ( A X ) j S bj+ Aj i=l, ..., i , (7)

bi-AjS(AX)jSbj i=i.+ 1 , . . ., K. (9)

19811 Hannan

For each set of constraints of the form (6),

b'
Since XS -
( A x ) i after dividing (7) by Ai, let
Ai 4

b. (AX),
This is identical to (4)when 2 5 -.
Ai Ai
Also, from (10) and (1 l),

Aj+dlF I1.

Thus, Aj + d; + d: I1 since d: = 0.
Hence, all feasible values of X and Aj in (6) and (7) are also feasible in Prob-
lem 1. In a similar manner it can be shown that all feasible values of X and A, in
(8) and (9) are feasible in Problem 1. Thus, A* I X " .
To show that A o I A * , suppose the optimal decision variables in Problem 1
are X " = (Xf, . . . ,Xi).Then for each value of i, either ( A X " ) i lbi or (AX"),.>bp
If ( A X 0 ) , s b i ,then

from (4) and d: = 0. Also,

from (1 2) and (5).

Similarly if (AX")j> b ,

and dlr =O. Also,

526 Decision Sciences [Vol. 12

We note that (13) is identical to (a), and (IS) is identical to (6). Now locate
the subproblem (sayj,) having constraints of the form (9) and (8) for all i for
which (AX")isbi and having constraints of ,the form (6) for all i for which
(AX0),>bi.The solution X=X", X = X " is feasible in subproblem j , , and thus
Xo5X*.
Since X * s X " and X o s X * , it follows that X * = X o . Q.E.D.
In general, A is sign-unrestricted in Problem 1 where X"<O if and only if the
set of constraints bi- Ai5(AX),5 bi+ Ai (i= 1, 2, . . ., K ) is infeasible. To facili-
tate the solution process, we can employ the nonnegativity restriction X r O , since
a feasible solution exists if and only if X r O .
In the problem presented in [3, pp. 329, 3321, the membership functions for
the three goals are

/ 0 if 80X, +40X2s620
80x1 + 40x2 - 620
if 620 5 8OX, + 40X2I630
10
p1 (profit goal) -
640- (80x1+ 40x9

I
\ o if
10
if 630 s 80Xl + 40X25 640

OX, + 4 0 ~ ~ r 6 4 0

OifXIs4

x,-4
if 4 s X l 5 6
2
p2 (sales goal for product 1) =

p3 (sales goal for product 2) =

if 4 s X 2 s 6

\ 0 if X 2 r 6 .
19811 Hannan 527

On the basis of the discussion above, this problem may be represented as

Max A (16)

s.t. 8x1 -+ 4X2+ d i - d [ = 63 (17)

A, all Xj, d,:, d; 2 0 j = 1,2; i = 1,2,3. (23)

The solution to this problem is identical to the solution provided in [3]:

namely, A = .96, Xl = 5.92, and X 2 = 3.92, yielding a profit of $630.40.

GP PROBLEMS WITH FUZZY PRIORITIES

The extension of the fuzzy GP problem to a problem containing fuzzy

priorities as well as fuzzy goals entails the introduction of linguistic variables such
as “important,” “not very important,” and “very important.” Narasimhan de-
fines p w,(p,(AX)) as the weighted contribution of the ith goal to the overall objec-
tive, where p represents the membership function corresponding to the fuzzy
priority associated with the ith goal.
The maximizing decision is given by [3, p. 3321

MaxpD(x) = Max Min pwWi(p,(AX)).

xro xr0 i

Thus, the weight attributed to a given goal is dependent upon the degree to
which the goal has been achieved.
The previous example is extended by Narasimhan after defining the follow-
ing membership functions for the fuzzy priorities [3, p. 333):
528 Decision Sciences [Vol. 12

( 0 otherwise

( 0 otherwise.
Again, eight subproblems are defined. Only one of the subproblems is found
to be feasible, and the solution given is X , = 5.6, X2= 4.55, X = 1. This solution is
incorrect and stems from some errors in defining the subproblems. Actually ( X i ,
Xd = (5.6, 4.55) only yields X = .625. The correct solution is (Xl, Xd = (5.6,
68/15) with X = -667. This is obtained by solving the problem:

Problem 2: Max X

X, X i , X2, d c , d: Z 0.

The Problem I formulation does not suffice for this extended problem be-
cause of the discontinuities in p w 2 and rw3.Each of these discontinuities gives
rise to two problems so four subproblems instead of eight need to be considered
after extending Problem 1. The only one of these four problems that has a feas-
ible solution is Problem 2.
The optimal profit for this extended problem, in which profit is supposed to
be more important than the other goals, is $629.33. In the previous problem,
where all goals were equally important, the profit was $630.40.This contradic-
tion exemplifies a basic problem with the definition of membership functions for
priorities as composite functions of the goal membership functions.
Conceptually it would be preferable if the weight attributed to each goal
were only dependent upon the importance of that goal relative to the other goals
and not upon the degree to which the goal can be achieved. The fuzziness should
apply to the decision maker’s imprecise estimation of the relative importance of
each goal.
The original formulation provided by Narasimhan, which gives rise to the
Problem 1 formulation, is predicated upon maximizing the minimum member-
ship value of the different goals rather than a weighted average of deviatioqs.
I98 I] Hunnan 529

If weights for the various goals are to be employed, a more natural way of intro-
ducing them would be to use the objective function

K
Min 2: wi(d,T + d:)
i= 1
K
in conjunction with (4) and (5) and the nonnegativity restrictions, where C wi
i= 1
= 1.
If this model is deemed to be more appropriate, the analyst must determine
values for the wi reflecting the decision maker’s fuzzy estimates. This can be ac-
complished by applying methods introduced by either Saaty [4] or Zeleny [ 5 ] .
Saaty’s method consists of eliciting pairwise comparisons from the decision
maker and then applying the theory of eigenvectors in order to obtain the set of
weights most consistent with the pairwise comparisons.
In the context of the problem considered earlier, suppose that the decision
maker feels that a deviation of one unit from the first goal (profit) is three times
more undesirable than a deviation of one unit from either the second goal (sales
for product 1) or the third goal (sales for product 2). Also, suppose that a devia-
tion of one unit from the second goal is twice as undesirable as a deviation of one
unit from the third goal. These elicited ratios can then be used to form the matrix

Here aij represents the number of times more important goal i is than goalj.
For instance, u12=3 because Goal 1 is three times as important as Goal 2. Also,
uji= 1 for all i and aij= l/uii. Thus, only n(n - 1)/2 comparisons need to be made
for a problem with n goals.
A is not consistent in this example (e.g., if Goal 1 is three times as important
as Goal 2 and Goal 2 is twice as important as Goal 3, Goal 1 should be 6 times as
important as Goal 3). Such inconsistencies, however, are typical of fuzzy valua-
tions provided by decision makers. The objective is to determine a set of weights
that is, in some sense, most consistent with the pairwise comparisons that have
been supplied.
530 Decision Sciences [Vol. 12

Consider the matrix equation A w = X w where

and A is an n x n matrix. Equivalently,

( A - x l ) w = 0.

If A is consistent, X has n values, n - 1 of which equal zero and the other,

Amax, equal to t i . If A is inconsistent, Amax>n,and the difference becomes more
pronounced as the degree of inconsistency increases. Saaty’s technique consists of
computing, , ,A and using it to compute the associated unit eigenvectors. These
serve as estimates of the relative weights wi.
For the above problem, w 1 = .59, w2= .25, and w 3 = .16. The new fuzzy goal
programming problem is

Min .59(d; + d [ ) + .25(dT + d i )+ .16(d; + d ; )

s.t. (17), (18), (19), and (23).

Zeleny’s method can be best applied when the information from the decision
maker is in linguistic form, such as “the weight is ‘very important,’ ‘not impor-
tant,’ etc.” Each of the linguistic variables describing a weight is given a discrete
membership function that describes membership values for a set of possible
values for the weight. A set of weights summing to one is then chosen such that
the minimum membership function is maximized.
For example, if there are three goals with associated weights w I , w2, w3, then
the membership grade for pwlApw2Apw3is

Thus,

Max ( p w l ~ r w 2 b w 3 )Max
= Min (pWl, pW2, p W 3 )

where w l + w2+ w 3 = 1.
As an example of this procedure, consider the extended problem given above
where the weight “very important’’ applies to p w and the weight “important”
applies to pw and pw3. Hypothetical membership kmctions are defined for these
weights as follows:
 1Y81] Hnnnan 53 1

~1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.0

PWl 0 0 .I .9 .7 .4 .2 0 0 0
~2 .I .2 .3 .4 .5 .6 .7 .8 .Y 1.0

cw2 .5 .9 .8 .2 0 0 0 0 0 0
~3 .I .2 .3 . 4 .5 .6 .7 .8 .9 1.0
gw3
S -9 .8 .2 0 0 0 0 0 0

These membership functions must be elicited from the decision maker with the
prior understanding that there are three goals and that the sum of the weights of
the goals is one.
Based upon the table above, the set of weights with the highest membership
function is ( w l , w2, w3)=(.4, .3, .3).
In this case, the new fuzzy goal programming problem is

Min .4(d; + d : ) + .3(dF + d i ) + .3(d; + d ; )

s.t. (17), ( I @ , (19), and (23).

SUMMARY

This pa er has demonstrated how the fuzzy goal programming problem con-
xp
sisting of 2 LP problems, each containing 3K constraints, may be reformulated
as a single LP problem with only 2K constraints.
With respect to the fuzzy G P problem with fuzzy priorities, it has been
shown that linking the membership functions for fuzzy weights to the degree to
which the associated goal is achieved can produce contradictory results. Two al-
ternative methods have been presented. Since the preference structures used to il-
lustrate these methods were hypothetical, it is meaningless to compare the results
of the sample problems to the Narasimhan solution. These methods can be con-
sidered superior to the method presented by Narasimhan in that when the weight
of a goal is increased, the resulting solution for that goal will always be at least as
good as the previous value for the goal. [Received: June 27. 1980. Accepted: Jan-
uary 6, 198 1 .]

REFERENCES

[I] Ignizio, J. P. God progrumming and extensions. Lexington, Mass.: Lexington. 1976.
I21 Lee, S. M. Cool progrumming for decision unalysis. Philadelphia, Penn.: Auerbach, 1972.
[3] Narasimhan, R. Goat programming in a fuzzy environment. Decision Scienws, 1980, 1 1 ,
325-336.
[4] Saaty, T. A scaling method for priorities in hierarchical structures. Journal oJ Ma//iema/ica/
PSyChdogy, 1977, /5(3),234-281.
[S] Zeleny, M. Compromise programming. In J . L. Cochrane& M . Zeleny (Fds.), Mulriple cri/eriu
decision moking. Columbia: University of South Carolina Press, 1973.