Professional Documents
Culture Documents
Yoon, K. A Reconciliation Among Discrete Compromise Solutions
Yoon, K. A Reconciliation Among Discrete Compromise Solutions
Yoon, K. A Reconciliation Among Discrete Compromise Solutions
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/2581948?seq=1&cid=pdf-reference#references_tab_contents
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.
Operational Research Society and Palgrave Macmillan Journals are collaborating with JSTOR to digitize, preserve and extend
access to The Journal of the Operational Research Society.
http://www.jstor.org
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
J. Opl Res. Soc. Vol. 38, No. 3, pp. 277-286, 1987 0160-5682/87 $3.00 + 0.00
Printed in Great Britain. All rights reserved Copyright ? 1987 Operational Research Society Ltd
The application of compromise solutions to discrete multi-objective problems brings about some technical
flexibilities, such as the selection of distance function for computing both normalized attribute ratings and
distances between two alternatives, and the choice between the ideal and negative-ideal alternatives for
implementing the axiom of choice. These flexibilities are undesirable, since the method may yield
conflicting preference-alternative rankings, depending on parameter choice. This paper introduces a
credibility measurement of distance function and takes a broader concept of the axiom of choice in order
to reconcile disagreement among compromise solutions.
Key words: axiom of choice, compromise solution, distance functions, multi-criteria decision-making,
preference ranking, value functions
Decision-makers (DM) are confronted with a situation in which they must choose from multiple
alternatives. Although it is very useful to use cost or profit as a measure of desirability, the DM
often discovers equal or more important goals, objectives or attributes. For example, the job one
chooses may depend upon its prestige, location, salary, advancement opportunities, working
conditions and so on. The car one buys may be characterized in terms of price, gas mileage, style,
safety, comfort and other considerations. Automobile manufacturers want to design a model which
maximizes fuel efficiency and riding comfort and minimizes production cost. Multiple objectives
usually conflict. In designing a car, for example, the goal of higher gas-mileage might degrade the
passenger-comfort rating because of resulting reduced passenger-space. A car is non-dominated if
there is no car which excels it in all attributes considered. Most cars on the market are in this class.
Hence a DM might experience non-trivial evaluation problems among the non-dominated
alternatives.
Multiple attribute decision-making (MADM) methods are management decision aids used in
evaluating m non-dominated alternatives which are characterized by n attributes. Let xi represent
the rating of alternative Ai with respect to attribute Xj. The elements xij can be integrated into an
(m x n) decision matrix whose rows indicate job performance of alternative Ai:
Our human cognitive power is limited such that most MADM methods have tried to reduce
attribute dimensions as much as possible. One popular approach is to transform n-dimensional
vector performance into a scalar performance. That is,
Vi= V(Ai) = V(xi, w) i = 1, ..., m, (2)
where V is the value function, and w = (wl, . . . , w, wJ) is the weight assigned to each attribute
which is assessed by a DM with the help of a decision analyst. Once a DM obtains a valid value
function (not an easy job), the DM should give the highest preference to an alternative whose value
is the maximum.
Technology has taught that it is impossible to manufacture the ideal car, that is an automobile
which possesses the best level or rating in all attributes considered. To be as close as possible to
such an ideal car is the rationale of human choice. Based on this axiom of choice,' Zeleny2'3
introduced the concept of compromise solutions for the resolution of conflicting objective-
functions. A compromise set contains alternatives which are closest to the ideal, as determined by
some measure of distance. Since compromise solutions form a greatly reduced non-dominated set,
they are very useful in multi-objective programming.4'5The principle of compromise solution is
equally applicable to MADM (or discrete multi-objective) problems.6'7In the MADM setting, the
277
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
Journal of the Operational Research Society Vol. 38, No. 3
One physical property of dpmeasurement is well known: when p increases, distance dpdecreases,
i.e. d > dp> ,59 and greater emphasis is given to the largest deviation in forming the total.
Ultimately, for p oo, the largest deviation completely dominates the distance determination, i.e.
dc = max { x) - x]
Distances with p = 1, 2 and so are especially operationally important: distances d, (the Manhattan
distance) and d2 (the Euclidean distance) are the longest and shortest distances in the geometrical
sense; distance do (the Tchebycheff distance) is the shortest distance in the numerical sense.
If we define the normalized rating as the ratio between individual and combined distance from
the origin 0 = (0, 0, . .. , 0), then the comparable rating of x1iis given as
M
m ) I/p
= Xj/l { XiiP} p = 1, 2
The above normalization equation has the property that when p increases, the effect of interference
among alternatives decreases. Finally, when p = oo, each attribute is compared by only its largest
attribute rating.
For non-monotonic attributes (e.g. blood pressure, and the sugar content in a cup of coffee,
where the most favourable value is located somewhere in the middle of an attribute range), they
become monotonic by taking the statistical z-score, i.e. (x,1- xJ)/5j, where x; is the most favourable
value, and aj is the standard deviation of alternative ratings with respect to attribute x;, respectively.
Then the comparable rating ri1(p) is obtained by using equation (4).
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
K. Yoon A Reconciliation Among Discrete Compromise Solutions
perceived ideal as one which has all the best attribute ratings attainable. The distance between an
alternative Ai and the perceived ideal alternative A * is given as
n II/p
where r * is the most favourable normalized rating for attribute x;. Then a set of discrete
compromise solutions is
C*= min{d*(i),i= 1,2,...,m}, p= 1,2,oo. (6)
The composite of the best attribute ratings attainable is the (positive) ideal alternative,
whereas the negative-ideal alternative is defined as that which is comprised of the worst
ratings only. The negative-ideal alternative is imagined to be opposite the ideal alternative.
Can we then use the anti-ideal as another reference point in the decision-making process? Yes.
Actually, we may have used the negative-ideal more often than the ideal since it is not easy to
set the proper ideal-level. For instance, we may try to make a new car (sub-ideal) which has
further distance from the existing car (negative-ideal) by increasing gas mileage and decreasing
the production cost, and so on. The counterpositive statements of the axiom of choice, alternatives
that are further from the negative-ideal are preferred to those that are closer, logically supports the
use of the negative-ideal. The distance between an alternative Ai and the anti-ideal alternative A -
is given as
( it I/P
d;(i)={ w'PIr1- rijP} , i= 1,2,..., m; p = 1,2, oo, (7)
where r is the least favourable normalized rating for x;. Then another set of discrete compromise
solutions based on the anti-ideal is
Cp =max{d; (i),i = 1,2,. ..,m}, p = 1,2, oo. (8)
Sometimes the compromise solution based on the ideal is not identical to that which is based
on the anti-ideal. For example, when we use d2 distance in evaluating the two alternatives Al
and A2 in Figure 1, AI is the choice based on A *; A2 is based on A -. This conflict can be resolved
IAn d
279
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
Journal of the Operational Research Society Vol. 38, No. 3
by considering dp*and d- simultaneously. Hwang and Yoon's TOPSIS (Technique for Order
Preference by Similarity to Ideal Solution)7',0defines the similarity to the ideal of alternative Ai as
It is clear that 0 < Tp< 1, where Tp= 1, if Ai= A*; Tp= 0 if Ai=A-.
The family of TOPSIS solutions may be difficult to illustrate graphically because they involve the
ratio of two distances, but the choice behaviour through such solutions can be explained by way
of indifference (value) curves.11'12A DM is assumed to give equal preference or value to any alter-
natives located on the same indifference curve. TOPSIS has the value function of d - /(d * + d -) = c,
where c is a given value. The value function can be rewritten as cd* - (1 - c)d- = 0. When the
d2distance measurement is employed, this expression indicates a variation of hyperbola where the
difference of two weighted [c and (1 - c)] distances from two fixed points (i.e. ideal and
negative-ideal points) is zero. Figure 2 shows some typical indifference curves from the TOPSIS.
Any curves with c > 0.5 are convex to the preference origin (i.e. ideal point), which indicates the
property of the diminishing marginal rate of substitution observed in most indifference curves;
whereas indifference curves with c < 0.5 are concave to the preference origin. This is an unusual
case, but it may be interpreted as a risk-prone attitude resulting from a pessimistic situation: when
a DM recognizes one's solution is closer to the negative-ideal than to the ideal, the DM is inclined
to take an alternative which consists of the best and worst attributes rather than one with two worse
attributes. For example, we might want to get one A grade and one F grade rather than two D
grades. This set of indifferencecurves also supports Zeleny's conjecture of the discriminatory power
of reference points: as alternatives deviate farther from a reference point (say, A *), the
discriminatory power of the reference point diminishes, and finally another reference (say, A -) is
needed.3 Shapes of indifference curves with c > 0.5 (convex to A *) show that they are using A *
as a reference. Also, indifference curves with c < 0.5 (convex to A -) reveal that they are under the
control of A -. When we use d, distance, the indifference curves are a set of straight lines with the
slope of -1. Figure 3 shows a set of indifference curves with dc, distance.
r2 1
A~~~~~~~~~~~~~~~A
280
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
K. Yoon A Reconciliation Among Discrete Compromise Solutions
.9
A C=.1
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
Journal of the Operational Research Society Vol. 38, No. 3
X2
parameter p increases. Then the credibility of a distance function dp decreases when parameter p
increases.
The credibility of a distance function is shown to be inversely related to the length of indifference
(distance) curve or the area of indifference hypersurface when there are more than two dimensions
in forming the distance. When there are no restrictions on the range of dimension, the ratio of class
sizes among distance functions is independent of the amount of distance. For instance, we compute
the class-size ratio of each distance by setting dp= c in Figure 4: 1= (li, 12, loo) = (4 2c;, 2itc, 8c)
= (2.83, 3.14, 4). But attribute ratings are populated within the finite range. Especially, normalized
attribute-ratings are located in the range of [0, 1]. Suppose we impose such constraints on the
dimension range as 0 < r1< 1 and 0 < r2< 1 in Figure 5. The class-size ratio is (2.83, 3.14, 4) with
1 .5 =_ ____________________ _ 1
_Feasible region \
FIG. 5. Indifference curves with d = 0.6 and dp=1.5 under range constraints.
282
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
K. Yoon-A Reconciliation Among Discrete Compromise Solutions
c = 0.60; it is (0.71, 0, 0) with c = 1.5. The class-size ratio under the dimension constraint renders
a local ratio which is dependent on the amount of c.
We may consider two criteria which have the global (unchangeable) ratio and play the role
of the combined individual class-sizes. One is the range of dp, which is the sum of various class-
sizes. The mean of dp, E(dp) becomes another criterion because the distance dp is expressed as
the larger amount with the smaller class-size. We take E(dp) to quantify the credibility of dp
because E(dp) is more sensitive to the number of dimensions than to the range. Since the credibility
of dp is proportional to E(dp), the relative--credibilityof each dp under the three-distance system
is written as
A NUMERICAL EXAMPLE
A customer wants to purchase a family car and is faced with the problem of choosing between
eight automobiles. The ratings based on 16 attributes are given in Table 2. The data is adopted
from Fleischer's text.14
Table 3 shows three sets of preference rankings based on d2 measurement in computing both
attribute-rating normalization and distance between two alternatives. The ranking based on d*,
R* differs from that which is based on d7-, R -. This disagreement can be resolved taking the D2
value, which becomes a TOPSIS ranking. It is interesting that the discriminatory power of the
283
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
Journal of the Operational Research Society Vol. 38, No. 3
s: O
=~~~~~~~~~~v
O'-O> v0-OOm r-
C>C
>t8$4O<^
C--00C
U
.~0 It
~0.1t0 -\0
, 0W o o
o00
CI
? m C$ > <
r- C> C>I'
a
U _
0
0
0 t r OO\ -
0
- ) C
$0i0$ C - t m >0 r
s 0.
0 U_
0 > _ =
C'
0 U 0U
00 000\0m000m-t 0 Uc
.,30
m O-
0
0 C
'IO 00 oot -N " Cq CI CI
06 $ ~ ~ ~ r-n ~a4oN ~oo-
00
~ o
~ * .
Q
CDI CD CDC0CDD CD CD , CD C>
D$$<
C'
U)- rlw0 C V0 r0 0 06 06 - Cl 00000t6
CdC'wis E
Cd
_0 C0 Nm st ms a o d 0 0 m v-C
0
284:4 C'S
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
K. Yoon A Reconciliation Among Discrete Compromise Solutions
reference point is revealed by those rankings: TOPSIS ranking is identical to R2 if D2 > 0.5, and
to R- if D2<0.5.
Instead of making eight other preference rankings, we try to obtain the unified TOPSIS solution
by way of the credibility of distance function. First, the unified normalized ratings are obtained
using
TABLE 4. The final TOPSIS solution with the unified distance measurement
d* d- D
Automobiles Value Ranking Value Ranking Value Ranking
A1 0.0429 2 0.1226 2 0.7407 2
A2 0.0606 3 0.1045 3 0.6327 3
A3 0.0657 5 0.1007 5 0.6051 5
A4 0.0947 6 0.0700 6 0.4252 6
A5 0.1224 7 0.0430 7 0.2599 7
A6 0.0371 1 0.1257 1 0.7722 1
A7 0.0635 4 0.1016 4 0.6155 4
Ax 0. 1328 8 0.0289 8 0. 1787 8
285
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions
Journal of the Operational Research Society Vol. 38, No. 3
interdependency between attributes which form an alternative. Again, the mutual effect among
attributes decreases as the parameter p increases. Particularly with p = 1, where there exists the
maximum dependency between attributes, the distance between A * and A - is always constant.
That is,
d* (i) + d (i) = constant, i = 1, 2,. . , m. (16)
Therefore, the minimization of d* is equivalent to the maximization of d-. Assuming that
A- = (0, 0, ... , 0) without losing generality, then the distance from the negative-ideal can be
written as
n
REFERENCES
1. C. H. COOMBS(1958) On the use of inconsistency of preferences in psychological measurement. J. exp. Psychol. 55,
1-7.
2. M. ZELENY(1973) Compromise programming. In Multiple Criteria Decision Making (J. L. COCHRANEand M. ZELENY,
Eds), pp. 262-301. University of South Carolina Press, Columbia, S.C.
3. M. ZELENY(1982) Multiple Criteria Decision Making. McGraw-Hill, New York.
4. C. L. HWANG and A. S. M. MASUD in collaboration with S. R. PAIDY and K. YOON (1979) Multiple ObjectiveDecision
Making. Lecture Notes in Economics and Mathematical Systems, Vol. 164. Springer, Berlin.
5. J. L. COHON(1978) Multiobjective Programming and Planning. Academic Press, New York.
6. K. R. MACCRIMMON(1968) Decision Making Among Multiattribute Alternatives:A Survey and ConsolidatedApproach.
RAND Memorandum, RM-4823-ARPA, The Rand Corporation, Santa Monica, Calif.
7. C. L. HWANG and K. YOON(1981) Multiple Attribute Decision Making. Lecture Notes in Economics and Mathematical
Systems, Vol. 186. Springer, Berlin.
8. P. NIJKAMPand A. VAN DELFT (1977) Multi-Criteria Analysis and Regional Decision-Making. Martinus Nijhoff Social
Sciences Division, Leiden, The Netherlands.
9. E. BECKENBACHand R. BELLMAN(1961) An Introduction to Inequalities. Random House, New York.
10. K. YOON and C. L. HWANG (1985) Manufacturing plant location analysis by multiple attribute decision making: Part
I-single plant strategy. Int. J. Prod. Res. 23, 345-359.
11. R. L. KEENEYand H. RAIFFA (1976) Decisions with Multiple Objectives. Wiley, New York.
12. K. R. MACCRIMMONand M. TODA (1969) The experimental determination of indifference curves. Rev. Econ. Stud. 36,
433-450.
13. V. CHANKONG,Y. Y. HAImEs and J. THADATHI (1985) Multiple criteria optimization: a state of the art survey. In
Decision Making with Multiple Objectives (Y. Y. HAImEs and V. CHANKONG, Eds), pp. 36-90. Lecture Notes in
Economics and Mathematical Systems, Vol. 242. Springer, Berlin.
14. G. A. FLEISCHER(1984) Engineering Economy. PWS Engineering, Boston, Mass.
15. P. C. FISHBURN(1965) Independence in utility theory with whole product sets. Opns Res. 13, 28-45.
16. W. EDWARDS(1977) Use of multiattribute utility measurement for social decision making. In Conflicting Objectives in
Decisions (D. E. BELL et al., Eds), pp. 247-276. Wiley, New York.
17. L. DUCKSTEINand S. OPRICOVIC (1980) Multiobjective optimization in river basin development. Wat. Resour. Res. 16,
14-20.
18. P. L. Yu (1973) A class of solutions for group decision problems. Mgmt Sci. 19, 936-946.
286
This content downloaded from 198.91.37.2 on Sun, 24 Jan 2016 15:11:32 UTC
All use subject to JSTOR Terms and Conditions