a comparison of two reference point methods in multiple objective mathematical programming
TRANSCRIPT
Continuous Optimization
A comparison of two reference point methods in multipleobjective mathematical programming
John Buchanan a,*, Lorraine Gardiner b
a Department of Management Systems, University of Waikato, Private bag 3105, Hamilton, New Zealandb Department of Accounting and Management Information Systems, College of Business, California State University,
Chico, CA 95929, USA
Received 8 March 2002; accepted 20 May 2002
Abstract
When making decisions with multiple criteria, a decision maker often thinks in terms of an aspiration point or levels
of achievement for the criteria. In multiple objective mathematical programming, solution methods based on aspiration
points can generate nondominated solutions using a variety of scalarizing functions. These reference point solution
methods commonly use a scalarizing function that reaches down from the ideal solution, in a direction specified by the
aspiration point. Conversely, a similar scalarizing function can push out from the nadir point toward a specified as-
piration point. These scalarizing functions are similar in structure but diametrically opposed in their reference points. In
this paper we examine how these approaches behave from a technical point of view and conduct an experiment to
understand better the human behaviour of users of these approaches. Insights we gain on the evenness of dispersion are
relevant when attempting to construct a representation of the nondominated set. Further, the technical characteristics
of the two formulations� solutions, combined with behavioural tendencies, allow us to comment on the implications fortheir use in interactive multiple objective methods.
� 2002 Elsevier Science B.V. All rights reserved.
Keywords: Multiple objective programming; Decision analysis
1. Introduction
Reference point methodology provides the
foundation for many interactive search procedures
in multiple objective programming (for exam-
ple, Wierzbicki, 1980; Steuer and Choo, 1983;
Korhonen and Wallenius, 1988; Steuer et al.,
1993). The concept of a reference point is consistentwith Simon�s (1957) description of satisficing deci-sion making where a decision maker tends to have
targets or goals in mind while proceeding toward a
decision. This satisficing approach is specifically
accommodated in the reference point methods of
Wierzbicki (1980) and Korhonen and Wallenius
(1988). However, reference point formulations have
also been used in ‘‘nonsatisficing’’ models such asSteuer and Choo (1983), where a reference point
formulation is used as a mechanism for implicitly
*Corresponding author. Tel.: +64-7-838-4470; fax: +64-7-
838-4270.
E-mail addresses: [email protected] (J. Buchanan),
[email protected] (L. Gardiner).
0377-2217/03/$ - see front matter � 2002 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0377-2217(02)00487-3
European Journal of Operational Research 149 (2003) 17–34
www.elsevier.com/locate/dsw
optimizing a decision maker�s value function.Reference point methods may ask a decision
maker to specify either aspiration points (desirable
levels of achievement), reservation points (levels of
achievement that should be attained, if at all pos-
sible) or worst outcome points (levels of achieve-
ment to be avoided). In this study, we focus our
attention on aspiration points. We acknowledge,
however, that procedures which use reservationpoints (e.g., Reeves andMacLeod, 1999) and worst
outcome points (e.g., Michalowski and Szapiro,
1992) represent important classes of reference point
methods.
The use of reference points appeared in the
early development of multiple objective program-
ming as part of the work by Charnes and Cooper
(1961) on goal programming. Wierzbicki (1980)produced seminal research on reference point
methods, including an investigation of the char-
acteristics of various achievement functions for
allowing the search for attractive nondominated
solutions to be controlled by reference points.
These achievement functions were designed to have
a significant advantage over goal programming by
producing only nondominated, or Pareto-optimal,solutions. An important class of achievement
functions considered in Wierzbicki�s paper hasformulations similar to the weighted Tchebycheff
distance metric.
In addition to their desirable structural features,
reference point methods have also appeared useful
from a methodological or operational perspective.
They provide the ‘‘shell’’ for simple, decisionmaker controlled solution methods that do not
require the decision maker to conform to any
particular requirements or axioms. The Pareto
Race method of Korhonen and Wallenius (1988),
where a decision maker simply moves over the
efficient surface in a relatively unstructured search
for a most preferred solution, is one such example.
Buchanan (1997) has proposed an even sim-pler solution method, also using a Tchebycheff
achievement function. This ‘‘GUESS’’ method
comprises successive guesses by a decision maker
and has been successfully used in experiments by
Buchanan and Daellenbach (1987), Corner and
Buchanan (1997) and Downing and Ringuest
(1998).
In this paper we compare two formulations,both using Tchebycheff-based achievement func-
tions and aspiration points, but having quite dif-
ferent philosophies. One seeks to ‘‘push out’’ from
an undesirable solution while the other ‘‘reaches
down’’ from a desirable but unachievable solution.
It is known that these two formulations produce
different solutions, given identical input (that is,
the same aspiration point) from the same decisionmaker (Hwang et al., 1993; Martinson, 1993).
These two formulations have been incorporated in
the TOPSIS solution method of Lai et al. (1994).
TOPSIS endeavors to find a compromise between
the positive ideal solution (reaching down) and the
negative ideal solution (pushing out).
From a behavioral perspective, the choice of
reference point can considerably affect a decisionmaker�s perception of the resulting solution.
Proverbs 27:7 states the point succinctly, ‘‘He who
is full loathes honey, but to the hungry even what is
bitter tastes sweet’’. Thus, if the reference point is
the nadir or worst solution, then any resulting
solution is a significant improvement. Conversely,
if the reference point is the ideal or best solution,
any resulting solution must be worse. Even if twosolutions are the same, any comparison with these
two reference points is likely to result in different
‘‘values’’ being placed on them. The two opposing
reference points for the formulations (pushing out
from an undesirable solution or reaching down
from a desirable solution) is similar to the concept
of framing, which has been discussed in the be-
havioral decision making literature. Image theory,a descriptive decision making model proposed by
Beach and Mitchell (1990), provides a useful il-
lustration of framing. A decision maker has im-
ages or schemata which represent, in some way, his
view or values. Within image theory, the first stage
of any decision is a compatibility test where each
alternative in turn is assessed to determine its ‘‘fit’’
with a decision maker�s images. Usually these im-ages will be, in some sense, ideal. More specifically,
Tversky and Kahneman (1981) show that how a
problem is framed (in their example, whether the
same problem is stated negatively or positively)
significantly affects decision maker choice. The
pushing-out and reaching-down formulations can
be viewed as different ways of framing the decision
18 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
problem, for the same aspiration point. Lai et al.(1994) illustratively allegorize these two reference
points as heaven and hell. To examine the behav-
ioral effects of this framing, we also present the
results of an experiment on how the same human
decision makers respond to solutions from these
two formulations, given the same aspiration point.
The paper is, firstly and principally, a sensitivity
analysis of the different solutions generated bythe two formulations, controlling for aspiration
points. Secondly, the paper provides recommen-
dations for decision support, particularly in the
context of interactive multiple criteria decision
methods. In this respect, the technical sensitivity
analysis provides a clear understanding of how
solution methods can be expected to behave (an
issue of controllability). Then the experiment addsto our understanding of how decision makers can
be expected to behave. A desired outcome is an
improved alignment of solution methods to deci-
sion makers. The paper is organized as follows. In
Section 2 we introduce necessary terminology and
present the two achievement function formula-
tions. Section 3 considers the discrepancy between
solutions generated by these two formulationsusing a q-dimensional hyperspherical approxima-
tion. Section 4 provides two illustrative examples
and Section 5 describes and summarizes results of
the behavioral experiment. A discussion of the
results and their implications for decision support
is presented in Section 6.
2. Terminology and definitions
The multiple objective mathematical program-
ming problem can be stated as:
MAX F ðxÞ ¼ ½f1ðxÞ; f2ðxÞ; . . . ; fqðxÞ�s:t: x 2 X
ð1Þ
where x is an n-dimensional vector of decision
variables; F is a vector-valued function comprising
q distinct objective functions of x; X � Rn is
the feasible set of constrained decisions; and Z ¼F ðxÞ � Rq is the image of X in objective function
(criterion) space. We further assume that set X
comprises m constraints of the form, X ¼ fx 2 Rn :gjðxÞ6 0; j ¼ 1; 2; . . . ;mg. The operator MAX
indicates that each objective function is to be
maximized over X.
The set of nondominated solutions N consists
of those criterion vectors z 2 Z which are not
dominated by another criterion vector in Z. A
criterion z1 is said to dominate solution z2 if
z2k 6 z1k for all k ¼ 1; 2; . . . ; qz2j < z1j for at least one j 2 f1; 2; . . . ; qg
ð2Þ
The set of weakly nondominated solutions consistsof those criterion vectors z 2 Z for which there isno other z0 2 Z such that z0 > z.In the remainder of the paper, two assumptions
are made about the preferences of the decision
maker. First, preferences are assumed to mono-
tonic; that is, given the MAX operator for each
objective, ‘‘more is always better than less.’’ Sec-
ond, a decision maker is presumed to want toconsider only nondominated solutions.
We define the maximum and minimum values
of each objective over the nondominated set N
as:
Maximum ¼ Uk ¼ maxz2Z
zk; k ¼ 1; 2; . . . ; q
Minimum ¼ Mk ¼ minz2N
zk; k ¼ 1; 2; . . . ; qð3Þ
The vector of Uk values, U, is usually referred to asthe ideal point while the vector of Mk values, M, isoften labelled the nadir point.
The Uk and Mk values are often used to rescale
the criterion vectors in order to provide commen-
surable measurement of all objectives over N(Steuer, 1986). The transformation is internal to
the procedure and in most solution methods the
decision maker is not asked to think in terms of
the transformed units. The function hðzÞ mapsany criterion vector z 2 N onto the range [0,1].
Using this transformation, hðMÞ ¼ ð0; 0; . . . ; 0Þand hðUÞ ¼ ð1; 1; . . . ; 1Þ and
hðzkÞ ¼zk �Mk
Uk �Mkð4Þ
The criterion space is therefore scaled and N is
contained within a q-dimensional unit hypercube
anchored at the origin of the nonnegative orthant.
The remaining discussion in this paper assumes the
transformation in (4).
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 19
2.1. PU and PM formulations
We now define the two different reference point
formulations which are compared in this paper.
They are hereafter referred to as Programs PU and
PM. PU uses a Tchebycheff-based achievement
function to minimize the maximum weighted de-
viation from the ideal solution U. Program PM, onthe other hand, seeks to maximize the minimum
weighted achievement from the nadir solution M.These two formulations use a simple form of the
Tchebycheff-based achievement function as op-
posed to an augmented form (Steuer, 1986). The
augmented form ensures that weakly nondomi-
nated solutions are not generated; however, for
our comparisons the simple form will suffice.For a given set of weights w, the PU and PM
formulations are:
PUðwÞ : min y
s:t: y Pwkð1� hðzkÞÞ; wk > 0;
k ¼ 1; 2; . . . ; qz 2 Z
ð5Þ
PMðwÞ : max y
s:t: y6wkhðzkÞ; wk > 0;
k ¼ 1; 2; . . . ; qz 2 Z
ð6Þ
The maxmin formulation (PM) was initially de-
veloped by Kaplan (1974) and subsequently dis-
cussed by Gupta and Arora (1977) and Posner andWu (1981). Corner and Buchanan (1995) used PM
in experimental work and the GUESS method
proposed by Buchanan (1997) is also based on the
PM formulation. The minmax (PU) formulation is
more commonly reported in the literature; varia-
tions of this formulation have been used in the
work of Benayoun et al. (1971), Nakayama and
Sawaragi (1984), and Steuer and Choo (1983), toname a few.
The PM formulation has M as its reference
point while for PU the reference point is U. Wedistinguish between a reference point such as either
M or U, which is an integral part of the problemformulation, and an aspiration point, which is an
input specified by the decision maker. The aspira-
tion point represents a target or goal and can beequivalently represented as a set of weights over the
criteria, contingent on the formulation being used.
Fig. 1 shows, in two dimensions, the relation-
ships among these different terms. The set of
nondominated solutions is represented by a hy-
persphere. hða1Þ and hða2Þ are aspiration points,transformed from the original a1 and a2. The linesegment between ð0; 1Þ and ð1; 0Þ is the set of pointswhere h1 þ h2 ¼ c ¼ 1, and other values for c arealso shown. h M1 is the PM solution for hða1Þ whileh U1 is the PU solution for hða1Þ; similarly for hða2Þ.From Fig. 1, it can be seen that the discrepancy
between PM and PU solutions is a function of the
shape of the nondominated set, the location of the
aspiration point relative to the nondominated set,
and the distance of the aspiration point from theline segment L ¼ hðMÞhðUÞ. We take the closenessof the aspiration point to L as a measure of the‘‘centrality’’ of the aspiration point. Even in the
two-dimensional case, it is obvious that the dis-
crepancy between h M and h U can be substantial.Fig. 1 shows PM as ‘‘pushing out’’ from hðMÞ, andPU as ‘‘reaching down’’ from hðUÞ.The relationship between aspiration points and
weights is as follows. For a given aspiration point
a ¼ ða1; a2; . . . ; aqÞ, the weights w ¼ ðw1;w2; . . . ;wqÞare:
Fig. 1. Comparing PM and PU in two dimensions.
20 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
For PU : wk ¼1
1� hðakÞ; k ¼ 1; 2; . . . ; q
For PM : wk ¼1
hðakÞ; k ¼ 1; 2; . . . ; q
ð7Þ
These weights can be normalized so that they sum
to unity although it is not necessary to do so. In
this paper, however, our investigation is under-taken using aspiration points, not weights. The use
of aspiration points rather than weights better re-
flects the intent of most reference point methodo-
logies.
Using aspiration points, then, the intent of PU
and PM formulations can be stated. PU seeks to
minimize the maximum weighted deviation, but so
that the proportional deviations are equal; that is:
1� h U11� hða1Þ
¼ 1� h U21� hða2Þ
¼ � � � ¼1� h Uq1� hðaqÞ
;
ð1� hðakÞÞ > 0; k ¼ 1; 2; . . . ; q ð8Þ
PM seeks to maximize the minimum weighted
achievement so that, if possible, proportional
achievements are equal; that is:
h M1hða1Þ
¼ h M2hða2Þ
¼ � � � ¼h MqhðaqÞ
; hðakÞ > 0;
k ¼ 1; 2; . . . ; q ð9Þ
There is, of course, no guarantee that the equalities
of (8) and (9) will hold. Rather, these equations
reflect the intent of the different formulations. In
almost every case, for a given aspiration point,
a ¼ ða1; a2; . . . ; aqÞ, PU and PM result in different
solutions. Further analysis of the discrepancy be-
tween PM and PU solutions for the two-dimen-sional polyhedral case is provided in Buchanan
and Gardiner (2001).
2.2. Aspiration point location
We assume that all aspiration points have been
scaled according to (4) and are contained within
the unit hypercube. The location of aspiration
point hðaÞ can be described relative to N, the set ofnondominated solutions. First, hðaÞ is either on Nor not. If not, hðaÞ belongs to one of three mutu-ally exclusive categories: under N, over N or neither
as shown in Fig. 2 and defined in Appendix A. Thebold portion of the boundary represents N. The set
of aspiration points that are neither under nor over
N is contained in the interior of the rectangle la-
belled ‘‘Neither’’. The aspiration points over and
under N are shown relative to this rectangle and N.
Over, under and neither can be related to for-
mulations (5) and (6) of PU and PM respectively
in terms of the optimal solution y , as shown inTable 1.
Fig. 2. Location of aspiration points.
Table 1
Relating y to aspiration point location
PU PM
y < 1 All h Uk > hðakÞ; hðaÞ is under hðaÞ is not on or under; hðaÞ could be over or neithery ¼ 1 All h Uk P hðakÞ with at least one h Uk ¼ hðakÞ; either
h U ¼ hðaÞ or h U dominates hðaÞ; hðaÞ is either on orunder
All h Mk P hðakÞ with at least one h Mk ¼ hðakÞ; eitherh M ¼ hðaÞ or h M dominates hðaÞ; hðaÞ is either on or under
y > 1 hðaÞ is not on or under; hðaÞ could be over or neither All h Mk > hðakÞ; hðaÞ is under
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 21
For either formulation (PU or PM), if solutionh is a proportional solution as defined in (8) and(9), then the aspiration vector hðaÞ will be eitheron, over or under N. When h U and h Mð6¼ hðaÞ)are both proportional and hðaÞ is over N, h M will
be closer to L than h U . The reverse holds whenhðaÞ is under N. A proof is given in Appendix A.
3. Analysis of the hyperspherical case
We apply our analysis of discrepancy to the
case where the set of nondominated solutions, N,
is described by the boundary of the nonnegative
section of a unit hypersphere centered at the ori-
gin. In two dimensions, this is the intersection of
the boundary of a unit circle with the nonnegativequadrant, where the equation for N is h21 þ h22 ¼ 1.This simplification of representing N as the
boundary of the nonnegative portion of a q-
dimensional spheroid aids the ensuing analysis
without any loss of generality, especially in higher
dimensions. A hyperspherical approximation is
one (symmetric) limit of a general polyhedral set,
as the number of facets tends to infinity. Conse-quently, the analysis of differences between the PM
and PU formulations is less affected by the ‘‘noise’’
of the nondominated surface. In addition to being
smooth and symmetric, the hyperspherical ap-
proximation shares with the unit hypercube the q
extreme points which individually maximize each
objective function over the unit hypersphere. For
example in 3-dimensions, these extreme points are(1,0,0), (0,1,0) and (0,0,1). Also, since the dis-
crepancy associated with the hyperspherical ap-
proximation tends to be less than in the linear case,
it provides a conservative estimate of discrepancy.
We define dh as the Euclidean distance betweenthe PM and PU solutions for any given hðaÞ anduse it as the measure of discrepancy. In order to
explore how the amount of discrepancy dependson the location of hðaÞ, we examine how close hðaÞis to the line segment L ¼ hðMÞhðUÞ and charac-terize the hyperplane containing hðaÞ. We definedaL as the Euclidean distance from hðaÞ to theclosest point on L and let c ¼ h1 þ h2 þ � � � þ hq.Fig. 1 illustrates these measures in two dimensions.
Note that for c6 1, all corresponding aspiration
points are on or under N. As c approaches 0,the aspiration points approach hðMÞ. For cP2=
ffiffiffi2
pffi 1:414, all corresponding aspiration points
are on or over N. As c approaches 2, the aspiration
points approach hðUÞ.
3.1. Discrepancies in two dimensions
Figs. 3 and 4 show, for two dimensions, thediscrepancy between PU and PM solutions
for different aspiration points, characterized by
their c value, their distance from the line segment
L ¼ hðMÞhðUÞ, and whether the aspiration pointis over or under N.
Fig. 3 shows discrepancy curves for selected
values of c that correspond to aspiration points,
hðaÞ, over N. As c increases, the aspiration points
Fig. 3. Discrepancy between PU and PM solutions for aspi-
ration points on the line h1 þ h2 ¼ c, when the aspiration pointis over N.
Fig. 4. Discrepancy between PU and PM solutions for aspi-
ration points on the line h1 þ h2 ¼ c, when the aspiration pointis under N.
22 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
are closer to the ideal point and the amount ofdiscrepancy increases sharply with daL. For hðaÞunder N, as c increases and a is closer to thenondominated set, the discrepancy decreases. Note
that for aspiration points over N the discrepancies
are convex, while they are concave for aspiration
points under N. Choosing aspiration points very
near to the ideal point U will result in considerablediscrepancy, as also will aspiration points whichare near M.Let us put the discrepancy measure in context.
The maximum theoretically possible discrepancy
in two dimensions, with a hyperspherical N, is
0.7654. In this (unlikely) situation, for a given as-
piration point, PU would give solution (1,0) while
PM would give solution ð1=ffiffiffi2
p; 1=
ffiffiffi2
pÞ, or vice
versa. Consider a ‘‘middle of the road’’ discrep-ancy of 0.2. What does this mean? For the case of
c ¼ 1; that is, for all aspiration points on the hy-perplane h1 þ h2 ¼ 1, the median discrepancy is0.21, with a mean of 0.19. A discrepancy of 0.2 is,
therefore, quite realistic. An example of a 0.2 dis-
crepancy is shown in Table 2.
Thus, the solution is such that h U1 is 42% larger
than h M1 while h U2 is 89% of h M2 . Since the ob-jective value has been normalized over the range
[0,1], we have that h U1 is 17.3% further along the
range (i.e., 17.3% better achievement) than h M1while h M2 is 9.9% further along the range than
h U2 Clearly, very different solutions result from the
same aspiration point––depending entirely on
the framing; that is, whether it is deviation from
the ideal or achievement from the nadir.
3.2. Discrepancies in q dimensions
We now extend this analysis and demonstrate
that the results described visually in two dimen-
sions are no less applicable in higher dimensions.
In Fig. 5, we present the discrepancy for the case
where the aspiration points are on the hyperplaneh1 þ h2 þ � � � þ hq ¼ 1, and the number of dimen-
sions ranges from 2 to 10 million. Note that all
aspiration points on this hyperplane are either
under or on N and all resulting PU or PM solu-tions are proportional.
The discrepancy ðdhÞ is a function of the dis-tance of the aspiration point from L ðdaLÞ and q, asfollows
dh ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2�
2qd2aL þ 2 1� d2aL � 1q
� � ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiq� 2þ d2aL þ 1
q � qd2aLq
q� 2þ d2aL þ 1q
� � ffiffiffiffiffiffiffiffiffiffiffiffiffiffid2aL þ 1
q
qvuuut
The discrepancy increases as q increases, but not
linearly, with limq!1 dh ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2� 2daL
p.
The maximal distance from L for q objectives isffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1� ð1=qÞ
pat which dh ¼ 0. It is interesting to
note that as q increases, the maximal discrepancy
shifts toward daL ¼ 0.
3.3. Volume of aspiration points under N
As the number of dimensions q increases, the
likelihood of a given aspiration point being above
N increases significantly. Let Vq be the volume of aq-dimensional hypersphere with radius r (see, e.g.,
Kendall (1961, p. 35)); Weisstein (1999, pp. 876–
878) and Vq-cube the volume of a q-dimensionalhypercube with side r. Since we are concerned onlywith the portion of Vq located in the nonnegativeorthant, the volume of interest is Vq=2q. The ratioVq=ð2qVq-cubeÞ describes the proportion of aspira-tion points under N, for any value of q. Table 3
Table 2
Discrepancy example where dh ¼ 0:2hða1Þ hða2Þ h U1 h U2 h M1 h M2
0.310 0.690 0.583 0.813 0.410 0.912
Fig. 5. Discrepancy for aspiration points on h1 þ h2 þ � � �þhq ¼ 1; q dimensions.
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 23
tabulates the proportion of aspiration points
under N for given values of q.
For a random allocation of aspiration points, at
least 83% of aspiration points will be over N forfive or more objectives, which affects the behavior
of the two reference point formulations, PU and
PM.
3.4. Dispersion of PU and PM solutions
Related to the discrepancy discussed above, the
behavior of these two formulations can also beexamined from the perspective of dispersion. That
is, for a given set of aspiration points, what does
the pattern of solutions produced by PU and by
PM look like? To examine this, we consider the
three-dimensional case with equally spaced aspi-
ration points within the unit cube. By projecting
the unit cube onto the plane, a visual picture of the
dispersion patterns for PU and PM can be ob-tained. Fig. 6 illustrates this projection; the per-
spective is that of looking straight down the ray L,from above the ideal point hðUÞ. The spherical,nondominated set is also shown.
Using a 0.05 grid, this represents 11� 11�11 ¼ 1331 aspiration points within the unit cube.Figs. 7 and 8 show the resulting dispersion of so-
lutions over the spherical nondominated set forPU and PM.
The dispersion of, and by implication the dis-
crepancy between, PU and PM solutions are dif-
ferent. Overall, PM provides a more even pattern
of solutions. PU solutions are either toward the
centre or to the absolute extremes. Although it
cannot be seen from Fig. 7, the three extreme point
solutions (1,0,0), (0,1,0) and (0,0,1) each comprises100 PU solutions in contrast to 10 PM solutions in
Fig. 8. For example, under PU there were 100
different aspiration points which all gave the same
final solution of (1,0,0). For the three-dimensional
sphere, 52% of all potential aspiration points are
under N. Figs. 9 and 10 show the patterns of dis-
persion depending on whether the aspiration point
is over N or under N.
As Figs. 9 and 10 show, PU solutions grouptowards the center for under aspiration points and,
similarly, PM solutions group towards the center
for over aspiration points. Thus, the position of
the aspiration point relative to N, combined with
the choice of reference point method, can result in
either greater or lesser dispersion of solutions. The
differences in dispersion are quantified in Table 4.
The PU solutions, however, are less uniformlydistributed than the PM solutions and have larger
gaps near the extremes in both cases. From the
perspective of controllability, this means that (in
some parts of the solution space) a minor change
in aspiration point location can result in a major
change in solution location––a result that may
prove less than intuitive for a decision maker.
We use the determinant of the variance–covariance matrix as a general measure of dis-
persion for each group of solutions. Using this
measure, for hðaÞ under N, the amount of disper-sion among the PM solutions is 28.5 times that of
Table 3
Tabulation of aspiration points under N
Number of objectives (q) Proportion of aspiration points
under N
2 0.785398
3 0.523599
4 0.308425
5 0.164493
10 0.002490
15 0.000012
Fig. 6. Two-dimensional projection of aspiration points in the
unit cube.
24 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
the PU solutions. Similarly the amount of PU
dispersion for solutions generated using hðaÞ overN is 22.2 times the dispersion among PM solu-
tions. It should be noted, however, that the dis-
persion measures in Table 4 reflect the large
number of extreme points for PU.
4. Illustrative examples
We now further examine the nature of the PM
PU discrepancy in the context of two examples.
Example 1. This example (modified from Zionts
and Wallenius, 1976) illustrates some results for a
simple MOLP problem. The problem formulation
is shown below, along with a two-dimensional
projection of the nondominated set.
MAX f1ðxÞ ¼ 3x1 þ x2 þ 2x3 þ x4f2ðxÞ ¼ x1 � 2x2 þ 2x3 þ 4x4f3ðxÞ ¼ �x1 þ 8x2 þ 1:6x3 þ 3x4
s:t: g1ðxÞ ¼ 2x1 þ x2 þ 4x3 þ 3x46 60g2ðxÞ ¼ 3x1 þ 4x2 þ x3 þ 2x46 60 xi P 0;
i ¼ 1; 2; 3; 4:
The nondominated set comprises facets ABDC,
DFB and the edge EC as shown in Fig. 11. At only
one point of the entire nondominated set (E) is oneobjective maximized at a value of 1 with all other
objectives at 0, unlike the hyperspherical non-
dominated set where each objective can attain a
value of 1 with all others at 0. While the PM and
PU formulations again produce different patternsof dispersion, they are similar to those already
shown for the hyperspherical nondominated set.
Here we consider all aspiration points inside the
unit cube, again using a 0.05 grid. The aspiration
points corresponding to hðMÞ and hðUÞ have,however, been excluded, leaving a total of 1329
aspiration points.
The patterns of dispersion for PU and PM areshown in Fig. 12. Both formulations give a good
coverage of solutions, except that PM provides
more solutions in the extreme areas of the non-
dominated set.
We now consider in the following tables how
the position of the aspiration point affects the re-
sulting solution. From Table 1, we use two cate-
gories of aspiration point position with respect toN-under/on and over/neither.
Fig. 8. Two-dimensional projection of PM dispersion.Fig. 7. Two-dimensional projection of PU dispersion.
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 25
Table 5 shows that about half the aspiration
points were dominated by or equal to the resulting
solution; that is, they were under/on N. Further-more about half of the solution points were such
that PMwas closer to the center (that is, ray L) thanPU. However, this depended almost entirely on theposition of the aspiration point. If hðaÞ was underN, then PU was almost always closer to the center.Consideration of proportional and nonpropor-
tional solutions is provided in Tables 6 and 7.
Table 4
Dispersion statistics
PU PM
hðaÞ is under NCentroid (0.55, 0.55, 0.55) (0.48, 0.48, 0.48)
Dispersion 8.490E�06 2.422E�04
hðaÞ is over NCentroid (0.43, 0.43, 0.43) (0.52, 0.52, 0.52)
Dispersion 1.079E�03 4.862E�05
Fig. 10. PU (left) and PM (right) dispersion for aspiration points over N.
Fig. 9. PU (left) and PM (right) dispersion for aspiration points under N.
26 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
Solution proportionality has been shown here
for two reasons. It is surmised that decision
makers would prefer a solution to be propor-tional to their guess in any interactive MCDM
solution method. Further, this tabulation gives
some indication of the relative magnitude of
proportional and nonproportional solutions, for
each formulation. For PM, only 38.83% of as-
piration points result in proportional solutions,
compared with 50.64% for PU. Considering only
the nonproportional solutions for PM, almost60% of them were under N; that is, the rayfrom hðMÞ through hðaÞ did not intersect
with the efficient set. For PU, almost 80% of
the nonproportional solutions resulted from
the case of over/neither; only 21.80% of non-
proportional PU solutions occurred when hðaÞwas under N.
Example 2. In this three-dimensional example
the nondominated set simply consists of a line
from (1,1,0) to (0,0,1). Again 1329 aspiration
points from within the unit cube are considered
and the resulting solutions are categorized in
Table 8.
Fig. 12. Example 1: Dispersion for PU (left) and PM (right).
Fig. 11. Two-dimensional projection of the simple MOLP
problem.
Table 5
Example 1: Categorization of aspiration points by closeness to center
Position of hðaÞ PM closer to center than PU PM ¼ PU PU closer to center than PM Total Percent
Under/on 10 33 632 675 50.79
Over/neither 602 36 16 654 49.21
Total 612 69 648 1329
Percent 46.05 5.19 48.76
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 27
With this nondominated set, a tidy result can be
seen. PM is closer to the center than PU 55.76% of
the time. Not surprisingly, only 9% of the solu-
tions are proportional; intuitively, this occurs only
when the ray through the aspiration point inter-sects with the nondominated line. In summary, if
hðaÞ is under or on, then PU is at least as close tothe center as PM. If hðaÞ is over or neither, thenPM is at least as close to the center as PU.
We include the two examples in this section to
illustrate how PU and PM solutions compare in
nonhyperspherical cases. The patterns of disper-
sion are similar to the hyperspherical results.Further, these examples have significant num-
bers of nonproportional solutions. The centrality
properties proven for proportional solutions in
Appendix A are also generally the case for the
mixture of proportional and nonproportional so-lutions found in these examples.
5. Experiment
The previous sections provide a sensitivity
analysis of PU and PM solutions, controlling for
aspiration point location, and demonstrate howthe differences between PM and PU solutions are
largely a function of the location of the aspiration
point. Clearly, in any practical use of reference
point methods, decision makers do not randomly
choose aspiration points evenly throughout the
unit hypercube. Perhaps in practice with actual
aspiration points, these differences are not as sig-
nificant. We therefore designed a simple experi-ment to examine this question and to also see
which of the two types of solutions (PM or PU)
decision makers preferred.
A production scheduling decision problem was
used for the experiment. This simple problem has
been used successfully in other experimental con-
texts by one of the authors (e.g., Corner and
Buchanan, 1997). The problem concerned themanufacture of electrical components for lamps in
a medium size New Zealand company where most
of the company�s problems were attributable topoor production scheduling. The goal was to de-
velop a production schedule which minimized the
three conflicting objectives of operating costs,
stockouts, and labor temporarily laid off. A more
complete problem description can be found inCorner and Buchanan (1997). The nondominated
set of the three objective production scheduling
problem was chosen to be spherical.
A naive solution method, similar to the GUESS
method (Buchanan, 1997), was used to solve the
Table 8
Example 2: Categorization of aspiration points by closeness to center
Position of hðaÞ PM closer to center than PU PM ¼ PU PU closer to center than PM Total Percent
Under/on 0 144 361 505 38.00
Over/neither 741 83 0 824 62.00
Total 741 227 361 1329
Percent 55.76 17.08 27.16 100.00
Table 6
Example 1: Categorization of aspiration points by propor-
tionality of PM solution
Position of hðaÞ PM solution is not
proportional
PM solution is
proportional
Under/on 473 58.18% 202 39.15%
Over/neither 340 41.82% 314 60.85%
Total 813 516
61.17% 38.83%
Table 7
Example 1: Categorization of aspiration points by propor-
tionality of PU solution
Position of hðaÞ PU solution is not
proportional
PU solution is
proportional
Under/on 143 21.80% 532 79.05%
Over/neither 513 78.20% 141 20.95%
Total 656 673
49.36% 50.64%
28 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
problem whereby the decision maker guesses asolution (the aspiration point) between M and Uand the method finds a solution using both the PM
and PU formulations. The method presents the
PM and PU solutions in random order to elimi-
nate any order effects. The decision maker chooses
between the PM and PU solutions (or indicates
indifference) and the method proceeds until a sat-
isfactory solution is found. Although it was thegoal of the participants to derive a good produc-
tion schedule, we were only interested in the choice
between the PM and PU solutions at each itera-
tion. The solution method was programmed in
Cþþ to run in a Windows environment. Fifty-
eight students participated in the experiment which
was treated as a case study and became part of
their assessable work for a course; 38 were MBAstudents at Auburn University (from three differ-
ent classes), while the remaining 20 were students
from the University of Waikato (18 from an
undergraduate course in Operations, 2 from a
graduate course in decision making).
The hypothesis was that participants would
prefer PM to PU because, we assumed, decision
makers are more ‘‘achievement-oriented’’ than‘‘deviation-oriented’’. Because the production
scheduling problem was formulated with a spher-
ical nondominated set, PM will always produce a
proportional solution. This is not the case for PU.
We therefore eliminated all iterations where the
PU solution was not proportional. The data was
further reduced by eliminating all responses of
indifference. The final data set contained aspira-
tion points which resulted in a proportional PM orPU solution, between which participants expressed
a clear preference.
5.1. Results
The essential results are presented here, with a
full discussion reserved for Section 6. Aspiration
points were distributed such that almost 75% wereover N. This is not surprising––participants simplywanted greater achievement than was often feasi-
ble. If aspiration points were randomly distrib-
uted, then only 47.64% of aspiration points would
be expected to be over N. The distribution ofguesses, therefore, is not random ðp ¼ 0:0000Þ.With the raw data (before nonproportional PU
solutions and indifference were excluded), 82% ofaspiration points were over N.Table 9 provides a contingency table showing
the location of hðaÞ and solution preference.
Overall, PM is preferred at 57.0%. However, since
aspiration point location and solution prefer-
ence are not independent ðp ¼ 0:0006Þ, this overallpreference for PM cannot be interpreted in isola-
tion. If hðaÞ is over N, PM is clearly preferred at61.0% ðp ¼ 0:0000Þ. If hðaÞ is under N, PU is
preferred in this sample at 55.2%, but the result is
not statistically significant ðp ¼ 0:2131Þ.Recall from Fig. 1 and Appendix A that gen-
erally, if the aspiration point is over N, then PM is
closer to the center than PU; the converse is true
for an aspiration point under N. The data abovesuggest that decision makers prefer solutions
Table 9
Categorization of aspiration points and solution choices
Prefer PM solution Prefer PU solution Total
hðaÞ under 19.3% 31.5% 24.5%
65 80 145
44.8% 55.2% 100.0%
hðaÞ over 80.7% 68.5% 75.5%
272 174 446
61.0% 39.0% 100.0%
Total 100.0% 100.0% 100.0%
337 254 591
57.0% 43.0% 100.0%
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 29
which are closer to the center, which representmore of a compromise. Certainly the achievement
hypothesis, where decision makers are assumed to
prefer an achievement-oriented approach, is not
confirmed. Rather the position of the aspiration
point (under N or over N) and the apparent desireof decision makers for ‘‘centered’’ solutions would
seem to explain the choice of solution.
The dispersion of PM and PU solutions pre-sents no surprises. PU shows greater dispersion
than PM (a factor of 2) because 75% of the aspi-
ration points are over N, a result consistent withTable 4.
We now turn to a discussion of these results and
consider the implications for decision support.
6. Discussion and implications
The motivation for this paper came from an
examination of reference point solution methods;
specifically, what is the effect of the choice of ref-
erence point? We limited our study to a simple
Tchebycheff formulation and considered both
ideal and nadir reference points. These two for-mulations (PU––reaching down and PM––pushing
up) typically give different solutions for the same
aspiration point (and decision maker). The choice
of reference point influences the structure of any
interactive solution model, resulting solutions and,
perhaps, the behavior of a decision maker.
To summarize, the two formulations have dif-
ferent philosophies. PU seeks to minimize themaximum deviation from the ideal vector; i.e., to
make all weighted deviations equal, if possible.
PM seeks to maximize the minimum achievement;
i.e., to make all weighted achievements equal, if
possible. Our study shows that different solutions
result from the same aspiration point, and that this
difference can be significant. The magnitude of this
difference depends on the location of the aspira-tion point, the number of objectives and the shape
of the nondominated surface. Particularly, if as-
piration points are located near the nadir or ideal
reference points and they are somewhat away from
the center, then the two solutions are often sub-
stantially different. We also compare the non-
dominated set coverage patterns for the two
formulations. The PM formulation generally pro-duces more evenly spaced solutions with smaller
gaps near the boundaries of N.We develop a classification of aspiration points
as under N, over N or neither. In general, if an
aspiration point is under N then the PU solution
will be closer to the center than the PM solution.
Conversely, if the aspiration point is over N then
the PM solution will be closer to the center thanthe PU solution. We also demonstrate that as the
number of objectives increases, the proportion of
potential aspiration points located over N tends toincrease dramatically. Further, evidence from our
experiment and other experiences with the refer-
ence point methods suggests that decision makers
typically want more than they can get, specifying
more aspiration points over N than would be ex-pected given the number of objectives.
These results enable us to comment further. If
the intent is to obtain relatively evenly spaced
samples of solutions from N, the PM formulation
appears preferable. The sensitivity of solutions is
much greater under the PU formulation when the
aspiration points approach the boundary of the
hypercube. Thus, the congruence between the as-piration point and the resulting solution becomes
an important consideration, particularly in an in-
teractive reference point solution method. An even
coverage of solutions is also important when
generating a representation of the nondominated
set. Steuer (1986, pp. 328–330) addresses this issue
through a choice of weights derived from uniform
and Weibull distributions. In contrast, we haveused a uniform distribution of weights and con-
sidered the effect of the different reference point
methods. Thus, if the aspiration point is over Nand near the ideal point, there is likely to be a
considerable difference between the aspiration
point and the resulting solution when PU is used.
The difference will also occur with PM, if the as-
piration point is under N and near the nadir point;however this appears less likely to happen because
of decision makers� persistent and understandablechoice of aspiration points near the ideal point. If
some congruence is desired, this suggests that PM
should be used so that the resulting solution will
generally be similar to the aspiration point. This
congruence, or local controllability (Wierzbicki,
30 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
1986), is a desirable feature of a reference pointmethod.
We hypothesized that decision makers may be
more comfortable with achievement, rather than
deviation, oriented solution methods. We hypoth-
esized that decision makers would prefer PM so-
lutions for this reason. While the experiment
showed evidence of preference for one type of so-
lution, this is not simply a preference for the resultsof one formulation over another. The preference
appears, rather, to be related to the centrality of
the resulting solution, suggesting that decision
makers tend to seek compromise solutions, when
given the opportunity. Practically speaking, then,
what should a decision analyst do? Recommend a
method that will give a centralising tendency (as-
suming most aspiration points are over N) andthereby support a decision maker in this direction?
If so, then PM should be recommended, given that
most aspiration points are over N. Or should ananalyst recommend a method which counteracts
this tendency of decision maker toward compro-
mise in the hope that new and better solutions will
be found? In this case, then, PU should be con-
sidered, again since most aspiration points are overN. However, as shown, the divergence obtainedfrom using PU may be extreme.
The sensitivity analysis and the behavioral
study underscore the importance and influence of
the choice of aspiration point. If a decision maker
is aware of how a particular reference point for-
mulation ‘‘works’’, he can choose aspiration points
accordingly. We recommend that a decision makermove away from the ideal point and choose aspi-
ration points more widely. Such a strategy should
generate a greater spread/variety of solutions, re-
gardless of the formulation.
Both the PM and PU methods have their ad-
vantages, although the above discussion suggests
that the PM formulation would appear to be a
better choice. However, since both methods exist,we should encourage decision makers to make use
of both methods; as Russo and Schoemaker (1989)
suggest, it can be advantageous to use more than
one frame of reference. The value comes not from
choosing the ‘‘best’’ or most appropriate frame of
reference, but from considering more than one
perspective or frame.
Appendix A
Location of aspiration point and solution cen-
trality.
Let domination cone D ¼ Rqþ n f0g ¼ fh 2 Rq 3
hk P08 k¼1;. . . ; q; 9 k¼1; . . . ; q 3 hk > 0g whereRq
þ is the nonnegative orthant of q-dimensionalspace. Then we define:
under N: Aspiration point hðaÞ is under N if
fhðaÞ þ Dg \ N 6¼ £. Thus hðaÞ is consideredto be under N if there is at least one h0 2 Nwhich dominates it. When hðaÞ is under N,h M and h U will both dominate hðaÞ. If not,the optimality of y M and y U is contradicted.
over N: Aspiration point hðaÞ is over N if there
exists some h0 2 N such that hðaÞ 2 fh0 þ Dg.Therefore hðaÞ is defined to be over N if it dom-inates at least one h0 2 N . An aspiration vectorhðaÞ over N may or may not dominate the re-sulting h M and h U solutions.
neither: Aspiration points may also be neitherunder nor over N. When this is the case, h M
and h U will neither dominate nor be dominatedby hðaÞ.
For proportional solutions, the location of theaspiration vector a will be either on, over or underN. If hðaÞ is on N, h U ¼ h M ¼ hðaÞ. Let us ex-amine the case where aspiration vector hðaÞ is noton N and both h U and h Mð6¼ hðaÞÞ are propor-tional. In this case, hðaÞ is either over or under N.We first consider hðaÞ over N and show that h M ismore central (closer to L) than h U . From this it
can be seen that when hðaÞ is under N, h U is themore central solution.
A.1. h(a) over N
When hðaÞ is over N, h M is on the line seg-
ment from hðMÞ to hðaÞ. In other words, h M ¼hðMÞ þ ahðaÞ for a < 1. With the transformationin (4), hðMÞ is at the origin, so h M ¼ ahðaÞ fora < 1. Similarly, h U is the end point of a line seg-ment extending from hðUÞ through hðaÞ. So h U ¼hðUÞ þ bðhðaÞ � hðUÞÞ for b > 1. Since hðUÞ is aunit vector, h U ¼ 1þ bhðaÞ � b, for b > 1. For
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 31
any hðzÞ, the point on L closest to hðzÞ using theEuclidean metric is
e ¼Pq
i¼1 hðziÞq
; . . . ;
Pqi¼1 hðziÞq
�
Consider kh M ; e Mk2 and kh U ; e Uk2, the mini-mal Euclidean distances between the PM and
PU solutions, respectively, and L. Let Q M ¼ðkh M ; e Mk2Þ
2and Q U ¼ ðkh U ; e Uk2Þ
2.
Q M ¼Xq
k¼1h Mk
�Pq
i¼1 h Mi
q
�2
Since h M ¼ ahðaÞ,
Q M ¼Xq
k¼1ahðakÞ
� a
Pqi¼1 hðaiÞq
�2
¼ a2Xq
k¼1hðakÞ
�Pq
i¼1 hðaiÞq
�2
¼ a2ðkhðaÞ; eak2Þ2
where ea is the point on L closest to hðaÞ.Similarly,
Q U ¼Xq
k¼1h Uk
�Pq
j¼1 h Ui
q
�2
Since h U ¼ 1þ bhðaÞ � b,
Q U ¼Xq
k¼1ð1
þ bhðakÞ � bÞ
�Pq
i¼1 ð1þ bhðaiÞ � bÞq
�2
¼ b2Xq
k¼1hðakÞ
�Pq
i¼1 hðaiÞq
�2
¼ b2ðkhðaÞ; eak2Þ2
Thus, kh M ; e Mk2 ¼ akhðaÞ; eak2 and kh U ; e Uk2 ¼
bkhðaÞ; eak2. Since a < 1 and b > 1, kh M ; e Mk2 <kh U ; e Uk2.
A.2. h(a) under N
Using the logic above, it can be shown that h U
is closer to L when hðaÞ is under N.
Appendix B
Terminology and definitions
a aspiration vector; a ¼ fa1; a2; . . . ; aqg;Mk < ak < Uk 8 k ¼ 1; 2; . . . ; q
daL Euclidean distance between given hðaÞand the point closest to it on L, ea
dh Euclidean distance or discrepancy between
the h M and h U solutions for any given hðaÞD domination cone; D¼Rq
þnf0g¼fh2 Rq3hkP08k¼1;2;...;q;9k¼1;...; q 3 hk > 0gwhere Rq
þ is the nonnegative orthant of q-dimensional space
e point on L closest to any given hðzÞ;e ¼ ð
Pqi¼1 hðziÞ=q; . . . ;
Pqi¼1 hðziÞ=qÞ
ea point on L closest to hðaÞ; ea ¼ðPq
i¼1 hiðaÞ=q; . . . ;Pq
i¼1 hiðaÞ=qÞe M point on L closest to h M ; e M ¼
ðPq
i¼1 h Mi =q; . . . ;
Pqi¼1 h
Mi =qÞ
e U point on L closest to h U ;e U ¼ ð
Pqi¼1 h
Ui =q; . . . ;
Pqi¼1 h
Ui =qÞ
fkðxÞ kth objective function of x 2 X � Rn,
k ¼ 1; 2; . . . ; qFðxÞ F ðxÞ ¼ ff1ðxÞ; f2ðxÞ; . . . ; fqðxÞg � Rq
gjðxÞ jth constraint on set of alternatives,
j ¼ 1; 2; . . . ;mhðzÞ transformation function; hðzkÞ¼ðzk�MkÞ=
ðUk �MkÞ; k ¼ 1; 2; . . . ; qhk transformed criterion values; hk ¼ hðzkÞ;
k ¼ 1; 2; . . . ; qh vector of transformed criterion values; h ¼
fh1; h2; . . . ; hqgh M an optimal solution to PM(w); h M ¼
fh M ;1 ; h M ;
2 ; . . . ; h M ;q Þ
h U an optimal solution to PU(w); h U ¼fh U ;
1 ; h U ;2 ; . . . ; h U ;
q ÞhðaÞ transformed aspiration vector; 0< hðakÞ<
1 8 k ¼ 1;2; . . . ;qhðMÞ transformed nadir vector; hðMÞ ¼ f0;
0; . . . ; 0ghðUÞ transformed ideal vector; hðUÞ ¼ f1;
1; . . . ; 1gL line segment hðMÞhðUÞMk nadir value for kth criterion; Mk ¼
minz2Nfzkg; k ¼ 1; 2; . . . ; qM nadir vector; M ¼ fM1;M2; . . . ;MqgN set of nondominated criterion vectors;
N ¼ fz : 9= z0 2 Z where z0 P z; z0 6¼ zg
32 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34
PMðwÞ multiple objective mathematical program
with weighted achievement function
based on MPUðwÞ multiple objective mathematical program
with weighted deviation function based
on Uq number of objective functions
Q M squared Euclidean distance between h M
and the point on L closest to it, e M
Q U squared Euclidean distance between h U
and the point on L closest to it, e U
Uk ideal value for kth criterion; Uk ¼maxz2Zfzkg; k ¼ 1; 2; . . . ; q
U ideal vector; U ¼ fU1;U2; . . . ;UqgVq hypervolume of a q-dimensional hyper-
sphere with radius rVq-cube hypervolume of a q-dimensional hyper-
cube with side rw vector of weights; w ¼ fw1;w2; . . . ;wqgX set of feasible alternatives; X ¼ fx 2
Rn : gjðxÞ6 0; j ¼ 1; 2; . . . ;mgy minimax or maximin variable in the PU
and PM formulations
zk criterion values; zk ¼ fkðxÞ; k ¼ 1; 2; . . . ; qz vector of criterion values; z¼fz1;z2;...;zqgZ image of X in criterion space; Z ¼
F ðxÞ � Rq
References
Beach, L.R., Mitchell, T.R., 1990. Image theory: A behavioral
theory of decisions in organisations. In: Staw, B.M.,
Cummings, L.L. (Eds.), Research in Organisation Behavior,
vol. 12. JAI Press, Greenwich, CT.
Benayoun, R., deMontgolfier, J., Tergny, J., Larichev, O., 1971.
Linear programming with multiple objective functions: Step
method. Mathematical Programming 3, 366–375.
Buchanan, J., 1997. A naive approach for solving MCDM
problems: The GUESS method. Journal of the Operational
Research Society 48 (2), 202–206.
Buchanan, J.T., Daellenbach, H.G., 1987. A comparative
evaluation of interactive solution methods for multiple
objective decision models. European Journal of Operational
Research 29, 353–359.
Buchanan, J.T., Gardiner, L.R., 2001. A comparison of two
reference point methods in multiple objective mathematical
programming, Department of Management Systems Re-
search Report Series, University of Waikato.
Charnes, A., Cooper, W.W., 1961. In: Management Models
and Industrial Applications of Linear Programming, vol. 1.
John Wiley, New York.
Corner, J.L., Buchanan, J.T., 1995. Experimental consider-
ation of preference in decision making under certainty.
Journal of Multi Criteria Decision Analysis 4 (2), 107–
121.
Corner, J.L., Buchanan, J.T., 1997. Capturing decision maker
preference: Experimental comparison of decision analysis
and MCDM techniques. European Journal of Operational
Research 98 (1), 85–97.
Downing, C.E., Ringuest, J.L., 1998. An experimental evalu-
ation of the efficacy of four multi-objective linear program-
ming algorithms. European Journal of Operational
Research 104 (3), 549–558.
Gupta, R., Arora, S.R., 1977. Programming problem with
maximin objective function. Opsearch 14, 125–130.
Hwang, C.-L., Lai, Y.-J., Liu, T.Y., 1993. A new approach for
multiple objective decision making. Computers and Oper-
ations Research 20 (8), 888–899.
Kaplan, S., 1974. Application of programs with maximin
objective functions to problems of optimal resource alloca-
tion. Operations Research 22, 802–807.
Kendall, M.D., 1961. A Course in the Geometry of n Dimen-sions. Charles Griffin & Company Limited, London.
Korhonen, P.J., Wallenius, J., 1988. A Pareto race. Naval
Research Logistics 35 (6), 615–623.
Lai, Y.-J., Liu, Y.-T., Hwang, C.-L., 1994. TOPSIS for
MODM. European Journal of Operational Research 76,
486–500.
Martinson, F.K., 1993. Fuzzy vs. minmax weighted multiob-
jective linear programming illustrative comparisons. Deci-
sion Sciences 24 (4), 809–824.
Michalowski, W., Szapiro, T., 1992. A bi-reference procedure
for interactive multiple criteria programming. Operations
Research 40 (2), 247–258.
Nakayama, H., Sawaragi, Y., 1984. Satisficing trade-off
method for multiobjective programming. In: Grauer, M.,
Wierzbicki, A.P. (Eds.), Interactive Decision Analysis.
Springer-Verlag, Berlin, pp. 113–122.
Posner, M.E., Wu, C.-T., 1981. Linear max–min programming.
Mathematical Programming 20, 166–172.
Reeves, G.R., MacLeod, K.R., 1999. Some experiments in
Tchebycheff-based approaches for interactive multiple ob-
jective decision making. Computers and Operations Re-
search 26, 1311–1321.
Russo, J.E., Schoemaker, P.J.H., 1989. Decision Traps. Dou-
bleday, New York.
Simon, H.A., 1957. Models of Man. Macmillan, New York.
Steuer, R.E., 1986. Multiple Criteria Optimization: Theory,
Computation, and Application. John Wiley, New York.
Steuer, R.E., Choo, E.-U., 1983. An iteractive weighted
Tchebycheff procedure for multiple objective programming.
Mathematical Programming 26, 326–344.
Steuer, R.E., Silverman, J., Whisman, A.J., 1993. A combined
Tchebycheff/aspiration criterion vector interactive multiob-
jective programming procedure. Management Science 39
(10), 1255–1260.
Tversky, A., Kahneman, D., 1981. The framing of decisions
and the psychology of choice. Science 211, 453–458.
J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34 33
Weisstein, E.W., 1999. The CRC Concise Encyclopedia of
Mathematics. CRC Press, Boca Raton, Florida.
Wierzbicki, A.P., 1980. The use of reference objec-
tives in multiobjective optimization. Lecture Notes in
Economics and Mathematical Systems 177, 468–
486.
Wierzbicki, A.P., 1986. On the completeness and constructive-
ness of parametric characterizations to vector optimization
problems. OR Spektrum 8, 73–87.
Zionts, S., Wallenius, J., 1976. An interactive programming
method for solving the multiple criteria problem. Manage-
ment Science 29, 519–529.
34 J. Buchanan, L. Gardiner / European Journal of Operational Research 149 (2003) 17–34