Math1 Comput. Modding, Vol. I I, pp. 206-209, 1988 0895.7177,88 $3.00 + 0.

00
Printed in Great Britain Pergamon Press plc

TOPICS IN THEORY OF THE ANALYTIC HIERARCHY PROCESS

OPPORTUNITIES AND LIMITATIONS OF AHP IN MULTIOBJECTIVE PROGRAMMING

David L. Olson

Department of Business Analysis & Research

Texas AbM University, College Station, TX 77843 USA

Ab_SILrat:l:. The Analytic Hierarchy Process is a useful tool for

multiobjrctive decision making in its own right. In addition, it has the
potential for expediting multiple objective programming analyses. Multiple
objective proqramming techniques face the problem of a large (if not
inEinite) number of alternatives. Generally, the problem involves
optimizing some unexpressed utility function over a feasible region. AHP
provides a Iuseful means of obtaining an initial linear approximation of
t.his unexpressed utility function, with the potential of expediting the MOP
analysi:;. Othl:r benefits include enhancinq decision maker learning through
use of the consistcsncy measure. Problems discussed are the impact of
comparing many objectives, the sensitivity of the consistency index, and
l.he use of the eigen vector as a literal estimator of utility.

@Y~OL’&. Analytic Hierarchy Process; multiple objective programming.

alternative investment choices, or means of

balancing return with risk, or balancing product
The Analytic Hierarchy Process ( AHP - Saaty, quality with short run profit measures.
1977, 1980) has proven t.0 5(~ vrry useful in
assisting decision maker selection from a finite Consideration of multiple objectives would yield
set of alternatives. This idea has been a model:
extended to decisions under conditions of
Max 2, - 2 c (21
xJ
multiple objectives as well. The purpose of
this paper is to discuss the benefits and ,=I lJ
problems involved in using AHP as a means of
developing a linear approximation of utility Max 2, - f: CZJ XJ

J=1
under conditions of multiple objectives in
linear programming models.

Management science is concerned with applying Max 2, - % ck, x,

the scientific approach to business problems J=l
through development of a mathematical model of a
st the same constraint set In (1)
problem, gathering sound data, and use of the
model to aid decision making. The intent is to
support human judqement. Many business decision A common way to consider these multiple
problems can be supported by linear programming objective functions simultaneously is through an
models. An obvious objective would be to additive linear utility function:
Mximize profit within the constraints of
feasibility. A well known general formulation
would be: Max wI 2, + w2 2, + . . + wlr 2, (31

st the same constraint set in (11

Max 2. - 2
J_, cJ xJ

Utility. The concept of needing to balance

st Jg, a11 xJ s bl economic values has long been considered.
Jeremy Bentham and other economists discussed
for i=l,..m utility. Utility as an operational theory has
x, 2 0 for j-l&l been questioned by Georghescu-Roeqen (1954) and
Morgenstern (1972). Further, Simon (1979)
reported the phenomenon of satisficing, where
Multiple Objectives. Many business decisions many decision makers were observed to accept
involve conflicting objectives, requiring less than optimal solutions to many decision
decision makers to reconcile tradeoffs among problems. March (1965) noted the bounded
these objectives (Cohen, 1978; Zeleny, 1983; rationality with which decision makers must
Steuer, 1986). Such tradeoffs often become often cope. The direct development of utility
necessary because of the need to balance is the objective of a strain of active research.
different rates of return to such scales as cash That approach, however, has not resulted in an
flow, after tax profit, or net present value of easy means of identifying specific decision

206
 Proc. 6th Int. Conf on Mathematical Mmodeiling 207

maker utility. Regardless, utility does this set. This information is used as a clue
provide a useful construct to explain how to the relative weights for the combined
decision makers balance multiple, conflicting objective function (3). A pattern for weights
objectives. focusing around the last selection is used to
generate the next set of alternatives presented
While utility could theoretically take on any to the decision maker. This set is also
shape, DeBreu (19751 defined a utility function filtered to provide maximum dispersion. The
as having the properties that more of a good is process continues for a predetermined number of
better than less (an increasing function), that steps, or until decision maker satisfaction is
more of a good contributes to utility but at a obtained.
decreasing rate (convex), and that there are no
satisfaction levels (continuous). This process is quite workable in practice, and
Steuer has presented a number of applications.
Multinle Objective Prooramminq. Many decision There are, however, two features which are
problems supported by linear programming unattractive with this method. First, the
involve multiple objectives. Multiple technique requires that a set of solutions be
objective programming is a class of techniques filtered to obtain maximum dispersion (find the
which seek solutions yielding optimal decision n most diverse solutions of m candidates).
maker utility. This is done in a variety of This is done automatically by Steuer’s code.
ways. However, that code is not widely available.
Another unattractive feature is that the
A key concept in multiple objective analysis is technique is entirely linear programming based,
that of nondominated solutions. For any number meaning that only extreme point solutions are
of objectives, a solution is nondominated if capable of being produced by the technique.
improvement on any one objective would While this is not a problem in single objective
necessitate diminished attainment on any other 1 inear programming analyses, it does not
objective. On the other hand, a dominated reflect the expected behavior of utility
solution would be one in which you could obtain functions.
a feasible solution which improved at least one
objective without diminishing any others. Analytic Hierarchy Process in MOLP. AHP has
been applied as a means of identifying a linear
Different modeling approaches to multiple approximation of utility for selection among a
objective problems can be viewed as different set of alternatives. What is proposed is that
views of utility. A single objective linear AHP be used as a means of identifying a working
programming model often views profit as a utility function approximation for a
linear measure of utility. Multiattribute mathematical programming analysis. The eigen
utility theory is concerned with identifying vector obtained from AHP would be the source
and defining utility. Multiple objective for the weights in (31. This has been applied
approaches (Hwang, et al. 1980; Evans 1984; in O-l multiple objective programming analyses
Rosenthal 1985) have taken a variety of by Mitchell and Bingham (19861 and Bard(19861,
approaches to utility. Generating nondominated and in a continuous multiple objective analysis
solution sets and allowing decision makers to by Olson, et al. (19861.
select from that set infers maximum decision
maker utility through choice. Unfortunately, The general procedure outlined in Olson, et al.
this may involve a rather large number of was to begin by development of a payoff table,
nondominated solutions for decision maker presenting decision makers with a view of the
inspection. Interactive techniques seek to optimal solutions available for each objective
incorporate decision maker utility through in turn. For k objectives, this would yield k
directed search for improved solutions assisted solutions (barring duplications). The payoff
by decision maker selection from among a table would measure the attainment of each
smaller subset of nondominated solutions. objective for each solution.

Desired Features of MOLP. A multiple objective PAYOFF TABLE

linear programming analysis should have a
measures 21 Z2 .Zk
number of features.
1) It should reflect utility, resulting in optimal
improved decision making.
2) Undue burden upon decision makers should be 21 z: z: .z, 1
avoided.
3) Decision maker learning should be enhanced, 2
Z2 z: z; .z,
especially through development of tradeoffs
among objectives.
4) Algorithmic support should be available,
ideally without the need for custom Zk t I z: . z;
computer code. Further, computer run time
should be reasonable.
5) The technique should work for models with where Pi{ is the attainment on the iN
more than two objectives. objective for the solution optimizing the J&
Steuer ‘5 ob]ective, Z; is the optimal solution for
Method One interactive MOLP
technique which’performs relatively well on obJectlve i.
this set of desireable features is Steuer’s
method (Steuer 19771. The general operation
of that techniqueis to generate a predetermined The decision maker can learn of the tradeoffs
number of diverse nondominated solutions for by viewing this table. Adjustments to
decision maker consideration. The decision eliminate the different scales of the k
maker is asked for the preferred solution from objective functions can be intuitively adjusted
208 Proc. 6th Znt. Conf.on Mathematical Modelling

for by the decision maker, or more a more Bingham h Mitchell, or in Bard, both of whom
formal means of adjusting scale as given in dealt with large numbers of objectives. An
Olson, et al. can be used. added complication is the need to reflect
relative importance to objectives, when the
AHP can then be applied to obtain the weights scales used to measure each objective in (31
for (31. Without these weights, an additive can be quite diverse. This point was
combined objective function is liable to lead addressed above in the section on the payoff
to misleading solutions. Further work on this table.
point can be obtained from the author. AHP
provides not only the approximated utility Another factor is the sensitivity of the
weights, but in addition provides the valuable consistency coefficient given by AHP. Saaty
feature of a measure of consistency (see Saaty (19771 developed the measure:
1980).
blllU - nMn-11
With the weights obtained, an initial solution
to the MOLP can be obtained with linear as a measure of consistency, where $,,= is
programming. This yields a solution with the maximum elgen value for the paired
attainments: comparison matrix and n is the number of
objectives compared.

To be consistent, this measure should be less

In addition, 2k new solutions can be obtained than .l. Manipulation of objective weights
by forcing improvement on each objective in indicates that if all objectives are near egual
turn, developing a search pattern of two steps. in importance, the consistency measure will
For each of k objectives, a constraint is require relatively consistent rankings of all
added : objectives. On the other hand, if one
objective is much more important than the
z, z z; l * others, very little consistency among the other
objective ratings is needed to pass the test
+ can be varied to some range (for example limit. This is at first sight a problem.
511 and 25X) from the current attainment to However, it also reflects the concept that
the optimal attainment for objective i in lower rated objectives are not as crucial.
order to generate 2k new solutions.
Consider the following paired comparison matrix
The intent of this step is to give the decision where the relative ranking of objective 2 to
maker new alternatives which improve one objective 3 is allowed to vary:
objective by controlled amounts. The 2ktl
solutions (including the initial solution) are Objective 1 2 3
presented to the decision maker for preference Objective
selection. 1 1 2 3
2 1 V
This procedure can be continued for a 3 1
predetermined number of iterations, using the
last selected alternative as the current The relative importance of objective 2 to
solution for each step as in Steuer’s method, objective 3 would have to be 1:l to 1:4 to pass
or until convergence (the decision maker the consistency test.
selects the last current solution). If further
investigation was desired, smaller grids around Next, consider:
the current solution can be incorporated.
Objective 1 2 3
The benefits expected from this approach are Objective
that special computer codes are not required 1 1 2 8
(generally available linear programming codes 2 1 V
can be used), decision maker learning is 3 1
enhanced (both through use of the payoff table
In this case, any relative I ranking of objective
and AHP), and this technique has the capability
2 to objective 3 greater than 1:l would pass
of producing nondominated solutions that are The point might be that
the consistency test.
not original constraint set corner points. This
minor details may not be that important.
Utility would be reflected through decision leads to the inference that consideration of
maker selection without the need to develop the
many objectives is not necessary.
full tradeoff function.
Both tiitchell & Bingham and Bard dealt with
Potential Problems W& &HJ. There are large number of
analyses involving a
potential difficulties with the AHP portion of Both grouped the objectives,
objectives.
this analysis. Some potential difficulties Bard, for
combining similar areas of interest.
with the rest of the analysis are addressed in reduced the paired comparison matrix
instance,
Olson, et al. This has the
from 10 objectives to 5.
favorable feature of reducing the number of
The first difficulty is in the potential number paired comparisons from 45 to 10. This would
of compar lsons needed on the part of the considered.
at first glance reduce the factors
decision maker. That number of paired of the original 10 objectives
However, many
comparisons is: If they
were of a similar nature to others.
contained the same functional tendencies (were
$ (n-1) highly correlated), statistical theory would
indicate that one representative would suffice.
While this number is potentially large, making
paired comparisons has not proven difficult in A last point is that AHP provides a linear
 Proc. 6th Int. Conf. on Mathematical Modelling 209

estimator of utility. It should be viewed as a Hwanq, C. L., A. S. M. Masud, S. R. Paidy and

gradient estimator of a nonlinear underlying K. Yoon (1960). Mathematical programming
utility function, and as such, can be expected with multiple objectives: A tutorial.
to change as the perspective of the decision Comnuters and Onerations Rm, 1,
maker changes. This point is addressed by 5-31.
Forman elsewhere in this proceedings. March, J. G. (1978). Bounded rationality,
ambiguity, and the engineering of choice.
Conclusions. A consistent means of explaining The Bell Journal of Economics, 2, 587-608.
the tradeoffs among objectives is a major need Mitchell, K. H. and G. Blngham (1986).
in multiobjective analysis. AHP provides a Maximizing the benefits of Canadian Forces
workable means of obtaining this tradeoff. In equipment overhaul programs using multl-
addition, a valuable indicator of relative objective optimization. INFOR.&
consistency is provided. 251-264.
Horgenstern, 0. (1972). Thirteen critical
Steuer’s method is a well developed, workable points in contemporary economic theory:
interactive multiple objective technique. An interpretation. Journal of Economic
However, the search pattern provided by that Literature, & 1163-1169.
technique is incapable of obtaining noncorner Olson, D. L., M. Venkataramanan and J. Mote
point efficient solutions due to the method (19861. A technique using analytical
used. Use of bounds on objectives through hierarchv nrocess in multiobjectlve
constraints would alleviate that problem, planning-mbdels. Socio-Economic Plannlnq
giving a multiple objective analysis the Sciences, 20, 361-368.
ability to reflect nonlinear utility. AHP can Rosenthal, R. E. (1985). Principles of
support this approach by providing a set of multiobjective optimization. Decision
objective function weights which expedite the Sciences
PI 16 133-152.
-I
search in the direction of improved utility. Saaty, T. L. (1977). A scaling method for
Used directly as a linear estimator of utility, priorities in hierarchical structures.
AHP would have the same feature as Steuer’s Journal of Mathematical Psycholoqv, II,
method, in that only original model corner 234-281.
points would be considered. Saaty, T. L. (1980). The Analytic Hierarchy
Process McGraw-Hill, New York.
A primary difficulty in applying AHP to Simon,‘A.)(1979). Rational decision making
multiobjective analyses is the potentially in business organizations. The American
large number of paired comparisons asked of Economic Review, & 493-513.
decision makers. However, paired comparisons Steuer. A. E. (1977). An interactive multlole
have been demonstrated to be relatively easy in objective-linear programming procedure:
real applications. Further, f ecus upon TINS Studies in the Manaoement Sciences,
important objectives would reduce the number of b 225-239.
paired comparisons needed, and focus decision Steuer, R. E. (1986). MUltiDle Criteria
maker attention upon more important objectives. Ontlmization: Theory, Computation and
AnDllcation. John Wiley C Sons, New York.
Probably more important is the need to Zeleny, H. (1983). Multinle Criteria Decision
eliminate the impact oE scale. It is difficult Makinq. McGraw-Hill, New York.
to give a relative importance of profit
measured in millions oE dollars, with reduction
of liability measured in lawsuits expected.
This scalar complication can be reduced by
focusing upon the concept of profit versus the
concept of risk, and eliminating differences in
scale by other means, such as the use of the
range in objective values obtained from the
payoff table as given in Olson, et al.

AHP provides a valuable means of supporting

multiple objective programming. It has proven
very useful in comparing discrete alternatives.
It should prove just as useful in supporting
multiple objective programming analyses.

REFERENCES

Bard, J. F. (1986). A multiobjective

methodology for selecting subsystem
automation options. Hanaqement Science,
12, 1628-1641.
Cohon, J. L. (1978). Multiobjective
Prooramminq and Planninq. Academic Press,
New York.
DeBreu, G. (1975). The Theory of Value. The
Colonial Press, Clinton, MA.
Evans, G. W. (1984). An overview of techniques
for solving multiobjective mathematical
programs. Manaqement Science, 30, 1268-
1282.
Georghescu-Roegen, N. (1954). Choice,
expectations and measurability. Quarterly
Journal of Economics, 68, 503-541.