Вы находитесь на странице: 1из 6

Socio-Econ. Plann. Sci. Vol. 20, No. 6, pp. 355-360, 1986 0038-0121/86 $3.00 + 0.

00
Printed in Great Britain Pergamon Journals Ltd

EXPLORING OPTIMIZATION THROUGH HIERARCHIES


AND RATIO SCALES

THOMAS L. SAATY
Graduate School of Business, University of Pittsburgh, Pittsburgh, PA 15260, U.S.A.

Abstract-This paper explores the concept of optimization by solely using the AHP and compares
outcomes with those obtained in traditional optimization theory without and with constraints. The
difference is essentially in the absence of the traditional black box involving complex manipulations in
algebra or the calculus on an assumed linear or nonlinear mathematical structure. This paper is an
exploratory attempt to use an individuals understanding together with a way to convert his judgments
to ratios to deal with optimization. At first sight it may go contrary to the readers (learned) intuition,
but in the end he must face the question of whether the magic of traditional manipulations gives rise to
better answers than ones actual and complete understanding in which one does not abdicate judgmental
control of the solution, and why. This idea is ripe for deeper and more detailed exploration.

1. INTRODUCTION and measurable quantity (such as dollar expenditure),


Optimization is a goal directed activity. Goals depend
but rather with the optimization of vague and intan-
gible concepts such as military worth, loss of life, or
on the perceptions and preferences of the people. It
trust in the business. How to deal with these is
follows that in the final analysis optimization is an
certainly not by starting with a simple exercise, and
attempt to satisfy peoples goals. Ordinarily we use
Cartesian geometry to solve an optimization problem
eventually coming to regard it as the real situation
because of the complexity and hard work involved in
by assuming independent variables represented or-
formulating the problem.
thogonally. If some variables are more important, we
take functions of them, e.g. we multiply them by It is useful to consider how rare it is that an
appropriate constants and raise them to powers to individual attempts to maximize or minimize any-
reflect this importance in terms of larger or smaller thing. Early in life we are trained for balance, follow
numbers than represented by other variables. Out of the happy medium, to be satisfied. The concept of
habit and convenience the same basic model is used sufficiency can be found not simply in some modem
to represent a wide variety of problems interpreted to works in economics and operations research but also
in ethics and philosophy. Even risk avoidance is only
fit the model. This is particularly useful when the
number of variables is very large, as it saves time, but minimized in theory. In practice people take varying
could destroy relevance. But is this justified? How degrees of risk to deal with the world.
valid are our highly structured models and their
solutions? How should the layman regard the black
box manipulations of these models, particularly when 2. COMMENTS ON THE ANALYTIC
they have no corresponding real life meaning? HIERARCHY PROCESS
In the book, Optimization in Integers and Related
The Analytic Hierarchy process (AHP) serves as a
Extremal Problems, [l], there is contained a compre-
framework for people to structure their own prob-
hensive framework for optimization, organized to
lems and provide judgements based on knowledge,
include formulations using equations and in-
reason or feelings, to derive a set of priorities consid-
equalities. Classical optimization which has its ori-
ered as an optimal solution to a decision problem,
gins in physics and engineering is concerned with the
perhaps involving resource allocation. The AHP pro-
maximization or minimization of a function (or a
vides the opportunity for the following kinds of
functional) subject to equality and/or inequality con-
considerations:
straints expressed in algebraic or, more generally, in
functional form, sometimes using derivatives and (i) While in classical optimization one is restricted
integrals. In programming theory the constraints also to measures of money, time, weight, temperature and
include the nonnegativity of the variables. other measurables, with the AHP one can use these
The purpose of optimization is to identify a best tangibles along with relative measurements of intan-
set of alternatives, but it need not be tied to a gibles derived through comparisons.
particular structure. For a large class of practical (ii) The structure of a problem is not set in
problems it is only with a great deal of simplification advance, but is largely developed by those who
that a problem is formulated in a standard and experience it.
familiar form so that manipulations can be carried (iii) No assumptions of orthogonality and inde-
out to derive a solution. For example, at times pendence need to be made as the variables may be
military or business problems are not concerned with interdependent and hence the use of a Cartesian
the maximization (or minimization) of a well defined coordinate system may not be appropriate.
355
356 THOMAS L. SAATY

(iv) One is not restricted to the use of a euclidean


metric.
(v) Different weights may be attached to different
constraints or criteria.
(vi) As in real life the objective function is so
intermeshed with the constraints that optimization is
treated as a process of interaction between objectives
and constraints without separating them.
(vii) One can optimize any number of objectives,
and not simply one.
(viii) We need models that improve our under- Fig. 1. Hierarchy for maximizing profit.
standing and ability to interpret the real world
through interaction so that the line which separates
what we think and what we experience disappears Let us assume that the farmer has not studied
and reality becomes knowledge itself. calculus, nor does he know an expert consultant who
could set up the problem for him. Even if he does, he
3. TWO ROLES FOR THE AHP IN has doubt that his crop will increase in such a
OPTIMIZATION systematic manner, or that the price predicted for the
future would be as expressed in the statement of the
There are two ways in which the AHP can be used problem. Of course, he can solve his problem using
in optimization. The first is natural for anyone con- pure arithmetic if he buys into these clear cut assump-
cerned with including the measurement of intangibles tions, and has no need of calculus anyway. The
in a model, for the AHP is simply a theory of complexity of his operations cannot be brought easily
generating and combining ratio scales. The resulting into a simple algebraic formula from which a precise
scale values or priorities are incorporated as answer is then obtained. The object is to assist him
coefficients of an objective function, of constraints or by unfolding the problem according to its uncer-
of a payoff matrix. This kind of application extends tainties, enabling the farmer to get a good estimate
our reach in making applications. for when he should sell his potatoes. It is hoped that
The other is direct use of the AHP in optimization. for this simple exercise the answer would be close to
It is useful when our existing framework of the the calculus answer by way of validating such an
optimization of an objective function subject to sep- approach which may be useful for other complex
arately written constraints is not easy to identify and problems he faces. Let us note that he may have
construct because the two cannot be neatly separated several questions in mind and we can only assist him
in reality but are in the mind of the modeller. We said to deal with them one or at best a few at a time, but
earlier that optimization is finding a best solution to without simplifying the original version to obscurity.
a problem no matter how that problem is structured. To solve the problem we construct a hierarchy (not
The process avoids imposing a prior mathematical necessarily unique) illustrated below, whose goal is to
structure on the problem with specified parameters maximize the farmers profit. Profit depends on yield
and variables and generates a numerical answer in a and on the price per bushel which he can get when
step by step fashion using judgment and concrete he harvests the crop. In the end profit is determined
information. by the number of weeks that the farmer waits to sell
Questions of independence and dependence among his crop. The number of weeks he waits to harvest
the entities being treated become academic in the and sell, a period of no more than five weeks,
AHP if we can treat the criteria and objectives as represent his alternatives (0, 1, 2, 3, 4, 5) at the
different but intimately related decompositions of the bottom or fourth level of the hierarchy. The other
reality we experience. We must define them carefully two levels below the Goal of profit are respectively,
to ensure that they capture what we intend them to yield, whose elements are low, medium, high, and the
mean and are related in desired ways. There is no price per bushel whose elements are small, medium,
harm in finding that there is overlap among them. We large. The alternatives at the bottom, indicating
can still study both their independent and inter- weeks will be compared with respect to the likelihood
dependent characteristics. Although a theory is avail- of obtaining a small, medium or large price per
able for this purpose (21, we need not go into such bushel.
detail in this paper. Clearly, to obtain the maximum profit, crop size
will have to be relatively large. However, eventually
4. AN EXAMPLE OF A MAXIMIZATION a very large harvest could be worthless, both because
PROBLEM the price is low and because the crop is subject to the
A calculus problem goes as follows: A farmer hazards of nature for a longer time.
wishes to sell his crop of 120 bushels of potatoes. If According to this observation, the pairwise com-
he sells it now he would get $1.00 per bushel. parisons among the different degrees of production
However, if he waits his crop will increase, due to can now be given by answering the following ques-
growth, by 20 bushels per week, but the price would tion: Comparing crop sizes, which one produces more
decrease by $0.10 per week. When should he sell to profit and how strongly? We use the following scale
maximize his profit? to make absolute comparisons as to how many times
The calculus answer to his problem is obtained by more one element dominates the other according to
maximizing (120 + 20x) (1 - 0. lx), whose solution is a common property. The judgment is first generated
x = 2 weeks. qualitatively and then the corresponding numerical
Exploring optimization 357

value is entered: Equal, 1; Moderate, 3; Strong, 5; The answer, a composite vector of all the priorities
Very Strong, 7; Extreme, 9: Intermediate Strengths, obtained so far, is interpreted as the likelihood of
2, 4, 6, 8; reciprocals of foregoing. selling in the given week for maximum profit:

Profit LOW Medium High Weights


Week number 0 1 2 3 4 5
Low 1 l/3 3 0.258
Likelihood of
Medium 1 5 0.637 selling: 0.183 0.160 0.226 0.214 0.120 0.097
High 1;3 l/5 I 0.105
Consistency ratio 0.033
The farmer has two options. The first is to select the
For brevity we will not give arguments to justify these week with the likelihood of higher yield which is the
judgements. The thoughtful reader should have no second week. Week 2 agrees exactly with the answer
difficulty in understanding these judgements, al- from calculus. The second option, which is perhaps
though he may not agree with all of them. It has been the better one because of the uncertainties, is to take
shown in dealing with hierarchic structures that the the expected value:
outcome is relatively stable to anything but the most
drastic change in the strategy of assigning judge- (0 x 0.183 + 1 x 0.160 + 2 x 0.226 + 3 x 0.0214
ments. + 4 x 0.120 + 5 x 0.097) = 2.22
Now compare the price per bushel for each yield
size. We answer the question: Given two category This answer recommends selling at week 2.22.
prices per bushel, which category is more likely to Had the question been concerned with price, the
obtain under that level of yield? We have: fourth level of the hierarchy would have considered
Low
different prices per bushel, as it also would, had the
S M L Weiehts
number of bushels (total yield) been the desired
s 1 l/3 115 0.109
outcome.
M 3 1 112 0.309
L 5 2 1 0.582
Consistency ratio 0.003
5. COMPLEX CHOICE
Medium S M L Weights
Intuition and understanding of a complex situation
S I l/5 l/2 0.128 should derive partly from numerical measurements
M 5 I 2 0.595
associated with that situation. Exactly how feelings
L 2 l/2 1 0.276
Consistency ratio 0.005 are shaped by numbers varies from individual to
individual. Optimization becomes a relative matter.
High S M L Weights Here is a simple illustration of how a professor of
S 1 4 5 0.691 mathematics, a friend of the writer, has explained
M l/4 I 1 0.160 how she allocates her money to shop for food. Several
L l/5 1 1 0.149 factors are involved, some of which are intangible
Consistency ratio 0.005
such as taste, on which one cannot place a bound as
It remains to compare the number of weeks in the in linear programming. Her hierarchy follows:
fourth level with respect to the price per bushel in the
third level. The question asked here is: of two sale
weeks which provides more profit at the indicated
level of price per bushel? We obtain the following
three paired comparison matrices and their eigen-
vector weights:
Small 0 I 2 3 4 5 Weights
0 1 I l/2 l/3 l/4 l/5 0.064
I 1 1 l/2 113 l/4 0.084
2 : 1 1 1 l/2 113 0.120
3 3 2 1 I I 112 0.173
4 4 3 2 I I 1 0.243
5 5 4 3 2 I I 0.315
Consistency ratio 0.015 Fig. 2. Hierarchy for shopping for food.
Medium 0 1 2 3 4 5 Weights
She assigned preference weights of 0.4 to meat and
0 1 113 115 l/5 l/3 1 0.053
I 3 1 l/3 l/3 I 2 0.122 0.6 to vegetables in the second level. Her matrices for
2 5 3 I 1 3 5 0.321 the elements in the third level with respect to those in
3 5 3 1 1 3 0.321 the second and those in the fourth with respect to
4 3 2 1 1 1 2 0.122
those in the third, follow. The meats and the veg-
5 1 l/2 I 1 1 I 0.060
Consistency ratio 0.008 etables listed in the fourth level are treated as two
separate clusters with respect to each attribute in the
Large 0 I 2 3 4 5 Weights third level.
0 I 3 4 5 6 I 0.426
I 1 3 4 6 0.254 Meat C T N CC Weinhts
2 1:2 I 1 3 : 5 0.150
3 l/3 l/2 I I 3 4 0.088 C I l/3 l/3 l/3 0.099
4 l/4 113 l/2 1 1 3 0.052 T 3 I 2 2 0.413
5 l/5 114 l/3 l/2 1 I 0.031 N 3 l/2 I 1 0.244
Consistency ratio 0.076 cc 3 l/2 1 1 0.244
358 THOMAS L. SAATY

Veg c T N CC Weights according to the priorities (e.g. by maximizing the


c I II5 II3 II3 0.079 product of the cost per unit of each food and its
T 5 I 2 3 0.477 priority together with the unknown amount to be
N 3 l/2 I 2 0.270
purchased) subject to these constraints. There are
cc 3 I13 l/2 I 0. I74
other possibilities for allocation mentioned in Saaty
Note that there are no comparisons necessary for cost and Kearns [3] investigated by the first author and
as actual store prices are available. J. P. Bennett. In the next section we illustrate this
The weights in the two matrices below were ob- kind of allocation subject to constraints.
tained by taking the reciprocals of the Per Serving
Store Price and normalizing it. We take the recip-
rocal of the price to measure relative benefits because
cheaper is better. 6. OPTIMIZATION SUBJECT TO
CONSTRAINTS
Per serving
cost store price Weights Consider an individual who has available three
B $1.04 0.152 types of food: meat (beef), bread, and vegetables
P $0.55 0.292 (broccoli). He must determine the optimal mixture of
F $0.995 0.161
PY 60.404 0.395 these foods to eat, minimizing cost and satisfying
A&B $0.99 0.088 minimum daily requirements for vitamins A, B, and
C&P $0. I26 0.694 the amount of calories intake.
s $0.402 0.218 Suppose that the cost of each food is as follows:
Taste
meat, $2.50/lb (%0.0055/g); bread, $055/lb
Taste B P F PY Weights ($0.0012/g); vegetable, $0.62/lb (%0.0014/g). The in-
B I 3 5 2 0.466 dividual knows, approximately, the amount of vita-
P l/3 I 3 I13 0.160 min A, B,, and calories per gram of each food. They
F l/5 l/3 I I13 0.08 are:
Taste A&B C&P S Weights Food Vit. A (I.U.) Vit. B, (mg) Calories (kcal)
A&B 1 4 3 0.623 Meat 0.3527/g 0.002 1/g 2.86/g
C&P l/4 1 II2 0.137 Bread 0.0000 0.0006 2.76
S l/3 2 I 0.24 Veg 25 0.002 0.25

Nutrition The minimum daily requirements are: Vitamin A:


N&t B P F PY Weights 7500 (I.U.), Vitamin B,: This amount varies from
B I 3 l/5 II3 0.125 individual to individual, and it is measured in mg/kg.
P l/3 I l/7 II4 0.061 The minimum requirement for an individual who
F 5 7 I 3 0.563 weighs 147 lb is 1.6338 mg.; Calories: 2050 kcal for
PY 3 4 l/3 I 0.251
the 147 lb individual.
Nutrit A&B C&P S Weights We have the linear programming problem
A&B I 3 II2 0.334 Minimize z: (5.5 x, + 1.2 x2 + 1.4 x3) x 10m3 dol-
C&P l/3 I l/3 0.142 lars
S 2 3 I 0.525
Subject to
Caloric content
0.3527 x, + + 25.000 x3 > 7.5000
CC B P F PY Weights
B 3 II4 II3 0.141 0.0021 x, + 0.0006 x2 + 0.02 x3 > 1.6338
P I)3 I II5 II4 0.071
F 4 5 I 3 0.52 2.86 x, +2.76 x,+ 0.25 x3 > 2.050
PY 3 4 l/3 I 0.268
XI, x2, x3 2 0

cc A&B C&P S Weights


Its solution is given by x, (meat) = 0, x2
A&B I 4 2 0.557 (bread) = 687.44 g, x3 (vegetables) = 610.67, at a cost
C&P l/4 I I13 0.123
I 0.32
per day of z = $1.67.
S I/2 3
Let us assume that an individual expresses his food
preferences in terms of the amount he desires to eat
On composing weights we find the following prior- per day as follows:
ities: A&B, 0.292; S, 0.197; F, 0.123; PY, 0.114; B,
x, (meat > 140 g (5 ozs)
0.108; C&P, 0.108; P, 0.051. Although the mathe-
matician considered this outcome as her optimal x2 (bread) < 56 g (2 slices)
choice, the problem developed as to how to allocate
x,(vegetables), no constraint
dollars to these categories of food. In general in
allocation problems, it may be that more dollars than What is wrong with this example? There is no
indicated by the highest priority are needed or that consideration of preference for the foods nor for the
the priority would lead to the allocation of too much importance of vitamins and calories to ones well
money beyond a certain point. What is a basis for being. They are all treated as equally important and
distributing a resource according to priority? One cost is the overriding factor. However, if both nour-
may sometimes need to identify minimum and max- ishment and cost are considered to be important but
imum amounts of each category of food to be used not equally, how would one incorporate such infor-
over a period, and allocate the money available mation in the objective function and in the con-
Exploring optimization 359

straints, maintaining a contrast or a relationship The following composite vector of priorities is ob-
between the two? tained by multiplying the above matrix by the
Let us now give an alternative approach to this weighted vector of nourishment and adding the re-
problem with the AHP. We have the following sulting vector to the vector of costs weighted by 0.25
hierarchy: which is the priority of cost.
Foods Meat Bmad Vegetables
Priorities 0.24 0.23 0.53
The total amount of food that an individual must eat
to minimize cost and satisfy the daily requirements
obtained earlier from the linear programming prob-
lem is x, + x, + x3 = 1298.11 g or 2.86 lb. If we take
this amount, which in practice may be high, and
distribute it as a total among the three foods accord-
ing to the above priorities the number of grams of
each food eaten per day is given by:
Fig. 3. Hierarchy for choosing a diet.
Foods Meat Bread Vegetables
Amount (g) 308.43 297.66 692.02
For a person with an average income, to satisfy his
nourishment needs is more important than the cost of The cost of this diet is $3.0224, however if we check
the food. Here we have for nourishment (N), and cost to see if the daily requirements are satisfied by
0: writing:

Daily
Meat Bread Veg Requirements
Vit. A 0.3527 0.0
Vit. B2 0.0021 0.0006
Cals 2.86 2.76

Daily requirements N C Weights we find that the caloric intake would not be met and
N I 3 0.75 hence a larger diet would be necessary. However, if
C l/3 1 0.25 we change the nourishment factors so they are
equally as important as cost, consumption would be
Vitamin A and B, are considered to be equally
290.80 g of meat, 419.30 g of bread and 588.50 g of
important, but both are more important than cal-
vegetables at a cost of $2.92 and the daily require-
ories. We have
ments would be met because we have:
Nourishment Vit. A Vit. B, Gals Weights
Vit. A 14,815.08
Vit. A I 1 2 0.4
Vit. B, 1 2 0.4 Vit. Bz
Cals 112 112 1 0.2 kcal 213:::
The amount of Vitamin A, B2 and calories per By way of sensitivity analysis we change the rela-
gram of meat, bread and vegetables are obtained tive importance of cost and nourishment so that cost
from the original data and then normalized in each is considered to be extremely more important than
column. We have: nourishment. We have:
Nourishment Cost Weights
Vit. A Vit. B, kcal
N 1 119 0.1
Meat 0.0139 0.4468 0.4872
C 9 1 0.9
Bread 0.0000 0.1277 0.4702
Vegetables 0.9861 0.4255 0.0426 If we also assume that Vit. A, Vit. Bz and kcal are
According to cost we have: equally important, we obtain:
Cost Foods Meat Bread Vegetables
Priorities 0.12 0.46 0.42
Meat 0.68
Grams of food
Bread 0.15 598.04 543.13
eaten per day 157.07
Vegetables 0.17
with the following contribution to daily require-
However, since we wish to minimize cost, we use the ments:
normalized reciprocals of the amounts given above. Vit. A 13,633.65
We have: Vit. B, 1.7749
kcal 2235.59
Cost-
Meat 0.105 and again the requirements are met at a cost of $2.34.
Bread 0.476 One conclusion that can be drawn from the com-
Vegetables 0.419 parison of the two approaches is that as we impose
360 THOMASL. SAATY

a preference structure on the diet problem the min- hierarchic model and the results make at least as
imum cost obtained from linear programming goes much sense, and are justified on rigorous mathe-
up from $1.67 to $3.80. The costs obtained by using matical ground, as those obtained in traditional
the Analytic Hierarchy Process fell between these two optimization. However, they have the advantage of
values as is to be expected since this formulation greater realism, by relying on more data to derive
includes additional variables. Another conclusion is their results and by avoiding elegant manipulations
that the AHP solution was obtained directly by of quantities which give the impression that the
expressing preferences and without the intervention solution is related to abstract operations that give the
of a complex algebraic structure that is forbidding to appearance of a magicians wand in producing an
the layman. Still the AHP has not been tested for a answer. In sum, our ability to handle more complex
large number of variables to determine its feasibility problems with multiple objectives, and diverse and
as a model in this case. intangible criteria depends on our flexibility to con-
sider adopting new methods without prejudice in
7. CONCLUSION
favor of old ones merely because of our habituation
and commitment.
The Analytic Hierarchy Process is a multi-criterion
decision method that is based on ratio scales and
relative comparisons. The success of its use rests with REFERENCES
the ability of the decision maker to express his 1. T. L. Saaty. Optimization in Integers and Related Ex-
preferences within an accessible hierarchic structure. tremal Problems. McGraw-Hill, New York (1970).
The outcome of the process is a best or optimum 2. T. L. Saaty. The Analytic Hierarchy Process. McGraw-
choice, considering all the necessary objectives, crite- Hill, New York (1980).
ria and alternatives in a single interrelated frame- 3. T. L. Saaty and K. P. Kearns. Analytical Planning.
work. Optimization problems can be represented in a Pergamon Press, Oxford (1985).

Вам также может понравиться