Sie sind auf Seite 1von 13

WHAT IS THE ANALYTIC HIERARCHY PROCESS?

Thomas L Saaty
University of Pittsburgh
1. Introduction
In our everyday life, we must constantly make choices
concerning what tasks to do or not to do, when to do them, and
whether to do them at all.
Many problems such as buying the most cost effective home
computer expansion, a car, or house; choosing a school or a
career, investing money, deciding on a vacation place, or even
voting for a political candidate are common everyday problems
in personal decision making. Other problems can occur in
business decisions such as buying equipment, marketing a
product, assigning management personnel, deciding on inventory
levels or the best source for borrowing funds. There are also
local and national governmental decisions like whether to act
or not to act on an issue, such as building a bridge or a
hospital, how to allocate funds within a department or how to
vote on a city council issue.
All these are essentially problems of choice. In addition they
are complex problems of choice. They also involve making a
logical decision. The human mind is not capable of considering
all the factors and their effects simultaneously. People solve
these problems today with seat of the pants judgments or by
mathematical models based on assumptions not readily verifiable
that draw conclusions that may not be clearly useful.
Typically individuals make these choices on a reactive and
frequently unplanned basis with little forethought of how the
decisions tie together to form one integrated plan. This whole
process of deciding what, when, and whether to do certain tasks
is the crux of this process of setting priorities. The
priorities may be long-term or short-term, simple or complex.
NATO AS! Series, Vol. F48
Mathematical Models for Decision Support
Edited by G. Mitra
Springer-Verlag Berlin Heidelberg 1988
110
some organization is needea. This organization can be obtained
through a hierarchical representation. But that is not all.
Judgments and measurements have to be included and integrated.
A procedure which satisfies these requirements is the Analytic
Hierarchy Process (AHP). The mathematical thinking behind the
process is based on linear algebra. Until recently its
connection to decision making was not adequately studied. With
the introduction of home computers basic linear algebra
problems can be solved easily so that it is now possible to use
the AHP on personal computers. The AHP differs from
conventional decision analysis techniques by requiring that its
numerical approach to priorities conform with scientific
measurement. By this we mean that if appropriate scientific
experiments are carried out using the scale of the AHP for
paired comparisons, the scale derived from these should yield
relative values that are the same or close to what the physical
law underlying the experiment dictates according to known
measurements in that area. The Analytic Hierarchy Process is
of particular value when subjective, abstract or
nonquantifiable criteria are involved in the decision.
With the AHP we have a means of identifying the relevant facts
and the interrelationships that exist. Logic plays a role but
not to the extent of breaking down a complex problem and
determining relationships through a deductive process. For
example, logic says that if I prefer A over Band B over C,
then I must prefer A over C. This is not necessarily so
(consider the example of soccer team A beating soccer team B,
soccer team B beating soccer team C and then C turning around
and beating A, and not only that, the odds makers may well have
given the advantage to C prior to the contest based on the
overall records of all three teams) and the AHP allows such
inconsistencies in its framework.
A basic premise of the AHP is its reliance on the concept that
much of what we consider to be "knowledge" actually pertains to
our instinctive sense of the way things really are. This would
seem to agree with Descartes' position that the mind itself is
the first knowable principle. The AHP therefore takes as its
premise the idea that it is our conception of reality that is
crucial and not our conventional representations of that
reality by such means as statistics, etc. With the AHP it is
possible for practitioners to assign numerical values to what
are essentially abstract concepts and then deduce from these
values decisions to apply in the global framework.
The Analytic Hierarchy Process is a decision making model that
aids us in making decisions in our complex world. It is a
three part process which includes identifying and organizing
decision objectives, criteria, constraints and alternatives
into a hierarchy; evaluating pairwise comparisons between the
relevant elements at each level of the hierarchy; and the
synthesis using the solution algorithm of the results of the
pairwise comparisons over all the levels. Further, the
algorithm result gives the relative importance of alternative
courses of action.
111
To summarize, the AHP process has eight major uses. It allows
the decision maker to: 1) design a form that represents a
complex problem; 2) measure priorities and choose among
alternatives; 3) measure consistency; 4) predict; 5) formulate
a cost/benefit analysis; 6) design forward/backward planning;
7) analyze conflict resolution; 8) develop resource allocation
from the cost/benefit analysis.
For the pairwise comparison judgments a scale of 1 to 9 is
utilized. This is not simply an assignment of numbers. The
relative intensity of the elements being compared with respect
to a particular property becomes critical. The numbers
indicate the strength of preference for one over the other.
Ideally, when the pairwise comparison process is begun,
numerical values should not be assigned, rather the comparative
strengths should be verbalized as indicated in the table below
of the fundamental scale of relative importance that is the
basis for the AHP judgments.
2. The Scale
Pairwise comparisons are fundamental in the use of the AHP.
We must first establish priorities for the main criteria by
judging them in pairs for their relative importance, thus
generating a pairwise comparison matrix. Judgments used to
make the comparisons are represented by numbers taken from the
fundamental scale below. The number of judgments needed fQr a
particular matrix of order n, the number of elements being
compared, is n(n-1)/2 because it is reciprocal and the diagonal
elements are equal to unity. There are conditions under which
it is possible to use fewer judgments and still obtain accurate
results. The comparisons are made by asking how much more
important the element on the left of the matrix is perceived to
be with respect to the property in question than the element on
the top of the matrix. It is important to formulate the right
question to get the right answer.
The scale given below can be validated for its superiority over
any other assignment of numbers to judgments by taking one of
the two illustrative matrices given below and inserting instead
of the numbers given numbers from another scale that is not
simply a small perturbation or constant multiple of our scale.
It will be found that the resulting derived scale is markedly
different and does not correspond to the known result.
112
TABLE 1
THE FUNDAMENTAL SCALE
Intensity of
Importance
on an Absolute
Scale Definition
1 Equal importance.
3 Moderate importance
of one over another.
5 Essential or strong
importance.
7 Very strong impor-
tance.
9 Extreme importance.
2,4,6,8 Intermediate values
between the two ad-
jacent judgments.
Reciprocals
Rationals
If activity i has
one of the above
numbers assigned to
it when compared
with activity j,
then j has the
reciprocal value
when compared with i.
Ratios arising from
the scale.
Explanation
Two activities
contribute equally to
the objective.
Experience and judgment
strongly favor one
activity over another.
Experience and judgment
strongly favor one
activity over another.
An activity is strongly
favored and its
dominance demonstrated
in practice.
The evidence favoring
one activity over
another is of the
highest possible order
of affirmation.
When compromise is
needed.
If consistency were to
be forced by obtaining n
numerical values to span
the matrix.
When the elements being compared are closer together than
indicated by the scale, one can use the scale 1.1, 1.2, ... , 1.9.
If still finer, one can use the appropriate percentage refinement.
113
There are various ways of carrying out measurement and in
particular pairwise comparisons. When the elements being
compared share a measurable property such as weight, they can be
measured directly on an absolute scale and pairwise comparisons
become unnecessary. However if one were to use the AHP and forms
ratios of these weights (resulting in a consistent matrix) then
solving for the principal eigenvector, gives the same result
obtained by normalizing the numbers. It is rare that numbers
should be used in this way. It is almost always the case that
measurements only represent some kind of arithmetical accuracy
which does not reflect the actual judgment or value that an
individual would assign to the numbers to reflect the
satisfaction of his needs. In some situations habituation and
familiarity may cause people to use the numbers as they are. It
must be understood that they should be able to justify how these
numbers correspond to their own judgment of the relative
intensity of importance.
There are times when people use their judgments to estimate
numerical magnitudes. They should do so by comparing them.
Otherwise, numbers or ranges of numbers can be arranged into
intensity equivalence classes and then compared either by
directly using representative numbers from each range, or
indirectly by comparing them qualitatively according to
intensity. For example, numbers in the millions, billions or
trillions may mean the same thing to an individual who is either
unfamiliar with very large numbers or does not know what these
values apply to or if known, there may still be no way to
incorporate in one's judgment or understanding the significance
of their magnitudes. Here we need to remember that the reason
for a hierarchic structure is to decompose complexity in stages
to enable us to scale its smallest elements gradually upwards in
terms of its largest elements. If it were possible to assign
meaningful numerical values to the smallest elements, there would
be no need for the elaborate process of careful decomposition.
Thus the question remains as to how responsive can individual
judgments be to make it possible to discriminate between elements
sharing a property or properties sharing a higher property. One
of the axioms of the AHP relates to how disparate the elements
are allowed to be.
In making paired comparisons,
hence also the derived scale
elements according to rank in
addition one may compare the
one first, or alternatively,
basis of the comparisons.
the accuracy of the judgments and
may be improved by ordering the
descending or ascending order. In
largest element with the smallest
take the middle element as the
One problem then in teaching people to use the AHP is first to
specify qualitative intensities of judgment and feeling that
facilitate spontaneous response without elaborate prior training
and second to associate appropriate linguistic designations with
114
these expressions. Finally, numerical scale values must in turn
be associated with these verbal expressions that lead to
meaningful outcomes and particularly in known situations can be
tested for their accuracy. Small changes in the words (or the
numbers) should lead to small changes in the derived answer.
Finally we use consistency arguments along with the well-known
work of Fechner in psychophysics to derive and substantiate the
scale and its range.
In 1860 Fechner
increasing stimuli.
considered a sequence of just
He denotes the first one by s .
o
just noticeable stimulus by
s = s + t,s = s
1 0 0 0
s (l+r)
o
having used Weber's law.
Similarly
2 2
s = s +t,s = s (l+r) = s (l+r) =s
2 1 1 1 0 0
In general
n
s = s = s (n = 0,1,2, ... )
n n-1 0
noticeable
The next
Thus stimuli of noticeable differences follow sequentially in a
geometric progression. Fechner felt that the
sensations should follow each other in an arithmetic sequence
occurring at the discrete points at which just noticeable
differences occur. But the latter are obtained when we solve for
n. We have
n = (log s - log s )/log
n 0
and sensation is a linear function of the logarithm of the
stimulus. Thus if M denotes the sensation and s the stimulus,
the psychophysical law of Weber-Fechner is given by
M = a log s + b, a -; 0
We assume that the stimuli arise in making pairwise
comparisons
interested
of ratios.
of relatively comparable activities. We are
in responses whose numberical values are in the form
Thus b = 0, from which we must have log s = 0 or s
a
The next
a
= 1, which is possible by calibrating a unit stimulus.
noticeable response is due to the stimulus
s =s a=a
1 0
This yields a response log / log = 1. The next stimulus is
2
s = s a
2 a
which yields a response of 2. In this manner we
sequence 1,2,3,.... For the purpose of consistency
obtain the
we place the
115
activities in a cluster whose pairwise comparison stimuli give
rise to responses whose numerical values are of the same order of
magnitude. In practice, qualitative differences in response to
stimuli are few. Roughly, there are five distinct ones as listed
above with additional ones that are compromises between adjacent
responses. The notion of compromise is particularly observable
in the thinking judgmental process as opposed to the senses.
This brings the total up to nine which is compatible with the
order of magnitude assumption made earlier.
Now we examine the impact of consistency on scaling. The
conscious mind absorbs new ideas by contrasting them through
scanning or through concentration and analysis to understand how
they are similar to familiar ideas. They are also related to
current or future activities and applied to concrete situations
to test their compatibility with what is already believed to be
workable.
The ideas may be accepted as a consistent part of the existing
understanding or they may be inconsistent with what is already
known or accepted. In that case the system of understanding and
practice is extended and adjusted to include the new ideas.
Growth implies such expansion. If the adjustment of old ideas to
accommodate a new one is drastic, then the inconsistency caused
by the new idea is great and may require considerable adjustments
in the old ideas and beliefs whose old relations may no longer be
intuitively recognizable and may call for a new theory or
interpretation if at all possible. But such major changes cannot
be made every hour, every day or even every week because it takes
time to interpret and assimilate relations. Thus inconsistency
from exposure to new ideas or situations can be
threatening, unsettling and painful. Our biology has recognized
this and developed for us ways to filter information in such a
way that we usually make minor adjustments in what we already
know when we are exposed to a new idea or better that we absorb
new ideas by interpreting them from the vantage point of already
established relations. Thus our emphasis on consistency exceeds
our desire for exposure and readjustment. As a result
maintaining consistency is considered to be a higher priority of
importance than changing. Yet the latter is considered to be also
an important concern. One conclusion is that our preoccupation
with consistency differs by one order of magnitude from our
preoccupation with inconsistency - a split of 90% and 10%.
In addition, in order to maintain their identity, the
significance of old ideas must be visibly greater than the
adjustments we bring into them because of exposure to the new.
In other words, the 90% effort to maintain consistency can at
best be divided among a few entities (at most 9) each of whch
would receive emphasis or priority of the order of 10% in the
understanding so that slight readjustment would not change the
old relations significantly. If one were to compare
homogeneous objects one would not need a scale whose range
extends beyond 9.
116
Before moving on to elaborate decision examples let us
illustrate the use of the scale in two elementary examples.
These are neither decision nor hierarchical examples. Those
will come later below. These examples are to demonstrate that
the AHP scheme for assigning measures works. It will be seen
that even though a person may not have an idea of the final
numerical value by making the comparisons according to his
everyday knowledge the derived answer appears to be very
reasonable. The first illustration below asks the individual
to provide judgments as to which of the seven drinks listed in
the matrix is consumed more in the United States, and how
strongly more. In the second example he is asked to compare
seven foods according to his idea of the relative amount of
protein in each. In both cases what is desired is the relative
percentage that each has as part of the total for the seven
drinks or the seven foods. In both cases we have given the
estimate and the actual results taken from statistical
sources. What is important here is that the judgments were
provided by very average people who had no idea of the true
answer.
Drink
Consumption
in the u.s. Coffee Wine Tea Beer Sodas Milk Water
Coffee 1 9 5 2 1 1 1/2
Wine 1 1/3 1/9 1/9 1/9 1/9
Tea 1 1/3 1/4 1/3 1/9
Beer
1 1/2 1 1/3
Sodas
1 2 1/2
Milk 1 1/3
Water 1
Note that the lower triangular half of this matrix is not
given. Those entries are the reciprocals of the entries in the
upper triangular half and it is not necessary to show them,
although of course they enter in the calculations. The derived
scale representing the priorities (or relative values) is
obtained by solving for the principal eigenvector of .the
eigenvalue problem Aw = A w where A is the matrix of judgments.
max
The derived scale based on the judgments in the matrix is:
Coffee Wine Tea Beer Sodas Milk Water
.177 .019 .042 .116 .190 .129 .327
with a consistency ratio of .022.
117
The actual consumption (from statistical sources) is:
.180 .010 .040 .120 .180 .140 .330
In the second example an individual gives judgments as to the
relative amount of protein in each food.
Protein in
Foods
A B C D E F G
A: Steak 1 9 9 6 4 5 1
B: Potatoes 1 1 1/2 1/4 1/3 1/4
C: Apples 1 1/3 1/3 1/5 1/9
D: Soybean 1 1/2 1 1/6
E: Whole Wheat 1 3 1/3
Bread
F: Tasty Cake 1 1/5
G: Fish 1
Here the derived scale and actual values are:
Steak Potatoes Apples Soybean Whole Wheat Tasty Fish
Bread Cake
.345 .031 .030 .065 .124 .078 .328
.370 .040 .000 .070 .110 .090 .320
with a consistency ratio of .028.
3. Absolute and Relative Measurement
cognitive psychologists [1] have recognized for some time
that there are two kinds of comparisons, absolute and relative.
In the former an alternative is compared with a standard in
memory developed through experience; in the latter alternatives
are compared in pairs according to a common attribute. The AHP
has been used to carry out both types of comparisons resulting
in ratio scales of measurement. We call the scales derived
from absolute and relative comparisons respectively absolute
and relative measurement scales. Both relative and absolute
measurement are included in the IBM PC compatible software
package Expert Choice [2].
Let us note that relative measurement is usually needed to
compare criteria in all problems particularly when intangible
ones are involved. Absolute measurement is applied to rank the
alternatives in terms of the criteria or rather in terms of
ratings or intensities of the criteria such as excellent, very
good, good, average, below average, poor and very poor. After
setting priorities on the criteria (or subcriteria, if there
are some) pairwise comparisons are also performed on the
ratings which may be different for each criterion (or
subcriterion). An alternative is evaluated, scored or ranked
by identifying for each criterion or subcriterion, the relevant
118
rating which describes that alternatives best. Finally, the
weighted or global priorities of the ratings, one under each
criterion corresponding to the alternative, are added to
produce a ratio scale score for that alternative. If desired,
in the end, the scores of all the alternatives may be
normalized to unity.
Absolute measurement needs standards, often set by society for
convenience, and sometimes has little to do with the values and
objectives of a judge making comparisons. In completely new
decision problems or old problems where no standards have been
established, we must use relative measurement to identify the
best one among the alternatives by comparing them in pairs.
It is clear that with absolute measurement there can be no
reversal in the rank of the alternatives if a new alternative
is added or another one deleted. This is desirable when the
importance of the criteria, although independent from the
alternatives according to function, meaning or context does not
depend on their number and on their priorities as it does in
relative measurement. In the latter if for example, the
students in a certain school perform badly on intelligence
tests, the priority of intelligence which is an important
criterion used to judge students, may be rescaled by dividing
by the sum of the paired comparison value of the students, a
transformation carried out through normalization. The priority
of each other criterion is also rescaled according to the
performance of the students. Thus the criteria weights are
affected by the weights of the alternatives.
It is worth noting that although rank may change when using
relative measurement with respect to several criteria, it does
not change when only one criterion is used and the judgments
are consistent. It can never happen that an apple which is
more red than another apple becomes less red than that apple on
introducing a third apple in the comparisons. It would be
counter-intuitive were that to happen. However, when judging
apples on several criteria, each time a new apple is
introduced, a criterion that is concerned with the number of
apples being compared changes as does another criterion
concerned with the actual comparisons of the apples. Such
criteria are called structural. The way they participate in
generating the final weights differs from the traditional way
in which the other criteria, called functional, do [4].
Let us now illustrate both types of measurement in decision
making.
4. Examples of Relative and Absolute Measurement
Relative Measurement:Reagan's Decision to Veto the Highway Bill
Before President Reagan vetoed the Highway Bill, a high budget
bill to repair roads and provide jobs in the economy in the
U.S., we predicted that he would veto it by constructing two
hierarchies, one to measure the benefits and the other the
costs of the possible alternatives of the decision and taking
that with the highest benefit to cost ratio. The public image
119
of Ronald Reagan had been tarnished since the "Iran-Contra"
affair. His leadership abilities for the remainder of his term
were being questioned. He looked at this Bill (a very popular
bill with the public) as an opportunity to improve his own and
his party's image. We imagined that he subconsciously went
through something organized along the lines shown here to make
the decision. Because of space limitation we will simply
present the hierarchies and the results.
BENEFITS HIERARCHY
Goal: Best Decision
Criteria: Political
Image
Efficiency Employ-
ment
Convenience
( .634) ( .157)
Alternatives: Veto for Modification
( .607)
( .152)
Sign Bill
( .393)
( .057)
In the benefits hierarchy Mr. Reagan's political image has by
far the greatest weight followed by employment, the creation of
800,000 jobs if the bill were to pass.
Goal:
Criteria: Political
Image
( .524)
COSTS HIERARCHY
Best Decision
Energy
( .062)
Construc-
tion
( .312)
Alternatives: Veto for Modification
( .518)
Maint-
enance
(.041)
Sign Bill
( .482)
Safety
( .061)
In the costs hierarchy Mr Reagan has his political image
foremost on his mind with the construction cost of $86 billion
a fairly distant second. The benefit to cost ratios of the
alternatives from the two hierarchies are: Veto Bill for
Modification .607/.518 = 1.172, Sign Bill .393/.482 = .815.
Thus Veto for Modification was the preferred outcome. In order
to modify the bill, President Reagan vetoed it, then sent it
back to Congress with a strong message to them to modify it and
resubmit.
One thing we need to mention here is that the two alternatives
must be compared in a separate matrix with respect to each
criterion and the resulting derived scales are each weighted by
120
the priority of the corresponding criterion and summed for each
alternative to obtain the overall priority shown for that
alternative. For example, in the benefits hierarchy the
derived scales and the weights of the criteria may be arranged
and composed as follows.
Political Efficiency Employ- Convenience Composite
Image ment Weights
(.634) (.157) (.152) (.057)
Modify .89 .10 .10 .20 .607
Bill
Sign .11 .90 .90 .80 .393
Bill
Note for example that the composite weight:
.607 = .89x .634 +.10 x .157 + .10 x .152 + .20 x .057
A final comment here is that the AHP has a more elaborate
framework to deal with dependence within a level of a hierarchy
or between levels [3], but we will not go into such details
here.
Absolute Measurement:Employee Evaluation
The problem is to evaluate employee performance. The hierarchy
for the evaluation and the priorities derived through paired
comparisons is shown below. It is then followed by rating each
employee for the quality of his performance under each
criterion and summing the resulting scores to obtain his
overall rating. The same approach can be used for student
admissions, giving salary raises, etc. The hierarchy can be
more elaborate, including subcriteria, followed by the
intensities for expressing quality.
121
Goal:
Employee Performance Evaluation
Criteria:
Intens-
ities:
Tech-
nical
( .061)
Excello
( .604)
Abv.Avg.
( . 24S)
Average
(.10S)
Bel. Av.
( .046)
Alternatives:
Maturity
(.196)
Very
( .731)
Accep.
(.188)
Immat.
( .081)
1) Mr. X Excell Very
2) Ms. Y Average Very
3) Mr. Z Excell Immat.
Writing
skills
( .043)
Excello
( .733)
Average
( .199)
Poor
( .068)
Average
Average
Average
Verbal
skills
( .071)
Excello
(.7S0)
Average
(.171)
Poor
( .078)
Timely
work
(.162)
Nofollup
(.731)
OnTime
(.188)
Remind
( .081)
Potential
(personal)
(.466)
Great
( .7 SO)
Average
(.171)
Bel.Av.
(.078)
Excell. OnTime Great
Average Nofollup Average
Excell. Remind Great
Let us now show how to obtain the total score for Mr. X:
.061 x .604 + .196 x .731 + .043 x .199 + .071 x .7S0 +
.162 x .188 + .466 x .7S0 = .623
Similarly the scores for Ms. Y and Mr. Z can be shown to be
.369 and .478 respectively.
It is clear that we can rank any number of candidates along
these lines.
REFERENCES
1. Blumenthal, A.L., The Process of Cognition, Prentice-Hall,
Englewood Cliffs, 1977.
2. Expert Choice, Software Package for IBM PC, Decision
Support Software, 1300 Vincent Place, McLean, VA 22101.
3. Saaty, Thomas L., The Analytic Hierarchy Process, McGraw-
Hill, 1981.
4. Saaty, Thomas L., "Rank Generation, Preservation, and
Reversal in the Analytic Hierarchy Decision Process",
Decision Sciences, Vol. 18, No.2, Spring 1987.

Das könnte Ihnen auch gefallen