European Journal of Operational Research 48 (1990) 9-26 9

North-Holland

How to make a decision:

The Analytic Hierarchy Process
Thomas L. Saaty
Joseph M. Katz Graduate School of Business, University of Pittsburgh, Pittsburgh, PA 15260, USA

Abstract: This paper serves as an introduction to the Analytic Hierarchy Process - A multicriteria decision
making approach in which factors are arranged in a hierarchic structure. The principles and the philosophy
of the theory are summarized giving general background information of the type of measurement utilized,
its properties and applications.

Keywords: Decision, priority, rank, cost-benefit, scales, ratios

1. How to structure a decision problem One certainly cannot compare according to size a
football with Mt. Everest and have any hope of
Perhaps the most creative task in making a getting a meaningful answer. The football and Mt.
decision is to choose the factors that are important Everest must be compared in sets of objects of
for that decision. In the Analytic Hierarchy Proc- their class. Later we give a fundamental scale of
ess we arrange these factors, once selected, in a use in making the comparison. It consists of verbal
hierarchic structure descending from an overall judgments ranging from equal to extreme (equal,
goal to criteria, subcriteria and alternatives in moderately more, strongly more, very strongly
successive levels. more, extremely more) corresponding to the verbal
To a person unfamiliar with the subject there judgments are the numerical judgments (1, 3, 5, 7,
may be some concern about what to include and 9) and compromises between these values. We
where to include it. When constructing hierarchies have completed compiling a dictionary of hier-
one must include enough relevant detail to: archies pertaining to all sorts of problems, from
represent the problem as thoroughly as possi- personal to corporate to public.
ble, but not so thoroughly as to lose sensitivity to A hierarchy does not need to be complete, that
change in the elements; is, an element in a given level does not have to
consider the environment surrounding the function as an attribute (or criterion) for all the
problem; elements in the level below. A hierarchy is not the
identify the issues or attributes that contribute traditional decision tree. Each level may represent
to the solution; and a different cut at the problem. One level may
identify the participants associated with the represent social factors and another political fac-
problem. tor to be evaluated in terms of the social factors or
Arranging the goals, attributes, issues, and stake- vice versa. Further, a decision maker can insert or
holders in a hierarchy serves two purposes. It eliminate levels and elements as necessary to
provides an overall view of the complex relation- clarify the task of setting priorities or to sharpen
ships inherent in the situation; and helps the deci- the focus on one or more parts of the system.
sion maker assess whether the issues in each level Elements that have a global character can be
are of the same order of magnitude, so he can represented at the higher levels of the hierarchy,
compare such homogeneous elements accurately. others that specifically characterize the problem at
hand can be developed in greater depth. The task
Received November 1989 of setting priorities requires that the criteria, the

0377-2217 j90j$03.50 © 1990- Elsevier Science Publishers B.V. (North-Holland)

10 T.L. Saaty /The AHP: How to make a decision

properties or features of the alternatives being wheeling and dealing, it is essential that the deci-
compared, and the alternatives themselves are sion theory itself used to assist the group in arriv-
gradually layered in the hierarchy so that it is ing at a decision be invariant to politics and
meaningful to compare them among themselves in behavior. It should be a science of scaling based
relation to the elements of the next higher level. on mathematics, philosophy and psychology.
Finally, after judgments have been made on the Among the various number crunching prac-
impact of all the elements and priorities have been tices, the most objectionable one is to assign any
computed for the hierarchy as a whole, sometimes, set of numbers to judgments on alternatives under
and with care, the less important elements can be a particular criterion, and then normalize these
dropped from further consideration because of numbers (by multiplying them by a constant that
their relatively small impact on the overall objec- is the reciprocal of their sum). Generally, different
tive. The priorities can then be recomputed sets of numbers are used to scale the judgments
throughout, either with or without changing the for the alternatives under different criteria. All the
remaining judgments. new normalized sets now lie in the interval [0, 1],
no matter what scale they originally came from,
and can be passed off to the uninitiated as com-
2. Scales of measurement- Avoiding mere num- parable. The appealing part of this practice is that
ber crunching the numbers have an apparently uniform underly-
ing structure and look not unlike probabilities.
One might argue the whole process of decision Thus they go unchallenged by the decision maker
making is so unstructured and so amorphous that and are then manipulated by the consultant or
it is no use trying to be precise. One is then facilitator who weights and adds them to find the
tempted to go further and conclude that "par- most preferred alternative.
ticipant satisfaction" is the main objective of deci- Note that when the initial numbers assigned
sion making. Were this the case, then multicriteria before normalization are ordinals, arbitrary num-
decision making would be a simple matter of bers that preserve order but carry no information
using ingenuity, supported with mathematical about differences or ratios of relative magnitudes,
terminology, to improve numbers that please peo- the resulting transformation produces a new set of
ple. But different sets of arbitrary numbers are ordinals lying between zero and one and you can
likely to result, producing different decisions, and be sure only that it preserves the same order. The
we are right back where we started. One set of operations of weighting and adding cannot then
numbers pleases a group of people who might be be meaningful because it is very likely that differ-
equally pleased with another set of numbers that ent results will be produced for different choices
contradicts the recommendations of the first set. of the ordinal numbers. By a judicious choice of
This is mere number crunching. If a decision ordinals one can make an alternative that is domi-
support theory is to be trustworthy there must be nant on even one criterion, no matter how unim-
uniqueness in the representation of judgments, the portant that criterion may be, have the largest
scales derived from these judgments, and the scales value after weighting and adding and thus turn
synthesized from the derived scales. out to be the most preferred. Any method such as
Let us consider for a moment group interaction this is not a decision theory but an approach that
that often leads to certain expectations, sometimes can mislead people.
arising from the very heat of the debate. Such a The Analytic Hierarchy Process is rigorously
debate cannot always incorporate in its arguments concerned with the scaling problem and what sort
the refinements resulting from the mathematical of numbers to use, and how to correctly combine
tradeoffs to ensure drawing valid conclusions. the priorities resulting from them. A scale of
Sometimes participants accept a process and its measurement consists of three elements: A set of
outcome because the situation is so complex and objects, a set of numbers, and a mapping of the
the arguments so convoluted that whatever objects to the numbers. In a standard scale a unit
surfaces in the end appears plausible. Although is used to construct the rest of the numbers of the
convincing a group about its qualitative prefer- scale. Examples of such a unit are the inch, the
ences involves the politics of persuasion and of pound, the angstrom, and the dollar. A standard
 T.L. Saaty /The AHP: How to make a decision 11

scale can be used to measure object or events with again. Our values of comfort and desirability and
respect to the property for which a scale is design- other social effects have to be at the bottom of
ed to measure. Since the unit is arbitrary, one can every interpretation and depend on higher goals
have different numbers to which the objects are that we may have in mind.
mapped. Because a standard scale is not unique, it For a large number of scales used in physics, it
is important to interpret the meaning of the num- is implicitly assumed that the scale can be ex-
bers used in the scale. Thus, in general, the num- tended out to infinity and applied to every imagin-
bers obtained from such a scale are merely stimuli able circumstance. In other words. interpretation
for the memory (what it felt like the last time the in physics assumes events as homogeneous. no
temperature was -15 o C) and have no intrinsic matter how near or far from the origin they may
significance. However, most carefully designed fall. What is most astonishing is the assumption in
standard scales are helpful in that they preserve physics that objects yet unknown but with the
certain numerical relations in the measurements same dimensional characteristics of the known
(the mapping) of the objects, giving us a better objects being measured can in fact be measured in
way to interpret the stimuli they are measuring the same way. Realization of the different inter-
than arbitrary scales. pretations we can make of the same number in
A scale may or may not have zero for an origin. physics would indicate that when numbers fall
For example, a scale of ordinal numbers can begin outside the realm of experience it is logical to
with any number. A ratio scale, such as absolute suspend the extension of the truth we construct
temperature preserves origin. Interval scales, which from experience to a domain for which we have no
measure the same phenomena like temperature, knowledge and feeling. It would be mostly fictive
preserve linear relations, but may have different speculation.
origins. Zero on the Fahrenheit scale is a different In economics. the arithmetic value of a dollar is
temperature than zero on the Celsius scale. Both assumed to be the same no matter whether a
are interval scales. Again, the numbers on these person has only one or a million dollars. But in
scales mean nothing unless one can recall situa- reality it is not. To buy a new Mercedes, ten
tions associated with the numerical readings being dollars and one hundred dollars are nearly equally
considered. They are just a convenient means of inadequate or useless as down payments. On the
communicating characteristics of objects or situa- other hand. for buying groceries, a hundred dol-
tions without everybody having to experience lars is much more useful. practically ten times
them. more useful than ten dollars. The first thousand
It often happens that the interpretations of dollars earned is much more important than the
numerical stimuli from a given standard scale first thousand dollars earned after a million.
differ depending on the circumstances. There is no We must be constantly and carefully attentive
simple rule that can be applied to interpret read- to how we interpret data from scales. Standard
ings from even a single scale when it is applied to scales force on us a way of thinking that is not in
a natural phenomenon. Intensity of sun light has a complete harmony with the way we really feel
different significance for different purposes. It about what they are measuring.
may be useful for sunbathing, but too bright for There is a more general method of measure-
reading. Similarly, a monotonic relation between ment that does not make use of standard scales. It
successive readings from a standard scale do not is the method of relative measurements useful for
assure us that even higher readings will be better properties for which there is no standard scale of
(or worse). More (or less) temperature does not measurement (love. political clout. straightness).
necessarily correspond to more (or less) useful- These are known as intangible properties. The
ness. Low temperatures are uncomfortable. As the number of such properties is extremely large. We
readings rise, they become more comfortable and can scarcely hope to device standard scales for
as they rise higher they can again become uncom- them all. We are driven to relative scales. and a
fortable. On the other hand, to preserve some surprising thing is that they can serve as a stand-
foods, low temperatures are very desirable and as ard for how to handle the very few standard scales
the readings rise, they become undesirable, and as we have. and not the other way round. A remarka-
they rise still higher they could become desirable ble aspect of relative scales is that they can use
12 T.L Saaty / The AHP: How to make a decision

information from standard scales when there is a of each pair of stones is used as the unit, and the
particular need to do so. Measurements in a larger one is measured in terms of multiples of
standard ratio scale are transformed to meas- that unit. It is difficult to do the inverse compari-
urements in a relative ratio scale by normalizing son without again using the smaller stone as the
them. unit. This is a sort of bias in human thinking,
This conversion process gives us a hint about which leads to considering a nonsymmetrical out-
the difference between the two kinds of scales. A come and the inclination not to force symmetry
relative scale for a property is generated for a on it. We have the matrix equation:
specific set of entities or objects. A standard scale
for a property is always out there ready to be AI A2 An
called into use. More significantly, a relative scale AI wljwl wljw2 wl/wn
WI
is essential to represent priority or importance if A2 w2jwl w2jw2 w2
w2/wn
one is generating the scale by making direct ob-
servations and judgments about the property un-
der study. It is also useful when one is interpreting wn
An wn!wl wn!w2 wn/wn
what the data from a standard scale really signify.
Relative scales are always needed to represent WI
subjective understanding. More is said about w2
arithmetic relations between the two types of scales =n
in Sections 8 and 12.

The foregoing formulation has the advantage of

3. Paired comparisons as ratios giving us the solution. But it also gives rise to a far
reaching theoretical interpretation. We have multi-
When we measure something with respect to a plied A on the right by the vector of weights
property, we usually use some known scale for w=(w1, w2 , ... ,w0 )T. The result of this multipli-
that purpose. A basic contribution to the subject cation is nw. If n is an eigenvalue of A, then w is
of this paper, the Analytic Hierarchy Process the eigenvector associated with it. Now A has rank
(AHP) is how to derive relative scales using judg- one because every row is a constant multiple of
ment or data from a standard scale, and how to the first row. Thus all its eigenvalues except one
perform the subsequent arithmetic operation on are zero. The sum of the eigenvalues of a matrix is
such scales avoiding useless number crunching. equal to its trace, the sum of the diagonal ele-
The judgments are given in the form of paired ments, and in this case, the trace of A is equal to
comparisons [6,7,8]. One of the uses of a hierarchy n. Therefore, n is the largest, or principal, eigen-
is that it allows us to focus judgment separately on value of A.
each of several properties essential for making a The solution of Aw = nw, called the principal
sound decision. The most effective way to con- right eigenvector of A, consists of positive entries
centrate judgement is to take a pair of elements and is unique to within a multiplicative constant.
and compare them on a single property without To make w unique, we normalize its entries by
concern for other properties or other elements. dividing by their sum. It is clear that if we are
This is why paired comparisons in combination given the comparison matrix A, we can recover
with the hierarchical structure are so useful in the scale. In this case the solution is the normal-
deriving measurement. We also note that some- ized version of any column of A.
times comparisons are made on the basis of stand- The matrix A= (ai), aiJ = wjw1, i, j =
ards established in memory through experience or 1, ... , n, has positive entries everywhere and satis-
training. fies the reciprocal property a1i = 1/aiJ· Any ma-
Assume that we are given n stones, A 1 , ••• , An, trix with this property is called a reciprocal ma-
whose weights w1 , ••• , wn, respectively, are known trix. In addition, A is consistent because the fol-
to us. Let us form the matrix of pairwise ratios lowing condition is satisfied:
whose rows give the ratios of the weights of each
stone with respect to all others. Here the smaller (1)
 T.L. Saaty j The AHP: How to make a decision 13

We see that the entire matrix can be constructed Let au= (1 + 8i 1 ) wjw1 , 8i 1 > - 1, be a per-
from a set of n elements which form a chain (or turbation of wjw1 , where w is the principal eigen-
more generally, a spanning tree, in graph-theoretic vector of A.
terminology) across the rows and columns.
It is easy to prove that a consistent matrix must Theorem 1. '-max? n.
have the ratio form A= (wjw1 ), i, j = 1, ... , n. A
necessary condition for consistency is that A be Proof: Using a 1,=lja 1i, and Aw=Amaxw, we
reciprocal. We show below that a necessary and have
sufficient condition for consistency is that the
principal eigenvalue of A be equal to n, the order
of A. When A is inconsistent, these two observa- D (2)
tions serve to help us derive a ratio scale whose
ratios are close to those of an underlying scale
w = (w 1, ••• , wn). The reciprocal axiom of the AHP Theorem 2. A is consistent if and only if '-max= n.
ensures that perturbations of a ratio scale are
themselves reciprocal. The homogeneity axiom en- Proof. If A is consistent, then because of (1), each
sures for the inconsistent case that the perturba- row of A is a constant multiple of a given row.
tions would be small, and hence that the two This implies that the rank of A is one, and all but
principal eigenvalues are close, from which it one of its eigenvalues Ai, i = 1, 2, ... , n, are zero.
would follow by an argument given in [14, p. 67] However, it follows from our earlier argument that
that the derived scale is close to an underlying I:7~ 1 A, =Trace( A)= n. Therefore, '-max= n. Con-
ratio scale. In addition, this second axiom enables versely, '-max= n implies 8i 1 = 0, and ai 1 = w,jw1.
us to explore the improvement of some of the D
judgments, thus also the improvement of incon-
sistency and the scale approximation. But there For the consistency index (CI), we adopt the
still remains the question of order preservation. value (Amax- n)/(n- 1). It is the negative aver-
The method we use to derive the scale must not age of the other roots of the characteristic poly-
only yield a ratio scale, but also capture the order nomial of A. This value is compared with the
inherent in the judgments, a very strong require- same index obtained as an average over a large
ment indeed. number of reciprocal matrices of the same order
In a general decision-making environment, we whose entries are random. If the ratio (called the
cannot give the precise values of the wjw1 but consistency ratio CR) of CI to that from random
only estimates of them. Let us consider estimates matrices is significantly small (carefully specified
of these values given by an expert who may make to be about 10% or less), we accept the estimate of
small errors in judgment. It is known from eigen- w. Otherwise, we attempt to improve consistency.
value theory [14], that a small perturbation around The reader may know about the experimental
a simple eigenvalue, as we have in n when A is findings of the psychologist George Miller in the
consistent, leads to an eigenvalue problem of the 1950's [4]. He found that in general, people (such
form A w = Amax w where Amax is the principal as chess experts) could deal with information in-
eigenvalue of A where A may no longer be con- volving simultaneously only a few facts, seven plus
sistent but is still reciprocal. The problem now is: or minus two, he wrote. With more, they become
to what extent does w reflect the expert's actual confused and cannot handle the information. This
opinion? Note that if we obtain w by solving this is in harmony with the stability of the principal
problem, and then form a matrix with the entries eigenvalue to small perturbations when n is small
( w,/w1 ), we obtain an approximation to A by a [7,14], and its central role in the measurement of
consistent matrix. consistency.
We now show the interesting, and perhaps Vargas [12] studied the case where the coeffi-
surprising result that inconsistency throughout cients of the matrix are random variables. He
the matrix can be captured by a single number focused his attention on gamma distributed coeffi-
'-max- n, which measures the deviation of the cients and derived a Dirichlet distribution for the
judgments from the consistent approximation. components of the eigenvector when the matrix is
14 TL. Saaty /The AHP: How to make a decision

consistent. When the matrix is inconsistent, the This will become clear in the second example
10% consistency bound is a sufficient measure to below. There is no reason why forcing standards
ensure that the eigenvector follows the Dirichlet on a problem should produce the same outcome
distribution with given parameters which can be obtained through relative measurement. These are
computed from the corresponding consistent ma- two different descriptive (what can be) and
trix. The gamma assumption is a powerful one normative (what should be) settings.
because of the inherent density of linear combina-
tions of these distributions.
4.1. Relative measurement: Choosing the best house
to buy
4. Two examples
When advising a family of average income to
The AHP is used with two types of measure- buy a house, the family identified eight criteria
ment, relative and absolute, the latter having to do which they thought they had to look for in a
with memory standards mentioned above. In both, house. These criteria fall into three categories:
paired comparisons are performed to derive prior- economic, geographic and physicaL Although one
ities for criteria with respect to the goaL In rela- may have begun by examining the relative impor-
tive measurement, paired comparisons are per- tance of these clusters, the family felt they wanted
formed throughout the hierarchy including on the to prioritize the relative importance of all the
alternatives in the lowest level of the hierarchy criteria without working with clusters. The prob-
with respect to the criteria in the level above. In lem was to decide which of three candidate houses
absolute measurement, paired comparisons are to choose. The first step is the structuring of the
also performed through the hierarchy with the problem as a hierarchy.
exceptions of the alternatives themselves. The level In the first (or top) level is the overall goal of
just above the alternatives consists of intensities or 'Satisfaction with house'. In the second level are
grades which are refinements of the criteria or the eight criteria which contribute to the goal, and
subcriteria governing the alternatives. One pair- the third (or bottom) level are the three candidate
wise compares the grades themselves under each houses which are to be evaluated in terms of the
criterion by answering questions such as: How criteria in the second leveL The definitions of the
much better is a student applicant with excellent criteria follow and the hierarchy is shown in Fig-
grades than one with very good grades? and how ure 1.
much better is a student applicant with average The criteria important to the individual family
letters of recommendation than one with poor were:
ones? and so on. The alternatives are not pairwise (1) Size of house: Storage space; size of rooms,
compared, but simply rated as to what category in number of rooms; total area of house.
which they fall under each criterion. A weighting (2) Location to bus lines: Convenient, close bus
and summing process yields their overall ranks. service.

Figure 1. Decomposition of the problem into a hierarchy

 T.L. Saaty j The AHP: How to make a decision 15

Table 1
The fundamental scale

Intensity of importance Definition Explanation

on an absolute scale
Equal importance Two activities contribute equally to the objective

3 Moderate importance of Experience and judgment strongly favor one activity over another
one over another

5 Essential or strong impor- Experience and judgement strongly favor one activity over another
tance

7 Very strong importance An activity is strongly favored and its dominance demonstrated in
practice

9 Extreme importance The evidence favoring one activity over another is of the highest
possible order of affirmation

2,4,6,8 Intermediate values be- When compromise is needed

tween the two adjacent
judgments

Reciprocals 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

Rationals Ratios arising from the scale If consistency were to be forced by obtaining n numerical values to
span the matrix

(3) Neighborhood: Little traffic, secure, nice The second step is the elicitation of pairwise
view, low taxes, good condition of neighborhood. comparison judgments. Arrange the elements in
(4) Age of house: Self-explanatory. the second level into a matrix and elicit judgments
(5) Yard space: Includes front, back and side, from the people who have the problem about the
and space from neighbors. relative importance of the elements with respect to
(6) Modern facilities: Dishwashers, garbage dis- the overall goal, Satisfaction with House. The
posals, air conditioning, alarm system, and other scale to use in making the judgments is given in
such items possessed by a house. Table 1. This scale has been validated for effec-
(7) General condition: Repairs needed, walls, tiveness, not only in many applications by a num-
carpet, drapes, cleanliness, wiring. ber of people, but also through theoretical com-
(8) Financing available: Assumable mortgage; parisons with a large number of other scales.
seller financing available, or bank financing. The questions to ask when comparing two

Table 2
Pairwise comparison matrix for level 1

2 3 4 5 6 7 8 Priority
vector
1 5 3 7 6 6 l I
0.173
.1 4
1 I I
2 1 5 3 3 0.054
3 1 3 ' 6 3 4 '
6
7
1
~ 0.188
I 1 1 I I I
4 7 5 6 1 "l 4 O.Dl8
5
I I I
3 1 I I '
I
0.031
6
6
I
;; '
I
}
'
4
1
4 2
2
' 6
I
6 0.036
1
7 3 5 6 7 5 5 0.167
8 4 7 5 8 6 6 2 0.333
,\max = 9.669, CI = 0.238, CR = 0.169
16 TL. Saaty I The AHP: How to make a decision

criteria are of the following kind: of the two compared pairwise are the houses with respect to
criteria being compared, which is considered more how much better one is than the other is satisfying
important by the family buying the house with each criterion in level 2. Thus there will be eight
respect to the overall goal of family satisfaction 3 X 3 matrices of judgments since there are eight
with the house? elements in level 2, and 3 houses to be pairwise
When the elements being compared are closer compared for each element. Again, the matrices
together than indicated by the scale, one can use contain the judgments of the family involved. To
the scale 1.1, 1.2, ... , 1.9. If still finer, one can use understand the judgments, a brief description of
the appropriate percentage refinement. the houses follows.
The matrix of pairwise comparisons of the House A. This house is the largest of them all.
criteria given by the homebuyers in this case is It is located in a neighborhood with little traffic
shown in Table 2, along with the resulting vector and low taxes. Its yard space is comparably larger
of priorities. The vector of priorities is the prin- than houses B and C. However, the general condi-
cipal eigenvector of the matrix. It gives the relative tion is not very good and it needs cleaning and
priority of the criteria measured on a ratio scale. painting. Also, the financing is unsatisfactory be-
In this case financing has the highest priority with cause it would have to be bank-financed at high
33% of the influence. interest.
In Table 2, instead of naming the criteria, we House B. This house is a little smaller than
use the number previously associated with each. House A and is not close to a bus route. The
Next we move to the pairwise comparisons of the neighborhood gives one the feeling of insecurity
elements in the lowest level. The elements to be because of traffic conditions. The yard space is

Table 3
Comparison matrices and local priorities

Size of house A B c Priority Yard space A B c Priority

vector vector
A 1 6 8 0.754 A 5 4 0.674
I I I
B 6 4 0.181 B 5 3 0.101
c 8
I I
4 0.065 c I
4 3 0.226
Amax= 3.136, CI = 0.068, CR=0.117 Amax= 3.086, CI = 0.043, CR = 0.074

Transportation A B c Priority Modem A B c Priority

vector facilities vector
I
A 1 7 5 0.233 A 8 6 0.747
I I I I
B 7 8 0.005 B 1i 1 5 0.060
c 5 8 1 0.713 c I
6 5 1 0.193
Amax = 3.247, CI = 0.124, CR =0.213 Amax= 3.197, CI = 0.099, CR =0.170

Neighborhood A B c Priority General A B c Priority

vector condition vector
I I
A 1 8 6 0.745 A 2 2 0.200
I I
B 1 4 0.065 B 2 1 1 0.400
c "I
6 4 0.181 c 2 1 1 0.400
Amax= 3.130, CI = 0.068, CR = 0.117 Amax = 3.000, CI = 0.000, CR =0.000

Age of house A B c Priority Financing A B c Priority

vector vector
I I
A 0.333 A 7 5 0.072
B 0.333 B 7 3 0.650
c 1 1 1 0.333 c 5 I
3 1 0.278
Amax= 3.000, Cl=O.OOO, CR =0.000 Amax= 3.065, CI = 0.032, CR =0.056
 TL. Saaty 1 The AHP: How to make a decision 17

Table 4
Local and global priorities

2 3 4 5 6 7 8
(0.173) (0.054) (0.188) (0.018) (0.031) (0.036) (0.167) (0.333)

A 0.754 0.233 0.754 0.333 0.674 0.747 0.200 0.072 0.396

B 0.181 0.055 0.065 0.333 0.101 0.060 0.400 0.650 0.341
c 0.065 0.713 0.181 0.333 0.226 0.193 0.400 0.278 0.263

fairly small and the house lacks the basic modern vectors by the priority of the corresponding crite-
facilities. On the other hand, the general condition rion and add across each row which results in the
is very good. Also, an assumable mortgage is desired vector of the houses in Table 4. House A
obtainable which means the financing is good with which was the least desirable with respect to fi-
a rather low interest rate. nancing (the highest priority criterion), contrary to
House C. House C is very small and has few expectation, had the largest priority. It was the
modern facilities. The neighborhood has high house that was bought.
taxes, but is in good condition and seems secure.
The yard space is bigger than that of House B, but 4.2. Absolute measurement: Employee evaluation
is not comparable to House A's spacious sur-
roundings. The general condition of the house is Absolute measurement is applied to rank alter-
good and it has a pretty carpet and drapes. natives in terms of ratings, intensities or grades of
The matrices of comparisons of the houses with the criteria. These grades may take the form:
respect to the criteria and their local priorities are excellent, very good, good, average, below average,
given in Table 3. poor and very poor. After establishing a scale of
The third step is to establish the composite or priorities for the criteria (or subcriteria, if there
global priorities of the houses. We lay out the are some) through paired comparisons, the grades
local priorities of the house with respect to each which may be different for each criterion or sub-
criterion in a matrix and multiply each column of criterion, are in turn pairwise compared according

Table 5
The hierarchy of employee evaluation

Goal: Employee performance evaluation

Criteria: Technical Maturity Writing Verbal Timely Potential
skills skills work (personal)
(0.061) (0.196) (0.043) (0.071) (0.162) (0.466)

Intensities: Ex cell. Very Ex cell. Excell. Nofollup Great

(0.604) (0.731) (0.733) (0.750) (0.731) (0.750)
Abv. avg. Accep. Avg. Avg. On time Averag.
(0.245) (0.188) (0.199) (0.171) (0.188) (0.171)

Avg. Immat. Poor Poor Remind Bel. avg.

(0.105) (0.181) (0.068) (0.078) (0.081) (0.078)
Bel. avg.
(0.046)

Alternatives:
(1) Mr. X Excell. Very Avg. Ex cell. On time Great
(2) Ms. Y Avg. Very Avg. Avg. Nofollup Avg.
(3) Mr. Z Excel!. lmmat. Avg. Ex cell. Remind Great
18 TL. Saaty I The AHP: How to make a decision

to their parent criterion. An alternative is values. An extension of this philosophy in prob-

evaluated, for each criterion or subcriterion, by lems where order relations between the variables
identifying the grade which best describes it. Fi- are important, is that on small perturbations of
nally, the weighted or global priorities of the grades the variables, the procedure produces close, order
are added to produce a ratio scale for the alterna- preserving results. The procedure described here
tive. Absolute measurement needs standard to has this characteristic.
make it possible to judge whether the alternative is Because of the natural way in which a matrix of
acceptable or not. Absolute measurement is useful ratios and small perturbations of that matrix lead
in student admission, faculty tenure and promo- to a principal eigenvalue problem, one may start
tion, employee evaluation, and in other areas where with this and generalize to positive matrices that
there is fairly good agreement on standard which need not be reciprocal. The German mathemati-
are then used to rate alternatives one at a time. cian Oskar Perron proved in 1907 that, if A =
Let us consider an abbreviated version of the (a;), a; 1 > 0, i, j = 1, ... , n, then A has a simple
problem of evaluating employee performance. The positive eigenvalue :X. max (called the principal ei-
hierarchy for the evaluation and the priorities genvalue of A) and :\max> I Ak I for the remaining
derived through paired comparisons are shown eigenvalues of A. Furthermore, the principal ei-
below. It is then followed by a rating of each genvector w = ( w 1, ... , wn) T that is a solution of
employee for the quality of performance under Aw = Amaxw has W; > 0, i = 1, ... , n. We can write
each criterion and summing the resulting scores of the norm of the vector w as II w II = eTw where
obtain his overall rating. The hierarchy in Table 5 e = (1, 1, ... , 1)T and we can normalize w by divid-
can be more elaborate, including subcriteria, fol- ing it by its norm. For uniqueness, when we refer
lowed by the intensities for expressing quality. to w we mean its normalized form. Our purpose
Let us now show how to obtain the total score here is to show how important the principal eigen-
for Mr. X (see Table 5): vector is in determining the rank of the alterna-
tives through dominance walks.
0.061 X 0.604 + 0.196 X 0.731 + 0.043 X 0.199
We have seen that ratio scale estimation has a
+ 0.071 X 0.750 + 0.162 X 0.188 natural setting in principal eigenvalue formula-
+ 0.466 X 0.750 = 0.623. tion. We will now show that in addition, the
Similarly the score for Ms. Y and Mr. Z can be principal eigenvector also has the order preserving
shown to be 0.369 an 0.478 respectively. It is clear properties we seek.
that we can rank any number of candidates along When A is consistent, one way to define the
these lines. Here the vector of priorities of the order of the alternatives is to require that one row
criteria has been weighted by the vector of relative of the matrix dominate elementwise another row.
number of intensities under each criterion and But when A is inconsistent it is no longer possible
then renormalized. We call this a structural rescal- to define dominance in this manner. Instead, we
ing of the priorities. borrow the concept of dominance from graph
theory where the sum of the coefficients in each
row of A is used. This concept carries over in a
5. Theoretical considerations (7) natural way to the inconsistent case. But we must
look at a different way to capture dominance by
There is a well known principle in mathematics considering further possibilities not simply from
that is widely practiced, but seldom enunciated the matrix itself or some arbitrary power of it, but
with sufficient forcefulness to impress its impor- from all its powers.
tance. A necessary condition that a procedure for The matrix A captures only the dominance of
solving a problem be a good one is that if it one alternative over every other in one step. But
produces desired results, and we perturb the varia- an alternative can dominate a second by first
bles of the problem in some small sense, it gives us dominating a third alternative and then the third
results that are 'close' to the original ones. This is dominates the second. Thus, the first alternative
precisely the use of continuity and uniform con- dominates the second in two steps. It is known
tinuity-to assure that after transforming a varia- that the result for dominance in two steps is
ble, originally nearby values go over to nearby obtained by squaring the pairwise comparison rna-
 T.L. Saaty / The AHP: How to make a decision 19

trix. Similarly, dominance can occur in three steps, where w is the normalized principal right eigen-
four steps and so on, the value of each is obtained vector of A, we have
by raising the matrix to the corresponding power.
The rank order of an alternative is the sum of the (7)
relative values for dominance in its row, in one
step, two steps and so on averaged over the num-
ber of steps. The question is whether this average The solution of the eigenvalue problem is ob-
tends to a meaningful limit. It is easy to see that it tained by raising the matrix A to a sufficiently
does when A is consistent because Ak = nk- 1A. large power then summing over the rows and
We can think of the alternatives as the nodes of normalizing to obtain the priority vector w =
a directed graph. With every directed arc from ( w1 , ... , wn) T. The process is stopped when the
node i to node j (which need not be distinct), is difference between components of the priority
associated a nonnegative number aiJ of the domi- vector obtained at the k-th power and at the
nance matrix. In graph-theoretic terms this is the ( k + 1)st power is less than some predetermined
intensity of the arc. Define a k-walk to be a small value.
sequence of k arcs such that the terminating node In reference [7] we gave at least five different
of each arc except the last is the source node of ways of deriving the priorities from the matrix of
the arc which succeeds it. The intensity of a k-walk paired comparisons. Besides the eigenvector solu-
is the product of the intensities of the arcs in the tion, these include the direct row sum average, the
walk. With these ideas, we can interpret the matrix normalized column average, and methods which
Ak: the (i, j) entry of Ak is the sum of the minimize the sum of the errors of the differences
intensities of all k-walks from node i to node j. between the judgments and their derived values
such as the methods of least squares and logarith-
Definition. The dominance of an alternative along mic least squares. We pointed out that the loga-
all walks of length k < m is given by rithmic least squares solution coincides with the
1 m Ake principal right eigenvector solution for matrices of
(3) order n = 3, which is the first value of n for which
m L
k~l
eTAke ·
inconsistency is possible and left and right eigen-
vectors are reciprocals of each other which is not
Observe that the entries of Ake are the row sums always the case for larger values of n. Since the
of Ak and that e TAke is the sum of all the entries appearance of the Analytic Hierarchy Process in
of A. the literature, other methods have been proposed
[7,15]. All methods yield the same answer when
Theorem 3. The dominance of each alternative along the matrix is consistent. The combined use of a
all walks k, is given by the solution of the eigenvalue measure of inconsistency which can be derived in
problem Aw = Amaxw. terms of both left and right eigenvectors, along
with the right eigenvector solution which captures
Proof. Let the dominance expressed in the judgments, is an
effective way to look at the problem. We argue
(4) that so long as inconsistency is tolerated, domi-
nance is the basic theoretical concept for deriving
and a scale and no other method qualifies. In addition,
a counterexample has been provided in which the
(5) method not only does not generate a good ap-
proximation, but also reverses rank. Some have
The convergence of the components of tm to the even resorted to artificial axiomatization thinking
same limit as the components of sm is the stand- that it gives a method the appearance of rigor,
ard Cesaro summability. Since although axioms are assumptions not proofs. It
may be that the arithmetic of a method is simpler
Ake than that used to obtain the eigenvector, but that
sk = TT ---> w as k ---> oo, (6)
eAe no longer matters, because of the widespread use
20 TL. Saaty /The AHP: How to make a decision

of the computer. It is reasonable to argue that a in determining rank (recall that if a> b then
theory for making sound decisions must stand on a+ c > b + c and c does not reverse order). An
clearly justifiable grounds. alternative that is a copy of another can dilute the
The software package Expert Choice, useful in priority of a decisive criterion so that it is no
teaching and in real applications, can handle both longer the controlling one in determining the final
relative and absolute measurement, as well as hav- rank. We have:
ing special capabilities such as structural rescaling,
combining group judgments, sensitivity analysis Theorem 4. LiJaiJ = n 2 , if and only if all alterna-
and dependence among the decision alternatives tives have equal dominance, i.e., aiJ = 1, i, j =
[2]. The reader interested in pursuing the subject 1, ... , n, with respect to a criterion.
further should consult references [3,7,8,13,15].
It is easy to see that n 2 is the minimum value
of possible total dominance by a paired compari-
6. Nonnalization - Scarcity and abundance son of n alternatives. The maximum value is,
using 9 as the upper range of the scale of paired
Normalization in the AHP is not just a mecha- comparisons, [in(n- 1)](9 + ~) + n = ~(41n 2 -
nical operation. It contains information on the 32n).
total dominance of the alternatives being com- Our discussion of absolute and relative meas-
pared which enables us to apportion the priority urement may be framed in economic terms. In
of the criterion to each alternative according to absolute measurement, value and need are identi-
the relative dominance of the alternative. cal; 'the more value, the better the need is satis-
Normalization can also be associated with the fied'. In relative measurement, value is assessed in
idea of scarcity and abundance of the presence of terms of need. Here surplus value may or may not
a criterion in the alternatives such as redness in satisfy more need. In fact, there are instances
fruit. Too many fruits that are red make red where satiation takes place and abundance can
abundant and unimportant in differentiating be- lead to a decline in the satisfaction of need.
tween individual fruits. Conversely, if redness oc-
curs intensely in some fruits but not in others, it is
thus scarce and can be used as a criterion to 7. Clustering
differentiate in making a decision among fruits.
Thus, the greater the contrast among the alterna- Comparisons of elements in pairs requires that
tives, the more useful is the priority value of the they be homogeneous or close with respect to the
criterion allotted to each. Conversely, when for common attribute; otherwise significant errors
example, the contrast of the alternatives is smal- may be introduced into the process of measure-
lest (when the total dominance is equal to n 2 , see ment. In addition, the number of elements being
below) and hence all the alternatives are alike, the compared must be small (not more than 9) to
portion of the criterion priority assigned to each is improve consistency and the corresponding accu-
equal. Thus a criterion with respect to which there racy of measurement. For example, we may clus-
is greater contrast and dominance among the al- ter apples in one way according to size, in another
ternatives is 'scarce' and consequently more in- way according to color and in still another way
fluential in determining rank than another crite- according to age. The question then is how to
rion on which there is no distinction among the perform clustering of homogeneous elements in an
alternatives, i.e., their matrix of paired compari- efficient way to facilitate paired comparisons.
sons has more 1's in it. In being scarce, more of Clustering is a process of grouping elements with
the criterion is allotted to the more dominant respect to a common property. One can then
alternative and hence it affects the final ranking of decompose the set of ordered elements with re-
that alternative more. spect to an attribute into clusters of, for example,
An 'abundant' criterion contributes less to rank seven elements each, from largest to smallest. The
determination because it contributes an equal or smallest element of the largest cluster is included
nearly equal priority to each alternative and when as one of the seven elements of the next cluster.
the sum is taken over the criteria it has little effect The relative weights of all the elements in this
 T.L. Saaty / The AHP: How to make a decision 21

second cluster are divided by the weight of the order on that attribute. They can then be clustered
common element and then multiplied by its weight into small groups as described above and pairwise
in the first cluster in this manner both clusters compared.
become commensurate and are pooled together. One reason why absolute measurement may not
The process is then repeated to the remaining be desirable is that it is strongly subjective. In
clusters. Sometimes one may need to introduce paired comparisons, measurement is based on ob-
hypothetical elements in order to preserve the servation of the relative intensity of a property
gradual descent from large to small. between two elements. Absolute measurement is
We discuss three ways as to how to perform based on observations stored in memory which
clustering on the alternatives of a decision prob- depend on experience and on the ability to recall
lem whose number may be very large and needs a it. For many problems, it is useful to first carry
different sorting for each of several attributes. out absolute measurement to sort and cluster the
They are ordered in the following discussion from elements, and then follow that with relative meas-
the least to the most efficient way. urement for greater accuracy. This is particularly
relevant in predicting most likely outcomes which
7.1. The elementary approach involve synergy among the alternatives.

Given n elements in a level of the hierarchy,

one may first make a pass through them by com- 8. Combining relative and absolute measurement -
paring one element with another, dropping it and Cost-benefit analysis
picking another if that one is perceived to be
larger and continuing the comparison. Thus, the An easy pitfall for an individual who has just
largest element is selected in n- 1 such compari- learned about the AHP is to take two or more
sons. The process is repeated for the remaining criteria on which alternatives are measured on the
n - 1 elements to identify the second largest ele- same existing standard scale, such as dollars or
ment and so on. In the end, the elements would be kilograms, normalize each set and then compose
arranged in descending order of size or intensity with respect to the criteria. One quickly discovers
according to an attribute and are sequentially that the answer is not the same as that obtained
clustered into groupings of a few elements each through the usual arithmetic and hastily concludes
from the largest to the smallest. This process is that the AHP is at fault. To avoid this problem,
highly inefficient and requires the astronomical one must exercise caution in converting measure-
number of (n- 1)! comparisons. ments on a standard scale to relative values when
several such criteria are involved [9].
7.2. Trial and error clustering For the arithmetic to conform with what one
ordinarily does, assign each criterion a priority
The alternatives may be put into groups of that is the sum of the measurements of the alter-
large, medium and small. Then the elements in natives with respect to it, divided by the sum of
each group are put into several clusters of a few the measurements of all the alternatives under all
elements each, and a first pass at comparisons is the criteria measured with that unit. To find the
used to identify misfits which are then taken out composite priority for each alternative multiply
and put into the appropriate one of the other two the criterion weight by its corresponding normal-
categories. Reclustering is then performed and ized weight under that criterion and sum over the
comparisons are carried out. If elements are found criteria. The result may be considered as the weight
not to fit, they are again moved to the appropriate of the alternative with respect to one super crite-
category. This process is repeated for each attri- rion composed of all the criteria with the unit of
bute. measurement. That super criterion may then be
compared with other intangible criteria and other
7.3. Clustering by absolute measurement super criteria with different units of measurement.
Alternatively, one can perform priority compari-
Each alternative is evaluated by absolute meas- sons on all the data available without ascribing
urement for an attribute, and thus in descending linearity to them as one ordinarily does with num-
22 TL. Saaty I The AHP: How to make a decision

hers. That is to say, one interprets what the num- measured in dollars along with intangible factors,
bers mean and uses judgment rather than cranking they must first be composed according to the
the number mechanically. With rare exception, the dollars and then combined with the intangible
judgement approach is by far the more effective criteria as described earlier. Marginal benefit cost
procedure using the AHP to deal with the underly- analysis along traditional lines can be carried out
ing complexity of a decision problem, and one by arranging the costs in increasing order, and
need not be afraid to do it. It is known that scales then forming ratios of successive differences of
are invented to facilitate communication, and must benefits and costs in that order. The very first
be experienced for a long time before they can be ratio is that of the alternative with the smallest
associated with our value system, during which cost. Then one forms the ratio of differences with
time they almost always are modified. that between the next highest cost and the smallest
The AHP is a descriptive theory. Therefore, it one in the denominator, and the corresponding
is not an automatic set-up for accommodating any difference in benefits in the numerator. Whenever
normative approach such as utility maximization. a difference in a numerator is negative, that suc-
It needs to be interpreted and adapted for that ceeding alternative is dropped from consideration.
purpose. It is even more difficult when a norma- In this manner, the alternative yielding the highest
tive theory has many exceptions so the AHP inter- marginal ratio is chosen.
pretation would not be universal. In utility maxi- The next question is how to find a reasonable
mization it is assumed that it is always the case way to combine benefits (B) and cost (C), when a
that the more utility or more money the better, ratio rather than a single hierarchy are used [10].
but there are many instances when this is not true. This is useful in considering conflict problems
For example, a government agency that is left involving more than one value system. Since the
with more money at the end of the year, gets a hierarchy of costs leads to a vector of values
smaller allocation in a future budget. A rich indi- indicating relative maximum costs, the reciprocals
vidual in a poor country under revolt by the poor of the entries of this vector could be thought of as
is likely to get killed. The availability of money is the minimum costs incurred in jointly maximizing
often an incentive to use money where other alter- benefits and minimizing costs. In general, the out-
natives could be more effective. Thus, AHP adap- comes from a hierarchy of benefits and a hierarchy
tation for utility calculation purposes needs to of costs are two ratio scales whose corresponding
take such caveats into consideration. One idea to ratios lead to a meaningful ratio scale. Their dif-
keep in mind with the AHP is that the montonic- ferences would not be meaningful.
ity of utilities need not be preserved and may be To maximize BjC is equivalent to maximizing
contradicted. log BjC, which if B and C are close, may be
The foregoing has bearing on benefit-cost approximated by (B/C)- 1 or (B- C)/C
analyses. Expected utilities are used for repeated known as return on investment, ROI. If B and C
decision making. In that case, benefit cost analysis are not comparable, it would be initially clear that
from two hierarchies is applied, and from it one only the benefits or only the costs determine
can also calculate marginal benefit to cost ratios. whether the allocation should be made or not.
Resource allocation may be made by using benefit Some people have attempted to compare the
to cost ratios thus derived. Short range decisions foregoing with what is traditionally done in a
often include low costs as benefits in a single single criterion choice problem. An example where
hierarchy. This is a useful approach when a deci- B - C alone gives rise to misleading results is
sion to spend money on one of several options has given by my colleague L.G. Vargas: You have $1
already been made. It is one way to combine million to invest to get $1.101 million or you have
benefits and costs as one does with dollars by $500 000 to invest to get $600 000. Which is a
taking differences. In the AHP one does not use better investment?
differences. In such one-time decisions, costs may
be regarded as inverse benefits so that benefits, B- C analysis gives: 1.101 -1 = 0.101,
low costs and other inconveniences are used to 0.6- 0.5 = 0.1,
establish priorities for the alternatives.
If, on the other hand, benefits and costs are and the first alternative would be chosen. ROI
 T.L. Saaty / The AHP: How to make a decision 23

analysis yields 10.1% and 20%, respectively, and earlier mathematical principle of small perturba-
the second alternative which minimizes risk would tions we gave in which small changes in input lead
be correctly chosen. to small changes in outcomes. In catastrophes,
there is a part of the problem in which small
changes in conditions at some critical value can
9. The semiotic connection give rise to very large changes in outcome. All we
can do is either to anticipate and take strong
The Analytic Hierarchy Process is a tool of preventive action or prepare ourselves in advance
information communication and signification. In a with contingency plans together with emotional
general sense, it belongs to the study of language acceptance for how to deal with the emergency to
or semiotics. Semiotics or semiology is a coding in pay a smaller penalty if it occurs, or be prepared
which one considers signals related by a set of to pay the full penalty having accepted that it
rules or syntax; a set of states or contents called a could happen. Because catastrophes are usually
semantic system, a set of possible behavioral re- unseen surprises, it gives us a feeling of control to
sponses independent of the content system, and a think we can account for catastrophes in our daily
rule associating signals with contents or with be- lives. According to the dictionary. a catastrophe is
havioral responses. The idea of a code covers all a sudden disaster or happening that causes great
four phenomena mentioned above. Eco [1] uses harm or damage; a calamity. Even the idea of
s-code for the first three and code for the fourth. what constitutes a catastrophe is relative. Some-
According to Morris [5], the creator of semi- one whose life is tormented by an opponent may
otics, semiosis is the process in which something regard it as a blessing if a catastrophe befalls the
functions as a sign. He mentions that both man tormentor. In other terms, a catastrophe may be
and animals do respond to certain things as signs regarded as a very strong discontinuity in think-
of something else, but their complexity in man is ing. How do we represent a catastrophe in terms
found in speech, art, writing, and even medical of the AHP?
diagnoses. Science and signs are inseparably inter- In a catastrophe the element of surprise arises
connected. Semiotics, a step in the unification of from a shift in ranking of the outcomes. Thus, if a
science, supplies the foundations for any science situation is ongoing a certain choice of alternatives
of signs; linguistics, logic, mathematics, rhetoric may be indicated, but a slight change in the situa-
and aesthetics. tion may cause a sudden shift to another choice
In the Analytic Hierarchy Process, words are that would have been previously very undesirable.
used for concepts involved in decisions. In a sense, How can we allow for catastrophic occurrences in
the purpose of all information is to decide on the AHP?
something-even if it is an imaginary hypothesis. One way to allow for a catastrophe is to always
The hierarchy is the syntax, the subject of a deci- include a criterion for the unknown that repre-
sion is the semantic, prioritization is the rule asso- sents a cluster of unforseen threats. It may itself
ciating signals with content, and behavior com- have subcriteria. The alternatives are carefully
prises the possible decision alternatives. This way prioritized with respect to this criterion for the
of looking at the AHP needs to be highlighted and unexpected and its descendants. The criterion it-
emphasized to drive home the universality of the self is assigned a low priority. Now we can imag-
decision process. ine that at each instant the judgments in the
criteria are flashed in their totality on a screen
followed by the best choice of alternatives. With
10. Catastrophe and the AHP some changes in judgments, in general, choice is
stable and the changes in relative priority of the
One of the most burning quests we undertake alternative slight. After a certain lapse of time,
using the knowledge we acquire is explaining and there is a sudden change in judgment in the direc-
forecasting, and in particular, forecasting catas- tion of this criterion making it more important.
trophes, or, mathematically speaking, sudden dis- The corresponding choice of alternatives is also
continuities that affect the order of the alterna- suddenly changed surfacing a very undesirable
tives. This is a situation that is at odds with the alternative. Thus, one might include among the
24 T.L. Saaty / The AHP: How to make a decision

OLDWRECK
(0.002)
CHEVY
(0.009)
SIZE HONDA
(0.051) (0.013)
TOYOTA PRIORITIES
(0.026) TOYOTA
(0.408)
OLD WRECK

--t~
HONDA
(0.005) (0.345)
CHEVY CHEVY
OPBRCOST (0.019) (0.197)
(0.135)
HONDA OLDWRECK
(0.044) (0.049)
TOYOTA
(0.067)
~ NORMAL ____,.1
1 (0.990) -1
OLD WRECK
(0.010)
CHEVY
(0.086)
HONDA
(0.086)
TOYOTA
(0.086)

OLDWRECK

LOWm~-{
(0.026)
CHEVY
(0.081)
(0.535) HONDA
(0.201)
TOYOTA
(0.227)
GOAL OLDWRECK

·~-f
(0.()()4)
CHEVY
(0.001)
PRIORITIES
HONDA
(0.006) TOYOTA
(0.000)
(0.142)
TOYOTA
HONDA
(0.001)
(0.115)
CHEVY
OLD WRECK

o•n=-{
(0.166)
(0.001)
CHEVY OLD WRECK
(0.576)
(0.000)
(0.001) HONDA
(0.000)
TOYOTA
(0.000)
4 ACCIDENT
(0010)
----1
OLDWRECK
(0.001)

=ce---1
CHEVY
(0.000)
HONDA
(0.002) (0.000
TOYOTA
(0.000)

OLDWRECK

w-~-{
(0.000)
CHEVY
(0.000)
(0.001) HONDA
(0.001)
TOYOTA
(0.001)

Figure 2. The catastrophe hlerarchy

 T.L. Saaty 1 The AHP: How to make a decision 25

alternatives some catastrophic ones. This type of cause of the dependence of the measurements of
thinking would apply in hierarchies where a real the alternatives on each other. However, relative
wrong choice is extremely undesirable and costly. measurement will preserve rank with respect to a
We can also speak of chaos in the AHP: a single criterion when the comparisons are con-
situation where the possibility of exerting control sistent. Most people understand the dependence
by creating a hierarchy and setting judgment no of alternatives in light of the notions of scarcity
longer exists, for it is not clear what takes prece- and abundance discussed earlier. There is no need
dence over what. Usually catastrophe produces to improvise notions of relevant and irrelevant
chaos. It is possible to represent both catastrophe alternatives, as is done in utility theory, because
and chaos in the continuous setting of the AHP. with relative measurement everything being com-
Both catastrophe and chaos are relative. A rev- pared is by definition relevant. Utility theory is
olution may be regarded as an undesirable obsessed with the reversal of rank because in that
catastrophe that uproots a society or as a neces- theory it can happen even with respect to a single
sary kind of action to produce a desired end. In criterion. This is a phenomenon that is strongly
planning, strong out-of-the-ordinary action may counterintuitive and can never be made mathe-
be crucial for affecting change. matically right. Utility theorists, however, try to
A simple illustration of how to represent a make it right by philosophical arguments about
catastrophe is given in Figure 2. An individual what is or is not a relevant alternative.
buying a car makes an innocent choice by assum-
ing that normal conditions will prevail over acci-
dents in the ratio of 0.99 to 0.01. He chooses a 12. Summary of principles
nice new Toyota, although he knows that there is
a chance that an accident can happen. If he were The AHP generates relative ratio scales of
to buy the car under the assumption of a high measurement. The measurements of a set of ob-
likelihood of an accident, 0.99 to 0.01, and if he jects on a standard scale can be converted to
were to act in his best self interest, he would relative scale measurements through normaliza-
choose an old wreck to buy. One answer to this tion. Only in a very localized way can a relative
dilemma is neither to buy the old wreck nor the set of measurements have a unit, obtained by
Toyota, but to buy a fairly good used car. But dividing the entire set by the smallest measure-
most of us are idealists who do not think that ment. The normalization and composition of
accidents will happen to us, and so we buy new weights of alternative with respect to more than a
cars. single criterion measured on the same standard
scale leads to nonsensical numbers, because nor-
malizing separate sets of numbers destroys the
11. What affects rank (11) linear relation among them. The weights must first
be composed with respect to all such criteria and
Although in catastrophes there is a sudden shift then normalized for AHP use. We can interpret
in rank, due to a change in the importance of the such composition as we did in Section 8 as a
criteria, what happens to the rank of the alterna- special kind of weighting of the particular criteria.
tives in the more mundane event that their num- Thus, the AHP, with its relative measurement
ber is changed by adding new ones or deleting old offers no guide on the outcome of manipulations
ones? The traditional rule is that rank reversal is based on combining different measurements from
not acceptable if, given that the alternatives them- a standard scale such as a criterion of benefits and
selves are independent of each other, a new alter- a criterion of costs, both measured in dollars, and
native does not introduce a new criterion or change used to select a best alternative.
the weights of the existing criteria. We tend to If we do not insist that the linearity of a scale
treat rank in a possessive manner by sometimes needs to be preserved (an old habit from when we
insisting that it stay the same no matter what logic did not have an effective way to interpret the
says. The absolute mode of measurement of the information content of readings from a standard
AHP complies with this normative inclination, but scale), we can then treat every criterion as an
the relative mode of measurement does not, be- intangible. In that case we must bear in mind that
26 TL. Saaty /The AHP: How to make a decision

the weights of the criteria do not derive from some tional Encyclopedia of Unified Science, Vols. I and II:
underlying standard scale. If they must depend on Foundations of the Unity of Science, Vol. 1, No. 2,
University of Chicago Press, 1938.
such a scale, we are back to the need to compose [6] Saaty, T.L., Decision Making for Leaders: The Analytic
before normalization. Hierarchy Process for Decisions in a Complex World, RWS
The moral is that we are sometimes led into Publications, Pittsburgh, PA, 1986; Original version pub-
developing blind expectations for that to which we lished by Lifetime Learning Publications, 1982.
are accustomed out of habit, and not necessarily [7] Saaty, T.L., Multicriteria Decision Making: The Analytic
Hierarchy Process, 1988; Revised and published by the
because its truth is something written in granite. author; Original version published by McGraw-Hill, New
We believe that our own tempered understanding York, 1980.
should produce closer results to experience than [8] Saaty, T.L., "Scaling method for priorities in hierarchical
simply following tradition, which has possibly structures", Journal of Mathematical Psychology 15/3
rutted our thinking, and induced us to forego (1977) 234-281.
[9] Saaty, T.L., "A note on the AHP and expected value
change in search of better ways that give better theory", Socio-Economic Planning Sciences 20/6 (1986)
answers. 397-398.
[10] Saaty, T.L., and Kearns, K.P., Analytical Planning: The
Organization of Systems, International Series in Modern
References Applied Mathematics and Computer Science, Vol. 7, Per-
gamon Press, New York, 1985.
[11] Saaty, T.L., and Vargas, L.G., "Inconsistency and rank
[1] Eco, U., A Theory of Semiotics, First Midland Book
preservation", Journal of Mathematical Psychology 28/2
Edition, Indiana University Press, IN, 1976.
(1984).
[2] Forman, E., and Saaty, T., Expert Choice Software
[12] Vargas, L.G., "Reciprocal matrices with random coeffi-
Package for IBM PC, Expert Choice, Inc., Pittsburgh, PA
cients", Mathematical Modelling 3 (1982) 69-81.
1983-1990.
[13] Vargas, L.G. (ed.), "Special Issue on The Analytic
[3] Harker, P.T. (ed.), "Special Issue on The Analytic
Hierarchy Process", Mathematical Modelling 9/3-5
Hierarchy Process", Socio-Economic Planning Sciences
(1987).
20/6 (1986). [14] Wilkinson, J.H., The Algebraic Eigenvalue Problem,
[4] Miller, G.A., "The magical number seven, plus or minus
Clarendon Press, Oxford, 1965.
two: some limits on our capacity for processing informa-
[15] Zahedi, F., "The analytic hierarchy process- A survey of
tion", The Psychological Review 63 (1956) 81-97. the method and its applications", Interfaces 16!4 (1986).
[5] Morris, C.W., Foundation of the Theory of Signs, Interna-