Advanced Decision and Game Theory

Chapter 2 : Analytic Hierarchy Process

Dr. Sonia Rebai

Tunisian Business School

University of Tunis
ü AHP was developed by Thomas L. Saaty during the 1970s.

ü AHP has been applied in a wide variety of areas such as decision

making for facility layout and health care.

ü The output of AHP is a prioritized ranking of the decision alternatives

based on the overall preferences expressed by the DM.

2
To use AHP the user needs to complete four steps

ü First the problem has to be structured.

ü Second, scores have to be calculated based on pairwise comparisons delivered

by the user. The decision maker does not need to provide a numerical
judgement; instead a relative verbal appreciation, more familiar to our daily live,
is sufficient.

ü Third, a consistency check is carried out

ü Forth, a sensitivity analysis is performed. This steps is optional but

recommended as confirmation of the robustness of the results.

3
ü AHP is based on the principle “divide and conquer”. To reduce the complexity of
decision problems, it is advantageous to break them down and solve one ‘sub-
problem’ at a time.
ü The problem is structured according to a hierarchy where the top element is
the goal of the decision. The second level represents the criteria, and the
lowest level represents the alternatives.
ü In more complex hierarchies, more levels can be added. These additional levels
represent the sub-criteria.
4
Example 1
A businessman wants to open a new sports shop in one of three different locations:

a) A shopping center. It has a high concentration of a variety of shops and restaurants. It is

a busy area, with a mix of customers and people walking around. Shops regularly use
large displays and promotions to attract potential customers. As demand for these retail
units is low, the rental costs are reasonable.

b) The city center. It is a busy area, and a meeting point for both young people and
tourists. Attractions such as dance shows, clowns and market stalls are often organized,
which attract a variety of visitors. The city center has several small shops located at
ground level in historical buildings, which suggests high rental costs. These shops have a
high number of customers and are often in competition. 5
c) A new industrial area. It is in the suburbs of the city, where several businesses have
recently been set up. Some buildings have been earmarked for small shops, but on the
whole it has been difficult to attract tenants, which means that rental costs are
currently low. Customers of the existing shops mainly work in the area and only few
customers come from the surrounding towns or cities to shop here.

Given this problem description, four criteria will be considered in making the final decision.

Criterion Explanation
Visibility Probability that a random passer-by notices the shop
Competition Level of competition in the area
Frequency Average number of customers in similar shops in the area
Rental cost Average rental cost by square meter
The hierarchical representation is as follows.

7
As the figure contains redundant information at the lowest level, the alternatives in the
hierarchy are often not repeated and are presented as follows:

8
Three types of priorities have to be calculated:

ü Criteria priorities. Importance of each criterion with respect to the top goal.

ü Local alternative priorities. Importance of an alternative with respect to one

specific criterion.

ü Global alternative priorities. The global alternative priorities rank

alternatives with respect to all criteria and consequently to the overall goal.
9
ü The criteria and local alternatives priorities are calculated using the same
technique. Instead of directly allocating performances to alternatives (or
criteria), AHP uses pairwise comparisons.

ü Pairwise comparisons form the fundamental building blocks of AHP. In

establishing the priorities, AHP will require stating how important each
criterion relative to each other criterion.

ü Pairwise comparisons are generally evaluated on the Saaty fundamental 1–9

scale. In practice, there is no fixed rule and other scales have been
proposed.
10
11
ü Each lower level is prioritized according to its immediate upper level.

ü For instance with regard to example 1,

§ to prioritize the criteria of level 2, an appropriate question would be:

‘Which criterion is most important for choosing the location of the
sports shop and to what extent?’

§ the alternatives in level 3 must be prioritized with regard to each

criterion in level 2. An appropriate question would be: ‘Which
alternative is preferable to fulfil the given criterion and to what
extent?’
12
ü The comparisons are collected in pairwise matrices.
ü The comparisons on the main diagonal are 1 because a given element is
compared with itself.
ü Pairwise matrices are reciprocal.
ü Consider a matrix filled by the pairwise comparison aij

§ If aij > 1 then i is more important than j

§ If aij < 1 then i is less important than j
§ If aij = 1, then both criteria are equally important,
§ aii = 1
13
ü For example, the following pairwise matrix shows that visibility is 1/4 as important
as competition and competition is 4 times as important as visibility.

ü From these comparison matrices, we calculate the local and criteria priorities;
then we aggregate these two priorities to establish the global priority. Priorities
only make sense if they are derived from consistent or near-consistent matrices.
14
Various methods have been proposed to calculate priorities from a pairwise
comparison matrix. In the following, three methods will be introduced:

ü an approximate method,

ü the eigenvalue method, and

ü the geometric mean.

Different methods might lead to different priority vectors, except in the case
of perfect consistency.
15
 16

Approximate method
Many variants have been developed as an approximate method. One of these is
detailed below.

This method is based on the following three-step procedure.

1. Sum the values in each column of the pairwise comparison matrix.

2. Normalize the pairwise comparison matrix by dividing each element in the

matrix by its column total.

3. Calculate the average of the elements in each row of the normalized matrix;
these averages provide the local priorities.
16
Example 2
Let’s consider a car purchasing decision problem. After a preliminary analysis, we
narrowed the list of decision alternatives to three cars. We decided that the following
criteria were relevant for the car selection decision process: Price, Miles per gallon
(MPG), Comfort, and Style.

Data regarding the Price and MPG are provided. However, measures of the Comfort
and Style cannot be specified so directly. Factors such as the car’s interior, type of
audio system, ease of entry, seat adjustments and driver visibility are considered in
order to determine the comfort level of each car. The style criterion is based on
subjective evaluation of the color and the general appearance of each car.
17
ü Note that even when a criterion can be easily measured, subjectivity
becomes an issue whenever a DM indicates his or her personal preference
for the decision alternatives.

ü For instance, the price difference might represent a great deal of money to
one person, but not much of a difference to another person. Thus, whether a
given car is considered ‘extremely more expensive’ than another one or
perhaps only ‘moderately more expensive’ depends upon the financial status
and the subjective opinion of the DM.
ü An advantage of AHP is that it can handle situations in which the unique
subjective judgements of the DM constitute an important part of the
decision-making process. 18
Table below summarizes the information collected about the selected cars.

Characteristics Car 1 Car 2 Car 3

Price 13 100 11 200 9 500
Colour Black Red Blue
MPG 29 33 38
Interior Deluxe Above Average Standard
Body type Four-door midsize Two-door sport Two-door compact
Sound and Navigation System CD + GPS CD + GPS CD

19
The Figure below displays the hierarchy for the car selection problem.

Select the
best car

Price MPG Comfort Style

Car 1 Car 2 Car 3

20
The complete pairwise comparison matrix for the car selection criteria is
as follows:

Price MPG Comfort Style

Price 1 3 2 2
MPG 1/3 1 1/4 1/4
Comfort 1/2 4 1 1/2
Style 1/2 4 2 1

21
The preferences for each decision alternative based on the for criteria are
given below.

Price Car 1 Car 2 Car 3 MPG Car 1 Car 2 Car 3

Car 1 1 1/3 1/4 Car 1 1 1/4 1/6
Car 2 3 1 1/2 Car 2 4 1 1/3
Car 3 4 2 1 Car 3 6 3 1

Comfort Car 1 Car 2 Car 3 Style Car 1 Car 2 Car 3

Car 1 1 2 8 Car 1 1 1/3 1/6
Car 2 1/2 1 6 Car 2 3 1 7
Car 3 1/8 1/6 1 Car 3 6 1/7 1
22
Step 1. Sum the values in each column for the criteria pairwise comparison
matrix.

Price MPG Comfort Style

Price 1 3 2 2
MPG 1/3 1 1/4 1/4
Comfort 1/2 4 1 1/2
Style 1/2 4 2 1
Sum 2.333 12.000 5.250 3.750

23
Step 2. Divide each element of the matrix by its column total.

Price MPG Comfort Style

Price 0.429 0.250 0.381 0.533
MPG 0.143 0.083 0.048 0.067
Comfort 0.214 0.333 0.190 0.133
Style 0.214 0.333 0.381 0.267

24
Step 3. Average the elements in each row to determine the priority of
each criterion.

Price MPG Comfort Style Priority

Price 0.429 0.250 0.381 0.533 0.398
MPG 0.143 0.083 0.048 0.067 0.085
Comfort 0.214 0.333 0.190 0.133 0.218
Style 0.214 0.333 0.381 0.267 0.299

25
ü The AHP synthetization procedure provides the priority of each criterion
in terms of its contribution to the overall goal of selecting the best car.

ü Using the criteria pairwise comparisons, AHP determines that

§ Price with a priority of 0.398 is the most important criterion in the car
selection process.

§ Style with a priority of 0.299 ranks second in importance and is closely

followed by Comfort with a priority of 0.218.

§ MPG is the least important criterion with a priority of 0.085.

26
In order to determine the priority of each car using each criterion, a
similar synthetization is conducted for each previously described pairwise
comparison matrix using the three-step procedure. The computations
provide the four sets of priorities shown below.

Price MPG Comfort Style

Car 1 0.123 0.087 0.593 0.265
Car 2 0.320 0.274 0.341 0.656
Car 3 0.557 0.639 0.065 0.080

27
ü Using this table, we see that Car 3 is the preferred alternative based on
Price (0.557) and on MPG (0.639). Car 1 is the preferred alternative
based on Comfort (0.593). Car 2 is the preferred alternative based on
Style (0.656). At this point, no car is the clear, overall best.

ü To calculate the overall priority we weight each car’s priority by the

corresponding criterion priority and then add the obtained values. This
calculation is as follows:

28
 Price MPG Comfort Style Priority (Pi)
Overall Priority
0.398
Car 1 0.123 0.087 0.593 0.265 0.265
0.085
Car 2 0.320 0.274 0.341 0.656 X = 0.421
0.218
Car 3 0.557 0.639 0.065 0.080 0.314
0.299

The AHP priorities show that Car 2 is preferred.

29
Eigenvalue Method
ü The most popular method to estimate a priority vector is that proposed by Saaty
himself, according to which the priority vector should be the principal eigenvector
of the matrix.

ü In the eigenvalue method, the vector of the priorities 𝑤 is calculated by solving the
equation 𝐴𝑤 = 𝜆𝑤

where 𝑤 = (𝑤! , … , 𝑤" ) and 𝜆 are the eigenvector and eigenvalue of A.

ü If the maximum eigenvalue 𝜆#$% = 𝑛 (the dimension of A), then the matrix is
perfectly consistent. Otherwise, the difference (𝜆#$% − 𝑛) gives an idea about the
inconsistency measure of A . 30
ü To find the eigenvalues and the eigenvectors of a given matrix, we solve
(𝐴 − 𝜆𝐼)𝑝 = 0
we need to find the roots of the characteristic polynomial of A: det(𝐴 − 𝜆𝐼) = 0

Example 3
Find the eigenvalues of the following matrix

det(𝐴 − 𝜆𝐼) = 0 ⟺ 𝜆& (3 − 𝜆) = 0. The eigenvalues are then 0 and 3.

ü Note that, in general if A is consistent, then one eigenvalue, 𝜆#$% is equal to n

and the other is equal to 0, with multiplicity (n−1).
31
Example 4

Find the eigenvalues of

!!"
det 𝐴 − 𝜆𝐼 = 0 ⟺ + 3𝜆! − 𝜆% = 0.
#$

We find that 𝜆&'( ≈ 3.9167.

The other two roots are conjugate complex and we are not interested in them.
Such solution can be easily found by any mathematical software.

Note that 𝜆&'( > 𝑛. Thus, A is inconsistent.

In general, if A is inconsistent then 𝜆&'( is strictly greater than 𝑛. 32

Now, we need to solve the equation system

to avoid the problem of infinitely solutions we add the condition

𝑤# + 𝑤! + 𝑤% = 1

we obtain 𝑤 = (0.660761, 0.131112, 0.208127)) .

33
ü The eigenvalue method has a major drawback: the right–left inconsistency.

ü This drawback was discovered two years after the publication of the
original AHP method.

ü If all comparisons are replaced by their reciprocal values (e.g., 5 becomes

1/5), then the derived ranking should logically also be reversed; which
should lead to a rank reversal phenomenon after an inversion of the scale.

ü For inconsistent matrices of rank n ≥ 4, rankings can be different after a

scale inversion. The following example illustrates this phenomenon.

34
Example 5
Consider the following pairwise matrix, asking “Which alternative is most
economical?”.

If the question is inverted: “Which alternative is most expensive?”, then the

comparisons are simply inverted.

35
The calculated priorities give for the first question the following ranking of the
alternatives: D >B > C > A > E.

However, for the second question the calculated priorities give the following ranking of
the alternatives: B > D > C > A > E. In this case, alternative B is preferred, but before
it was alternative D.
Geometric Mean Method
ü In order to avoid the left–right rank reversal, other methods have also been
proposed for obtaining a vector of weight w from a comparison matrix A.
ü The geometric mean method, proposed by Crawford and Williams in 1985,
is very appealing and is a widely used for practical applications.
ü Accordingly, each component of w is obtained as the geometric mean of
the elements on the respective row divided by a normalization term so
that the components of w eventually add up to 1.

37
Example 6
Let’s consider the following matrix

By using the previous expression, we obtain

𝑤 = (0.119, 0.208, 0.454, 0.219))

38
ü Priorities only make sense if they are derived from consistent or near-
consistent matrices. If the degree of consistency is unacceptable, the DM
should review and revise the pairwise comparisons before proceeding with
the AHP analysis.

ü A consistency check permit to detect possible contradictions in the entries of

a pairwise matrix.

ü The reasons for these contradictions could be vaguely defined problems, a

lack of sufficient information, uncertain information, or lack of concentration.
39
Suppose, for example, that the decision maker gives the following pairwise
comparisons:

ü The shopping center is 2 times more visible than the city center.

ü The city center is 3 times more visible than the industrial area.

ü The industrial area is 4 times more visible than the shopping center.

The third assertion is inconsistent. Based on the two first assertions; the
industrial area should be 6 times more visible than the shopping center.
40
ü A matrix filled by the pairwise comparison aij is called consistent if the
transitivity and the reciprocity rules are respected.

Transitivity Rule:

aij = aik x akj

Reciprocity Rule:

aij = 1/aji

where i, j and k are any alternatives of the matrix.

41
ü The threshold for defining an intolerably inconsistent matrix is not clear.
The most commonly used method was developed by Saaty (1977), who
proposed a consistency index (CI), which is related to the eigenvalue
method.

𝜆&'( − 𝑛
𝐶𝐼 =
𝑛−1

where λmax is the maximal eigenvalue and n is the matrix order.

ü The higher the CI, the more inconsistent the judgments and vice versa.
ü The consistency ratio (CR) is given by

CR = CI/RI

where RI is a random index obtained by averaging the CI of 500 randomly

filled matrices.

ü Saaty (1977) calculated the random indices shown in the Table below

n 1 2 3 4 5 6 7 8 9 10
RI 0 0 0.58 0.9 1.12 1.24 1.32 1.41 1.45 1.49

ü If CR is less than 10%, then the matrix is of an acceptable consistency.

43
ü In addition to the exact mathematical calculation of λmax, to compute
the consistency ratio, an approximation method in 5 steps may be
performed:

ü Step 1. Multiply each value in the ith column of the pairwise

comparison matrix by the priority of the ith item; Then, sum the
values across the rows.

44
Priority 0.398 0.085 0.218 0.299

Price MPG Comfort Style Price MPG Comfort Style Sum

Price 1 3 2 2 0.398 0.255 0.436 0.598 1.687
MPG 1/3 1 1/4 1/4 0.133 0.085 0.054 0.075 0.347
Comfort 1/2 4 1 1/2 0.199 0.340 0.218 0.149 0.907
Style 1/2 4 2 1 0.199 0.340 0.436 0.299 1.274
Step 2. Divide the elements of the weighted sum vector obtained in step 1
by the corresponding priority for each criterion.

Priority (Pi) Price MPG Comfort Style Sum Sum/Pi

0.398 0.398 0.255 0.436 0.598 1.687 4.236
0.085 0.133 0.085 0.054 0.075 0.347 4.077
0.218 0.199 0.340 0.218 0.149 0.907 4.163
0.299 0.199 0.340 0.436 0.299 1.274 4.264

46
Step 3. Calculate the average of the values found in step 2; this average
estimated an approximative value of 𝛌max.

Price MPG Comfort Style Sum Sum/Pi

0.398 0.255 0.436 0.598 1.687 4.236
0.133 0.085 0.054 0.075 0.347 4.077
0.199 0.340 0.218 0.149 0.907 4.163
0.199 0.340 0.436 0.299 1.274 4.264
Average = 4.185

47
Step 4. Calculate the consistency index (CI) as follows:

𝜆&'( − 𝑛 4.185 − 4
𝐶𝐼 = = = 0.0616
𝑛−1 4−1
Step 5. Compute the consistency ratio:

𝐶𝐼 0.0616
𝐶𝑅 = = = 0.068 < 0.10
𝑅𝐼 0.9

Thus, the degree of consistency of the pairwise comparisons is acceptable

48
Example 5
We want to choose one of two candidates to appoint a supervisor in an emergency
department of a hospital.
Two criteria were selected: technical skills and supervisory skills.
The second criterion is assumed to be twice as important as the first.
By studying the CVs of the two candidates it was judged that:
• according to the 1st criterion, the 2nd candidate is 4 times more competent than the
1st.
• according to the 2nd criterion, the 1st candidate is 3 times more competent than the
2nd.
Use the approximate method to make a decision. 49