Interval Estimations of Global Weights in

AHP by Upper Approximation

Tomoe Entani a Hideo Tanaka b

a
Kochi University
2-5-1 Akebono Kochi 780-8520, JAPAN
TEL: +81-88-844-8212, FAX: +81-88-844-8225
Email: [email protected]
b
Hiroshima International University
Gakuendai 555-36,Kurose, Hiroshima 724-0695, JAPAN
TEL: +81-823-70-4886, FAX: +81-823-70-4852
Email: [email protected]

Abstract

In Analytic Hierarchy Process (AHP) structured hierarchically as several criteria

and alternatives, the priority of an alternative is obtained by using the pairwise
comparisons based on decision maker’s intuition. Thus, the given comparisons are
uncertain and inconsistent each other. We use the interval approach for obtaining
interval evaluations which are suitable for handling uncertain data. Since the given
comparisons are ratio measures and too large intervals are not useful information,
the intervals should be normalized and their redundancy should be reduced. We
introduce interval probability which fills the role of interval normalization instead
of crisp normalization in the estimations at each hierarchy. Then, as a final decision,
the interval global weights reflecting a decision maker’s uncertain judgements as
their widths without redundancy are obtained.

Key words: Analytic hierarchy process, Interval local and global weights, Interval
referenced priority weights, Upper approximations

1 Introduction

Analytic Hierarchy Process (AHP) is a useful method in multi-criteria decision

making problems [4]. It is structured hierarchically as criteria and alternatives
and proposed to determine the priority weights of alternatives which are called
the global weights. From a pairwise comparison matrix for criteria, the ref-
erenced priority weights are obtained by the eigenvector method [4]. In the

Preprint submitted to Elsevier Science 20 March 2007

same way, from a pairwise comparison matrix for alternatives under each cri-
terion, the local weights for the criterion are obtained. The elements of the two
types of comparison matrices are relative measurements given by a decision
maker. The obtained weights from the matrix can reflect his/her attitude in
the actual decision problem. The weights obtained by the conventional AHP
lead to a linear order of alternatives. Uncertainty of an order of alternatives
in AHP is discussed in [5]. As shown in [5], there exists a problem that pair-
wise comparisons might be uncertain and inconsistent with each other because
they are based on human intuition. The approaches for dealing with interval
and fuzzy comparisons have been proposed in [1] and [3], respectively. It is
easier for a decision maker to give interval or fuzzy comparisons than crisp
ones, since they are suitable to represent uncertain human judgements. The
approach with interval weights [1] is rather complex comparing to our ap-
proach which can handle interval data, because the approach [1] is based on
solving problems on all vertices for obtaining interval weights. In case of fuzzy
comparisons [3], the optimal weight pattern is obtained as crisp by maximiz-
ing the degree of consistency in a sense of fuzzy comparisons. Although we
propose the model to obtain interval weights from the crisp comparisons, the
model can be extended to interval comparisons and the obtained weights are
intervals. In the similar setting to [1], the approaches for dealing with decision
maker’s preference statements instead of pairwise comparisons are proposed in
[6]. This seems to be very practical, but obtaining the upper and lower bounds
of interval weights has been proposed without defining the interval weights.

In this paper, it is assumed that the estimated referenced priority and local
weights are intervals to reflect inconsistency of the given pairwise comparisons.
The widths of the obtained interval weights can be regarded as the index of
inconsistency among the given comparisons. If the consistent comparisons are
given, we can obtain crisp weights by the proposed approach and they are the
same as ones obtained by eigenvector method [7]. With the obtained interval
referenced priority and local weights, the global weights of all alternatives are
also obtained as intervals. However, their widths tend to be large if we use
interval arithmetic. Thus, the obtained intervals by interval arithmetic can not
be a meaningful as long as unnecessary widths of intervals are not removed.
The optimization problems to obtain global (aggregated) evaluation as fuzzy
number is proposed in [3], where the utilities under criteria representing the
local weighs are crisp and the optimal weight pattern representing the refer-
enced priority weights is determined from the given fuzzy comparisons. We
propose to obtain the global weights of alternatives as follows: first the crisp
referenced priority weight within the obtained interval one is multiplied by
the interval local weight under the criterion and then they are added over all
criteria. The crisp referenced priority weights are selected such that their sum
becomes one. The interval global weights are also obtained as interval prob-
abilities so that they are normalized to reduce redundancy. The normalized
interval global weights are useful information for a final decision in the sense

2
that they reflect a decision maker’s uncertain judgements as intervals without
redundancy.

2 Crisp Weights by Conventional Eigenvector Method

AHP is a method to deal with the weights with respect to many alternatives
and to determine the priority weight of each alternative [4]. When there are n
alternatives, a decision maker compares a pair of alternatives for all possible
pairs to obtain a pairwise comparison matrix A as follows.
 
 1 ··· a1n 

 .. .. 

A = [aij ] =  . aij . 
 
 
an1 · · · 1

where aij shows the priority ratio of alternative i comparing to alternative

j and they satisfy the following relations so that the decision maker gives
n(n − 1)/2 comparisons.

Diagonal elements aii = 1

(1)
Reciprocal elements aij = 1/aji

where it follows from the reciprocal relation that the assumed model is a ratio
model.

From the given comparison matrix, the crisp priority weights w ∗i (i = 1, ..., n)
are obtained by eigenvector method. The eigenvector problem is formulated
as follows.

Aw = λw (2)

where λ is the eigenvalue and w is the eigenvector. Solving (2), the eigenvector
w ∗ = (w1∗ , . . . , wn∗ )t corresponding to the principal eigenvalue is obtained as
the priority weight vector. It is noted that the sum of the obtained weights
w ∗i (i = 1, ..., n) is normalized to be one.

3
3 Interval Weights by Approximation Models

The given pairwise comparison aij is approximated by the ratio of priority

weights, wi and wj , symbolically written as aij ≈ wi /wj .

Assuming the priority weight wi as an interval, we obtain the interval priority

weights denoted as Wi = [wi , wi ]. Then, the approximated pairwise compari-
son with the interval weights is defined as the following interval.
 
Wi w i wi
= ,
Wj w j wj

where the upper and lower bounds of the approximated comparison are defined
as the maximum range.

3.1 Interval Normalization

First it should be noted that the sum of weights obtained by the conventional
AHP is normalized to be one. Therefore, we consider interval probability pro-
posed in [9] so as to normalize the interval weights. Their conditions are defined
as follows.

Definition 1 Interval weights (W1 , ..., Wn ) are called interval probabilities if

and only if

i=j wi + wj ≥ 1 ∀j
(3)
i=j wi + wj ≤ 1 ∀j

where Wi = [wi , wi ].

It can be said that the conventional normalization which makes the sum of
values is one is extended to the interval normalization by using the above
conditions. The sums of the bounds of intervals are constrained by one in
two ways. These conditions make the intervals meaningful in the sense that
there are elements in the intervals so as to make the sum of them be one. In
order to make intervals be normalized, their redundancy should be reduced
by (3) as the constraint conditions in the optimization problem described in
the subsection 3.2.

4
3.2 Upper Approximations of Crisp Pairwise Comparisons

The proposed approach is based on the view that the interval weights are
obtained so as to include the given interval [7]. The obtained interval weights
satisfy the following inclusion relations.
 
Wi wi wi
aij ⊆ = , ∀(i, j)
Wj wj w j

which means
wi wi
≤ aij ≤ ∀(i, j). (4)
wj wj

The approximated interval weights should be estimated as closely as possible

to the given comparisons subject to the above inclusion relations. The concept
of the least upper approximation [8], that is the width of each interval weight
must be minimized, is applied. In the following LP problem, simply the sum
of widths of all weights is minimized under the constraints.

min i (w i − wi )
s.t. i=j w i + wj ≥ 1 ∀j
w i + wj ≤ 1 ∀j
i=j (5)
wi wi
≤ aij ≤ ∀(i, j)
wj wj
wi ≥ wi ≥  ∀i

where constraint conditions consist of the interval normalization (3) and in-
clusion relations (4). (5) is called the upper approximation model.

The interval weights by (5) include the given inconsistent comparisons data.
The width of the interval represents uncertainty of each weight and the least
uncertain weights are obtained by solving (5).

3.3 Interval Global Weights

The decision problem in AHP is structured hierarchically as criteria and al-

ternatives as in Fig.1.

A decision maker gives pairwise comparison matrix for alternatives Ai (i =

5
 decision

C1 ･･ Ck ･･ Cm

A1 ･･ Ai ･･ An

Fig. 1. Structure of decision problem in AHP

1, ..., n) under each criterion and also comparison matrix for criteria Ck (k =
1, ..., m) comparing alternatives and criteria importance, respectively. In Fig.1,
the criteria are at one layer, however, it is possible to construct several layers
of criteria. By the proposed approximation model, the local weight of alter-
native Ai under criterion Ck is denoted as Wki = [w ki , w ki ] and the referenced
priority weight of criterion Ck is denoted as Pk = [pk , pk ]. The global weight of
alternative Ai is obtained as Wi = k Pk Wki by interval arithmetic. It is the
sum of multiplications of the referenced priority weights and corresponding
local weights so that it represents the priority of an alternative considering all
criteria. In case of several layers of criteria, the global weights can be obtained
by repeating the similar calculation. The local and referenced priority weights
are intervals so that the global weights tend to be obtained as large intervals
because of interval arithmetic. Therefore, we consider that the global weight is
obtained by the crisp referenced priority weights, whose sum is one, within the
interval weights. The upper bound of interval global weight of the alternative
Ai is obtained as follows.

wi = max k pi∗
k w ki

s.t. k pi∗
k = 1
(6)
pk ≤ pi∗
k ≤ pk ∀k

where pi∗
k (k = 1, ..., m) are decision variables for the crisp referenced priority
weights of criteria that maximize the upper bound of the interval global weight
of the alternative Ai .

Similarly, the problem to obtain the lower bound of interval global weight of

6
the alternative Ai is formulated as follows.

wi = min k pik∗ w ki
s.t. k pik∗ = 1 (7)
pk ≤ pik∗ ≤ pk ∀k

The interval global weight of the alternative Ai can be denoted as Wi =

[w i , wi ]. The bounds satisfy pi∗ i
k w ki ≤ pk∗ w ki because of maximizing and min-
imizing the objective functions, respectively. It represents the possible range
under the condition that the sum of referenced priority weights is one.

The interval global weights reflect inconsistency in the given comparisons for
criteria and alternatives under each criterion. The crisp referenced priority
weights are determined by maximizing and minimizing the global weight of
each alternative Ai , respectively. The selected crisp weights for the upper and
lower bounds of interval global weight are different each other. (6) and (7) are
formulated for each alternative Ai so that the crisp referenced priority weights
are also different among alternatives. They depend on local weights of the
alternative under the criteria. The obtained interval global weights satisfy the
conditions in (3). This fact is proved in Appendix. By the proposed models for
global weights (6) and (7), the interval global weights are obtained as interval
probabilities, that is, they are normalized.

4 Numerical Example

There are 6 alternatives (A1, A2, A3, A4, A5, A6) and 6 criteria (C1, C2,
C3, C4, C5, C6). A decision maker gives pairwise comparison matrices for 6
criteria and 6 alternatives under each criterion. Then, the global weights of all
alternatives are calculated from them by the proposed models.

The decision maker compares all the pairs of criteria and gives the compar-
isons. The pairwise comparison matrix for 6 criteria followed by (1) is as

7
follows.
 
 1 1/9 1/8 1/9 1/9 1/9 
 
9 1 2 1/5 1/6 6 
 
 
 
8 1/2 1 1/7 1/6 5 
 
Pairwise comparison matrix for criteria =  
 
9 5 7 1 2 7 
 
 
9 6 6 1/2 1 5 
 
 
9 1/6 1/5 1/7 1/5 1

The interval referenced priority weights obtained from the above matrix by
the upper approximation model (5) are shown in Table 1.
Table 1
Interval referenced priority weights of criteria
Criterion Referenced priority Width
C1 [0.008, 0.056] 0.048
C2 [0.042, 0.101] 0.059
C3 [0.042, 0.072] 0.030
C4 0.505 —
C5 0.253 —
C6 [0.014, 0.072] 0.058

The decision maker also gives 6 pairwise comparison matrices for alternatives.
As an example, the pairwise comparison matrix under C1 followed by (1) is
as follows.

Pairwise comparison matrix for alternatives under the criterion C1

 

1 1 1/6 2 1 1/7 
 
 1 1 1/6 2 1 1/7 
 
 
 
 6 6 1 7 6 1/2 
 
=  
 
 1/2 1/2 1/7 1 1/2 1/8 
 
 
 7 1 1/6 2 1 1/8 
 
 
7 7 2 8 8 1

The interval local weights obtained from the above matrix is

W 1 = ([0.042, 0.072], [0.042, 0.072], 0.253, [0.032, 0.063], [0.042, 0.063], 0.507)t.

8
In the same way, the decision maker gives pairwise comparison matrices under
other criteria and the obtained interval local weights are shown in Table 2.
Table 2
Interval local weights of alternatives under each criterion
Local weight
Alternative C1 C2 C3
A1 [0.042,0.072] 0.267 0.222
A2 [0.042,0.072] 0.133 [0.111,0.222]
A3 0.253 0.133 [0.056,0.111]
A4 [0.032,0.063] [0.067,0.167] 0.222
A5 [0.042,0.063] 0.133 [0.111,0.222]
A6 0.507 [0.167,0.267] 0.111

Local weight
Alternative C4 C5 C6
A1 0.308 [0.071,0.143] 0.118
A2 0.154 [0.143,0.286] 0.235
A3 0.154 0.286 [0.235,0.275]
A4 [0.077,0.154] 0.143 [0.059,0.118]
A5 [0.077,0.154] [0.071,0.143] [0.137,0.235]
A6 0.154 0.143 0.118

In Table 3 and Fig.2, the interval global weights obtained by (6) and (7) are
shown. From the widths of global weights, uncertainties of A3 and A6 are
smaller than those of A2,A4 and A5. The widths of interval global weights are
denoted as index of inconsistency in the given information. The interval global
weights are interval probabilities since they satisfy the conditions in (3). They
are normalized so that the intervals are not too large or too small and the sum
of crisp weights within the intervals is one. Intervals without redundancy are
useful and helpful information for a final decision.

Fig.3 illustrates the partial order relation of alternatives based on the obtained
interval global weights. It should be noted that interval order relation is de-
fined in [?], where A = [a, a] ≤ B = [b, b] holds if and only if a ≤ b and a ≤ b.
A1 is superior and A4 and A5 are not superior to the other alternatives. The
conventional crisp global weights lead the linear order relation, however, the
decision maker’s evaluations are more likely partial order because of uncertain
judgements. The interval global weights are suitable to represent the uncer-
tainty in the decision maker’s judgements, because they lead the partial order

9
Table 3
Interval global weights of alternatives
Alternative Global weight Width
A1 [0.213,0.233] 0.020
A2 [0.141,0.187] 0.046
A3 [0.191,0.197] 0.007
A4 [0.096,0.150] 0.054
A5 [0.083,0.156] 0.073
A6 [0.169,0.179] 0.010

relation of alternatives.

In Table 4, the selected crisp referenced priority weights of C1,C2,C3 and C6

for the upper bound of interval global weights of A2 and A6 by (6) are shown.
As for A2, its local weights under C3 and C6 are large among all alternatives
from Table 2. The crisp referenced priority weights of the two criteria are equal
to the upper bound of intervals in order to maximize the global weight. As
for A6, on the other hand, local weights under C1 and C2 are large among
all alternatives, their crisp referenced priority weights are the same as the
upper bound of intervals. The referenced priority weights are determined by
reflecting each alternative’s local weights.

0.25

0.20 A1
A3
0.15 A6
A2
0.10
A4
A5
0.05

0.00

Fig. 2. Interval global weights

Table 4
Selected crisp referenced priority weights for upper bounds of interval global weights
of alternatives A2 and A6 by (6)
C1 C2 C3 C6
A2 0.008 0.090 0.072* 0.072*
A6 0.056* 0.101* 0.042 0.043
* illustrates that it equals to the upper bound of interval referenced priority weight.

10
 1A
3A
2A 6A
4A 5A
Fig. 3. Partial order relation of alternatives
5 Conclusion
In the decision problem in AHP, the global weights of alternatives considering
criteria are obtained as a final decision. In each aspect, such as comparing crite-
ria or alternatives under each criterion, a decision maker gives intuitive judge-
ments as pairwise comparisons. The given comparisons in each aspect might
be inconsistent each other. Since the global weights are calculated through
these aspects, this paper emphasizes that the obtained global weights should
reflect the uncertainty of all the given information.
In this paper, the interval weights, which are suitable for representing un-
certain information, are obtained from the given crisp pairwise comparison
matrix. The interval weights are determined so as to include the given com-
parisons and minimize the widths according to the least upper approximation
approach. The widths of the obtained interval weights reflect inconsistency
in the given pairwise comparison matrix based on a decision maker’s intu-
itive judgements. When the pairwise comparison matrices for criteria and al-
ternatives under each criterion are given, the referenced priority and local
weights are obtained as intervals, respectively, by the proposed upper approx-
imation model (5). The crisp referenced priority weights are selected such
that their sum becomes one. For the upper and lower bounds of the inter-
val global weights, they are determined by maximizing and minimizing the
global weights for each alternative in (6) and (7), respectively. Therefore, the
selected crisp weights depend on the bounds of interval global weights and
alternatives. They reflect the local weights of each alternative. The interval
global weights reflect decision maker’s uncertain judgements and their widths
can be regarded as index of inconsistency. By the proposed global weights
models (6) and (7), the interval global weights are also obtained as interval
probabilities. Such normalized interval global weights are useful information
for a final decision in the sense that they reflect a decision maker’s uncertain
judgements as intervals without redundancy.
11
Appendix

The interval global weight of alternative Ai is obtained as

Wi = [ k pik∗ w ki , k pi∗ i i∗
k w ki ] where pk∗ and pk (k = 1, ..., m) are the optimal
solutions of (6) and (7) for each alternative, respectively. The first condition
of interval probabilities (3) is verified as follows.

i=j k pi∗
k w ki + k pjk∗ wkj
≥ i=j k pjk∗ w ki + k pjk∗w kj (i)
= k pjk∗ ( i=j wki + wkj ) (ii)
≥ k pjk∗ (iii)
=1

(i) The referenced priority weights pjk∗ (k = 1, ..., m) are the optimal solutions
of (7) for alternative Aj . They can be the possible solutions of (6) for alterna-
tive Ai , since the constraint conditions of (6) and (7) are the same. It follows
j
that k pi∗ k w ki ≥ k pk∗ w ki .

(ii)The interval local weights Wki = [w ki , wki ](k = 1, ..., m) for each criterion
Ck are interval probabilities so that it holds i=j w ki + w kj ≥ 1 from the
condition of interval probabilities (3).

(iii)The sum of the referenced priority weights pjk∗ (k = 1, ..., m) are normalized
to be one in the constraint condition of (6).

The second condition in (3) is verified in the similar way as follows.

i=j k pik∗ w ki + k pj∗

k w kj

≤ i=j k pj∗
k w ki + k pj∗
k w kj

= k pj∗
k ( i=j wki + wkj )
≤ k pj∗
k = 1

Then, the interval global weights Wi (i = 1, ..., n) satisfy both of the two
conditions in (3). Thus, they are interval probabilities. (Q.E.D.)

12
References

[1] A.Arbel (1989) “Approximation Articulation of Preference and Priority

Deviation, ” European Journal of Operational Research, 43:317-326.

[2] D.Dubois and H.Prade (1980) “Systems of Linear Fuzzy Constraints,” Fuzzy
Sets and Systems, 3:37-48.

[3] S.Ohnishi, D.Dubois, H.Prade and T.Yamanoi (2006) “A Study on the Analytic
Hierarchy Process,” Proc. of the 11th International Conference on Information
Processing and Management of Uncertainty in Knowledge-Based Systems
(IPMU’06), Paris, France.

[4] T.L.Saaty (1980) The Analytic Hierarchy Process, McGraw-Hill, NY.

[5] T.L.Saaty and L.G.Vergas (1987) “Uncertainty and Rank Order in the Analytic
Hierarchy Process”, European Journal of Operational Research, 32:107-117.

[6] A.A.Solo and R.P.Hamalainen (1995) “Preference Performing through

Approximate Ratio Comparisons,” European Journal of Operational Research,
82:458-475.

[7] K.Sugihara and H.Tanaka (2001) “Interval Evaluations in the Analytic

Hierarchy Process by Possibility Analysis,” International Journal of
Computational Intelligence, 17,3:567-579.

[8] H.Tanaka and P.Guo (1999) Possibilistic Data Analysis for Operation Research,
Physica-Verlag, A Springer Verlag Company.

[9] H.Tanaka, K.Sugihara and Y.Maedab (2004) “Non-additive Measures by

Interval Probability Functions,” Information Sciences, 164:209-227.

13