You are on page 1of 9

Spectral analysis of inconsistent interval comparison matrices in AHP

Ladislav Luk1

1 Introduction
AHP with weight generation for comparison matrices ~ one of the most popular DM methods in practice. Basic features pairwise comparisons of individual variants source: quantitative data available for each decision variant (e.g. quantitative measurement, expert assigned values, etc) => pairwise comparison matrix selection of proper scales for measures / ranges plays very important role.

2 Generation of weights in AHP


In developing AHP, Saaty [4] proposed a measure for inconsistency of a pairwise (n,n) comparison matrix A known as consistency index cI cI = (max - n)/(n - 1), (1) max the greates eigenvalue of Aw = w. However, cI depends upon dimension n ~ drawback ~ it complicates its versatility in many practical situations.
Univ.of West Bohemia, Faculty of Econ., Dept.of Econ. and Quantitative Methods, Husova 11, 30614 Pilsen, lukasl@kem.zcu.cz
1

Saatys revision a new measure proposed, i.e. consistency ratio cR cR = cI / rI(n), (4) rI(n) mean value of cI takled as a random variable calculated by (1) from random matrices formed by comparing fractions wi/wj, i.e. rI(n) = E(cI) => cR [0,1] Limit cases cR = 0 the ideal case with perfect transversality of comparison relations, cR = 1 a random matrix lacking any detectable ones.

Theory of AHP is rather rich ~ a lot of generalizations pairwise comparison matrix A will be an interval matrix aij a range of importance between i-th and j-th objectives, i.e. the i-th objective is from lij till uij times more important than the j-th one. A=
1 [l , u ] 21 21 M [l n1 , u n1 ]

[l12 , u12 ]
1 M

L L M L

[l1n , u1n ]
M 1

[l n 2 , u n 2 ]

(5)

[lij, uij] real interval number :: lij, uij > 0, lij uij , lij = 1/uji , and uij = 1/lji , i,j=1,2,...,n. aii diagonal element ~ convention: aii=lii=uii = 1, i=1,2,...,n ~ precisely aii= [1,1]. ~> A reciprocal interval matrix. Computational methods for generation of weights in case of interval comparison matrix A based upon Wang, Y-M., et al. (2005) NLP approach. (present paper) statistical sampling technique ~ {expressions (6) (8)} :: crucial step solving spectral problems of sampled matrices A from A.

3 Computational results
Two examples of interval comparison matrices given Wang, Y-M., et al. (2005) to investigate influence of interval consistency, role of reciprocity. Definition: an interval pairwise comparison matrix A (5) is interval consistent iff max (lik lkj) min (uik ukj) , for all i,j,k=1,,n.
k k

(9)

Suitable representation A is represented by a couple of crisp matrices (A1, A2).

3.1 First example consistent interval pairwise comparison matrix


A=
[2,5] 1 1 1 1 , 5 2 1 1 1 , ,1 4 2 3 1 ,1 1 ,1 3 2 [2,4] [1,3] [1,3] [1,2]
1 1 , 2 ,1 1
1 0 . 5 0 . 5 1 5 4 3 1 3 2 . 1 1 1 1 2 1

[1,2]

(10)

A1 =

2 1 0.2 1 . 0.25 0.3 3 . 0.3 3 0.25

1 1 1 , 1 0.5 1 1 2

A2 =

We have focused upon solving spectral problems of matrices sampled from A. Fig.1 presents eigenvalues of the matrices A1, A2 and A3 = (A1+A2)/2

we see max(A1) =3 < 4 strange value ~ as 4 =!= the theoretical limit provided by AHP theory, however, all matrices A1, A 2, A 3 violate the reciprocity, find that the averaged matrix A3 in spite of such defect yields rather anticipated value max(A3) =4.43 .

Fig. 1 Eigenvalues of matrices A1, A2 and A3 = (A1+A2)/2

Next step generate matrices A by sampling matrix A assuming uniform distribution of values within each given interval. 1. the first set of six matrices Am, m=4,..,9 simply generated without taking particular care on reciprocity, thus enabling a weaker interpretation of interval reciprocity since the matrix A (10) fulfils the interval reciprocity requirements. 2. the second set Am , m=14,..,19 on the contrary, was generated with taking care on reciprocity strictly. Note: the identical seeds for pseudorandom number generating processes were used in order to maintain the same sampling of corresponding matrices from the first and the second group, respectively, in order to investigate the role of neglecting/preserving reciprocity of sampled matrices. Fig.2 and Fig.4 eigenvalue clusters of the first and the second group of matrices Am, respectively. Fig.3 and Fig.5 comparisons of these eigenvalue clusters with the eigenvalues of matrices A1, A 2, A 3, as well.

Fig. 2 Eigenvalues of matrices Am, m=4,..,9

Fig. 3 Eigenvalues of matrices Am, m=1,..,9

Fig. 4 Eigenvalues of matrices Am, m=14,..,19

Fig. 5 Eigenvalues of matrices Am, m=1,2,3, 14,..,19

Results Table 1: max calculated by Mathematica module Eigensystem and corresponding values of cI (1), Table 2: means and standard deviations of calculated eigenvalues.
A1, A2, A3 cI Am, m=4,..,9 cI Am, m=14,..,19 cI 3.00961 -0.330131 4.32007 0.106691 4.10926 0.0364213 5.78399 0.594663 4.27582 0.0919394 4.07403 0.0246782 4.39638 0.132127 4.58198 0.193995 4.13811 0.046038 4.53713 0.179042 4.27164 0.0905468 4.48681 0.162271 4.06005 0.0200169 4.06211 0.122555 4.06211 0.0207049

Table 1 max and cI of matrices Am (1-st example)

E(max) Am, m=4,..,9 Am, m=14,..,19 4.42825 4.1192

Std.dev.( max) 0.12452 0.0806125

|max(max - E(max))| 0.153738 0.152437

Table 2 Means, standard deviations and the largest deviations of max of matrices Am (1-st example)

3.2 Second example inconsistent interval pairwise comparison matrix


The second example is processed exactly in the same way as the first one, and the interval matrix A is given by (11), now. The figures and tables of results obtained are issued so that to enable specific comparisons. Now, the interval matrix A is given by (11), and A is computationaly represented by a couple of crisp matrices (A 1, A 2) again
[1,2] [1,2] 1 1 [3,5] 1 ,1 2 1 1 1 1 ,1 , 2 5 3 1 , 1 1 , 1 1 , 1 3 2 5 4 8 6 [2,3] [4,5]
[6,8] , 1

A=

(11)
3 5 . 8 1

1 1 1 0 .5 3 A1= 0.5 012 1 . . 0.3 3 0.2 0.125

2 4 , 6 1

2 2 1 1 . A2= 1 0.1 3 5 3 1 & 0.5 0.25 0.16

Test of consistency (9) applied for upper triangular matrix elements ~> {T,F,F,F,T,F,F,T,F,T} => A (11) is not consistent, now.

Fig. 6 Eigenvalues of matrices A1, A2 and A3 = (A1+A2)/2

Fig. 7 Eigenvalues of matrices Am, m=4,..,9

Fig. 8 Eigenvalues of matrices Am, m=1,..,9

Fig. 9 Eigenvalues of matrices Am, m=14,..,19

Fig. 10 Eigenvalues of matrices Am, m=1,2,3, 14,..,19

The Fig.6 shows the max of A1, A2 and A3 = (A1+A2)/2 again. we inspect max(A1) < 4, too, however much less than in the previous example, which seems rather interesting. all matrices A1, A 2, A 3 violate reciprocity relations again, but we may observe shifts of their max to higher values except for the matrix A2. Sampled matrices 1. Am, m=4,..,9 generated optionally with neglecting reciprocity, 2. Am , m=14,..,19 preserving reciprocity, again. Note: the same sampling procedure was used as in the previous example. Fig.7 and Fig.8 eigenvalue clusters of Am, m=4,..,9, Fig.9 and Fig.10 similar results for Am , m=14,..,19. Results

Tables 3 and 4 are similar to Tables 1 and 2 giving the corresponding results of the second example, i.e. A given by (11).

A1, A2, A3 cI Am, m=4,..,9 cI Am, m=14,..,19 cI

3.80186 -0.0660466 4.5771 0.192365 4.51751 0.172503

5.45491 0.484969 4.54068 0.180226 4.48237 0.160791

4.62407 0.208023 4.62312 0.207706 4.56664 0.18888 4.72348 0.24116 4.57735 0.19245 4.65709 0.21903 4.44305 0.147684 4.57174 0.19058 4.53061 0.17687

Table 3 max and cI of matrices Am (2-nd example) E(max) Am, m=4,..,9 Am, m=14,..,19 4.61553 4.51959 Std.dev.( max) 0.0669853 0.0508309 |max(max - E(max))| 0.107945 0.0765371

Table 4 Means, standard deviations and the largest deviations of max of matrices Am (2-nd example)

Mathematica code snippets


(* build matrix A19, call Eigensystem[], print & plot results *) A19=Table[ raij[i,j],{i,4},{j,4}]; (* maintain the reciprocity by force ! *) For[i=2,i4,i++, For[j=1,j<i,j++,Part[A19,i,j]=1/Part[A19,j,i]] ] Print["def.Orange ~ color09, A19 ~ reciproc.rand.convex(A01,A02) := ",color09," , ",MatrixForm[A19]] {eigval19,eigvec19}=Eigensystem[N[A19]] Print["A19 eigenvalues:= ",eigval19] Print["A19 eigenvectors:= ",eigvec19] reEigval19=Table[Re[eigval09[[n]]],{n,4}] imEigval19=Table[Im[eigval09[[n]]],{n,4}] Print["A19 Re[eigenvalues]:= ",reEigval19] Print["A19 Im[eigenvalues]:= ",imEigval19] lpEigval19=ListPlot[Table[{reEigval19[[n]],imEigval19[[n]]}, {n,4}], AxesLabel{Re_axis,Im_axis}, PlotStyle{cList[[9]],PointSize[0.02]}]

4 Conclusion
We have analyzed two examples of interval pairwise comparison matrices of AHP and discussed their spectral characterictics in detail. Computational analysis shows the differences between both types of matrices. In particular, the inconsistent matrices yield higher consistency indeces, which corresponds to natural estimations.

Acknowledgements
Both supports of this work by the grant project no. 402/09/1536 of the GACR, and the reseach project no. LC06075 of the Cz MSMT, are highly acknowledged.
References
[1] Dadkhah, K.M. Zahedi, F.: A mathematical treatment of inconsistency in the analytic hierarchy process. Math.Comput.Modelling. Vol.17, No.4/5, pp.111-122, Pergamon Press 1993. [2] Hofman, J. Luk, L.: Multi-objective decision making in marketing management. (in czech), Proc. of 2-nd Seminar Computational Economics (Vpotov ekonomie), Publ. Univ. of West Bohemia, Pilsen 2004, pp.93-102, ISBN 80-7043-342-6. [3] Jablonsk, J.: Measuring the efficiency of production Mathematical and Computer Modelling 46 (2007), 1091-1098. units by AHP models.

[4] Saaty, T. L.: The Analytic Hierarchy Process. New York, Mc Graw Hill 1980. [5] Wang, Y-M. Yang, J-B. Xu.D-L.: Interval weight generation approaches based on consistency test and interval comparison matrices. Appl. Math. and Comput. 167 (2005) 252-273, Elsevier 2005.

You might also like