SDA 8: The Analytic Hierarchy Process The importances of the criteria could be approximated by the AHP using pairwise comparisons: • T. L. Saaty, The Analytic Hierarchy Process: Planning, Priority Setting, Resource Allocation, McGraw-Hill, New York, 1980. Suppose that the value function has the form 1 v(y) = q wiyi i=1 If wi = 0, the corresponding outcome yi can be deleted from consideration. Thus, we shall assume that wi > 0, i = 1, 2, . . . , q. Define the weight ratio by wi wij = . wj Note that, for any i, j, k indexes −1, wij = wji wij = wik wkj . 2 Define the matrix of weight ratios as W = [wij ]q×q : w1 w1 w w 1 2 w w 2 2 w1 w2 . . . . . wq wq w1 w2 w1 w1 ... w3 wq w2 w2 ... w3 wq . . . . . . . . wq wq ... w3 wq A matrix W is called consistent if its com−1, ponents satisfy the equalities wij = wji wij = wik wkj for any i, j and k. 3 Observe that: • Since each row of W is a multiple of the first row, the rank of W is one, and thus there is onlu one nonzero eigenvalue which is q. This due to the fact that wii = 1 and that the sum of all eigenvalues is equal q to the trace of W (i.e. i=1 wii = q). • We can easily chack that W w = qw therefore w must be the eigenvector of W corresponding to the maximum eigenvalue q. 4 As a living system, human perception and judgment are subject to change when the information inputs or psychological states of the decsion maker change. A fixed weight vector is difficult to find. Saaty proposed the following to overcome this difficulty: Estimate or elicit the weight ratio wij by aij and let A = [aij ]q×q be the matrix of components {aij }. Note that as each wij > 0, we expect and shall assume that all aij > 0. 5 −1, Saaty sugFurthermore, as wij = wji gested that in practice, only aij , j > i need to be assessed. Since A is found as an approximate for W , when the consistency conditions are almost satisfied for A, one would expect that the normalized eigenvector corresponding to the maximum eigenvector of A, denoted by λmax, will also be close to w. Theorem 1. The maximum eigenvalue, λmax, of A is a positive real number. Let ŵ be the normalized eigenvector corresponding to λmax of A. Then ŵi > 0 for all 1 ≤ i ≤ q. 6 Theorem 2. The maximum eigenvalue of A satisfies the inequality λmax ≥ q. Assume we have q objectives and we want to construct a scale, rating these objectives as to their importance with respect to the decision, as seen by the analyst. We ask the decision maker to compare the objectives in paired comparisons. If we are comparing objective i with objective j, we assign the values aij and aji as follows: 7 • aij = a−1 ji • If objective i is more important than objective j then aij gets assigned a number as follows: Note that the above observation is valid for any matrix which is consistent. 8 Intensity of relative importance Definition 1 equal importance 3 weak importance (of one over the other) 5 strong importance 7 demonstrated importance over the other 9 absolute importance 2,4,6,8 intermediate values between Saaty’s scale of relative importances. 9 Example 1. Let us consider the following matrix 1 9 7 1 1 1 A = 9 5 1 5 1 7 To find λmax we solve det[A − λI] = 0 that is, 10 1−λ 9 7 1 1 1 − λ det 9 5 1 5 1−λ 7 = (1 − λ)3 − 3(1 − λ) + 9/35 + 35/9 = 0 The maximum solution is λmax = 3.21. After normalization we get 11 ŵ1 = 0.77, ŵ2 = 0.05, ŵ3 = 0.17. We illustrate Saaty’s method on a job selection problem (3 alternatives compared on 6 criteria) overall satisfaction with the job level 1: focus criteria research level 3: alternatives growth benefits A colleauges B Choice of job. 12 location C reputation The question asked was, which of a given pair of criteria is seen as contributing more to overall satisfaction with a job and what is the intensity or strength of the difference? res. growth benefits coll. location reputation priority research 1 1 1 4 1 1/2 0.16 growth 1 1 2 4 1 1/2 0.19 benefits 1 1/2 1 5 3 1/2 0.19 colleaug. 1/4 1/4 1/5 1 1/3 1/3 0.05 location 1 1 1/3 3 1 1 0.12 reputation 2 2 2 3 1 1 0.30 Pairwise comparison matrix of criteria. The relative weights of criteria (priorities) 13 can be computed as normalized geometric means of the rows (which are very close to the eigenvector corresponding to the largest eigenvalue of the matrix) The geometric means are computed as m1 = 1 × 1 × 1 × 4 × 1 × 1/2 6 1 × 1 × 2 × 4 × 1 × 1/2 m2 = 6 m3 = 1 × 1/2 × 1 × 5 × 3 × 1/2 6 14 m4 = 6 1/4 × 1/4 × 1/5 × 1 × 1/3 × 1/3 1 × 1 × 1/3 × 3 × 1 × 1 m5 = 6 m6 = √ 6 2×2×2×3×1×1 So, the relative weight (priority) of the criterion research is obtained as m1 p1 = m1 + m2 + m3 + m4 + m5 + m6 15 Then we compare the alternatives on each of the criteria research A B A B C 1 1/4 1/2 4 1 3 2 1/3 1 growth A B A B C C priority C priority 1 1/4 1/2 4 1 3 2 1/3 1 16 0.14 0.63 0.24 0.10 0.33 0.57 benefits A B C priority A B C 1 3 1/3 1/3 1 3 1 1 colleaug. A A B C B C priority 1 1/3 5 3 1 7 1/5 1/7 1 17 location A A B C 1 1 7 1 1 7 1/7 1/7 1 reputat. A A B C B C priority B C priority 1 7 9 1/7 1 5 1/9 1/5 1 18 0.77 0.17 0.05 We obtain 0.14 0.10 0.16 × 0.63 + 0.19 × 0.33 + · · · 0.24 0.57 0.77 +0.30 × 0.17 0.05 0.40 A = 0.34 B 0.26 C So job A should be selected as the best alternative. 19