You are on page 1of 14

Does “Fuzzifying” AHP Improve the Quality of Multi-Attribute

Decision Making?

Li Lun and Poh Kim Leng


Department of Industrial & Systems Engineering
National University of Singapore
{u0706517 | isepohkl}@nus.edu.sg

Abstract

The Analytic Hierarchy Process (AHP) model was designed by Thomas L.Saaty (Saaty, 1980) in
1970s as a decision making aid. It is especially suitable for complex decisions which involve the
comparison of decision elements which are difficult to quantify. It is based on the assumption
that when faced with a complex decision the natural human reaction is to cluster the decision
elements according to their common characteristics. Along the way of its development, people
found that the uncertainty in assigning priorities for the selected decision elements is not being
addressed. This naturally led to the incorporation of the traditional AHP with fuzzy logic to
suggest the relative strength of factors in the corresponding criteria, thereby enabling the
construction of a fuzzy judgment matrix to facilitate decision making. Fuzzy Analytic Hierarchy
Process (FAHP) model became the solution to the problem. No doubt, FAHP has a better
performance in dealing with vague input data, but the trade-off is a more complicated calculation
it adds to the traditional AHP model. This paper aims to analyse the worthiness of using FAHP
instead of traditional AHP to improve the quality of multi-attribute decision making in
consideration of its complexity based on outranking method/ELECTRE 1.

Key words: Analytic Hierarchy Process (AHP), Fuzzy Analytic Hierarchy Process (FAHP),
ELECTRE 1, Multiple Attribute Decision Making (MADM)

1
1. Introduction

The Analytic Hierarchy Process (AHP) model was designed by Thomas L.Saaty in 1970s as a
decision making aid. It is especially suitable for complex decisions which involve the
comparison of decision elements which are difficult to quantify. It is based on the assumption
that when faced with a complex decision the natural human reaction is to cluster the decision
elements according to their common characteristics. Along the way of its development, people
found that the uncertainty in assigning priorities for the selected decision elements is not being
addressed. This naturally led to the incorporation of the traditional AHP with fuzzy logic to
suggest the relative strength of factors in the corresponding criteria, thereby enabling the
construction of a fuzzy judgment matrix to facilitate decision making. Fuzzy Analytic Hierarchy
Process (FAHP) model became the solution to the problem. No doubt, FAHP has a better
performance in dealing with vague input data, but the trade-off is a more complicated
calculation it adds to the traditional AHP model. One important thing that we have to take note
of is that the application of different techniques to a problem may lead to different rankings of
alternatives and hence different decisions. Therefore, it is crucial for the decision makers to
adopt the most appropriate method.

This paper aims to present a comparative study on AHP and FAHP based on the outranking
process of ELECTRE 1 (Roy, 1968). Different people have different ability to deal with the
problems. Here, only two types of people will be considered, namely, the real-world
businessmen and the mathematically-inclined people.

This paper is organized as follows: Section 2 the review of AHP & FAHP; Section 3 problem
solving using both AHP & FAHP; Section 4 the ELECTRE 1 method; Section 5 the
comparative study of AHP & FAHP. Finally, Section 6 concludes the paper.

2. Review of AHP & FAHP

The Analytic Hierarchy Process (AHP) was developed by Satty (1980). It is especially suitable
for complex decisions which involve the comparison of decision elements which are difficult to
quantify. It is based on the assumption that when faced with a complex decision the natural
human reaction is to cluster the decision elements according to their common characteristics.

AHP involves building a hierarchy (ranking) of decision elements and then making comparisons
between each possible pair in each cluster (as a matrix). This gives a weighting for each element
within a cluster (or level of the hierarchy) and also a consistency ratio (useful for checking the
consistency of the data).

Users of the AHP first decompose their decision problem into a hierarchy of more easily
comprehended sub-problems, each of which can be analyzed independently. The elements of the
hierarchy can relate to any aspect of the decision problem—tangible or intangible, carefully
measured or roughly estimated, well- or poorly-understood—anything at all that applies to the
decision at hand.

Once the hierarchy is built, the decision makers systematically evaluate its various elements,
comparing them to one another in pairs. In making the comparisons, the decision makers can use
2
concrete data about the elements, or they can use their judgments about the elements' relative
meaning and importance. It is the essence of the AHP that human judgments, and not just the
underlying information, can be used in performing the evaluations.

The AHP converts these evaluations to numerical values that can be processed and compared
over the entire range of the problem. A numerical weight or priority is derived for each element
of the hierarchy, allowing diverse and often incommensurable elements to be compared to one
another in a rational and consistent way. This capability distinguishes the AHP from other
decision making techniques.

In the final step of the process, numerical priorities are derived for each of the decision
alternatives. Since these numbers represent the alternatives' relative ability to achieve the
decision goal, they allow a straightforward consideration of the various courses of action.

As we know, the AHP assumes that the multi-criteria problem can be completely expressed in a
hierarchical structure. The data acquired from the decision makers are pair-wise comparisons
concerning the relative importance of each criterion. Since it is difficult to map qualitative
preferences to quantitative ones, a degree of uncertainty is associated with some or all pair-wise
comparison values in an AHP problem. Fuzzy AHP is there to solve this problem. The earliest
study of FAHP was done by Laarhoven (1983), where the fuzzy comparing judgment is
represented by triangular membership functions. According to the method of logarithmic least
squares, the priority vectors can be obtained. Chang (1996) introduced another approach for
handling FAHP. He used triangular fuzzy numbers for pair-wise comparison and extent analysis
for the synthetic extent values of the pair-wise comparisons. Besides these two methods, there
are also other approaches for FAHP, but the details are not presented in this paper.

3. Problem solving using both AHP and FAHP

In this section, two examples will be solved using both AHP and FAHP to demonstrate their
respective features. The experience learned will be used later for the comparative study. The
examples are small uncertainties in pair-wise comparison and large uncertainties in pair-wise
comparison.

3.1 Small uncertainties involved in pair-wise comparison

In this case, the decision makers are quite sure about their decisions.

⎛1 8 5⎞
⎜ ⎟
Consider a matrix A = ⎜ 3 1 4 ⎟
⎜5 7 1⎟
⎝ ⎠

Here the Row Geometric Mean Approximation method is adopted due to its better performance
compared to Column Normalization method. So the weight

3
n
n
∏a
i =1
ij

wi = (1)
n ⎛ n ⎞
∑ ⎜n

i −1 ⎝
∏ aij ⎟

i =1 ⎠

According to Equation (1), the normalized weights of each criterion are obtained,

Table 1 Normalized weights obtained from matrix A

Criterion X1 X2 X3
Symbols for normalized
w₁ w₂ w₃
weight
Normalized weights 0.381 0.255 0.364

Now we use triangular fuzzy numbers to represent matrix A and assume the uncertainties of
each element is rather small.

⎡ (1,1,1) (7,8,9) (4,5,6)⎤


A1 = ⎢⎢(2,3,4 ) (1,1,1) (3,4,5)⎥⎥
⎢⎣(4,5,6 ) (6,7,8) (1,1,1) ⎥⎦

⎛ 1 9 6⎞ ⎛1 7 4⎞
⎜ ⎟ ⎜ ⎟
A1U = ⎜ 4 1 5 ⎟ A1L = ⎜ 2 1 3 ⎟
⎜ 6 8 1⎟ ⎜4 6 1⎟
⎝ ⎠ ⎝ ⎠

Table 2 Normalized weights obtained from matrix A1

Criterion X1 X2 X3
Symbols for normalized
w₁ w₂ w₃
weight
Normalized weights (0.392,0.381,0.373) (0.235,0.255,0.268) (0.377,0.364,0.359)

3.2 Large uncertainties involved in pair-wise comparison

⎡ (1,1,1) (5,8,11) (2,5,8)⎤


A2 = ⎢⎢(0,3,6 ) (1,1,1) (1,4,7 )⎥⎥
⎢⎣(2,5,8) (4,7,10) (1,1,1) ⎥⎦

⎛ 1 5 2⎞ ⎛ 1 11 8 ⎞
⎜ ⎟ ⎜ ⎟
A2U = ⎜0 1 1⎟ A2 L = ⎜ 6 1 7 ⎟
⎜2 4 1⎟ ⎜ 8 10 1 ⎟
⎝ ⎠ ⎝ ⎠

4
Table 3 Normalized weights obtained from matrix A2

Criterion X1 X2 X3
Symbols for normalized
w₁ w₂ w₃
weight
Normalized weights (0.401,0.381,0.519) (0.284,0.255,0) (0.352,0.364,0.481)

4. The ELECTRE 1 Method

ELECTRE 1 is one of the outranking methods which originated from Europe. Alternatives are
compared on the basis of two matrices: concordance and discordance matrices. Concordance
matrix reflects those cases where the first alternative in a pair of alternatives is superior to the
second. Discordance matrix reflects those cases where the first alternative is inferior to the
second. Outranking relationships are developed based on weighted concordance and discordance
indices. It only gives us partial ranking of the alternatives (Nijkamp, 1975).

Outranking method uses various mathematical functions to indicate the degree of dominance of
one alternative or group of alternatives. The model permits any two alternatives to remain
incomparable with each other. For instance, in the context of a choice problem, if alternative A is
better than both the alternative B and C, leaving alternative B and C incomparable will not affect
the final decision made. However, there are only two candidates for comparison in this paper, so
incomparability is not a concern here. Outranking method can be branched in to the ELECTRE
(Elimination et Choice Translating Reality) family (Roy 1996) & TOPSIS (Technique for Order
Preference by Similarity to Ideal Solution) family, but in this paper, ELECTRE 1 is adopted.

5. Comparative study of AHP & FAHP

5.1 Scenario 1: For real-world businessmen

In previous studies (Liu & Poh, 2003) criteria such as ability, simplicity, compatibility, nature of
data and software support were used to evaluate the strengths and weaknesses of various MADM.
However, in consideration of the characteristics of AHP and FAHP, robustness, simplicity and
software support are chosen as the criteria for comparison in this paper. The definitions of the
three criteria are as follows:

Ability to handle uncertainty (C1): This criterion refers to the method’s ability to deal with the
uncertainties that may be involved in assigning weights during pair-wise comparisons.

Simplicity (C2): This refers to how easy it is to understand the operation and to interpret the
results. Apparently, the simpler the method is more desirable.

Software support (C3): The decision making process will be faster and more accurate with the
help of the advanced tools. The effect of using software is evident when dealing with larger sets
of data. Hence, the method with existing commercially available and affordable software is
preferred.

5
Complexity of input data (C4): This refers to how detailed the information of input data is
needed for the method.

Table 4 Weight of each criterion

Criterion Description Weight


C1 Ability 2
C2 Simplicity 4
C3 Software support 3
C4 Complexity of data 1

In the case of real business world, the decision makers are usually not very familiar with the
complicated mathematics, and this makes a decision making method that is easy to understand
and use a must. That is why the criterion “simplicity” has the highest weight, followed by the
criterion “software support”. This is because we have to consider the time value of money, and
we have to complete the work accurately and at the same time, efficiently. A method with
existing commercially available and affordable software is necessary to speed up the decision
making process.

However, for those decision makers who are mathematics inclined, the weights carried by
various criteria will be different. And this will be covered in the sensitivity analysis in the later
section of the paper.

Method
Effectiveness

C1 C2 C3 C4

Ability Simplicity Software Complexity of data


support

AHP FAHP

Figure 1 shows the hierarchy for this comparative analysis

6
Table 5 shows the qualitative scale and the corresponding quantitative ratings using the scoring
system introduced firstly by Guigou (1975).

Table 5 Rating scale

Qualitative scale Quantitative scale


Excellent 20
Above average 15
Average 10
Below average 5
Unsatisfactory 0

The criterion scores are set by the author based on her understanding of these two methods. For
C1 (Ability), FAHP is no doubt better than AHP in terms of dealing with uncertainties. For C2
(Simplicity), AHP beats FAHP due to its more straight-forward concept as well as calculations.
Since there is no commercial software available for FAHP at all, FAHP scores “0” for C3
(Software support). For C4 (Complexity of data), AHP gets a higher score because it requires
less information about the input data than FAHP.

Table 6 Criterion scores

C1 C2 C3 C4
AHP 10 20 20 20
FAHP 20 5 0 5
Weight 2 4 3 1

Table 7 is the standardized criterion scores derived from Table 3 where the highest score of each
column is set as “1” and the lowest as “0”.

Table 7 Standard criterion scores

C1 C2 C3 C4
AHP 0 1 1 1
FAHP 1 0 0 0
Weight 2 4 3 1

With these standardized criterion scores, it is now the time for the final step, i.e. to obtain the
concordance & discordance matrices. Concordance and discordance matrices are computed using
Table 4. The formulae for the computation are as follows:

Concordance index C (A, B) = ∑ (w⁺+0.5w⁼) / (w⁼+w⁺+w⁻)

Discordance index D (A, B) = MAX [ Z BK – Z AK ] / (1-0) for all k where B≥A

Using the above formulae, the detailed calculations are shown below,

C (AHP, FAHP) = [(4+3+1) +0] / [2+4+3+1] = 0.8

7
C (FAHP, AHP) = [2] / [2+4+3+1] = 0.2

D (AHP, FAHP) = {MAX [(1-0)]} / (1-0) = 1

D (FAHP, AHP) = {MAX [(1-0), (1-0), (1-0)]} / (1-0) = 1

⎛ x 0.8 ⎞
The concordance matrix C = ⎜⎜ ⎟⎟
⎝ 0.2 x ⎠

⎛ x 1⎞
The discordance matrix D = ⎜⎜ ⎟⎟
⎝1 x⎠

Then a conclusion can be reached by exploiting the significance/meaning of the concordance and
discordance matrices. In order to do this, the concordance threshold (p) and discordance
threshold (q) are chosen as the references for the classification of the kernel and not-kernel set.
Only the alternative with a concordance index larger than q and at the same time a discordance
index smaller than q is dominating and placed in the set of kernel. In this case, p is first set as 0.8,
and q is first set as 1. Since both the discordance indices are 1, any change in q will not affect the
outranking result and the result only depends on the value of p. With the help of Excel, we found
out that AHP outranks FAHP as long as p changes within the range of 0.2< p≤0.8.

5.2 Scenario 2: For mathematically-inclined people

Mathematically-inclined people are more familiar with the mathematics involved in the methods,
and thus, they do not value Simplicity as much as real-world businessmen do and value more for
the Ability of the method. Due to these differences, each step of the comparative study is
performed again to obtain a more comprehensive result for mathematically-inclined people.

Table 8 Weight of each criterion

Criterion Description Weight


C1 Ability 4
C2 Simplicity 2
C3 Software support 2
C4 Complexity of data 1

The rating scale remains unchanged.

Table 9 Rating scale

Qualitative scale Quantitative scale


Excellent 20
Above average 15
Average 10
Below average 5
Unsatisfactory 0

8
The criterion score and standard criterion score are also the same as that in scenario 1 because
these are the inherent features of the two methods. They do not vary with scenarios.

Table 10 Criterion score

C1 C2 C3 C4
AHP 10 20 20 20
FAHP 20 5 0 5
Weight 4 2 2 1

Table 11 Standard criterion score

C1 C2 C3 C4
AHP 0 1 1 1
FAHP 1 0 0 0
Weight 4 2 2 1

Then, the concordance and discordance matrices are calculated.

C (AHP, FAHP) = [(4+3+1) +0] / [2+4+3+1] = 0.556

C (FAHP, AHP) = [2] / [2+4+3+1] = 0.44

D (AHP, FAHP) = {MAX [(1-0)]} / (1-0) = 1

D (FAHP, AHP) = {MAX [(1-0), (1-0), (1-0)]} / (1-0) = 1

⎛ x 0.556 ⎞
The concordance matrix C = ⎜⎜ ⎟
⎝ 0.44 x ⎟⎠

⎛ x 1⎞
The discordance matrix D = ⎜⎜ ⎟⎟
⎝1 x⎠

Then we found out that AHP outranks FAHP when p varies from 0.44 to 0.556 exclusive.

6. Conclusion and Future Work

In this paper, we have studied and compared the AHP and FAHP method using ELECTRE 1
method. Considering the above two scenarios, we can conclude that in general, AHP outranks
FAHP in solving decision making problems. This means that it’s usually not worth it to carry out
so much calculation for FAHP just because it has better performance in dealing with
uncertainties. We should just adopt AHP instead to avoid the troubles.

9
According to the results obtained from the Excel, only if FAHP scores higher in C3 (Software
support), i.e. there is commercially available software for FAHP, can it outrank AHP. This
undoubtedly is an encouragement as well as a motivation for us to develop more user friendly
software for FAHP to be widely accepted by users in the future.

References
A. Ozdagoglu & G. Ozdagoglu. (2007). Comparisons of AHP and Fuzzy AHP for the multi-
criteria decision making processes with linguistic evaluations. Istanbul Ticret Universitesi Fen
Bilimleri Dergisi 6 (11):65-85.

B. Roy. (1968). Classment et choix en presence de points de vue multiples (la method
ELECTRE). Revue Francaise d’ Automatique Information et Recherche Operationelle (RIRO)
8:57-75.

D. Chang. (1996). Applications of the extent analysis method on fuzzy AHP. European Journal
of Operational Research 95:649-655.

J.L. Guigou. (1971). On French Location Models for Production Units. Regional and Urban
Economics 1(2):107-138.

P. Nijkamp. (1975). A Multi-attribute Analysis for Project Evaluation: Economic-Ecological


Evaluation of a Land Reclamation Project. Regional science association 35:87-111.

R.J. Liu, K.L. Poh and C.U.Lee (2003). An outranking analysis of MCDM methods. In Z. Jiang,
X. Qian and Z Jiang (editors), Proceedings of IE&EM'2003, The 10th International Conference
on Industrial Engineering & Engineering Management, Shanghai, pp 273, China Machine Press.

T. Saaty. (1980). A Scaling for Priorities in Hierarchical Structures. Journal of Mathematical


Psychology 15: 234-281.

10
Appendix

Detailed calculations for the AHP and FAHP example in section 3 are as follows,

⎛1 8 5⎞
⎜ ⎟
Consider a matrix A = ⎜ 3 1 4 ⎟
⎜5 7 1⎟
⎝ ⎠

Here the Row Geometric Mean Approximation method is adopted due to its better performance
compared to Column Normalization method. So the weight

n
n
∏a
i =1
ij

wi =
n ⎛ n ⎞
∑ ⎜n

i −1 ⎝
∏ a ij


i =1 ⎠

⎡ (1,1,1) (7,8,9) (4,5,6)⎤


A₁ = ⎢(2,3,4 ) (1,1,1) (3,4,5)⎥⎥

⎢⎣(4,5,6 ) (6,7,8) (1,1,1) ⎥⎦

⎛ 1 9 6⎞ ⎛1 7 4⎞ ⎛1 8 5⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
A 1U = ⎜ 4 1 5 ⎟ A 1L = ⎜ 2 1 3⎟ A 1M = ⎜ 3 1 4 ⎟
⎜ 6 8 1⎟ ⎜4 6 1⎟ ⎜5 7 1⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠

Consider the upper bound matrix, we have

3
9*6
w1 = = 0.373
3
9 * 6 + 3 4 *5 + 3 6 *8

3
4*5
w2 = = 0.268
3
9 * 6 + 4 *5 + 3 6 *8
3

3
6 *8
w3 = = 0.359
3
9 * 6 + 3 4 *5 + 3 6 *8

Consider the lower bound matrix, we have

3
7*4
w1 = = 0.392
3
7*4 + 2*3 + 3 4*6
3

11
3
2*3
w2 = = 0.235
3
7*4 + 2*3 + 3 4*6
3

3
4*6
w3 = = 0.377
3
7 * 4 + 2*3 + 3 4*6
3

⎡ (1,1,1) (5,8,11) (2,5,8)⎤


A2 = ⎢(0,3,6 ) (1,1,1) (1,4,7 )⎥⎥

⎢⎣(2,5,8) (4,7,10) (1,1,1) ⎥⎦

⎛1 5 2⎞ ⎛ 1 11 8 ⎞ ⎛1 8 5⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
A 2U = ⎜0 1 1⎟ A 2L = ⎜6 1 7⎟ A 2M = ⎜ 3 1 4 ⎟
⎜2 4 1⎟ ⎜ 8 10 1 ⎟ ⎜5 7 1⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠

Consider the upper bound matrix, we have

3
5*2
w1 = = 0.519
3
5 * 2 + 0 *1 + 3 2 * 4
3

3
0 *1
w2 = = 0
3
5 * 2 + 0 *1 + 3 2 * 4
3

3
2*4
w3 = = 0.481
3
5 * 2 + 0 *1 + 3 2 * 4
3

Consider the lower bound matrix, we have

3
8*5
w1 = = 0.401
3
8*5 + 3 3* 4 + 3 5* 7

3
3* 4
w2 = = 0.284
3
8*5 + 3* 4 + 3 5* 7
3

3
5*7
w3 = = 0.352
3
8*5 + 3 3* 4 + 3 5* 7

12
The detailed Excel worksheets for both scenario 1 and scenario 2 are presented below:

13
14

You might also like