Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Hierarchy Process
Maria Teresa Lamata*
Dep. Ciencias de la Computación e Inteligencia Artificial,
Universidad de Granada, Escuela Tecnica Superior Ingeniería Informática,
18071-Granada, Spain
Let us consider the Analytic Hierarchy Process in which the labels are structured as graded
numbers. To obtain the scoring that corresponds to the best alternative, or the ranking of the
alternatives, we need to use a total order for the graded numbers involved in the problem. In this
article, we consider a definition of such a total order, which is based upon two subjective aspects:
the degree of optimism/pessimism and the liking for risk /safety. As several operations, such as
product, quotient, and so forth, of fuzzy numbers do not preserve the triangularity, we also use
the graded numbers that are analogous to the fuzzy numbers; however, the operations with graded
numbers are carried out as a simple extension of operations with real intervals. © 2006 Wiley
Periodicals, Inc.
1. INTRODUCTION
In many situations, we use measures or quantities that are not exact but approx-
imate. In those cases, the concept of fuzzy number is more adequate than that of
real number. Since its foundation in 1965 by Zadeh,1 the Theory of Fuzzy Sets has
experienced a great growth and has obtained great importance. This fact is due,
fundamentally, to the ability of fuzzy sets to express the ambiguity and vagueness
that are inherent in the human language. Based on the usual operations with real
numbers, Zadeh’s extension principle gives rise to arithmetic with fuzzy numbers,
which is commonly used.
The Analytic Hierarchy Process ~AHP! proposed by Saaty 2 is a very popular
approach of multicriteria decision making ~MCDM! that involves qualitative data
and has been applied during the last 20 years in many situations of decision mak-
ing. The problem can be abstracted as how to derive weights, rankings, or impor-
tance for a set of activities according to their impact on the situation and the
objective of the decision to be made.
The method uses a reciprocal matrix of decisions obtained by pair compari-
sons such that the information is given in a linguistic form. The eigenvector is
*e-mail: mtl@decsai.ugr.es.
A i and A j are equally important. 1 Two elements contributed equally to the property.
A i is moderately more important than A j . 3 Experience and judgment slightly favor one
element over another.
A i is strongly more important than A j . 5 Experience and judgment strongly favor one
element over another.
A i is very strongly more important than A j . 7 An element is strongly favored and its dominance
is demonstrated in practice.
A i is extremely more important than A j . 9 The evidence favoring one element over the other
is of the highest possible order of affirmation.
The scales 2, 4, 6, and 8 are also used and 2, 4, Compromise is needed between two adjacent
represent compromises among the 6, 8 judgments.
tabulated scale.
冢 冣
1 a 12 J a 1j J a 1n
1/a 12 1 J a 2j J a 2n
.
A⫽
1/a 1j 1/a 2j J a ij J a in
.
1/a 1n 1/a 2n J 1/a in J 1
The elements a ij are considered to be estimates of the ratios wi /wj , where w is the
vector of actual weights of the alternative, which is what we want to find. All the
ratios are positive and satisfy the reciprocity property: a ij ⫽ 1/a ji ∀i, j ⫽ 1,2, . . . , n.
Saaty’s eigenvector solution of Aw ⫽ l max w always exists if the consistency ~or
transitivity! condition a ij * a jk ⫽ a ik ~i, j, k ⫽ 1,2, . . . , n! is satisfied.
Then, from a theoretical point of view, the AHP is based on the following
four axioms:
~1! The decision maker can provide paired comparison a ij of two alternatives i and j
corresponding to a criterion/subcriterion on a ratio scale that is reciprocal, that is,
a ij ⫽ 1/a ji ∀i, j ⫽ 1,2, . . . , n.
~2! The decision maker never judges one alternative to be infinitely better than another
corresponding to a criterion, that is, a ij ⫽ ` ∀i, j ⫽ 1,2, . . . , n.
~3! The decision problem can be formulated as a hierarchy.
~4! All criteria/subcriteria that have some impact on the given problem, and all the rele-
vant alternatives, are represented in the hierarchy in one go.
Aw ⫽ l max w
• The least-squares method ~LSM !,16 which minimizes the objective function:
((
i⫽1 j⫽1
n
冉 冊 a ij ⫺
wi
wj
2
~1!
subject to
n
( wi ⫽ 1,
i⫽1
wi ⬎ 0, i ⫽ 1,2, . . . , n
But, this nonlinear optimization problem has multiple solutions. To eliminate this prob-
lem Chu et al.16 modify the objective function ~1! in the form
min ( ( ~wi ⫺ wj a ij ! 2
i j
• The geometric means method ~also known as the logarithmic least-squares method
~LLSM !!, where the approximating vector w has elements of the form
wi ⫽ 冉) 冊
i⫽1
n
a ij
1/n
, i ⫽ 1,2, . . . , n
The vector w is usually normalized so that the sum of the elements is one. It can be
proved that the vector w with elements defined by the above equations is the solution of
the problem of minimizing the sum of squares
n n
vi ⫽ exp 冋 1
n
n
( log a ij
j⫽1
册 n
⫽ ) a ij1/n
j⫽1
~4!
( a ij
j⫽1
wi ⫽ n n
( ( a ij
i⫽1 j⫽1
it is obvious that the product of two triangular fuzzy numbers is not a triangular
fuzzy number, but verifies this result for a graded number.
Zadeh’s graded numbers are defined as nonincreasing families of bounded
real intervals, indexed by the unit interval @0,1#. Its operations are defined as a
natural extension of the interval operations. So, analogous results to those, which
are commonly used for the families of a-cuts of fuzzy numbers, are easily obtained
for the graded numbers.
A detailed study of such definitions and results appears in Refs. 3 and 4. Here,
we only consider the definition of Zadeh’s graded number ~using compact inter-
vals! and the operations, which are used to solve an AHP decision problem.
A Zadeh’s graded number is defined as any mapping c : @0,1# r $@A, B# : A ⱕ
B 僆 R% that assigns to each a 僆 @0, 1# the interval @a~a!, b~a!# such that ∀a, b 僆
@0,1# , @a ⬍ b ] a~a! ⱕ a~ b! ⱕ b~ b! ⱕ b~a!# .
Hereafter, we shall only use this kind of number, which will be called graded
numbers for brevity.
Obviously, each graded number is determined by two functions a, b : @0,1# r R
that satisfy the following three conditions:
The set of graded numbers is denoted by GZ ~R!. Obviously this set extends
the real line R, because each real number P can be identified with the graded num-
ber given by the constant function 僆 R a~a! ⫽ b~a! ⫽ P, ∀a 僆 @0,1# .
In the whole set GZ ~R!, the functions a~a! and b~a! that represent each graded
number are too arbitrary. To consider graded numbers that will be really used as
linguistic terms in AHP, we shall restrict ourselves to small subsets of GZ ~R!.
For convenience, we use in this article specific kinds of graded numbers: the
“triangular graded numbers.” This kind of number is well used as membership
International Journal of Intelligent Systems DOI 10.1002/int
432 LAMATA
functions associated with linguistic labels. More precisely, we have the following
definition.
• Triangular graded number to any graded number determined by two linear functions
a, b : @0,1# r R, whose graphs describe a triangle. More precisely, for any three real
numbers A ⱕ M ⱕ B, we have the triangular graded number @a~a!{b~a!# determined by
the following functions:
• The L-R-graded numbers, fixing some functions L and R as a basis for obtaining the
function a~a! and b~a!, respectively. The triangular graded numbers would result as
the particular case where a~a! and b~a! are linear, a~0! ⫽ A, a~1! ⫽ b~1! ⫽ M, and
b~0! ⫽ B. Obviously, this kind of graded number would be analogous to the known
L-R-fuzzy numbers defined by Dubois and Prade.31
• 3-graded number to Zadeh’s graded number determined by two step functions a, b:
I r R, which are constant on the intervals @0, 13_ #, ~ 13_ , 23_ #, and ~ 23_ , 1#. Thus, for any real
numbers A ⱕ A1 ⱕ A2 ⱕ B 2 ⱕ B 1 ⱕ B, we have the 3-graded number determined by the
functions
a~a! ⫽ A, b~a! ⫽ B, if 0 ⱕ a ⬍ 1
_
3
Operations. For any positive graded numbers c~a! ⫽ @~a!, b~a!# , c1 ~a! ⫽
@a 1 ~a!, b1 ~a!# , and c2 ~a! ⫽ @a 2 ~a!, b2 ~a!# and any real number P, we are inter-
ested here in the operations related to AHP problems whose values are between
1/9 and 9, and which are defined as follows:
~a! Inverse:
~c1 /c2 !~a! ⫽ $x/y : x 僆 c1 ~a!, y 僆 c2 ~a!% ⫽ @a 1 ~a!/b2 ~a!, b1 ~a!/a 2 ~a!#
~e! Root
In the same way, the problem of ranking Zadeh’s graded numbers is similar to
the problem of ranking intervals. For this reason, based upon an order between
intervals @A, B# ~with ∀A ⱕ B 僆 R! we shall consider finally, via the integration
on I, an order between graded numbers.
We must extend the usual order of the real line R. Therefore, it is fully rea-
sonable to take I1 ⫽ @A 1 , B1 # ⱕ @A 2 , B2 # ⫽ I2 when A 1 ⱕ A 2 and B1 ⱕ B2 . ~See
Figure 2.!
Nevertheless, the situation is not so obvious when one interval is included in
the other one; in this case, A 1 ⬍ A 2 ⱕ B2 ⬍ B1 . ~See Figure 3.!
Figure 4. Representation of the different regions associated with the different orders.
Let us note that this is a lexicographic order; however, the second part is
equivalent to the following: The interval with the greatest amplitude B ⫺ A ~resp.
with the lowest amplitude B ⫺ A! is preferred.
A similar criterion must be expected for values of l near zero ~resp. one!.
Nevertheless, for other values of l, the preference about the greatest or the lowest
value of the amplitude B ⫺ A can be considered as another subjective parameter,
which enables us to rank different intervals with the same value of M~l! ~I!. A
broad interval represents a high variation in the set of possible rewards. Con-
versely, a narrow interval represents a low variation of the possible rewards, whose
values may be obtained with a higher degree of certainty.
For this reason, we think that the preference for a broad interval is associated
with a “risky” or “gambler” individual, whereas the preference for a narrow inter-
val is associated with a “safe” one. All this leads to the following definition.
_I1 ⱕlr I2 ? M~l! ~I1 ! ⬍ M~l! ~I2 ! or @Ml ~I1 ! ⫽ Ml ~I2 ! and B2 ⫺ A 2 ⱖ B1 ⫺ A 1 #
_I1 ⱕls I2 ? M~l! ~I1 ! ⬍ M~l! ~I2 ! or @Ml ~I1 ! ⫽ Ml ~I2 ! and B2 ⫺ A 2 ⱕ B1 ⫺ A 1 #
Because the safety is associated with a more extensive interval whereas the
risk is associated with a smaller interval,
• For l ⫽ 0 we first rank vertical half-lines ~preferring the right lines to the left ones! and
second, inside each half-line, we rank the points ~preferring the high points to the low
ones!.
• For l ⫽ 1 we first rank horizontal half-lines ~preferring the high lines to the low ones!
and second, inside each half-line, we rank the points ~preferring the right points to the
left ones!.
• For 0 ⬍⬍ l ⬍⬍ 1, we first rank the parallel half-lines ~1 ⫺ l!x ⫹ ly ⫽ m ~according to the
values of m! and second, we rank the points inside each half-line ~which corresponds to
different intervals with the same value of Ml !. The two values of k give the two ways to
rank the points.
A~c! ⫽ 冕 0
1
a~a!d~a! B~c! ⫽ 冕
0
1
b~a!d~a!
In addition, and taking into account the last paragraph, for each l 僆 @0,1#, we
consider the value Ml ~c! :⫽ lB~c! ⫹ ~1 ⫺ l!A~c!.
First, we can partially rank the graded numbers, ranking its associated values
Ml ~c!. This would be analogous to a method given by de Campos and González 11
and González 12 to rank fuzzy numbers, which generalizes Yager’s 25 method ~cor-
responding to the case l ⫽ _12 !. Let us remark that Gonzalez uses the Stieltjes inte-
gral, with respect to any additive measure defined on @0,1#.
Second, we must rank the different graded numbers with the same Ml ~c!. If
they have associated different intervals @A, B# , then we rank these intervals using
Definition 6. This gives rise to the following partial order in GZ ~R!.
Definition 8. For each l 僆 @0,1# and each k 僆 $r, s% (assuming that k ⫽ r for
l ⬇ 0 and k ⫽ s for l ⬇ 1), we define the following partial order ⱕlk in GZ ~R!:
• If @A~c!, B~c!# ⫽ @A~f!, B~f!# , then c ⱕlk f is equivalent to @A~c!, B~c!# ⱕlk
@A~f!, B~f!#.
• If @A~c!, B~c!# ⫽ @A~f!, B~f!#, then the graded numbers c and f are not to be compared.
Definition 9. For each l 僆 @0,1# and each k 僆 $r, s% (assuming that ak exists,
k ⫽ r for l ⬇ 0 and k ⫽ s for l ⬇ 1!, we define the following total order ⱕlk :
• If @A~c!, B~c!# ⫽ @A~f!, B~f!#, then we use these intervals to rank the graded numbers
c and f, according to the total order ⱕlk given in Definition 2.
• If @A~c!, B~c!# ⫽ @A~f!, B~f!#, then we use the intervals c~ak ! and f~ak ! to rank c
and f (using also the total order ⱕlk given in Definition 2).
5. CASE STUDY
Figure 5. Hierarchical representation with values for the energy allocation problem.
International Journal of Intelligent Systems DOI 10.1002/int
438 LAMATA
Table II. Pairwise comparison matrix for the three criteria.
EG EQ NS LLMS
Table III. Pairwise comparison matrix for the three alternatives related to the criteria.
C1 C2 C3 ~LLMS!
EG
C1 @1, 1# @2.5, 3.5# @4.5, 5.5# @0.506, 0.800#
C2 @7/24, 5/12# @1, 1# @1.5, 2.5# @0.180, 0.305#
C3 @11/60, 9/40# @5/12, 3/4# @1, 1# @0.100, 0.167#
EQ
C1 @1, 1# @1.5, 2.5# @6.5, 7.5# @0.465, 0.726#
C2 @5/12, 3/4# @1, 1# @4.5, 5.5# @0.269, 0.438#
C3 @15/112, 13/84# @11/60, 9/40# @1, 1# @0.063, 0.089#
NS
C1 @1, 1# @1.5, 2.5# @2.5, 3.5# @0.415, 0.740#
C2 @5/12, 3/4# @1, 1# @1.5, 2.5# @0.228, 0.329#
C3 @7/24, 5/12# @5/12, 3/4# @1, 1# @0.101, 0.121#
A @0.388, 0.962#
B @0.163, 0.407#
C @0.077, 0.181#
We obtain three graded numbers, one for each alternative; the upper value is
in accordance with the characteristic of optimism l ⫽ 1, whereas the lower values
represent the pessimism l ⫽ 0.
We see that the alternative C1 dominates the other two alternatives and C2
dominates C3 , and then it is possible to see, in this case, that the preference order
is the following: C1 Ɑ C2 Ɑ C3 . ~See Figure 6.!
6. CONCLUSIONS
We present a new approach for the analytic hierarchy process in the general
case in which the linguistic labels are know as graded numbers. We have demon-
strated that the model is more general than the model proposed by Saaty.
Taking into account the definition of division given in Section 3, it is natural
that the final graded numbers have deviation toward the right; they are not
symmetric.
In the particular case in which we work only with the central value of the
solution, the result is the same as the corresponding in the original work of Saaty,
but this central value is not for the value l ⫽ 0.5, because, as we have said, the
final results are not symmetric.
We have the possibility to introduce the OWA operator 32 ; in this case the
three final values are taking account, introducing for this the degree of optimism.
We exhibit the advantages for the graded numbers, whose arithmetic is a sim-
ple extension of the Interval Analysis, whereas the fuzzy numbers need additional
conditions in order to be operated via their a-cut.
References
1. Zadeh LA. The concept of a linguistic variable and its application to approximate reason-
ing. Inform Sci 1975;8:199–249.
2. Saaty TL. The analytic hierarchy process. New York: McGraw-Hill; 1980.
3. Herencia JA. Graded sets and points: A stratified approach to fuzzy sets and points. Fuzzy
Set Syst 1996;77:191–202.
4. Herencia JA. Graded numbers and graded convergence of fuzzy numbers. Fuzzy Set Syst
1997;88:183–194.
5. Herencia JA, Lamata MT. A total order for the graded numbers used in decision problems.
Int J Uncertainty Fuzziness Knowl Based Syst 1999;7:267–276.
6. Negoita CV, Ralescu D. Applications of fuzzy sets to systems analysis. Basel: Birkhäuser
Verlag; 1975.
7. Moore RE. Interval analysis. Englewood Cliffs, NJ: Prentice-Hall; 1966.
8. Moore RE. Methods and applications of interval analysis. Philadelphia: SIAM; 1979.
9. Bouchon-Meunier B, Kosheleva O, Kreinovich V, Nguyen HT. Fuzzy numbers are the
only fuzzy sets that keep invertible operations invertible. Fuzzy Set Syst 1997;91:155–163.
10. Ralescu D. A survey of the representation of fuzzy concepts and its applications. In: Gupta
MM, Ragade RK, Yager RR, editors. Advances in fuzzy sets theory and applications.
Amsterdam: North-Holland; 1979. pp 77–91.
11. de Campos LM, González A. A subjective approach for ranking fuzzy numbers. Fuzzy Set
Syst 1989;29:145–153.
12. González A. A study of the ranking function approach through mean values. Fuzzy Set
Syst 1990;35:29– 41.
13. Fechner GT. Elements of psychophysics, vol 1. New York: Holt, Rinehart & Winston;
1965.
14. Thurstone LL. A law of comparative judgments. Psych Rev 1927;34:273–286.
15. Saaty TL. A scaling method for priorities in hierarchical structure. J Math Psych
1977;15:234–281.
16. Chu A, Kabala R, Springarn K. A comparison of two methods for determining the weights
of belonging to fuzzy sets. J Opt Theor Appl 1979;27:531–541.
17. Adamo JM. Fuzzy decision trees. Fuzzy Set Syst 1980;4:207–219.
18. Baas SM, Kwakernaak H. Rating and ranking of multiple-aspect alternatives using fuzzy
sets. Automatica 1977;13:47–58.
19. Baldwin JF, Guild NCF. Comparison of fuzzy sets on the same decision space. Fuzzy Set
Syst 1979;2:213–233.
20. Bortolan G, Degani R. A review of some methods for ranking fuzzy subsets. Fuzzy Set
Syst 1985;15:1–19.
21. Delgado M. Verdegay JL, Vila MA. A procedure for ranking fuzzy numbers using fuzzy
relations. Fuzzy Set Syst 1988;26:49– 62.
22. Matarazzo B, Munda G. New approaches for the comparison of L-R fuzzy numbers: A
theoretical and operational analysis. Fuzzy Set Syst 2001;118:407– 418.
23. Wang X, Kerre EE, Ruan D. Consistency of judgement matrix and fuzzy weights in fuzzy
analytic hierarchy process. Int J Uncertainty Fuzziness Knowl Based Syst 1995;3:35– 46.
24. Wang X, Kerre EE. Reasonable properties for the ordering of fuzzy quantities. Fuzzy Set
Syst 2001;118:375–385.
25. Yager RR. A procedure for ordering fuzzy subsets of the unit interval. Inform Sci
1981;24:143–161.