You are on page 1of 5

MULTICRITE DECISION MAKING USING FUZZY LOGIC

Heinrich J. Rommelfanger
J.W. Goethe-University Frankfurt am Main, Institute of Statistics and Mathematics
Mertonstr. 17-23, D-60054 Frankfurt an Main, Germany
Rommelfanger@wiwi.uni-fiankfiu-t.de

Abstract To underpin this statement we refer to the following


faults:
- Utility are a elegant theoretical tool for mapping a
In this paper a new method for solving multicriteria deci- multidimensional goal vector in a one-dimensional real
sion problems is represented. Similar to the Analytic number. But practice do not accept this trick; because
Hierarchy Process or the Utili@ Value Analysis the goals neither the proposal to determine utilities by using lot-
and subgoals are ordered in a hierarchical structure. The teries nor the method to reduce the dimensions step by
goals on the lowest level are valuated by means of lin- step by means of the transformation principle can con-
guistic variables based on cardinal or ordinal scales vince.
which are proposed by experts. The evaluations of the - The Utility Value Analysis reduce the problem of util-
subgoals are aggregated by using expert rules and fuzzy ity measurement by assigning a utility value to each
inference. goal separately. Moreover the scaling problem is also
reduced, as the non-metric scaled goals are divided in
subaspects which can be measured on a metric scale.
1. Introduction For getting a accumulated result all utility values
should be aggregated step by step using weighted alge-
Decision making is one of the most central and perva-
braic sum. The open question is how to get the correct
sive human activities. Everybody makes decisions all the
weights.
time. Unfortunately the majority of models in literature of
- The Analytic Hierarchy Process completely ignores
decision theory or operations research have concentrated
on single criterion decision making. But in our complex the scaling problem. Moreover, the procedure for
world, we need to set priorities on the solutions according deriving priorities using paired comparisons is not
convincing. I have my doubts about that artificial
to their effectiveness by considering their benefits and
method using a lot of mathematical calculation as
costs, risks and opportunities and the resources they need.
summing up ratios and calculating eigenvectors which
Le., in practice we have to consider several objectives
simultaneously. are not well-founded. Also further developments as
Analytic Network Process or Fuzzy Analytic Hierarchy
In literature, a lot of approaches are proposed for solv- Process do not eliminate these objections.
ing multicriteria decision problems. Well known methods - Only Outranking Methods based on satisfying solu-
are the Utility Theory based on the use of lottery compari- tions in the sense of Simon are convincing procedures
sons, Bayesian Theory based on probabilities, Outranking for selecting a best solution. But often we have the
Methods based on ordinal comparisons of concordance problem that all alternatives should be ranked.
and discordance, the Analytic Hierarchy Process (AHP)
based on hierarchical structures, paired comparisons and In this paper a new method for solving multicriteria
priority vectors and Goal Programming. All these proce- decision problems is represented. Similar to the Analytic
dures have the handicap that they map the ideas of the Hierarchy Process or the Utility Value Analysis the goals
decision maker about the real problem in a crisp mathe- and subgoals are ordered in a hierarchical structure. At
matical model, neglecting inaccuracies about the data and first, the goals on the lowest level are valuated by linguis-
the structure. Particularly the fact that some of the criteria tic variables based on cardinal or ordinal scales. The lin-
can only be measured on an ordinal scale is usually disre- guistic variables are modeled by experts using suitable
garded. Therefore, it often happens that a rational solution data banks. Then the evaluations of the subgoals are
of the mathematical model is calculated but not necessarily aggregated by using expert rules and fuzzy inference. The
of the real problem. results are hzzy evaluations. If required the fuzzy results
may be compressed using defuzzification methods.

0-78034453-7/98 $10.00 0 1998 IEEE

360

Authorized licensed use limited to: Universitas Indonesia. Downloaded on January 23,2021 at 14:26:33 UTC from IEEE Xplore. Restrictions apply.
2. Hierarchical structure Market Market Rate of
Sales
Empirical studies reveal that persons are not able to Share Growth Innovation
compare consistently more than three criteria. Therefore it P P P P-
is useful to work with hierarchical goal systems where at P P m P-
most three or four subaspects are aggregated at each step. P P g P
Moreover, complex and not directly measurable criteria, P m P P
as for example the credit solvency of a business firm,may P m m P
often be explained more transparent and more intelligible P m g P+
by a hierarchical system of subcriteria. P g P P+
P g m m-
For getting a overall ranking of an alternative, at first m
P g g
the aspects on the bottom level have to be evaluated. Then m P P P
those evaluations of the subattributes need to be aggre- m m m-
P
gated step by step until the top level of the hierarchical m m-
P g
concept is reached. m m P P+
Usually this aggregation process is accomplished by m m m m
means of aggregation operators depending on parameters m m g m
or weights which ask for further specification by the deci- m g P P+
sion maker. m 0 m m
An disadvantage of these operators is that they portray m g g g-
the complex conjunction mechanism of the human mind g P P P
only incompletely. In literature and practice there exist a g P m P+
lot of examples, that the weights within an objective sys- g P g m-
tem depend not only on the objectives but often change g m P m-
with the obtained values. E.g., a rule table for the evalua- g m m m
tion of adjustments of salability is presented in [3]. It g m g m+
seems impossible or at least very difficult to describe this
0
g P g-
aggregation pattern by means of operators. g g m g
LL L g+
Table 1: Evaluation of Sales
3. Aggregation of subcriteria by expert rules
In the artificial intelligence literature we can find many Market Growth (MG) Grade
models in which the human aggregation process is
described by means of rules formulated by experts. The MG < - 5 YO 6
Table 1 presents a rule map established by an expert team -5 % IMG < 0 % 5 (big risk)
of the Commerzbank AG Frankfurt am Main, see [ 11. 0 % 5 MG < 5 yo 4 medium (m)
Using a rule map for aggregating subaspects to take the 5 yo I MG < 7 % 3 (medium risk)
fact into account that the weights do not depend merely on
the different criteria but of the evaluation of these criteria 7?'oIMG<10% 2
too. Therefore, a rule based aggregation is a very flexible 1 (small risk)
10%5MG
tool for expressing an individual imagination. In the rule
Table 2: Evaluation of Market growth
matrix in Table 1, the evaluation of determinants is not
described by degrees of the interval [0, 11 but by the lin- To improve this situation we could try to enlarge the
guistic terms ,,poor", ,,medium" and ,,good". For every number of valuations for each criterion. This, however
possible situation on the lower hierarchy level an aggrega- would result in an explosive increase of the number of
tion rule is defined. In this study only three values per rules as for m aspects with r possible valuations there exist
criterion were distinguished, yet additional ratings r m rules. In the case of the valuation of ,,Sales" in
(- and/or +) were allowed for the aggregation results. Table 1, we have now 33 = 27 rules. An increase in the
Objections against those rules can be raised due to the number of valuations to 5 would result in
fact that the rules are very inaccurate; the terms ,,good", 53 =I25 rules.
,,medium", ,,poor" allow a comparably large interpretation Therefore it is necessary to restrain the number of
spectrum, see e.g. Table 2. Moreover the strong distinc- valuations for each aspect. Moreover if the rule map gets
tions between the three classes give the impression of an too large the expert team will not be able to guarantee a
arbitrary classification. conscious distinction of each situation.

361

Authorized licensed use limited to: Universitas Indonesia. Downloaded on January 23,2021 at 14:26:33 UTC from IEEE Xplore. Restrictions apply.
4. Description of expert rules by fuzzy sets
i good
As demonstrated in the example above different values
of the same interval portray a linguistic term not corre-
spondingly. This is a result of the fact that intervals only
allow Yes/No-statements. Therefore we propose to model
the linguistic terms by fuzzy sets; this theory makes it _. , ~,~
,
possible to describe the different membership degrees 13 25 i299 35
according to the categories of ,,poor", ,,medium" or
Figure 4: Valuation of Rate o f Innovation
,,good" credit solvency as precise as the credit expert can
express it. These membership functions must be specified
as carefully as possible, because they will decisively influ-
ence the valuation process. Nevertheless we will never
obtain membership functions which are accurate in every
detail, because a lot of data about similar firms and knowl-
edge of the trade must be collected by the expert team.
Therefore the form of membership functions in expert
systems will be very simple and the same design will be -3 -2 -1 0 1 2 3 X

used repeatedly. In practice it is sufficient to work with Figure 5: Valuation of Sales


fuzzy numbers or fuzzy intervals of the LR-type.
Whereas fuzzy control applications usually employ Presenting hzzy sets as categories of valuation pro-
very simple triangular or trapezoid fuzzy sets, we propose vides the opportunity to precisely inform users of expert
to use reference functions of the s-type, which take pattem systems about their valuation basis. Moreover by com-
to normal probability distributions in non-technical appli- paring data of different branches of business this pro-
cations. In the Figures 2, 3 , 4, 5 the reference function ceeding allows to resort to knowledge already stored in
exp(-u2) is used. To specify the form of a membership data bases.
function for small membership degrees is particularly A way for a practical transformation has been analyzed
difficult. To avoid errors we propose to neglect all in a pilot study at the Institute of Statistics and Mathemat-
membership degrees smaller than a minimum level E, in ics of the University of Frankfurt am Main. The intention
the following Figures this minimum level is fixed at E = was to design membership functions in the course of an
0.05. expert system for supporting the evaluation of the financial
and operating position of business firms as part of the
Membership function annual audit. In this case, three different types of member-
ship functions were constructed by using suitable data
banks and the knowledge of experts.
The limitation to a standardized procedure with three
types implies the advantage that the data of a f m ' s data
bank can be extensively used. By that fuzzy membership
.

3 4
1

7 12 I6 10 24'
'% functions can be defined for any reference number and any
branch of business with little effort - as long as the neces-
Figure 1: Valuation of Market Share sary information is available.

Membership function
poor * . medium . :. good 5. Rule based aggregation
-\I- .' f- When modeling the linguistic evaluation terms by fuzzy
sets, the aggregation rules are only applied to those cases,
; \ in which the evaluations produce a membership degree 1
for all subaspects. Then the corresponding rule of the rule
map is applied and leads to a distinct evaluation of the
upper-aspect with the membership degree 1, too. There-
fore the use of fuzzy sets helps users to understand the
Figure 3: Valuation o f Market Growth basic principles of the expert knowledge more easily. This
,,understanding" is an essential factor for the acceptance of
an expert system and related to that for its successful reali-
zation.

362

Authorized licensed use limited to: Universitas Indonesia. Downloaded on January 23,2021 at 14:26:33 UTC from IEEE Xplore. Restrictions apply.
For all the other cases, where at least one evaluation has Now all rules with positive DOF contribute to the
a membership degree smaller than 1, no special rules have valuation of Sales Potential in proportion to their DOFs.
been stated by the experts. We assume that the given rules When ,,turning down" the evaluations corresponding to
can be extended to situations in the vicinity. The rules are the individual rules we recommend the use of the Max-
softened with the consequence that now many rules can be Prod-inference which means that the membership values
used simultaneously in a weakened manner. are fixed in proportion to the corresponding DOF, see
For a real situation we denote the degree of fulfillment Figure 6. We are convinced that the Max-Prod-inferenceis
with the descriptions of state in the rule maps by DOF. more adequate than the Max-Min-inference which is often
According to the proceeding in fuzzy control, DOF is used in Fuzzy Control applications, because the elimina-
defined as the minimum of the membership degrees at- tion of membership values which go above the DOFs
tached to the ,,inputs" of this rule. Having examined vari- implies that rules with medium DOF get more influence.
ous operators regarding their ability to describe the human
goor medium minus medium
conjunction behavior in specific cases, we came to the A
conclusion that the minimum operator should be used. 0.47
Besides others the minimum operator has the advantage
that only few rules with positive DOFs exist, whereas by
using compensatory operators almost all rules will show
positive DOFs and we therefore have to expect inter-
mediate evaluation.
-3
.- ":_.::\-
-x;x:Ko322 -2 -1 0 1 -,
For example, a firm D with the characteristics Fig. 6. Valuation of sales using Max-Prod-Inference
(Market Share ,Market Growth, Rate of Innovation)
= (4 % , 5 % , 2 9 %) In this context we want to remark that there is an essen-
can be described by means of vectors containing member- tial difference between fuzzy control applications and non-
ship degrees: technical evaluation and decision problems. Usually tech-
(pg!r(4%)3 p:2dium(4%), pgood(4%)
MS nical control processes are rapidly repeated. Therefore it is
sufficient when an approximately correct action is carried
= (0,62,0,22,0) and out, because the correction will follow immediately. On
the contrary decision support systems require a definite
decision for every section which evidently has to be cor-
= (0,0,47,0,32) . rect. As a consequence the procedure in decision support
systems has to be handled more carefully than in fuzzy
This ,,fuzzification" of the crisp input values produces a
control. Not only the linguistic evaluation terms have to be
connection between the observed values and the linguistic
defined more carefully but also the calculation of the
valuations in the rule base.
In accordance with the Table 1, we get for the firm D the DOFs and the influence of the DOFs on the final result
require an exact empirical examination.
following four positive DOFs 2 E:
D°FRule 5 In fuzzy control the total result is calculated by apply-
ing the Maximum-operator for aggregating the evaluations
= in ( ~ k $ r ( 4 %,) 5 ~g-ji~m(29%)) of the rules with positive DOF. But in decision support
= Min (0,62 , 1 ,0,47) = 0,47 poor systems corrections should be considered. In our example
the valuation ,,medium" for the Sales can be found twice.
D°FRule 6 On the one hand we do not consider it to be right that the
rating ,,medium" only counts with a DOF
Max(0,47, 032) = 0,47 which means that the rule with
the second best positive DOF is completely neglected. On
the other hand it seems absurd to add the DOFs if two or
more rules turn up with the same ,,output", because it
would then be possible to get DOF-values greater then 1.
We propose to adopt a middle course and suggest the use
of the algebraic sum. By doing so we get - in the example
above - for the valuation ,,medium" the total DOF
0,47 + 0,32 - 0,47 x 0,32 = 0,64 which presents a more
= Min (0,22 , 1 ,0,32) = 0,22 medium
balanced valuation.
Then the sales are vaguely evaluated by means of the
membership function shown in Figure 7.

363

Authorized licensed use limited to: Universitas Indonesia. Downloaded on January 23,2021 at 14:26:33 UTC from IEEE Xplore. Restrictions apply.
trouble if we use more than three evaluation classes. As
the sum of the membership values can differ fiom 1, it is
recommended to normalize the membership degrees. In
the meanwhile their exist a practical procedure for reduc-
ing the number of classes if output results are used as
inputs in the next aggregation step.

6. Final remarks and future outlook.


-3 -2 -I 0 I
The application of linguistic variables and the employ-
Figure 7: Valuation of Sales for the firm D ment of fuzzy conjunction methods offer an appropriate
method to model the human decision process. By doing so
As expected we get a fuzzy evaluation. If it is intended we have no problem with ordinal scaled data and the
to ,,compress" the data to a unique valuation, the well aggregation procedure is no longer depend on constant
known defuzzification procedures can be used; the best- weights. Moreover, the knowledge of experts and the
known are the center of gravity method and the center of information in data basis can be used without any prob-
area method. But using these defuzzification procedures a lems. As proposed in [3] it is possible to use vague input
metric scale is necessary. On the other hand , ordinal data too.
scales are sufficient in the whole fuzzy inference proce-
dure.
References
But in hierarchical systems defuzzifying is not neces-
sary, because it is better to use directly the fuzzy valua- 1. Nolte-Hellwig, K.U.; Leins, H.; Krakl, J.: Die Steuerung von
tions as inputs for the next aggregation step, where the Bonitatsrisiken im Firmenkundengeschaft. In: Luthje, B
(Ed.): Risikomanagement in Banken - Konzeptionen und
corresponding DOFs are interpreted as membership val-
Steuerungssysteme. Verband offentlicher Banken, Bonn
ues. For evaluating the ,,Enterprise Performance" on the (Berichte und Analysen Bd. 13) 1991
higher level of the hierarchy the firm D get the ,,fuzzified"
valuation of Sales (S): 2. Rommelfanger, H.:Entscheiden bei Unsch&e.- Fuzzy Deci-
sion Support-Systeme. Springer-Verlag Berlin Heidelberg
@s,- (D), P;(D)Psm-(D)> P i (D), PLsm+(D)>P i ( D ) , PS,+(D)) 1983. second edition 1994
3. Rommelfanger, H.: Fuzzy Logic-Based Processing of Expert
= (0, 0,22, 0,22, 0,64, 0 , 0, 0). Rules Used for Checking the Credit Solvency of Small Busi-
ness Firms or for Supporting Analytic Procedures of
Here, we recognize the problem, mentioned in sector 3, Auditors. in Ribeiro R.R.; Yager R.R.; Zimmermann H.J.;
that we will get problems by using too many valuation Kacprzyk J. (Eds.): Studies in Fuzziness and soft computing.
grades. Only on the top level of the hierarchy there is no Physica-Verlag, Heidelberg 1998

3 64

Authorized licensed use limited to: Universitas Indonesia. Downloaded on January 23,2021 at 14:26:33 UTC from IEEE Xplore. Restrictions apply.

You might also like