Multiple-Criteria Fuzzy Evaluation:

The FuzzME Software Package

Jana Talasova1

Pavel Holecek2

1., 2. Faculty of Science, Palacky University Olomouc

tr. Svobody 26, 771 46 Olomouc, Czech Republic
Email: [email protected], [email protected]

Abstract This paper introduces a new software product

FuzzME. It was developed as a tool for creating fuzzy models of
multiple-criteria evaluation and decision making. The type of evaluations employed in the fuzzy models fully corresponds with the
paradigm of the fuzzy set theory; the evaluations express the (fuzzy)
degrees of fulfillment of corresponding goals. The FuzzME software
takes advantage of linguistic fuzzy modeling to the maximum extent.
In the FuzzME software, both quantitative and qualitative criteria
can be used. The basic structure of evaluation is described by a goals
tree. Within the goals tree, aggregation of partial fuzzy evaluations is
done either by one of fuzzified aggregation operators (fuzzy weighted
average, ordered fuzzy weighted average) or by a fuzzy expert system.
The paper contains an illustrative example of the software usage.
The application concerns a soft-fact-rating problem that was solved
in one of Austrian banks.
Keywords Fuzzy expert system, Fuzzy OWA operator, Fuzzy
weighted average, Multiple-criteria fuzzy evaluation, Normalized
fuzzy weights, Software.

1 Introduction
In practice, sophisticated models of multiple-criteria evaluation are required (e.g. rating of clients in banks, evaluation of
hospitals or universities, comparison of alternative solutions
to ecological problems). For creating the evaluating models,
setting some of their input data and interpreting their outputs,
experts knowledge is needed (e.g. evaluations of alternatives
according to qualitative criteria, partial evaluating functions
for quantitative criteria, a choice of a suitable type of aggregation, criteria weights, or eventually, a rule base describing
the relation between criteria values, the overall evaluation and
a linguistic description of obtained results). Because uncertainty is the typical feature of any expert information, the
fuzzy set theory is a suitable mathematical tool for creating
such models. For the practical use of the fuzzy models of
multiple-criteria evaluation, their user-friendly software implementation is necessary. But a good theoretical basis of the
used models is crucial, too. The clear and well-elaborated theory of multiple-criteria fuzzy evaluation makes it possible to
create an understandable methodics for the software user. And
a good methodics is essential for correct application of any
software to solving real problems.
There is a large number of papers and books dealing with
the theory and methods of multiple-criteria evaluation that
make use of the fuzzy approach (e.g. [1], [2], [3], [4]).
Multiple-criteria evaluation (as a basis of decision making)
was even one of the earliest applications of fuzzy sets (see

[1]). In more than 40 years of existence of the fuzzy sets theory, several software products for multiple-criteria decision
making, which use the fuzzy modeling principles in different degrees and in different ways, have been developed. In
practice, FuzzyTECH (see [5]) is probably the best-known of
these. It enables to use the specific instruments of the fuzzy
set theory for solving multiple-criteria evaluation and decision
making problems. Generally, FuzzyTECH is a universal software product which makes it possible to create and use fuzzy
expert systems (or fuzzy controllers). It also includes the possibility to derive fuzzy rule bases from given data by means
of neural network algorithms. In the book [6], there were described many successful applications of this software to solving multiple-criteria evaluation and decision making problems
in the areas of business and finance. Similarly, fuzzy toolboxes of general mathematical systems such as Matlab can be
used for multiple-criteria decision making.
The FuzzME software (Fuzzy models of Multiple-criteria
Evaluation), presented in this paper, is based on a theoretical concept of evaluation which is very close to the original
Zadehs ideas. Similarly to his paper [1], the evaluations of
alternatives according to particular criteria represent their degrees of fulfillment of the corresponding partial goals. Besides
evaluations expressed by real numbers in [0, 1], fuzzy evaluations modeled by fuzzy numbers on the same interval are employed in the software. They represent, analogously, the fuzzy
degrees of fulfillment of the partial goals which are connected
to the criteria. Resulting fuzzy evaluations, which are obtained
by aggregation, have a similar clear interpretation. This theoretical approach to (fuzzy) evaluation was published in the
book [7] and in the paper [8].
The predecessor of the FuzzME software package in terms
of the used theoretical basis was the NEFRIT software. This
software for multiple-criteria evaluation and decision making,
which is also based on fuzzy technologies, was developed
in about 2000 by the Czech software company TESCO SW
Inc. The fuzzy model of evaluation applied there is described
in detail in [8] and in the book [7] (a demo version of NEFRIT is enclosed in the book). NEFRIT makes it possible
to work with expert fuzzy evaluations of alternatives according to qualitative criteria. Values of the quantitative criteria
can either be crisp or fuzzy. Evaluating functions for quantitative criteria represent membership functions of corresponding partial goals. The main evaluation structure is expressed
by a goals tree. For aggregation of the partial fuzzy evaluations the weighted average method is used. The weights (crisp
only, not fuzzy) express the shares of partial evaluations in the

aggregated one. Fuzzy evaluations on all levels of aggregation express the fulfillment of the corresponding goals. The
NEFRIT software was originally developed for the Czech National Bank (decision making about granting a credit). Further, it was used e.g. by the Czech Tennis Association, the
Czech Basketball Association and in other institutions. Nowadays it is tested by the Supreme Audit Office of the Czech
Republic.
In contrast to NEFRIT, the FuzzME software makes it possible to use also uncertain weights in the aggregation by means
of the weighted average method. The theory of normalized
fuzzy weights, procedures for their setting, and an effective
algorithm for calculation of fuzzy weighted average are taken
from [9], [10] and [11]. Another fuzzy aggregation operator,
available in the FuzzME software, is a fuzzified OWA operator. Again, it works with normalized fuzzy weights. The fuzzy
OWA operator and the used algorithm for its calculation are
described in [12]. In the FuzzME software, multiple-criteria
evaluating functions can also be defined by means of fuzzy
rule bases. In accordance with [7], two algorithms are offered
for the approximate reasoning - the standard Mamdani algorithm and a modified Sugeno algorithm.
There are also software tools for multiple-criteria decision
making based on other mathematical methods. But they are
usually designed for solving a particular assignment. Our investigation by means of Internet did not result software fully
comparable to FuzzME. Its universality and comprehensiveness make it unique.

Preliminaries

A fuzzy set A on a universal set X is characterized by its membership function A : X [0, 1]. Ker A denotes a kernel of
A, Ker A = {x X | A(x) = 1}. For any [0, 1], A
denotes an -cut of A, A = {x X | A(x) }. A support of A is defined as Supp A = {x X | A(x) > 0}.
The symbol hgt A denotes a height of the fuzzy set A,
hgt A = sup {A(x) | x X}. An intersection and a union
of the fuzzy sets A and B on X are defined for all x X
by the following formulas: (A B)(x) = min {A(x), B(x)},
(A B)(x) = max {A(x), B(x)}.
A fuzzy number is a fuzzy set C on the set of all real numbers < which satisfies the following conditions: a) the kernel
of C, Ker C, is not empty, b) the -cuts of C, C , are closed
intervals for all (0, 1], c) the support of C, Supp C, is
bounded. A fuzzy number C is called to be defined on [a, b],
if Supp C [a, b]. Real numbers c1 c2 c3 c4 are
called significant values of the fuzzy number C if the following holds: [c1 , c4 ] = Cl(Supp C), [c2 , c3 ] = Ker C, where
Cl(Supp C) denotes a closure of Supp C.
Any fuzzy number C can be characterized by a pair of functions c : [0, 1] <, c : [0, 1] < which are defined by the
following formulas: C = [c(), c()] for all (0, 1], and
Cl(Supp C) = [c(0), c(0)]. The fuzzy number C is called to
be linear if both the functions c, c are linear. A linear fuzzy
number is fully determined by its significant values because
c() = (c2 c1 ) + c1 , c() = (c3 c4 ) + c4 . For that
reason, we can denote it as C = (c1 , c2 , c3 , c4 ).
An ordering of fuzzy numbers is defined as follows: a fuzzy
number C is greater than or equal to a fuzzy number D, if
C D for all (0, 1].

A fuzzy scale makes it possible to represent a closed interval of real numbers by a finite set of fuzzy numbers. Let
T1 , T2 , ..., Ts be fuzzy numbers defined on [a, b], forming a
fuzzy partition on the interval, i.e., for all x [a, b] the following holds
s
X
Ti (x) = 1,
(1)
i=1

then the set of the fuzzy numbers can be linearly ordered (see
[7]). If the fuzzy numbers T1 , T2 , ..., Ts are defined on [a, b],
form a fuzzy partition on the interval and are numbered according to their linear ordering, then they are said to form a
fuzzy scale on [a, b] .
An uncertain division of the whole into m parts can be modeled by normalized fuzzy weights. Fuzzy numbers V1 , ..., Vm
defined on [0, 1] are normalized fuzzy weights if for any
i {1, ..., m} and any (0, 1] it holds that for any vi Vi
there exist vj Vj , j = 1, ..., m, j 6= i, such that
vi +

m
X

vj = 1.

(2)

j=1,j6=i

The FuzzME software

The mathematical models of the FuzzME software are based

primarily on the theory and methods of multiple-criteria evaluation that were published in [7] and [8]. The theory of normalized fuzzy weights, the definition of fuzzy weighted average, and the algorithm for its computation were taken from
[9], [10] and [11]. The fuzzified OWA operator and the algorithm for its calculation published in [12] are also used in the
software.
In the FuzzME software, the basic structure of the fuzzy
model of multiple-criteria evaluation is expressed by a goals
tree. The root of the tree represents the overall goal of evaluation and each branch corresponds to a partial goal. The goals
at the ends of branches are connected either with quantitative
or qualitative criteria.
When an alternative is evaluated, evaluations with respect
to criteria connected with the terminal branches are calculated
first. Independently of the criterion type, each of the evaluations is described by a fuzzy number defined on the interval
[0, 1]. It expresses the fuzzy degree of fulfillment of the corresponding partial goal.
These partial fuzzy evaluations are then aggregated according to the defined type of the tree node. Three types of aggregation are available: a fuzzy weighted average (fuzzy WA),
an ordered fuzzy weighted average (fuzzy OWA) or aggregation by means of a fuzzy expert system. For aggregation by
fuzzy weighted average or ordered fuzzy weighted average,
normalized fuzzy weights must be set. The weights express
uncertain shares of the partial evaluations in the aggregated
one. For the fuzzy expert system, the fuzzy rule base must
be defined and an inference algorithm must be chosen (either
the Mamdani algorithm or the generalized Sugeno algorithm
of approximate reasoning).
The overall evaluation reflects the degree of fulfillment of
the main goal. A verbal description of the overall evaluation
can be obtained by means of the implemented linguistic approximation algorithm.

The overall evaluations can be compared within the frame

of a given set of alternatives. By this comparison the best
of the alternatives can be chosen. That is why the FuzzME
software can be also used as a decision support system.
The import and export of data is supported by the software,
too. The FuzzME software is available in the Czech and English versions.

Figure 1: The main window of the software

3.1

Goals tree

Goals trees represent the basic structure of fuzzy models of

multiple-criteria evaluation in the FuzzME software. When a
goals tree is designed, the main goal is consecutively divided
into goals of progressively lower levels. The process of division is stopped when such goals are reached whose fulfillment
can be assessed by means of some known characteristics of
alternatives (i.e. quantitative or qualitative criteria).
The design of a tree structure in the goals-tree editor is the
first step in forming a fuzzy evaluation model in FuzzME. In
the next step, the type of each node in the tree must be specified. For the nodes at the ends of tree branches the user defines
if the node is connected with a quantitative or qualitative criterion. For the other nodes he/she sets the type of aggregation
- fuzzy weighted average, ordered fuzzy weighted average or
fuzzy expert system. An example of a goals tree is illustrated
in Fig. 2.
3.2

Criteria of evaluation

In the models of evaluation created by the FuzzME software,

qualitative and quantitative criteria can be combined arbitrarily.
3.2.1

Qualitative criteria

According to qualitative criteria, alternatives are evaluated

verbally, by means of values of linguistic variables of special
kinds - linguistic scales, extended linguistic scales and linguistic scales with intermediate values.
A linguistic variable is defined as a quintuple (V, T (V), X,
G, M ), where V is a name of the variable, T (V) is a set of its
linguistic values, X is a universal set on which the meanings
of the linguistic values are defined, G is a syntactic rule for
generating values in T (V), and M is a semantic rule which
maps each linguistic value C T (V) to its mathematical
meaning, C = M (C), which is a fuzzy set on X.

A linguistic scale on [a, b] is a special case of the linguistic variable (V, T (V), X, G, M ), where X = [a, b], T (V) =
{T1 , T2 , ..., Ts } and the meanings of the linguistic values
T1 , T2 , . . . , Ts are modeled by fuzzy numbers T1 , T2 , . . . , Ts
which form a fuzzy scale on [a, b]. As the set of linguistic
values of the scale is defined explicitly, it is not necessary to
include the grammar G into the scale notation.
In the FuzzME software, the user defines a linguistic scale
for each qualitative criterion in the fuzzy-scale editor. For example, the linguistic scale quality of a product can contain
linguistic values poor, substandard, standard, above standard
and excellent. The evaluating linguistic scale is usually defined on [0, 1]; in other cases, it has to be transformed to this
interval.
The extended linguistic scale contains, besides elementary
terms of the original scale, T1 , T2 , . . . , Ts , also derived terms
in the form Ti to Tj , where i < j, i, j {1, 2, . . . , s}.
For example, the user can evaluate quality of a product by
the linguistic term standard to excellent. The meaning of
the linguistic value Ti to Tj is modeled by Ti L Ti+1 L
L Tj , where L denotes the union of fuzzy sets based
on the Lukasiewicz disjunction; e.g. (Ti L Ti+1 )(x) =
min {1, Ti (x) + Ti+1 (x)} for all x <.
The linguistic scale with intermediate values is the original
linguistic scale enriched with derived terms between Ti and
Ti+1 , i {1, 2, . . . , s 1}. The meaning of the derived term
between Ti and Ti+1 is modeled by the arithmetic average of
the fuzzy numbers Ti and Ti+1 .
In the FuzzME software, the user evaluates a given alternative according to a qualitative criterion by selecting a proper
linguistic evaluation from a drop-down list box. He/she can
choose the value from a standard linguistic scale, extended
scale or scale with intermediate values.
The three mentioned structures of linguistic values are also
applied when resulting fuzzy evaluations are approximated
linguistically.
3.2.2 Quantitative criteria
The evaluation of an alternative with respect to a quantitative
criterion is calculated from the measured value of the criterion by means of the evaluating function expertly defined for
the criterion. The evaluating function is the membership function of the corresponding partial goal. The FuzzME software
admits both crisp and fuzzy values of quantitative criteria. The
fuzzy values represent inaccurate measurements or expert estimations of the criteria values. In the case of a fuzzy value,
the corresponding partial fuzzy evaluation is calculated by the
extension principle.
In the FuzzME software, the evaluating function of a quantitative criterion is formally set by means of a fuzzy number.
For example, if the evaluating function is defined by a linear
fuzzy number F = (f1 , f2 , f3 , f4 ), then f1 is the lower limit
of all at least partly acceptable values of the criterion, f2 is the
lower limit of its fully satisfactory values, f3 is the upper limit
of the fully satisfactory values, and f4 is the upper limit of the
acceptable values.
For example, when a bank evaluates expected profitability
of projects, the evaluating function can be defined by a linear
fuzzy number with significant values 10, 30, 500, 500. In that
case, values lower than 10% are not satisfying at all (the client

would not be able to pay the money back to the bank). For the
values from 10% to 30% the satisfaction of the bank is growing linearly. Values greater than 30% are fully satisfactory
from the banks point of view. Values greater than 500% are
not supposed to occur. This way we can define a monotonous
evaluating function, which is the most common in the evaluating models, by a fuzzy number.

normalized fuzzy weights V1 , . . . , Vm , is a fuzzy number U on

[0, 1] whose membership function is defined for any u [0, 1]
as follows
U (u) = max{min {V1 (v1 ), ..., Vm (vm ), U1 (u1 ), ..., Um (um )}
m
m
X
X
|
vi ui = u,
vi = 1, vi , ui [0, 1], i = 1, ..., m}.
i=1

3.3

i=1

(3)

Methods of aggregation of partial evaluations

The calculated partial fuzzy evaluations are then consecutively

aggregated according to the structure of the goals tree. With
respect to the defined type of the tree node, the fuzzy weighted
average method, the ordered fuzzy weighted average method
or the fuzzy expert system method is used for the aggregation.
Each of the aggregation methods is suitable for a different situation:
The fuzzy weighted average is used if the goal corresponding with the node of interest is fully decomposed into disjunctive goals of the lower level. The normalized fuzzy weights
represent uncertain shares of these lower-level goals in the
goal corresponding with the considered node.
Again, the ordered fuzzy weighted average requires that the
goal corresponding with the given node is decomposed into
disjunctive goals of the lower level. In contrast to the fuzzy
weighted average, the usage of this aggregation operator supposes special users requirements concerning the structure of
partial fuzzy evaluations. The normalized fuzzy weights again
represent uncertain shares of the partial evaluations in the aggregated one. But the normalized fuzzy weights are not linked
to the individual partial goals; the correspondence between the
weights and the partial evaluations is given by the ordering of
partial evaluations of the alternative of interest. It means, evaluations with respect to the same partial goal can have different
weights for different alternatives.
If the relationship between the evaluations of the lower level
and the evaluation corresponding with the given node is more
complex (if neither of the two previous methods can be used),
and if expert knowledge about the relationship is available,
then the aggregation function is described by a fuzzy rule base
of a fuzzy expert system. The approximate reasoning is used
to calculate the resulting evaluation. In particular, evaluating
function described by a fuzzy expert system is used if the fulfillment of a goal at the end of a tree branch depends on several mutually dependent criteria (i.e., if combinations of criteria values bring synergic or disynergic effects to the resulting
multiple-criteria evaluation).
3.3.1 Fuzzy weighted average
If the fuzzy weighted average is used for aggregation of partial fuzzy evaluations, then the uncertain weights of the corresponding partial goals, which express their shares in the
superior goal, must be set. To define consistent uncertain
weights, a special structure of fuzzy numbers, normalized
fuzzy weights, must be used.
In the FuzzME software, both real and fuzzy normalized
weights can be used. Normalized real weights, i.e., real numm
P
bers v1 , ..., vm , vj 0, j = 1, ..., m,
vj = 1, represent a
j=1

special case of the normalized fuzzy weights.

The fuzzy weighted average of the partial fuzzy evaluations,
i.e., of fuzzy numbers U1 , ..., Um defined on [0, 1], with the

For an expert who sets the fuzzy weights, it is not so easy

to satisfy the condition of normality. That is why the FuzzME
software allows to set only an approximation to the normalized fuzzy weights - fuzzy numbers W1 , ..., Wm on [0, 1] satisfying the following weaker condition
wi Ker Wi , i = 1, ..., n :

n
X

wi = 1.

(4)

i=1

The software removes the potential inconsistence in W1 , ...,

Wm and derives the normalized fuzzy weights V1 , ..., Vm from
them.
The structure of normalized fuzzy weights and the fuzzy
weighted average operation are studied in detail in [9], [10]
and [11]. Conditions for verifying normality of fuzzy weights,
an algorithm for normalization of fuzzy weights satisfying the
condition (4), and an algorithm for calculating fuzzy weighted
average, which are all used in the FuzzME software, can be
found there. Let us notice, that the used algorithm of fuzzy
weighted average calculation is very effective.
3.3.2 Ordered fuzzy weighted average
The fuzzy OWA operator is used in case that the evaluator
has special requirements concerning the structure of the partial evaluation. For example, he/she does not want any partial
goal to be satisfied poorly. Then the weight of the minimum
partial evaluation of any alternative equals 1, and the weights
of all its other partial evaluations equal 0. The aggregated
fuzzy evaluations then represent the guaranteed fuzzy degrees
of fulfillment of all the partial goals (the fuzzy MINIMAX
method). Another example of the fuzzy OWA operator usage
could be the evaluation of subjects who can choose in which
of the three areas they will be mostly involved. The evaluation
algorithm should take into account their right of choice. Then,
e.g., the results in the area where the subject performs best
contribute to the overall evaluation by about one half, results
from the second area by one third and results from the area in
which the subject was least involved contribute to the overall
evaluation only by one sixth. A practical application of such a
fuzzy evaluation model could be the overall evaluation of the
academic staff with respect to their results in the areas of research, education, and management of education and science.
The ordered fuzzy weighted average represents a fuzzification of the crisp OWA operator by means of the extension principle. Uncertain weights are modeled by normalized fuzzy
weights as in the case of fuzzy weighted average.
The following notation will be used to define the ordered
fuzzy weighted average: if (x1 , ..., xm ) is a vector of real
numbers, then (x(1) , ..., x(m) ) is a vector in which for all
j {1, . . . , m}, x(j) is the j-th greatest number of x1 , ..., xm .
The ordered fuzzy weighted average of the partial fuzzy
evaluations, i.e., of fuzzy numbers U1 , ..., Um defined on

[0, 1], with the normalized fuzzy weights V1 , . . . , Vm , is a

fuzzy number U on [0, 1] whose membership function is defined for any u [0, 1] as follows
U (u) = max{min {V1 (v1 ), ..., Vm (vm ), U1 (u1 ), ..., Um (um )}
m
m
X
X
|
vi u(i) = u,
vi = 1, vi , ui [0, 1], i = 1, ..., m}.

The final fuzzy evaluation of the alternative is given as the

union of all the fuzzy evaluations that were calculated for the
particular rules in the previous step, i.e.,
0

U =

n
[

Ui .

(9)

i=1

Generally, the result obtained by the Mamdani inference algo(5) rithm need not be a fuzzy number. So, for further calculations
within the fuzzy model, it must be approximated by a fuzzy
The algorithm used to calculate the ordered fuzzy weighted number.
average in the FuzzME software was taken from [12], where
The advantage of the generalized Sugeno inference algofuzzification of the OWA operator is described in detail. The rithm (see [7]) is that the result is always a fuzzy number. In
used algorithm is an analogy to the one used for the fuzzy its first step, the degrees of correspondence hi , i = 1, . . . , n,
weighted average.
are calculated in the same way as in the Mamdani fuzzy inference algorithm.
3.3.3 Fuzzy expert system
The resulting fuzzy evaluation U is then computed as a
The fuzzy expert system is used if the relationship between the weighted average of the fuzzy evaluations Ui , i = 1, 2, . . . , n,
criteria (or the partial evaluations) and the overall evaluation which model the mathematical meanings of linguistic evaluais complicated. Theoretically, it is possible to model, with an tions on the right-hand sides of the rules, with the weights hi .
arbitrary precision, any Borel measurable function by means This is done by the following formula
of a fuzzy rule base (properties of Mamdani and Sugeno fuzzy
n
P
controllers, see e.g. [13]) In reality, the quality of the approxhi .Ui
imation is limited by the experts knowledge of the relationU = i=1
.
(10)
n
P
ship.
hi
If the fuzzy rule base models the relation between values of
i=1
criteria and the fulfillment of the corresponding partial goal,
3.4 Overall fuzzy evaluations, the optimum alternative
then the evaluation function is of the following form
The final result of the consecutive aggregation of the partial
If C1 is A1,1 and . . . and Cm is A1,m , then E is U1
(6) fuzzy evaluations is an overall fuzzy evaluation of the given
alternative. The obtained overall fuzzy evaluations are fuzzy
If C1 is A2,1 and . . . and Cm is A2,m , then E is U2
numbers on [0, 1]. They express uncertain degrees of fulfill.............................................
ment of the main goal by the particular alternatives.
If C1 is An,1 and . . . and Cm is An,m , then E is Un
The FuzzME software compares alternatives according to
the centers of gravity of their overall fuzzy evaluations. A
center of gravity of a fuzzy number U on [0, 1] that is not a
where for i = 1, 2, . . . , n, j = 1, 2, . . . , m, (Cj , T (Cj ), real number, is defined as follows
Vj , Mj ) are linguistic scales representing the criteria, Ai,j
R1
U (x).x dx
T (Cj ) are their linguistic values, (E, T (E), [0, 1], Me ) is a lintU = R0 1
.
(11)
guistic scale representing the evaluation of alternatives and
U (x) dx
0
Ui T (E) are its linguistic values.
In the FuzzME software, rule bases are defined expertly. If U = u and u <, then tU = u. In the FuzzME softThe user defines such a rule base by assigning a linguistic ware, the optimum alternative is the one whose overall fuzzy
evaluation to each possible combination of linguistic values evaluation has the largest center of gravity.
At present, the FuzzME software is aimed above all at solvof criteria.
For given values of criteria, a resulting fuzzy evaluation is ing multiple-criteria evaluation problems. To ensure high percalculated either by the Mamdani fuzzy inference algorithm formance in choosing the optimum alternative, it will be necessary to include in the software other methods of ordering
or by the generalized Sugeno inference.
In the case of the Mamdani fuzzy inference, the degree hi of the fuzzy evaluations in the future. Some approaches are
of correspondence between the given m-tuple of fuzzy values proposed in [7] and further research in this area is planned.
0
0
0
(A1 , A2 , . . . , Am ) of criteria and the mathematical meaning
4 Example of a practical application of the
of the left-hand side of the i-th rule is calculated for any i =
FuzzME software
1, . . . , n in the following way
The FuzzME software was tested e.g. on a soft-fact-rating
0
0
hi = min {hgt(A1 Ai,1 ), . . . , hgt(Am Ai,m )}. (7) problem of one of the Austrian banks. The problem was
solved in co-operation with the Technical University in Vi0
Then for each of the rules, the output fuzzy value Ui , i = enna (see [14]). The fuzzy model of evaluation represents a
1, . . . , n, corresponding to the given input fuzzy values, is cal- part of the creditability evaluation of companies carried out by
the bank - the evaluation according to soft (qualitative) data,
culated as follows
which complements the evaluation according to hard (quan0
y [0, 1] : Ui (y) = min {hi , Ui (y)}.
(8) titative) data. The previous practical experience of the bank
i=1

i=1

showed that it is not good to restrict the evaluation to hard

data only.
In total, 62 companies were evaluated by the fuzzy model.
The goals tree of the model contained 27 qualitative criteria.
During the testing, two approaches were compared - the
original soft-fact-rating model used by the bank and the fuzzy
models created in the FuzzME software.
The original evaluation model used simple discrete numeric
scales with intuitively set linguistic descriptors for the evaluation according to the particular qualitative criteria. The aggregation of partial evaluation was done by the standard weighted
average.
In testing by the FuzzME software, the applied linguistically described numeric scales were analyzed. It was found
that in some cases the correspondence between the linguistic
and numerical values was not perfect. Two new fuzzy models
were formed. The first one used uniform fuzzy scales representing a simple fuzzification of the original numeric scales.
The other worked with fuzzy values which tried to model, as
closely as possible, the linguistic descriptors used in the original evaluation model. The results of the two models were quite
different. At the same time, the normalized crisp weights were
replaced by normalized fuzzy weights which correspond better to the experts knowledge about the importance of criteria.
The subsequent discussion on results of the soft-fact-rating
showed that there exist criteria values and combinations of criteria values which signalize a substantial danger that the company will go bankrupt or at least will have problems acquitting
the debt. That is why, besides the evaluation of companies
based on fuzzy weighted average (average rating), a fuzzy
expert system was applied to calculate another evaluation (a
risk rate of the company). The particular rules of the base
identified the dangerous combinations of criteria values and
assigned to them the corresponding risk rates. The solely use
of the original fuzzy model without the fuzzy expert system
would have lead to a rating score, which may have underestimated the risk inherent to this company. The use of the fuzzy
expert systems offers the possibility to visualize and calculate
such additional risk combinations.
Finally, both evaluations were aggregated with the fuzzy
MINIMAX method. This method is a special case of a fuzzy
OWA operator.The resulting evaluation is the infimum of the
fuzzy numbers representing the partial evaluations.
The obtained results showed that the solid theoretical basis
of the evaluation fuzzy models formed in the FuzzME software improves the quality of evaluations. Positive experiences with such fuzzy models of evaluation could win over
the present-day opponents to the soft-fact-rating in the future.

Figure 2: The simplified structure of the used goals tree

5 Conclusion
The software product FuzzME is a result of many years of
research in the area of the theory and methods of multiplecriteria fuzzy evaluation. The type of evaluation consistently
used in the software corresponds well to the fuzzy sets theory paradigm; the evaluations of alternatives express the fuzzy
degrees of fulfillment of given goals. In the FuzzME software, several new methods, algorithms and tools of fuzzy
modeling were implemented, e.g.: a structure of normalized
fuzzy weights, fuzzy weighted average and ordered fuzzy
weighted average operations and algorithms for their calculation, linguistic scales and linguistic variables derived from
them. Well-elaborated theoretical basis of the FuzzME software provides a clear interpretation of all steps of the evaluation process and brings understanding of methodology to the
user. The FuzzME software is user-friendly. The positive features of the software product proved themselves by solving the
mentioned soft-fact-rating problem.
References
[1] R.E. Bellman and L.A Zadeh. Decision-making in fuzzy enviroment. Management Sci., 17 (4):141164, 2007.
[2] R.R. Yager. On ordered weighted averaging aggregation operators in multicriteria decision making. IEEE Trans.Systems Man
Cybernet, 3 (1):183190, 1988.
[3] Y. J. Lai and C. L. Hwang. Multiple Objective Decision Making.
Springer - Verlag, Berlin, Heidelberg, 1994.
[4] H. Rommelfanger. Fuzzy Decision Support Systeme. Springer
- Verlag, Berlin, Heidelberg, 1988.
[5] INFORM GmbH. FuzzyTECH home page. http://www.
fuzzytech.com/.
[6] C. Von Altrock. Fuzzy Logic and NeuroFuzzy Applications in
Bussiness and Finance. Prentice Hall, New Jersey, 1996.
[7] J. Talasova. Fuzzy methods of multiple criteria evaluation and
decision making (in Czech). Publishing House of Palacky University, Olomouc, 2003.
[8] J. Talasova. NEFRIT - multicriteria decision making based on
fuzzy approach. Central European Journal of Operations Research, 8(4):297319, 2000.
[9] O. Pavlacka. Fuzzy methods of decision making (in Czech).
PhD thesis, Faculty of Science, Palacky Univerzity, Olomouc,
2007.
[10] O. Pavlacka and J. Talasova. Application of the fuzzy weighted
average of fuzzy numbers in decision-making models. New
Dimensions in Fuzzy Logic and related technologies, Proceedings of the 5th EUSFLAT Conference, Ostrava, Czech Repub epnicka, V. Novak, U.
lic, September 11-14 2007, (Eds. M. St
Bodenhofer), II:455462, 2007.
[11] O. Pavlacka and J. Talasova. The fuzzy weighted average operation in decision making models. Proceedings of the 24th International Conference Mathematical Methods in Economics,
13th - 15th September 2006, Plzen (Ed. L. Lukas), pages 419
426, 2006.
[12] J. Talasova and I. Bebca kova. Fuzzification of aggregation operators based on Choquet integral. Aplimat - Journal of Applied
Mathematics, 1(1):463474, 2008. ISSN 1337-6365.
[13] B. Kosko. Fuzzy thinking: The new science of fuzzy logic. Hyperion, New York, 1993.
[14] K. Furst. Applying Fuzzy Models in Rating Systems. Department of Statistics and Probability Theory, Vienna University of
Technology, Vienna, 2008. Term paper.