1. Introduction
Fuzzy multiple criteria decision-making (FMCDM) methods are techniques used to find the trade-off option out of all feasible alternatives that are characterized by multiple criteria and when data cannot be measured precisely. In such situations the ratings of alternatives and the criteria weights, whose evaluations are based on unquantifiable, incomplete, or unobtainable information, are usually expressed by interval numbers [
1,
2,
3] or fuzzy numbers (convex fuzzy numbers—CFNs) [
4,
5,
6]. A new approach consists in using an FMCDM method based on ordered fuzzy numbers (OFNs) which are well suited to handle incomplete and uncertain knowledge and information.
This new approach to multiple criteria decision support has been considered, so far, in a few papers only [
7,
8,
9]. The first application of OFNs for the FMCDM method was presented at the Sixth Podlasie Conference on Mathematics in Bialystok, Poland [
7] by Kacprzak and Roszkowska, and discussed in detail in [
8]. In that paper, the authors evaluated alternatives with respect to criteria using linguistic expressions, in which linguistic terms were quantified on a scale given in advance. The scale was extended to include intermediate values, such as “more than 2” or “less than 3”, which were expressed by trapezoidal OFNs, together with “2” and “3”. The additional property of OFNs (i.e., orientation) was used to include information about the type of criteria (benefit or cost). The authors showed that the Fuzzy Simple Additive Weighting (FSAW) and Fuzzy Technique for Order of Preference by Similarity to Ideal Solution (FTOPSIS) methods based on OFNs can better distinguish alternatives as compared with classical Simple Additive Weighting (SAW) and Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) methods, which used crisp values, and with FSAW and FTOPSIS, which used classical fuzzy numbers. The FTOPSIS method with OFNs has also been used in solving a real-life problem of discrete flow control in a manufacturing system [
9]. The authors have tested the FTOPSIS method with OFNs in a flow control system and compared it to the classical TOPSIS method and to other simple control methods. As a result, they concluded that FTOPSIS with OFNs is better suited for the analysed case than the classical TOPSIS or than most other methods considered.
One of the main steps in FMCDM methods is the determination of the appropriate weights (the relative importance) of the criteria, because of their significant impact on the final result. Since each criterion has a different meaning, we cannot assume that all of them are equally important [
10,
11]. In the literature, various methods to determine criteria weights have been proposed. Most of them can be classified into two categories: subjective and objective weights, depending on the information provided. Subjective weights are determined according to the preferences or judgments of the decision-makers only and can reflect the subjectivity of their knowledge and experience. This category includes: the eigenvector method [
12], the weighted least square method [
13], the Delphi method [
14], and others. Objective weights are determined by solving mathematical models. They disregard subjective preferences or judgments of the decision-makers; instead, they are based on objective information (e.g., decision matrix). This category includes the entropy method [
15], multiple objective programming [
16], and others.
Most of FMCDM applications to real-life problems use only subjective weights. However, when it is not possible to obtain reliable subjective weights, objective weights become useful [
17]. One of the methods of obtaining objective weights applies the notion of Shannon entropy mentioned before. Moreover, let us note that it is reasonable and logical that when ratings of alternatives with respect to criteria are imprecise, as is the case in FMCDM methods, the weights of criteria should also be imprecise. Therefore, Shannon entropy should be extended to imprecise Shannon entropy.
Applications of Shannon entropy to the determination of weights in FMCDM problems have been discussed in the literature. One of the approaches is based on the characterization of a fuzzy number by its
-levels and by an extension of Shannon entropy to interval data [
2]. The main limitation of this approach is that the obtained weights do not have to preserve the property that their ranges belong to [0, 1] (but can be normalized) and they do not sum up to 1. Moreover, the empirical example with five levels {0.1, 0.3, 0.5, 0.7, 0.9} has shown that the rankings for different α-levels can be different. The authors have concluded that the overall ranking of the criteria cannot be easily determined. Another approach to the determination of weights in FMCDM problems uses the concept of defuzzification [
6,
18,
19,
20]. This relies on the conversion of fuzzy numbers into real numbers; afterwards, the classical Shannon entropy is used. In this approach, during defuzzification, we can lose some important information, such as symmetry, width of the support and kernel, location on the
axis of the fuzzy number, etc. Moreover, in the literature we can find many methods of defuzzification of fuzzy numbers and OFNs. Different defuzzification methods can, therefore, generate different rankings of criteria and their relative importance.
The main goal of this paper is to extend the concept of Shannon entropy, using OFNs, which avoid the aforementioned drawbacks. The proposed approach allows to obtain the weights of criteria in the form of OFNs which are normalized and sum to 1. Moreover, several theorems concerning important mathematical properties resulting from the proposed approach, as well as numerical simulations, are presented. An illustrative example will show that the proposed approach based on OFNs gives the same ranking of weights for various α-levels, while the approach from [
2] (with fuzzy numbers) results in a ranking that can change from one α-level to another.
The rest of the paper is organized as follows: In
Section 2 we introduce basic definitions and notations of OFNs. In
Section 3 we present the proposed method of determination of criteria weights using extended Shannon entropy based on OFNs. A simple numerical example is shown in
Section 4. In
Section 5 presents a comparison of the proposed approach with an approach using fuzzy numbers. Conclusions end the paper.
2. Ordered Fuzzy Numbers
In this section, some definitions related to OFNs used in the paper are briefly presented. The model of ordered fuzzy numbers (OFNs) was introduced and developed by Kosiński and his two co-workers, Prokopowicz and Ślęzak, in 2002 in a series of papers [
21,
22,
23,
24]. In our paper, we use the term “ordered fuzzy numbers” which was proposed by Professor Kosiński and his colleagues Prokopowicz and Ślęzak in 2002. After the death of Professor Kosiński, to commemorate and honour him, these numbers are often called “Kosinski fuzzy numbers” [
25]. Arithmetic operations in this model are similar to the operations on real numbers, which are a special case of OFNs.
Definition 1. [
23,
26].
An ordered fuzzy number is an ordered pair of continuous functions The set of all OFNs will be denoted by
. The elements of the OFN
are called:
, the up part, and
, the down part. To conform with the classical notation of fuzzy numbers, the independent variable of both functions
and
will be denoted by
, while their values, by
(
Figure 1a). The continuity of both functions
and
implies that their images are bounded intervals, called
and
, respectively (
Figure 1a). Their endpoints will be described as follows:
and
.
In general, the functions
and
of the OFN
need not be invertible as functions of the variable
; only continuity is required in Definition 1. However, if we assume, additionally, that [
24]: (A1)
is increasing and
is decreasing, (A2)
(pointwise), we can define the membership function of the OFN
as follows (
Figure 1b):
If the functions
and/or
are not invertible or assumption (A2) is not satisfied, we can obtain so-called improper OFNs (
Figure 2). Instead of the membership function, we can then define the membership curve (or relation) in the
-plane, consisting of the functions
and
(as functions of the variable
) and a segment of the line
over the interval
. Moreover, let us note that in general
does not have to be smaller than
(
Figure 2), in which case we can obtain improper intervals for
and/or
(discussed in the context of extended interval arithmetic by Kaucher [
27]).
Figure 1c shows an OFN as a fuzzy number in the classical meaning (as a convex fuzzy number). Its membership function has an extra arrow denoting the orientation of the OFN, e.g., the order of the inverse functions
and
. The pair of continuous functions
determines a different OFN than the pair
.
Figure 3 shows that although the two curves have an identical shape, the corresponding membership functions determine two different OFNs.
The set of OFNs can be divided into two subsets:
numbers with a positive orientation, if the direction of the OFN is the same as the
-axis (
Figure 3a); and
numbers with a negative orientation, if the direction of the OFN is opposite (
Figure 3b).
Orientation can be used to present additional information, for example:
if we describe an object’s speed, orientation tells us whether the object is moving away or towards us;
if we analyse the total revenue and/or the total cost of a company, orientation can describe the dependence of the current value on the reference value (increasing or decreasing);
if we apply FMCDM methods, orientation can show the type of the criteria (benefit or cost).
Basic arithmetic operations on OFNs are defined as pairwise operations on their functions
and
. Let
,
, and
be OFNs. The arithmetic operations: addition
, subtraction
, multiplication
, and division
on them are defined in
as follows:
where “
”
{
,
,
, /} and where
is defined when
and
for each
.
Since real numbers are a special case of OFNs, they can be represented in
as follows. Let
and let
be the constant function, i.e.,
for all
. Then
is the OFN which represents the real number
in
. Now we can define the multiplication of an OFN
by a real number
by the formula:
In Equation (1) of the membership function of the OFN
there are four characteristic real numbers:
,
,
, and
. If the functions
and
are linear, these four numbers uniquely describe
as follows (
Figure 4):
The number
is called a trapezoidal OFN if
(
Figure 4) and a triangular OFN if
which for simplicity is often written as follows:
A trapezoidal OFN
, where
is called normalized. The representation of Equations (4) or (5) allows us to quickly perform arithmetic operations on trapezoidal (triangular) OFNs using these characteristic points. Let
and
be trapezoidal OFNs. The arithmetic operations on these numbers are then defined by the formula:
where
.
Figure 5a illustrates addition of two triangular OFNs
and
, and the result
, whereas
Figure 5b illustrates multiplication of a trapezoidal OFN
by a real number
and the result
.
Defuzzification is a main operation in fuzzy controllers, fuzzy inference systems, and FMCDM methods, which allows to rank OFNs.
Definition 2. [
28].
A map from the space of all OFNs to reals is called a defuzzification functional if it satisfies:for any and , where represents a crisp number.
Let
be an OFN. In
, the most frequently used defuzzification methods for
are [
28]:
- -
- -
- -
- -
RCOM (random choice of maximum)
- -
- -
KKCOM (KK choice of maximum)
- -
3. Fuzzy Criteria Weights Based on Fuzzy Shannon Entropy
In this section, we propose an extension of Shannon entropy to a fuzzy environment which will be used to obtain the fuzzy criteria weights for an FMCDM method based on OFNs.
Let us consider a multi-criteria problem which consists of the set of alternatives
and the set of criteria
. In general, the criteria can be classified into two types: benefit (
) and cost (
). For a benefit criterion, a higher value is better, while for a cost criterion, a smaller value is better. A multi-criteria problem is usually expressed in matrix form, as follows:
where
is the rating of the
th alternative with respect to the
th criterion. Additionally, the relative importance of criteria is given by a vector of weights:
where
is the weight of criterion
, satisfying the condition
.
Most of the MCDM applications to real-life decision-making problems use only subjective weights defined by the decision-maker. However, when it is not possible to obtain reliable subjective weights, objective weights become useful [
17]. One of the methods of obtaining objective weights is the application of the concept of Shannon entropy.
Entropy is a term from information theory which is also known as the average (expected) amount of information [
29] contained in each criterion (each column of the decision matrix Equation (14)). The greater the value of entropy in a specific criterion, the smaller the differences in the ratings of alternatives with respect to this criterion. This, in turn, means that this criterion provides less information and has a smaller weight. It follows that this criterion becomes less important in the decision-making process.
Let us consider the decision matrix (Equation (14)), where . Then criteria weights can be determined as follows:
- 1.
Construct the normalized decision matrix
, where:
- 2.
Construct the vector of Shannon entropy
, where:
and
is defined as 0 if
.
- 3.
Calculate the vector of diversification degrees
, where:
The higher the degree
, the more important the corresponding criterion
.
- 4.
Calculate the vector of criteria weights
, where:
It is reasonable and logical that when the ratings of alternatives with respect to criteria are represented by OFNs, the weights of criteria should be also represented by OFNs. This means that the concept of Shannon entropy needs to be extended to fuzzy Shannon entropy based on OFNs.
We will present a method of determining the weights of criteria based on OFNs using triangular OFNs (the proposed method can be easily extended to trapezoidal OFNs). The orientation of OFNs will be used to distinguish between types of criteria. Namely, to represent the value of a benefit criterion we use positive triangular OFN , i.e., an OFN such that , while for a cost criterion, negative triangular OFN , i.e., an OFN such that . Then the calculation of criteria weights based on triangular OFNs can be described in the following steps.
- STEP 1:
Construct the fuzzy decision matrix
, where:
is the rating of alternative
with respect to criterion
represented by a triangular OFN.
- STEP 2:
Construct the normalized fuzzy decision matrix
, where:
If, for all we have or or , we define or or to be 0, respectively.
Remark 1: The orientation of OFNs is used for input data to distinguish between the types of criteria. Let us note that during the calculations using Equation (21) the orientation of the resulting OFNs can change (from positive to negative, and vice versa) and even improper OFNs can occur. This is shown in the numerical example presented below.
Let us consider the following three OFN-rated alternatives with respect to j-th benefit criterion: , , and . The normalized OFNs, calculated using Equation (14), are: , , and . Note that after normalization, has the same orientation as the input data, has the opposite orientation, and is an improper OFN.
- STEP 3:
Construct the vector of fuzzy Shannon entropy
, where:
and
or
or
is defined as 0 if
or
or
, respectively.
- STEP 4:
Calculate the vector of fuzzy diversification
, where:
- STEP 5:
Calculate the vector of fuzzy criteria weights
, where:
If for all we have or or , we define or or to be 0, respectively.
Theorem 1. The criteria weights (Equation (24)) are normalized.
Proof. Let us note that for all we have , and . Hence, it follows that , , and , which means that . ☐
Theorem 2. If all the ratings of alternatives with respect to criteria are crisp data then the proposed method of determining the weights of criteria using the fuzzy Shannon entropy based on OFNs leads to the classical method of determining the weights of the criteria using the classical Shannon entropy.
Proof. Let us note that if all the ratings of alternatives with respect to criteria are crisp data, then in Equation (20) we have and also in Equation (21). This means that from Equation (22) we have and, therefore, in Equation (23). Finally, from Equation (24) we have . ☐
Theorem 3. If at least one of the ratings of any alternative with respect to any criterion is an OFN, all the criteria weights are also OFNs.
Proof. Assume that for and we have , which means that is an OFN. Then from Equation (21) we have for all . Next, from Equation (22) we have and from Equation (23) . Finally, from Equation (24) we have for all . ☐
Theorem 4. The obtained weights from Equation (24) satisfy the condition .
Proof. From Equation (24) we have:
☐
Theoremf 5. If all alternatives are evaluated identically with respect to a criterion, for example, th criterion , then .
Proof. Assume that for each alternative the evaluations with respect to -th criterion are the same and equal to . Therefore, using Equation (21), after the normalization for all , we have . Since and using Equation (22), we have . From Equation (23) we have and from Equation (24) we obtain . ☐
Theorem 6. If all alternatives are evaluated identically with respect to a criterion, for example, k-th criterion and j-th criterion , e.g., for , then
Proof. Assume that for each alternative the evaluations with respect to -th criterion and j-th criterion are the same, e.g., for . Then, using Equation (21), we have for and when we calculate the entropy using Equation (22), we obtain . Hence, from Equation (23) we have and, finally, using Equation (24), we obtain . ☐
4. An Illustrative Example
In this section, we present a simple numerical example of the proposed method. Let us consider the multi-criteria problem of selecting a provider of medical equipment to a medical centre. Four bidders , , , and responded to the invitation to bid. They are rated with respect to the following criteria: : price, : length of warranty, : conditions of service, : multifunctionality of the equipment, that is, the capability for extension and modification, : payment term, : comprehensiveness of the offer, that is, training, delivery, installation, and possibility of leasing.
For the ratings of the alternatives with respect to the criteria the linguistic variables from
Table 1 are used. The results of the ratings are shown in
Table 2. The linguistic variables from
Table 2 are converted into triangular OFNs; the corresponding fuzzy decision matrix is presented in
Table 3. Next, this matrix is normalized using Equation (21) and presented in
Table 4. Based on the data from
Table 4 and using Equations (22)–(24), the vector of fuzzy Shannon entropy, the vector of fuzzy diversification, and the vector of fuzzy criteria weights are calculated and shown in
Table 5. To determine the ranking of the criteria we can use one of the defuzzification methods of Equations (7)–(13) presented in
Section 2. The simplest defuzzification method for triangular OFNs consists in selecting a central value which is equivalent to using one of the Equations (7)–(12).
Table 6 presents the results of defuzzification; the last row shows the rank of each criterion
.
5. A Comparison of the Proposed Approach with the Approach Based on CFN
In this section, the approach to the determination of fuzzy criteria weights presented in
Section 3 is compared with the approach based on fuzzy numbers described in [
2]. To show differences in these approaches we use the illustrative example presented in
Section 4 and the concept of α-levels. To compare the α-levels and create a ranking of criteria we will use the method proposed by Hu and Wang [
30]. This method used another characterization of an
-level
, using its centre
and its radius
. Thus, any
-level can be written in the form
. If
is another
-level, the method proposed by Hu and Wang can be written as:
and:
Table 7 presents weights and rankings of criteria for selected values of
, both for classical fuzzy numbers and for OFNs. Calculations have shown that for CFNs and
from 0 to
the ranking of criteria is the same, namely
. For greater values of
the ranking of criteria is different. Starting from
the ranking is as follows
, which is the same as the ranking presented in
Table 6. This shows that, for different values of
, different rankings are obtained. It is, therefore, difficult to determine the overall ranking of the criteria. On the other hand, the application of the approach proposed in
Section 3 gives the same ranking for
, which is compatible with the rankings obtained in
Table 7 for CFNs and
.
Remark 2. A fuzzy number can be approximated by a triangular fuzzy number using -levels. Then its central value is determined by setting ,
while its support is obtained by setting [
31].
Using the above remark and the approach proposed in [
2], fuzzy numbers representing weights of criteria can be determined.
Table 8 presents fuzzy criteria weights
, fuzzy criteria weights
defuzzified by the centre of gravity (Equation (13)), and the ranking of criteria
. The resulting ranking of criteria weights is:
; this is different from the rankings presented in
Table 7 for CFNs and for different values
.
Table 9 presents OFN-based fuzzy criteria weights
, fuzzy criteria weights
defuzzified by the centre of gravity (Equation (13)), and the ranking of criteria
. The obtained ranking of criteria weights is:
, compatible with the ranking obtained in
Table 6 and
Table 7 for OFNs.