method
Tom´aˇs Tal´aˇsek∗†, Jan Stoklasa‡, Mikael Collan‡and Pasi Luukka‡
∗LUT Graduate School, Lappeenranta University of Technology Skinnarilankatu 32, 53851, Lappeenranta, Finland
Email: tomas.talasek@gmail.com
†Department of Applied Economics, Palack´y University, Olomouc Kˇr´ıˇzkovsk´eho 12, 77180, Olomouc, Czech Republic
‡School of Business and Management, Lappeenranta University of Technology Skinnarilankatu 32, 53851, Lappeenranta, Finland
Email: jan.stoklasa@lut.fi, mikael.collan@lut.fi, pasi.luukka@lut.fi Abstract—Linguistic approximation is a suitable way of
trans-forming mathematical outputs into words that can be easily used and understood by laymen. The methods for linguistic approximation range from simple distance based ones to more complex methods aspiring on finding high semantic match be-tween the approximated output and its linguistic label. This paper builds on Bonissone’s proposal of a two step method for linguistic approximation based on a pattern recognition approach.
It suggests an algorithm for finding a partial ordering of fuzzy numbers utilizing the partial results from the two step method.
As such it proposes a means of finding a partial ordering of fuzzy numbers through linguistic approximation. The proposed algorithm is showcased on several numerical examples and its performance is briefly discussed.
I. INTRODUCTION
In practical applications the outputs of the mathematical models are frequently in the form of numbers or fuzzy sets. The users of mathematical models therefore must be trained how to interpret these outputs. Since both numbers and fuzzy sets are not a natural means of presenting information (knowledge of the mechanism that generated the results might be necessary for appropriate interpretation), there is a risk (especially for the new practitioners) that they misinterpret the output and react inappropriately. Reasonable way how to avoid the problem of misinterpretation is to present the output of the model in a more natural way for the users – in the linguistic form. The process of translation between fuzzy sets (or real numbers) and words is called linguistic approximation. This paper focuses on one particular linguistic approximation method proposed by Bonissone [1] and suggest a extension of this method that would allow direct comparisons of mathematical outputs through linguistic approximation (or using the information computed to obtain the linguistic approx-imation). Basic notions of the theory of fuzzy sets are defined in Section 2. Bonissone’s method for linguistic approximation is presented in Section 3. In the next Section, the extension of Bonissone’s method is proposed together with the ideas how this method could be used for the purposes of ordering fuzzy numbers. Behavior of the proposed algorithm for the ordering
of fuzzy numbers is studied on three numerical examples in Section 4.
II. PRELIMINARIES
LetU be a nonempty set (the universe of discourse). A fuzzy setAonUis defined by the mappingA:U→[0,1]. For eachx∈Uthe valueA(x)is called amembership degreeof the elementxin the fuzzy setAandA(.)is called a member-ship functionof the fuzzy setA. Ker(A) ={x∈U|A(x) = 1} denotes akernelofA,Aα={x∈U|A(x)≥α}denotes an α-cutofAfor anyα∈[0,1], Supp(A) ={x∈U|A(x)>0} denotes asupportofA.
A fuzzy number is a fuzzy setAon the set of real numbers which satisfies the following conditions: (1) Ker(A)6=∅(A is normal); (2) Aα are closed intervals for allα ∈ (0,1]
(this implies A is unimodal); (3) Supp(A) is bounded. A family of all fuzzy numbers onUis denoted byFN(U). A fuzzy number Ais said to be defined on [a,b], if Supp(A) is a subset of an interval[a, b]. Real numbersa1 ≤a2 ≤ a3 ≤a4 are calledsignificant valuesof the fuzzy number A if[a1, a4] =Cl(Supp(A))and[a2, a3] =Ker(A), where Cl(Supp(A))denotes a closure of Supp(A). Anunionof two fuzzy sets Aand B (based on Lukasiewicz disjunction) is a fuzzy set (A∪LB)defined as follows: (A∪LB)(x) = min{1, A(x) +B(x)},∀x∈U. Fuzzy setAonUis asubset of fuzzy setBonU(A⊆B) ifA(x)≤B(x),∀x∈U.
The fuzzy number Ais called linear if its membership function is linear on [a1, a2] and [a3, a4]; for such fuzzy numbers we will use a simplified notationA= (a1, a2, a3, a4).
A linear fuzzy numberAis said to betrapezoidalifa26=a3
andtriangular ifa2 =a3. We will denote triangular fuzzy numbers by ordered tripletA= (a1, a2, a4). More details on fuzzy numbers and computations with them can be found for example in [2].
A partial ordering of fuzzy numbers can be defined in the following way: LetAandBbe fuzzy numbers on[a, b]then A > BifAα> Bα∀α∈(0,1]that is if([infAα,supAα]>
[infBα,supBα]) ⇔ ((infAα > infBαandsupAα ≥
– 285 –
CINTI 2015 • 16th IEEE International Symposium on Computational Intelligence and Informatics • 19–21 November, 2015 • Budapest, Hungary
978-1-4673-8520-6/15/$31.00©2015 IEEE
a semantic rule which to every linguistic valueA ∈ T(V) assigns its meaningA = M(A) which is usually a fuzzy number onX.Linguistic approximation is the process that assign appropriate labels (known linguistic terms of a linguistic scale) to general fuzzy sets. From mathematical point of view it is a mapping from the set of all fuzzy sets onXtoT(V).
A linguistic variable(V,T(V), X, G, M)is called a lin-guistic scaleon[a, b]ifX= [a, b],T(V) ={T1, . . . ,Ts}and Ti=M(Ti), i= 1, . . . , s, form a fuzzy scale on[a, b]. Terms T1, . . . ,Ts are called elementary terms of linguistic scale.
Theextended linguistic scaleis linguistic scale, that besides elementary terms T1, . . . ,Ts contains also derived terms in the formTitoTj, where i < jand i, j ∈ {1, . . . , s}and M(TitoTj) =Ti∪LTi+1∪L. . .∪LTj. The extended linguistic scale thus contains linguistic values of different levels of uncertainty – from the possibly least uncertain elementary terms{T1, . . . ,Ts}to the most uncertain linguistic termT1s
(Uncertainty can be assessed by the cardinality of the meanings of these linguistic terms). Derived linguistic termsTitoTj
are calledlevelj−itermsand can be also denoted byTij. Elementary linguistic termsTiare calledlevel 1 termsand can be also denoted byTii(i.e.Ti=Tiito unify the notation).
More details on linguistic scales and extended linguistic scales can be found for example in [6].
Linguistic hedgeis a word (generally an adverb) that can be applied to a linguistic term to modify its meanings. From mathematical point of view a linguistic hedge is a function which modifies the membership functions of fuzzy sets -for example the linguistic hedgevery (using Zadeh’s [10]
definition) applied to a fuzzy setAonUresults is a fuzzy setveryAwith membership function defined(veryA)(x) = A2(x), x∈U.
We are frequently required to be able to represent fuzzy sets by real numbers, this procedure is calleddefuzzification.
In applications an approximation of a fuzzy setAby its center of gravity (COG)tAis frequently used. Thecenter of gravity of a fuzzy setAdefined on[a, b]is defined by the formula tA=Rb
aA(x)x dx/Rb
aA(x)dx. Other possible defuzzification methods are discussed in [3].
III. TWO STEP METHOD FOR LINGUISTIC APPROXIMATION
In 1979 Bonissone [1] introduced his two step pat-tern recognition approach for linguistic approximation. This method consist of two steps – in the first step the set ofm suitable linguistic terms of some linguistic variable (suitability
in the way, that m terms with characteristics most similar to the ones possessed by the output Outare preserved. For this purpose eachTi, i= 1, . . . , nis represented by four real numbers, that each represent one of the four features (selected by Bonissone [1]) of this fuzzy set: i) Cardinality of fuzzy set Ti(Card(Ti)); ii) Fuzziness of the fuzzy setTi(nonprobalistic entropy, Entropy(Ti)); iii) Center of gravity of fuzzy setsTi
(COG(Ti)); iv) Skewness of fuzzy setTi(SKEW(Ti)). These four features represent the fuzzy set (and the linguistic term Ti associated with it) in four-dimensional space. For pres-election of linguistic terms the weighted Euclidean distance d1(Ti, Out)is computed between the fuzzy setOutand each of the fuzzy setsTi, i= 1, . . . , nrepresented by a quadruplet of the numerical values of the chosen four features using Formula (1). A reordered set of linguistic terms{Tp1, . . . ,Tpn}
wherewiare normalized real weights1(i.e.P4 i=1wi= 1
In the second step, the linguistic approximationT∗∈ Tp
of the fuzzy setOutis computed. Fuzzy setT∗(which is a mathematical meaning of a linguistic termT∗) is computed
1The choice of weights is usually left with the user of the model and some features could be even optional. Wenstøp [7] for example proposed (in his method for linguistic approximation) to use only two features – cardinality and center of gravity.
– 286 –
as T∗ = arg minTpj∈Tpd2(Tpj, Out) using the modified
whereA∗(x) = A(x)/Card(A(x)). This way the linguistic termT∗is found as the closest linguistic approximation among the preselected linguistic terms.
IV. ORDERING OF FUZZY SETS THROUGH LINGUISTIC APPROXIMATION BASED ON THETWO STEP METHOD
Linguistic approximation is usually used in situations when the user of the mathematical model requires the results in an understandable form (i.e. in a linguistic form). To find a proper linguistic approximation lots of calculations may need to be done and substantial part of these result is not used for anything else (see the complexity of Bonissone’s method described in the previous Section). This raises a question, whether it is possible in situations when we need to approximate more outputs of mathematical models (e.g. outputs of model representing evaluation of various alternatives in decision making problems – for more information see e.g. [5]) to use linguistic approximation (or partial results obtained through calculation of linguistic approximation) for the purposes of decision making – for example to order the alternatives with respect to their evaluations. In this paper we are therefore focusing on the possibility of using the information obtained in the process of linguistic approximation using the two step method to order the outputs of a mathematical model (fuzzy numbers onX). For an overview of other possible methods for ordering fuzzy numbers and their reasonable properties see [8], [9] .
In this paper the structure of an extended linguistic scale (V,T(V), X, G, M)is used. We suppose that the user of the mathematical model specifies (or at least approves) the mean-ings of itsnthe elementary linguistic terms of this scale. Lin-guistic terms of this extended linLin-guistic scale can be ordered by the user directly (this could be difficult and time consuming) or can be partially ordered through the partial ordering of fuzzy numbers that represent the meanings of these linguistic terms, whereTijis preferred toTkl,Tij Tkl⇔M(Tij)> M(Tkl), whereTij,Tkl ∈ T(V),1≤i < j≤n,1≤k < l≤n.
Linguistic termsTij andTklare incomparable according to this ordering, ifi > kandj < l. In the case of incomparamble fuzzy numbers, we can use a different method to obtain the ordering of these fuzzy numbers (e.g. a method based on the centers of gravity, whereTijtTkl⇔tM(Tij)> tM(Tkl), see Fig. 1). However, from our point of view it is not reasonable to present ordering obtained through the center of gravity method to the user due to possible information loss. It is reasonable to present such terms as incomparable (and eventually let the ordering of these term on the user of the model).
Methods for linguistic approximation usually use complex structures of linguistic terms (linguistic hedges may even be applied to generate the set of linguistic terms). In these cases there is a potential risk, that the user may not understand all these linguistic terms correctly. Therefore in our proposed method the elementary linguistic terms form a linguistic scale
1
Fig. 1. Fuzzy numbersAandBthat are incomparable through natural ordering of fuzzy numbers and center of gravity method (top Figure); fuzzy numbersAandBthat are incomparable using the natural ordering method, but can be ordered based on their centers of gravityBtA(bottom Figure).
and the user must confirm, that he/she understands each elementary linguistic terms correctly and that its fuzzy number meaning is appropriately defined. This in combination with the construction of the linguistic term set of the extended linguistic scale ensures, that the user understands correctly all possible outputs of linguistic approximation (output can be either one elementary term or a derived term represented as two elementary terms connected by ”to”). Therefore the user works only with objects (linguistic terms, respectively fuzzy numbers representing their meaning) that he/she understands or with their Lukasiewicz union. The use of the extended linguistic scale ensures, that the output of linguistic approximation is a linguistic term, which meaning is modeled by a normal, unimodal fuzzy set (a fuzzy number).
Let us now take a closer look on the use of the two step method for linguistic approximation in ordering of fuzzy numbers. Let us consider r outputs of some mathematical modelOut1, . . . , Outrin the form of fuzzy numbers onX. Let us consider an extended linguistic scale(V,T(V), X, G, M) withnelementary termsT1, . . . ,Tn,. This linguistic scale will be used for the linguistic approximation of Out1, . . . , Outr and also to obtain an ordering of these outputs.
According to Bonissone in the first step of linguistic approximation it is our goal to reduce the set of all linguistic termsT(V)in the way, thatmterms with meanings seman-tically closest to the ”ideal” linguistic description of each ap-proximated output are found,TpOutq={TpOut1 q, . . . ,TpOutm q}, where TpOuti q ∈ T(V), i = 1, . . . , m, q = 1, . . . , r. Since the semantic context is provided by the universe on which the meanings of the linguistic values of the output linguistic variable are defined, it is sufficient to account for the position, fuzziness and shape of the fuzzy sets (that represent meanings of linguistic terms) to find a pair of semantically close ones.
In the second step the most appropriate linguistic approx-imation of each output is found from its preselected term set.
In the process the Bhattacharyya distanced2(M(Tij), Outq) between the meaning of each linguistic termTijfromTpOutq
and each output Outq, q= 1, . . . , ris computed. Therefore
– 287 –
CINTI 2015 • 16th IEEE International Symposium on Computational Intelligence and Informatics • 19–21 November, 2015 • Budapest, Hungary
1) Leti= 1.
2) We computeTappOut1andTappOut2.
3) If TapproxOut1 i TapproxOut2i then Out1 is preferred to Out2 or ifTapproxOut2 i TapproxOut1 i thenOut2 is pre-ferred to Out1. END (ordering has been found).
ELSE Go to 4.
4) IfTapproxOut1 iandTapproxOut2iare incomparable, leave the ordering of these two outputs to the user2. END (outputs are incomparable). ELSE Go to 5.
5) Ifi < mandTapproxOut1 1=TapproxOut21, then increasei and Go to 3. ELSE Go to 6.
6) Find a pair of adjacent linguistic terms of the same level TijOut1,Ti+1,j+1Out1 and beginning of six and choose another pair. END (ordering has been found).
V. EXAMPLES OF THE PROPOSED METHOD: Let(V,T(V),[0,1], G, M)be a extended linguistic scale with five elementary linguistic termsT1, . . . ,T5 with mean-ings specified in Table I. Meanmean-ings of elementary terms are depicted in Fig. 2. Derived linguistic terms are constructed in accordance with Chapter 2. Outputs of the mathematical model are fuzzy numbers on[0,1](in examples we consider only triangular fuzzy numbers as approximated outputs). For the preselection step the value ofmis equal to 5 (five linguistic terms are preselected).
In next three subsections we present three different exam-ples and stress the strong and weak properties of the presented model for ordering of the outputs.
2This situation is depicted in Fig. 1. These situations involve the compar-isons of fuzzy sets (one is a subset of the other) with different cardinalities.
The natural ordering based onα-cuts of these fuzzy sets is non-existent. Since we are looking for easily interpretable results, we prefer at this point to present both outputs to the decision maker as incomparable and leave the choice of the better one with him/her.
0.25
0
0 0.25 0.5 0.75 1[x]
Fig. 2. Meanings of the elementary linguistic terms of the extended linguistic scale used for linguistic approximation.
Fig. 3. Outputs of mathematical model used in Example 1 (AandB) and Example 2 (CandD).
A. Example 1
Let fuzzy numbers A = (0.1,0.2,0.3) and B = (0.2,0.3,0.4) (presented in Fig. 3) be two outputs of a mathematical model – evaluations of two different alternatives.
Our goal is to choose a better alternative using the algorithm proposed in this paper.
At first the preselection of linguistic terms fromT(V)is performed. Order of all the linguistic terms ofVwith their distances from the fuzzy setA(B) is presented in Table II.
Five terms are preselected for the second step –T1,T2,T12, T3andT23(these terms are the same for both outputsAand B). In the second step the terms are ordered with respect to Bhattacharyya distance and the results are presented in Table III.
As can be seen from the Table III, the ordering of the first three preselected terms is identical for bothAandB. However the fourth suggested term forB(T3) is preferred to the fourth suggested term forA(T1). HenceBis better thanA.
B. Example 2
Let fuzzy numbers C = (0.6,0.7,0.8) and D = (0.7,0.8,0.9)(presented in Fig. 3) be two outputs of a math-ematical model we need to order.
Again five terms are preselected for the second step –T4, T45,T34,T5andT3. These terms are the same for both outputs CandDand even their ordering is identical (see Table IV).
– 288 –
TABLE II. OUTPUTS OF THE FIRST STEP OF THE TWO STEP METHOD FOR FUZZY SETSAANDB(d1DISTANCES TO THE MEANINGS OF ALL
LINGUISTIC TERMS OFV);ORDERED. Preselection of linguistic terms for fuzzy numberA
Linguistic term T1 T2 T12 T3 T23
Preselection of linguistic terms for fuzzy numberB
Linguistic term T2 T1 T3 T12 T23 TABLE III. PRESELECTED TERM SETS FOREXAMPLE1AND THE
BHATTACHARYYA DISTANCES OF THEIR MEANINGS TOAANDB RESPECTIVELY. TABLE IV. PRESELECTED TERM SETS FOREXAMPLE2AND THE
BHATTACHARYYA DISTANCES OF THEIR MEANINGS TOCANDD RESPECTIVELY.
As Table IV suggests, Step 6 of the proposed algorithm needs to be applied in this case. That is first we need to find two adjacent linguistic values of the same level (pairsT3,T4;T4,T5
andT34,T45can be used). Based on the pairT4,T5outputD is considered better than outputC. The same result can be obtained for the pairT34,T45. PairT3,T4 does not provide information based on which the ordering can be found.
C. Example 3
Let fuzzy numbersA= (0.1,0.4,0.7),B= (0.3,0.4,0.5) andC= (0.3,0.6,0.9),D= (0.5,0.6,0.7)(presented in Fig.
4) be two pair of outputs of a mathematical model we need to order.
Second step results are presented in Table V. The results suggest, thatB is better thanA whileC is better thanD.
It is worth noting, that both cases are similar (B ⊆Aand D⊆C; the numerical values of the Bhattacharyya distance to the ordered preselected terms are identical forA andC and forBandD, although different terms were preselected).
However, in the first case the output with lower cardinality is considered better whereas in the second case the output with larger cardinality is preferred.
Fig. 4. Outputs of mathematical model used in Example 3.
TABLE V. PRESELECTED TERM SETS FOREXAMPLE3AND THE BHATTACHARYYA DISTANCES OF THEIR MEANINGS TOA,B,CANDD
RESPECTIVELY.
In the paper we have proposed a utilization of the infor-mation obtained by the two step method for linguistic ap-proximation by Bonissone for the ordering of fuzzy numbers.
This partial ordering is derived from the best candidates on appropriate linguistic approximation of the fuzzy numbers to be approximated. The extended linguistic scale has been proposed as a suitable linguistic variable for the approximation.
The performance of the proposed algorithm is showcased on three numerical examples.
ACKNOWLEDGMENT
The research presented in this paper was partially supported by the grant IGA FF 2015 014 of the internal grant agency of the Palack´y University, Olomouc.
– 289 –
CINTI 2015 • 16th IEEE International Symposium on Computational Intelligence and Informatics • 19–21 November, 2015 • Budapest, Hungary
Lappeenranta: Lappeenranta University of Technology, 2014.
[7] Wenstøp, F.: Quantitative analysis with linguistic values.Fuzzy Sets and Systems,4(2), (1980), 99–115.
[8] Wang, X. and Kerre, E. E.: Reasonable properties for the ordering of fuzzy quantities (I).Fuzzy Sets and Systems,118(3), (2001), 375–385.
[9] Wang, X. and Kerre, E. E.: Reasonable properties for the ordering of fuzzy quantities (II).Fuzzy Sets and Systems,118(3), (2001), 387–405.
[10] Zadeh, L. A.: The concept of a linguistic variable and its application to approximate reasoning I, II, III.Inf. Sci.,8, (1975), 199–257, 301–357, 9, (1975), 43–80.
– 290 –
Publication IV
Tal´aˇsek, T., Stoklasa, J. and Talaˇsov´a, J.
Linguistic approximation using fuzzy 2-tuples in investment decision making
Reprinted with the permission from
Proceedings of the 33rd International Conference on Mathematical Methods in Economics 2015,
pp. 817–822, 2015,
c 2015, University of West Bohemia, Plzeˇn