CIPRIAN COSTIN POPESCU
An improvement an optimization algorithm with application to LR and LL fuzzy numbers, is given. Two situations are presented: the case of LR fuzzy numbers and the special LL case.
AMS 2000 Subject Classification: 90C05.
Key words: LR-fuzzy number, LL-fuzzy number, weighted least squares.
1. INTRODUCTION
We discuss and extend an optimization method [2] based on the fuzzy least squares approach. The improvement consists in the introduction of the weighted model which substantially enlarge the range of the applications. In Section 2 we recall some basic definitions and results about fuzzy sets. Sec- tion 3 contains the general model setted up on n statistical units. In fact, our aim is to discuss the problems which occur when the weighted algorithm is applied to the particular case of LL fuzzy numbers, as in Section 5. Till then it is necessary to recall some results on more generall LR fuzzy numbers.
These preliminary results are presented in Section 4.
2. SOME DEFINITIONS AND PRELIMINARIES
We start with
Definition 2.1. Let X be a universal set [9, 11], a collection of objects denoted by x [1]. Then a fuzzy set A, A ⊂X is defined as a set of ordered pairs, namely, A = {(x, µA(x))|x∈X}. Here, µA(x) is a function called membership function of A inX with the properties
(i)µA(x) :X →[0,1];
(ii)µA assigns to each elementx∈X a numberµA(x)∈[0,1] called the grade of membership (or “degree of compatibility” or “degree of truth”, see [1]) of x inA.
MATH. REPORTS10(60),2 (2008), 197–203
Definition 2.2. Denote by Aα ⊂ A the subset of elements x ∈A whose degree of membership is at least α. Then Aα is called the α-level set [9] (or the α-level cut [1] and, clearly, Aα={x∈X |µA(x)≥α}.
The setA0α ={x∈X |µA(x)> α} is called the strongα-level cut [1].
Definition 2.3. A fuzzy setA⊂X is said to be convex if µA(λx1+ (1−λ)x2)≥min{µA(x1), µA(x2)}, for any x1, x2 ∈X and λ∈[0,1] [1].
Definition 2.4. IfA andB are two fuzzy sets, we say thatA⊂B if and only if µA(x) ≤ µB(x) for all x ∈ X. For a fuzzy set A, its complement A is defined as a fuzzy set with membership function µA(x) = 1−µA(x) for all x∈X see [9]).
Definition 2.5. A fuzzy number M is of LR type if there exist two functions L (left), R (right) and the scalars a, b > 0 such that µM(x) = L(m−x/a) if x ≤m, and µM(x) = L(x−m/b) if x ≥m, where m ∈R is the mean value ofM anda, bare called the left and right spreads respectively [1]. A triangular fuzzy number is a particular type of semisymmetricLRfuzzy number [1, 8, 13].
3. THE MODEL
We start with the model given by Coppi et al. [2]. Consider input crisp variables X1, . . . , Xm and a fuzzy output variable, Ye, on a sample of size n.
The data will be denoted by (yei,xi),i= 1, . . . , n.
Assume thatYe is anLRfuzzy variable: Ye = (m, a, b)LR,a, b >0, where µYe(y) =L(m−y/a) if y ≤m, and µ
Ye(y) =L(y−m/b) if y ≥m. Consider the theoretical values mt, at, bt and the errors ε, εA, εB. Thus, see [2], we can write
m=mt+ε, m−a= (mt−at) +εA, m+b= (mt+bt) +εB. We replace the parametersα,βA,βB,γA,γBwhich allow us to transform these equations into a regression model, namely,
mt=Fα, at=βAmt+γA1, bt=βBmt+γB1, where Fis a matrix with rows fiT = [f1(xi), . . . , fp(xi)] (see again [2]).
The theoretical values of the output variables are yei∗
= ((mt)i,(at)i,(bt)i)LR, i= 1, n.
Next, we will generalize the distanced2and the theoretical results from [2].
4. WEIGHTED LR CASE
Definition 4.1. Forw= (w1, w2, w3), where w1, w2, w3∈R∗+, define d2(w;y,eye∗) =d2(w; (m,a,b)LR,(mt,at,bt)LR) =
=w1km−mtk2+w2k(m−λa)−(mt−λat)k2+w3k(m+ρb)−(mt+ρbt)k2, where λ=R1
0 L−1(r) dr and ρ=R1
0 R−1(r) dr.
Theorem 4.1. The relation
d2(w;y,eey∗) = (w1+w2+w3)
(m−mt)T(m−mt)
−
−2w2λ(m−mt)T(a−at) +w2λ2(a−at)T(a−at) + +2w3ρ(m−mt)T(b−bt) +w3ρ2(b−bt)T(b−bt) holds.
Proof. We have
d2(w;y,eye∗) =w1km−mtk2+w2k(m−λa)−(mt−λat)k2+ +w3k(m+ρb)−(mt+ρbt)k2 =w1(m−mt)T(m−mt) +
+ h
w2(m−mt)T(m−mt)−2w2λ(m−mt)T(a−at) + +w2λ2(a−at)T(a−at)
i + +
h
w3(m−mt)T(m−mt) + 2w3ρ(m−mt)T(b−bt) + +w3ρ2(b−bt)T(b−bt)
i
=
= (w1+w2+w3)h
(m−mt)T(m−mt)i
−2w2λ(m−mt)T(a−at) + +w2λ2(a−at)T(a−at) + 2w3ρ(m−mt)T(b−bt) +
+w3ρ2(b−bt)T(b−bt).
We will discuss how the model given in [2] is affected after the introduc- tion of the weighted distance, d2(w;y,eye∗). For w1 =w2 =w3 = 1 we obtain the results from [2].
A procedure of finding α, βA, βB, γA, γB is to minimize the weighted distance d2(w;·,·) between the experimental measurements of the response variable yei,i= 1, n and the theoretical valuesyei∗. In other words, we have to solve the problem
α,βA,βminB,γA,γB
d2(w;ey,ey∗) = min
α,βA,βB,γA,γB
D2LR(w;α, βA, βB, γA, γB).
Theorem 4.2. The problem min
α,βA,βB,γA,γB
D2LR(w;α, βA, βB, γA, γB)ad- mits local solutions which can be improved using an iterative estimation algo- rithm as follows:
α=
"
X3
i=1
wi
−w2λβA(2−λβA) +w3ρβB(2 +ρβB)
#−1
×
× FTF−1
FT[(w1+w2+w3)m−w2λ(mβA+a−1γA) + +w2λ2(aβA−1βAγA) +w3ρ(mβB+b−1γB) +w3ρ2(bβB−1βBγB)
, βA=λ−1 αTFTFα−1
λ αTFTa−αTFT1γA
− αTFTm−αTFTFα , βB=ρ−1 αTFTFα−1
ρ αTFTb−αTFT1γB
+ αTFTm−αTFTFα , γA= (nλ)−1
λ1T(a−FαβA)−1T(m−Fα) , γB = (nρ)−1
ρ1T(b−FαβB) +1T(m−Fα) . Proof. We have
DLR2 (w;α, βA, βB, γA, γB) = (w1+w2+w3)
(m−Fα)T(m−Fα)
−
−2w2λ(m−Fα)T(a−FαβA−1γA) + +w2λ2(a−FαβA−1γA)T(a−FαβA−1γA) +
+2w3ρ(m−Fα)T(b−FαβB−1γB) + +w3ρ2(b−FαβB−1γB)T(b−FαβB−1γB) =
= (w1+w2+w3) mTm−2mTFα+αTFTFα
−
−2w2λ mTa−mTFαβA−mT1γA−αTFTa+αTFTFαβA+αTFT1γA + +w2λ2 aTa−2aTFαβA−2aT1γA+αTFTFαβ2A+2αTFT1βAγA+nγA2
+ +2w3ρ mTb−mTFαβB−mT1γB−αTFTb+αTFTFαβB+αTFT1γB
+ +w3ρ2 bTb−2bTFαβB−2bT1γB+αTFTFαβB2 + 2αTFT1βBγB+nγB2
. Equaling to zero the partial derivatives of DLR2 (w;α, βA, βB, γA, γB):
and obtain the equations FTFαT
(w1+w2+w3)−2w2λβA+w2λ2βA2 + 2w3ρβB+w3ρ2βB2
=
= (w1+w2+w3)FTm−w2λ FTmβA+FTa−FT1γA
+ +w2λ2 FTaβA−FT1βAγA
+w3ρ FTmβB+FTb−FT1γB
+ +w3ρ2 FTbβB−FT1βBγB
, w2
αTFTm−αTFTFα−λ αTFTa−αTFTFαβA−αTFT1γA
= 0, w3
−αTFTm+αTFTFα−ρ αTFTb−αTFTFαβB−αTFT1γB
= 0, w2
αTFT1−mT1+λ aT1−αTFT1βA−nγA
= 0,
w3
αTFT1−mT1−ρ bT1−αTFT1βB−nγB
= 0.
An iterative solution is given by the relations
α= [(w1+w2+w3)−w2λβA(2−λβA) +w3ρβB(2 +ρβB)]−1×
× FTF−1
FT[(w1+w2+w3)m−w2λ(mβA+a−1γA) + +w2λ2(aβA−1βAγA) +w3ρ(mβB+b−1γB) +w3ρ2(bβB−1βBγB)
, βA=λ−1 αTFTFα−1
λ αTFTa−αTFT1γA
− αTFTm−αTFTFα , βB=ρ−1 αTFTFα−1
ρ αTFTb−αTFT1γB
+ αTFTm−αTFTFα , γA= (nλ)−1
λ1T(a−FαβA)−1T(m−Fα) , γB = (nρ)−1
ρ1T(b−FαβB) +1T(m−Fα) , and the proof is complete.
5. THE WEIGHTED LL SYMMETRICAL CASE
In this case L = R, a = b. Consequently, we have Ye = (m, a)LL. Moreover, ρ = λ, βA = βB and γA = γB (see [2]). We have the following results.
Theorem5.1. In the special LL symmetrical case the distance becomes d2(w;y,eey∗) = (w1+w2+w3)
(m−mt)T(m−mt) +
+2 (−w2+w3)λ(m−mt)T(a−at) + (w2+w3)λ2(a−at)T(a−at). Proof. We have
d2(w;y,eye∗) =w1km−mtk2+
+w2k(m−λa)−(mt−λat)k2+w3k(m+ρb)−(mt+ρbt)k2=
=w1km−mtk2+w2k(m−λa)−(mt−λat)k2+ +w3k(m+λa)−(mt+λat)k2 =
=w1km−mtk2+w2k(m−mt)−λ(a−at)k2+ +w3k(m−mt) +λ(a−at)k2 =w1(m−mt)T(m−mt) + +
h
w2(m−mt)T(m−mt)−2w2λ(m−mt)T(a−at)+w2λ2(a−at)T(a−at) i
+ +
h
w3(m−mt)T(m−mt)+2w3λ(m−mt)T(a−at)+w3λ2(a−at)T(a−at) i
=
= (w1+w2+w3)h
(m−mt)T(m−mt)i +
+2 (−w2+w3)λ(m−mt)T(a−at) + (w2+w3)λ2(a−at)T(a−at).
Thus we have the problem
α,βminA,γA
d2(w;y,eye∗) = min
α,βA,γA
D2LL(w;α, βA, γA). Theorem5.2. The problem min
α,βA,γA
DLL2 (w;α, βA, γA)admits local solu- tions which can be improved using an iterative estimation algorithm as follows:
α=
(w1+w2+w3) + 2 (−w2+w3)λβA+ (w2+w3)λ2β2A−1
×
× FTF−1
FT[(w1+w2+w3)m+ (−w2+w3)λ(mβA+a−1γA) + + (w2+w3)λ2(aβA−1βAγA)
, βA= αTFTFα−1
αTFT(a−1γA)+λ−1(−w2+w3)
w2+w3 αTFTm−αTFTFα
, γA= 1
n
1T(a−FαβA) + λ−1(−w2+w3) w2+w3
−αTFT1+mT1
. Proof. We have
DLL2 (w;α, βA, γA) = (w1+w2+w3) h
(m−Fα)T(m−Fα) i
+ +2 (−w2+w3)λ(m−Fα)T(a−FαβA−1γA) + + (w2+w3)λ2(a−FαβA−1γA)T(a−FαβA−1γA) =
= (w1+w2+w3) mTm−2mTFα+αTFTFα +
+2(−w2+w3)λ mTa−mTFαβA−mT1γA−αTFTa+αTFTFαβA+αTFT1γA + (w2+w3)λ2 aTa−2aTFαβA−2aT1γA+αTFTFαβA2+2αTFT1βAγA+nγA2
. Similarly to Theorem 4.2, we obtain the equations
FTFαT
(w1+w2+w3) + 2 (−w2+w3)λβA+ (w2+w3)λ2βA2
=
= (w1+w2+w3)FTm+ (−w2+w3)λ FTmβA+FTa−FT1γA
+ + (w2+w3)λ2 FTaβA−FT1βAγA
, (−w2+w3) −αTFTm+αTFTFα
+ + (w2+w3)λ −αTFTa+αTFTFαβA+αTFT1γA
= 0, (−w2+w3) αTFT1−mT1
+ (w2+w3)λ −aT1+αTFT1βA+nγA
= 0.
Thus,
α= [(w1+w2+w3) + +2 (−w2+w3)λβA+ (w2+w3)λ2βA2−1
FTF−1
FT[(w1+w2+w3)m+
+ (−w2+w3)λ(mβA+a−1γA) + (w2+w3)λ2(aβA−1βAγA) ,
βA= αTFTFα−1
αTFT(a−1γA) + λ−1(−w2+w3) w2+w3
αTFTm−αTFTFα
, γA= 1
n
1T(a−FαβA) +λ−1(−w2+w3)
w2+w3 −αTFT1+mT1
, and the proof is complete.
Forw1=w2 =w3 = 1 we obtain the solutions from [2].
REFERENCES
[1] I.N. Chalkidis, Ch.D. Tzimopoulos, Ch.H. Evangelides, M. Sakellarioy and St.
Yannopoulos, Fuzzy logic model for soil water balance problem. In: Proc. 2nd IASME/WSEAS Internat. Conf. on Water Resources, Hydraulics and Hydrology, (Portoroz, Slovenia), May 15–17, 2007.
[2] R. Coppi, P. D’Urso, P. Giordaniand and A. Santoro, Least squares estimation of a linear regression model with LR fuzzy response. Comput. Statist. Data Anal.51(2006), 1, 267–286.
[3] R. Coppi and P. D’Urso,Regression analysis with fuzzy informational paradigm: a least- squares approach using membership function information. Int. J. Pure Appl. Math.8 (2003),3, 279–306.
[4] R. Coppi and P. D’Urso,Three-way fuzzy clustering models for LR fuzzy time trajecto- ries. Comput. Statist. Data Anal.43(2003),2, 149–177.
[5] R. Coppi, M.A. Gil and H.A.L. Kiers,The fuzzy approach to statistical analysis. Com- putat. Statist. Data Anal.51(2006),1, 1–14.
[6] C. Dou, W. Woldt, I. Bogardi and M. Dahab,Steady state groundwater flow simulation with imprecise parameters. Water Res.31(1995),11, 2709–2719.
[7] P. D’Urso and P. Giordani, Fitting of fuzzy linear regression models with multivariate response. Int. Math. J.3(2003),6, 655–664.
[8] A. Kaufmann and M. Gupta, Introduction to Fuzzy Arithmetic, Theory and Applica- tions. ITP, USA, 1991.
[9] B. Liu,Uncertain Programming. Wiley, New York, 1999.
[10] V. Preda,Teoria deciziilor statistice. Ed. Academiei Romˆane, Bucure¸sti, 1992.
[11] L.A. Zadeh,Fuzzy Sets. Information and Control8(1965), 338–353.
[12] L.A. Zadeh,Generalized theory of uncertainty(GTU)-principal concepts and ideas. Com- put. Statist. Data Anal.51(2006), 15–46.
[13] H. Zimmermann,Fuzzy Set Theory and its Applications. Kluwer, Dordrecht, 1996.
Received 15 July 2007 Academy of Economic Studies
Department of Mathematics Calea Dorobant¸ilor 15-17, Sector 1
010374 Bucharest, Romania cippx@yahoo.com