A Note on Linear Regression Model Using
Non-Symmetric Triangular Fuzzy Number Coefficients
Dug Hun Hong1) ․ Kyung Tae Kim2)
Abstract
Yen et al. [Fuzzy Sets and Systems 106 (1999) 167-177] calculated the fuzzy membership function for the output to find the non-symmetric triangular fuzzy number coefficients of a linear regression model for all given input-output data sets. In this note, we show that the result they obtained in their paper is invalid.
Keywords : Fuzzy regression analysis, Fuzzy triangular coefficients, Minimization of fuzziness, Non-symmetric coefficients
The following model shows the dependence of the output variable on the inputs variables,
˜ = f( x, AY ˜ ) = A˜
0+ A˜
1x1+ …+ A˜
nxn, (1) where ˜ is the fuzzy output, x = [xY 1, x2, …, xn]T is the real-valued input vector, and A˜ = { A˜
0, A˜
1, …, A˜
n} is a set of fuzzy numbers.
The membership function for the set ˜ is defined by Zadeh's extension Y principle as follows:
μ ˜Y( y )=
{
0max ( a1,… , an) = f- 1( y, x){ minjμ ˜Aj(aj) } otherwise.if f - 1(y,x) ≠φ,Then the regression analysis problem is defined as: given a set of crisp data
1) First Author : Professor, Department of Mathematics, Myongji University, Kyunggi 449-728, Korea
E-mail : [email protected]
2) Professor, Department of Electronics and Electrical Engineering, Kyungwon University, Kyunggi 461-701, Korea
E-mail : [email protected]
points < x1, y1>, < x2, y2>, … , < xm, ym>, we want to find a set of fuzzy parameters ˜A
0, A˜
1, … , A˜
n for the Eq.(1) which is the best fit to the given data points, according to some criteria of goodness of fit.
In (1), ˜A
i is the fuzzy coefficient of the variable xi in the regression model of additive form. If ˜A
is have triangular membership functions, then each fuzzy number coefficient ˜Ai can be uniquely defined by
˜A
i= {aLi, aCi, aUi },
where aLi is the lower limit, aUi is the upper limit, and aCi is the point having the property that μ ˜A
i(aCi) = 1. The property of symmetry of the fuzzy coefficient ˜A
i enables us to establish the following two relations:
aCi = ( aLi + aUi )/2 and
aSi= aCi - aLi = aUi- aCi, where aCi is the center and aSi the spread of ˜A
i.
For symmetric triangular fuzzy number coefficients, the membership function μ ˜A
i for ˜A
i , i = 1,…,n can be described as
μ ˜A
i(ai) =
{
1 - ( a1 - ( a0 Cii- a- aCii)/a)/aSiSi, a, a otherwise.CCii≤a- aiSi≤a≤aCii+ a≤aCiSi,,Having established the membership function for each fuzzy coefficient ˜A
i, the fuzzy output from the linear model f( x, A˜ ) in (1) can be expressed according to the principle of extension and fuzzy arithmetic on fuzzy numbers [1] as
˜ = f( x, AY ˜ ) = (fC( x), fS( x)),
where fC(x) is the center of the fuzzy linear model f( x, A˜ ) and has the form fC( x) = aC0+ aC1x1+ … + aCnxn
and fS( x) is the spread of f( x, A˜ ) and defined as fS( x) = aS0+ aS1|x1| + … + aSn|xn|.
where [f( xj)]h= [ A˜0]h+ [ A˜1]hxj1+ … + [ A˜n]hx jn such that [∙]h represents the h-level set of a fuzzy number.
In regression , the goal is to find the fuzzy coefficients that minimize the above-mentioned spread of fuzzy output for all the data sets. The cost function, Z , to be minimized can be written as
Z = aS0+ ∑n
i = 1
[
aSi j = 1∑m |x ji|]
which can also be expressed as
Minimize Z =fS( x1) + fS( x2 )+ … + fS( xm) subject to the set of constraints
yj∈[ f( xj)]h.
However, if these triangles are not symmetric , we need minimum three parameters to uniquely describe each. For example, ˜A
i can be described by the triplet {aLi, aPi, aUi } or by {sLi, aPi, sRi}, where aPi is the point at which
μ ˜A
i(aPi) = 1, sLi is the left-side spread from the peak point aPi, and sRi represents the right-side spread. The membership function for each ˜A
i has the form
μ ˜A
i(ai) =
ꀊ
ꀖ ꀈ
︳︳
︳︳
︳︳
︳︳
1 - ai- aPi
sRi , aPi≤ai≤aPi+ sRi, 1 - aPi- ai
sLi , aPi- sLi≤ai≤aPi,
0 otherwise.
Following the principle of extension, Yen et al.[2] used the fuzzy membership function for the output by
μ ˜Y( y )=
ꀊ
ꀖ ꀈ
︳︳
︳︳
︳︳
︳
︳︳
︳︳
︳︳ 1 -
y -∑
i aPixi- aP0 sR0+∑
i sRi|xi| , aP0+∑aPixi≤y≤aP0+∑
i aPixi+
(
sR0+∑i sRi|xi|)
,1 -
aP0+∑
i aPixi- y sL0+∑
i sLi|xi| , aP0+∑
i aPixi-
(
sL0+∑i sLi|xi|)
≤y≤aP0+∑i aPixi,0 otherwise.
(2) i.e, ˜ = {SY L0+∑
i SLi|xi|, aP0+∑
i aPixi, SR0+∑
i SRi|xi| }.
But this result is wrong. Let A˜
i= {SLi, aPi, SRi} , i = 1, 2 and λ∈R be a real number. By the extension principle, the following rules for addition and real multiplication can be represented as
λ A˜ =
{
{λS{|λ|SL, λaR, λaP, λSP, |λ|SR} L} if λ≥0,if λ < 0,˜A
1+ A˜
2= {SL1+ SL2, aP1+ aP2, SR1+ SR2}.
This result can be generalized to linear combinations of fuzzy numbers as follows:
˜ = AY ˜
0+ A˜
1x1+ … + A˜
nxn
= {SL0+ ∑
xi≥0|xi|SLi+ ∑
xi< 0|xi|SRi, aP0+∑
i aPixi, SR0+ ∑
xi≥0|xi|SRi+ ∑
xi< 0|xi|SLi}, i.e.
μ ˜Y( y )=
ꀊ
ꀖ ꀈ
︳︳
︳︳
︳︳
︳︳
︳︳
︳︳
︳︳
︳
︳︳
︳︳
︳︳
︳︳
︳︳
︳︳
︳︳
︳ 1 -
y -∑
i aPixi- aP0 SR0+ ∑
xi≥0|xi|SRi+ ∑
xi< 0|xi|SLi , if aP0+∑aPixi≤y≤aP0+∑
i aPixi+
(
SR0+ x∑i≥0|xi|SRi+ ∑xi< 0|xi|SLi)
,1 -
aP +0 ∑
i aPixi- y SL0+ ∑
xi≥0|xi|SLi + ∑
xi< 0|xi|SRi , if aP0+∑
i aPixi-
(
SL0+ x∑i≥0|xi|SLi + x∑i< 0|xi|SRi)
≤y≤aP0+∑i aPixi,0, otherwise.
(3) So μ˜A( y ) in (2) is not true but μ ˜A( y ) in (3) is true.
We consider the following example.
Example. Let ˜A
0= {SL0= 1, aP0= 2, SR0= 3 },
˜A
1= {SL1= 1, aP1= 4, SR1= 6 }, and x1=- 2. Then by (3),
˜ = AY ˜
0+ ( - 2) A˜
1= { left - side spread = 1 + ( | - 2|)6 = 13, peak point
= 2 + ( - 2)4 =- 6, right-side spread = 3 + ( | - 2|)1 = 5 } = {13, - 6, 5 }. But by (2), we have ˜ = {3, - 6, 15 } which is not true.Y
From the expression of (3), we get the constraints of the regression as
1 -
y -∑
i aPixi- aP0 SL0+ ∑
xi≥0|xi|sLi+ ∑
xi< 0|xi|sRi ≥h and
1 -
aP0+∑
i aPixi- y SR0+ ∑
xi≥0|xi|sRi+ ∑
xi< 0|xi|sLi ≥h if we have taken membership function value at h-cut.
References
1. Tanaka, H. Vejima, S. and Asia ,K. (1982). Linear regression analysis with fuzzy model, IEEE Trans. Systems Man Cybernet. 12(6), 903-907.
2. Yen, K. K. Ghoshray, S. and Roig, G. (1999). A Linear regression model using triangular fuzzy number coefficients, Fuzzy Sets and Systems 106, 167-177.
[ received date : Dec. 2004, accepted date Apr. 2005 ]