Professional Documents
Culture Documents
= +
%
where a is the center, 0 > is the left spread and 0 > is the right spread of a% .
If = , then the triangular fuzzy number is called a symmetric triangular fuzzy
number and it is denoted by ( , ) a . A fuzzy number ( , , )
LR
a a = % of type L-R is a
function from real number into the interval ( 0 , 1) satisfying
,
( ) ,
0 ,
t a
R a t a
a t
a t L a t a
otherwise
| |
+
|
\
| |
=
|
\
%
where L and R are non increasing and continuous functions from (0,1) to (0,1) satisfying
L(0)=R(0)=1 and L(1)=R(1)=0.A binary operation T on the unit interval is said to be a
triangular norm [9] (t-norm) if and only if f T is associative, commutative, non-decreasing and
T(x,1)=x for each [0,1] x . Moreover, every t-norm satisfies the inequality,
( , ) ( , ) ( , ) min( , )
w M
T a b T x y T a b a b = where
, 1
( , ) , 1
0 ,
w
a if b
T a b b if a
otherwise
=
= =
The critical importance of min( , ), , max(0, 1) ( , )
w
a b a b a b and T a b + is emphasized from
a mathematical point of view in Ling [9]. The usual arithmetic operations on real numbers can be
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
75
extended to the arithmetic operations on fuzzy numbers by means of extension principle by Zadeh
[10], which is based on a triangular norm T. Let A
%
and B
%
be fuzzy numbers of the real line .
The fuzzy number arithmetic operations are summarized as follows:
Fuzzy number addition :
( )
x, y,x y z
A B (z) sup T A(x), B(y) .
+ =
( =
% % % %
Fuzzy number multiplication :
( )
x,y,x.y z
A B (z) sup T A(x), B(y) .
=
( =
% % % %
The addition (subtraction) rule for L-R fuzzy number is well known in case of T
M
based addition,
in which the resulting sum is again an L-R fuzzy number, i.e., the shape is preserved. Let
( , , ) , ( , , )
A A LR B B LR
A a B b = =
% %
. Then using T
M
in the above definition,
( , , )
M A B A B LR
A B a b = + + +
% %
It is also known that the
w
T based addition and multiplication preserves the shape of L-R fuzzy
numbers[11,12,13,14]. We know that T
M
based multiplication does not preserve the shape of L-
R fuzzy numbers. In this section, we consider
w
T based multiplication of L-R fuzzy numbers.
Let T=
w
T be the weakest t-norm and let ( , , ) , ( , , )
LR LR
A A B B
A a B b = =
% %
be two L-R
fuzzy numbers, then the addition and multiplication of ( , , ) , ( , , )
LR
A A LR B B
A a B b = =
% %
is defined as[15],
( , max( , ), max( , ))
W A B A B LR
A B a b = +
% %
( )
( )
( , max( , ), max( , )) , 0
( , max( , ), max( , )) , , 0
( , max( , ), max( , )) , 0, 0
0, , , 0, 0
0, , , 0, 0
(0, 0, 0) , 0, 0
A B A B LR
A B A B RL
A B A B LL
W
A A
LR
A A
RL
LR
ab b a b a for a b
ab b a b a for a b
ab b a b a for a b
A B
b b for a b
b b for a b
for a b
>
<
< >
=
= >
= <
= =
% %
In particular, if ( , ), ( , )
A B
A a B b = =
% %
are symmetric fuzzy numbers, then the multiplication of
A and B
% %
is written as, ( , max( , ))
w A B LL
A B ab b a =
% %
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
76
A distance between fuzzy numbers: Several metrics are defined on the family of all fuzzy
numbers[16]. The generalized Hausdorff metric fulfil many good properties, also easy to
compute in terms of generalized mid and spread functions. The concept of generalized mid and
spread may be useful in working with (fuzzy) convex compact sets [17].
) , (
B A d
H
is the Hausdorff metric between crisp sets
B and A given by,
)
`
b a b a B A d
B b
A a
A a
B b
H
inf sup , inf sup max ) , (
If [ ] [ ]
2 1 2 2 1 1
, , b b I and a a I = = are two intervals, then
( ) { }
2 1 2 1 2 2 1 1 2 1
, max , I spr I spr I mid I mid b a b a I I d
H
+ = =
Where
2
,
2
1 2
1
2 1
1
a a
sprI
a a
I mid
=
+
= [16].
The generalized Hausdorff metric between
T T
b B a A ) , (
~
, ) , (
~
= = is then,
+ = + =
b a B A D b a B A D )
~
,
~
( , 5 . 0 )
~
,
~
(
1
3. FUZZY LINEAR REGRESSION USING THE PROPOSED
APPROACH
In this section, we are discussing fuzzy linear regression about the proposed approach based on
Hausdorff metric using T
w
norm based operations with crisp input- fuzzy output data, in which
the coefficients of the models are also considered as fuzzy numbers.
Consider the set of observed data
{ }
( , ), 1,...,
i i
X Y i n =
% %
where ( , )
i
i i
X x =
%
and ( , )
i i i
Y y e =
%
are symmetric fuzzy numbers. Our aim is to fit a fuzzy regression model with fuzzy
coefficients to the aforementioned data set as follows:
0 1 1 i w w i w w p w ip w i
Y A ( A X ) ..... ( A X ) A X = =
% % % % % % % %
,
1,.... = i n
(1)
where
( )
, , 1,...
j j j
A a j p = =
%
are symmetric fuzzy numbers and the arithmetic
operations are based on the weakest T
w
norm.
Consider the least absolute optimization problem as follows:
Minimizes
1
( , )
w j
D Y A X
% % %
i.e.,
Min
( )
1 0 1 0
0.5 max ,
n k n k
i ij j i j ij ij j
i j i j
y x a e a x
= = = =
+
(2)
subject to
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
77
( )
( )
( )
1 1
1
0 1
1 1
1
0 1
1
( ) max , ( )
( ) max , ( )
0 1,
max , 0, 1, 2,...
n n
j ij j ij ij j i i
j p
j
n n
j ij j ij ij j i i
j p
j
j ij ij j
j p
a x L h a x y L h e
a x L h a x y L h e
h
a x i m
=
=
+
+
< <
=
(4)
Solving this optimization problem using LINGO, we can estimate the fuzzy coefficients
of the model. Several methods have been proposed to detect the presence of outliers,
relying on graphical representation and/or on analytical procedures. The box plot or box-
and-whisker plot describes the key features of data through the five-number summaries:
the smallest observation, lower quartile (Q1), median (Q2), upper quartile (Q3), and the
largest observation (sample maximum). A box plot indicates the abnormal observations.
Usually, outlier cut offs are set at 1.5 times the inter-quartile range [18, 19]. In a fuzzy
framework, we can draw box plots side by side to detect outliers in the distributions of
the centers, of the spreads and of the input variables. To overcome limitations in previous
approaches, Hung and Yang [20] consider the effect of each observation on the value of
the objective function in Tanakas approach. Let J
M
be the optimal value of the objective
function and J
M
(i)
is the corresponding value obtained deleting the i
th
observation. The
ratio
( ) i
M M
i
M
J J
r
J
=
x1 x2 x3 y eps
1 2 0.5 15.25 5.83 3.56 0.095626
2 0 5 14.13 0.85 0.52 0.163109
3 1.13 1.5 14.13 13.93* 8.5* 0.445439
4 2 1.25 13.63 4 2.44 -0.30979
5 2.19 3.75 14.75 1.65 1.01 -0.22509
6 0.25 3.5 13.75 1.58 0.96 0.077927
7 0.75 5.25 15.25 8.18 4.99 0.198405
8 4.25 2 13.5 1.85 1.13 -0.25154
* indicates outlier
Table 1. Dataset with crisp input and fuzzy output with outlier
Figure 1. Fuzzy regression model with outlier for the dataset in Table 1 using the proposed approach.
The fuzzy regression model obtained by the proposed approach,
1 2 3
(0, 3.8212) (0, 0) (0, 0) (0.3013, 0)
w w w w w w
Y X X X =
%
with
optimum value h=0.215 and the graph is given in Fig.1. In table 1, third data is an
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
79
abnormal data, after identifying the abnormal data using Hung and Yang [20] method, deleting
the data which yields a better result. The fuzzy regression model obtained by proposed approach
is
1 2 3
(2.482, 0) (0, 0) (0, 0) (0.2042, 0)
w w w w w w
Y X X X =
%
with optimum level h= 0.322 (shown in Figure 2.)
Figure 2. Using Hung and Yang [20] method of omission approach for the dataset with outlier in table1
using the proposed approach
Comparing the performance of the proposed with the some other existing methods, using the
Choi and Buckleys [5] method, the optimal model is
1 2 3
2.8273 0.3878 1.0125 0.6185 (0,1.0696, 2.0042) Y x x x =
%
Chen Hseuh [23] proposed a least square approach to fuzzy regression models with crisp
coefficients.
1 2 3
16.7956 1.0989 1.1798 1.8559 (0, 2.8888) Y x x x =
%
Hassanpour et al. [24] proposed least absolute regression method that minimizes the difference
between centers of the observed and estimated fuzzy responses and also between the spreads of
them, using goal programming approach. They took into account fuzzy coefficients for crisp
inputs in their model. Employing their model for the given example yields the following model,
1 2 3
( 2.8273, 0.0000) (0.3877, 0.0000) (1.0125, 0.000) (0.6185, 0.1790) Y x x x =
%
Using the graded mean integration representation, the similarity measure for the proposed and
above existing models is given in table 2.
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
80
Methods Proposed
Choi and
Buckley[5]
Chen and Hseuh
[23]
Hassanpour [24]
Mean Similarity
measure between
the observed and
the existing
methods
0.326389 0.307915 0.122202 0.209148
Table 2. Similarity measure between various models with multiple inputs
The table 2 illustrates that the mean similarity measure for the proposed model is 0.3264 which
has effective performance with other existing methods using Hung and Yang method of omission
approach with outlier.
EXAMPLE-2
We now apply this model to analyse the effect of the composition of Portland cement on heat
evolved during hardening. The data shown in Table 3 except
i
s are taken from [25]. The
values are assumed by R.Xu et al. [26].
Obs.No.
( , )
i i i
y y s = %
1 i
x
2 i
x
3 i
x
( , )
i i i
Y Y =
%
Goodness
of fit
1 (78.5,6.9) 7 26 6 (78.33,0.709) 0.9995
2 (74.3,6.4) 1 29 15 (71.99,0.709) 0.8997
3 (104.3,9.4) 11 56 8 (106.25,0.709) 0.9364
4 (87.6,7.8) 11 31 8 (88.96,0.709) 0.9748
5 (95.5,8.6) 7 52 6 (96.30,0.709) 0.9926
6 (109.2,9.9) 11 55 9 (105.75,0.709) 0.8997
7 (102.7,9.3) 3 71 17 (104.81,0.709) 0.9565
8 (72.5,6.2) 1 31 22 (74.75,0.709) 0.8994
9 (93.1,8.3) 2 54 18 (91.56,0.709) 0.9712
10 (115.9,10.6) 21 47 4 (116.20,0.709) 0.9993
11 (83.8,7.4) 1 40 23 (81.16,0.709) 0.8994
12 (113.3,10.6) 11 66 9 (113.35,0.709) 0.9999
13 (109.4,9.9) 10 68 8 (112.85,0.709) 0.8996
Table 3. Performance of the proposed model in Example 2.
By using the proposed method, we have
1 2 3
(47.299, 0.7093) (1.6963, 0) (0.6914, 0) (0.196, 0)
w w w w w w
Y x x x =
%
with h=0.675 (shown in Fig.3)
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
81
Figure 3. Fuzzy regression model with crisp multiple input fuzzy output in Example 2.
Furthermore, we calculate the goodness of fit using the fitted equation and the observed values.
The results are given in table3, the goodness of fit of every
i
y% and
i
Y
%
are all greater than 0.9,
which indicates that the fitted result of the model is very good.
EXAMPLE-3
Consider the following crisp input-fuzzy output data given by Tanaka et al. [27]
( ; ) ((8.0,1.8) ;1), ((6.4, 2.2) ; 2), ((9.5, 2.6) ;3), ((13.5, 2.6) ; 4), ((13.0, 2.4) ;5),
T T T T T
y x = %
By applying the proposed approach described in section 3, the fuzzy regression model is derived
as (4.079, 2.1) (1.8458, 0)
w w
Y x =
%
with h = 0.468. A Summary of results of some
other techniques, including their models as well as their performances, are given in Table 4.
To show the performance of fuzzy regression models we considered these five pairs of
observations listed above. The dataset do not have the level of detail and complexity than
those used in other studies, but in the literature these data have been considered by many
researchers for the experimental evaluation and comparison of their proposed methodology. Table
4 lists the regression models obtained by the methods proposed by other authors based on the five
pairs of observations considered above. The first fuzzy regression model based on fuzzy
observations was proposed by Tanaka et al.[27] (THW) (1982). Several authors pointed out that
this method has several disadvantages and modified it or developed their own methodologies to
prevent the problems. THW[26] regression model estimates the fuzzy regression coefficients by
linear programming. This model has a numeric slope and a fuzzy intercept. KB[28], DM[29],
WT[30] and HBS[31] models have a fuzzy slope and intercept. If the explanatory variables are
numeric values and dependent variables are symmetric fuzzy numbers, WT[30] model is the same
as DM[29] approach. NN[32] and CH[33]model have a numeric slope. In this case the proposed
regression model is given in Figure 4.
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
82
Figure 4. Fuzzy regression model for the data set in Example 3.
Methods
Goodness of fit
measure
Similarity
measure
Proposed method using Tw-norm based operations
(4.079, 2.1) (1.8458, 0)
w w
Y x =
%
with
h=0.468
0.9068 0.61116
Tanaka et al. method (THW)[27]
(0.385, 7.7) 2.1 Y x =
%
0.9424 0.43625
Kim and bishu model 1988 (KB)[28]
(3.11, 4.95, 6.84) (1.55,1.71,1.82) Y x =
%
0.9125 0.4987
Nasrabedi and Nasrabedi (2006) method (NN)[32]
(2.36, 4.86, 7) 1.73 Y x =
%
0.9149 0.51554
Diamond (1988)(DM)[29]
(3.11, 4.95, 6.79) (1.55,1.71,1.87) Y x =
%
0.9135 0.4865
Wu and Tseng method (WT)[30]
(3.11, 4.95, 6.79) (1.55,1.71,1.87) Y x =
%
0.9135 0.4865
Hojati et al 2005(HBS)[31]
(5.1, 6.75, 8.4) (1.1,1.25,1.4) Y x =
%
0.8915 0.62467
Chen and Hseuh method 2009(CH)[33]
1.71 (2.63, 4.95, 7.27) Y x =
%
0.9175 0.4865
Table 4.Performance of different models for crisp input and fuzzy output given
in Example 3.
Analyzing these results, we can observe that our technique is very easy to compute in practice
and gives a first and easy approach for the problem of analyzing regression problems with crisp
input and triangular fuzzy output data.
5. CONCLUSION
Based on the distance between the centers and spreads, a new method is proposed for fuzzy
simple regression using T
w
-norms. The models which are used here have the input and the
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
83
output data as well as the coefficients are assumed to be fuzzy. The arithmetic operations based
on the weakest t-norm are employed to derive the exact results for estimation of parameters. The
efficiency of the proposed approach is studied by similarity measure based on the graded mean
integration representation distance of fuzzy numbers. In addition the effect of the outlier is
discussed for the proposed approach. By comparing the proposed approach with some well
known methods, applied to three data sets, it is shown that the proposed approach using shape
preserving operations was more effective. Studying the effect of outlier in center value of
response variable using proposed method with T
W
- norm operation is our future work.
REFERENCES
[1] Chang P.T. & Lee E.S, (1994) fuzzy least absolute deviations regression based on the ranking of
fuzzy numbers, in Proc. IEEE world congress on computational intelligence, pp.1365-1369.
[2] Kim K.J, Kim D.H & Choi S.H, (2005) Least absolute deviation estimator in fuzzy regression, J.
Appl. Math. Comput. Vol.18, pp. 649-656.
[3] Torabi H, & Behboodian J, (2007) Fuzzy least absolutes estimates in linear regression models,
Communi. Stat. Theory methods, Vol. 36, pp. 1935-1944.
[4] Chen L. H. & Hsueh C.C, (2007) A mathematical programming method for formuation a fuzzy
regression model based on distance criterion, IEEE Trans. Syst. Man Cybernet. Vol.B, pp. 37 705-
712.
[5] Choi S. H. & Buckley J. J, (2008) Fuzzy regression using least absolute deviation estimators, Soft
Comput. Vol. 12, pp. 257-263.
[6] Taheri S. M. & Kelkinnama M, (2008) Least absolute regression, in Proc. 4
th
international IEEE
conference on intelligent systems, varna Bulgaria, , Vol. 11, pp. 55-58.
[7] Taheri S. M. & Kelkinnama M, (2012) Fuzzy linear regression based on least absolute deviations,
Iranian J. fuzzy system Vol. 9, pp. 121-140.
[8] Hong D.H., Lee S. & Do D.Y, (2001) Fuzzy linear regression analysis for fuzzy input-output data
using shape preserving operations, Fuzzy sets Syst. Vol.122, pp. 513-526.
[9] Ling C.H, (1965) Representation of associative functions, Publications Mathematicae Debrecen,
Vol. 12, pp. 189-212.
[10] Zadeh L.A, (1978) Fuzzy sets as a basis for a theory of possibility, Fuzzy sets syst. Vol. 1, pp. 3-
28.
[11] Hong D.H, (2001) Shape preserving multiplication of fuzzy intervals, Fuzzy sets and Syst. Vol.
123, pp. 81-84.
[12] Hong D.H,(2002) On shape preserving addition of fuzzy intervals, Journal of Mathematical
Analysis and applications, Vol. 267, pp. 369-376.
[13] Kolsevera A, (1994) Additive preserving the linearity of fuzzy intervals, Tata mountains Math.
Publi. Vol. 6 pp.75-81.
[14] Mesiar R, (1997) Shape preserving addition of fuzzy intervals, Fuzzy sets and Syst. Vol. 86 pp. 73-
78.
International Journal of Fuzzy Logic Systems (IJFLS) Vol.3, No2, April 2013
84
[15] Hong D. H. & Do H.Y, (1997) Fuzzy system reliability analysis by the use T
W
(the weakest norm)
on fuzzy arithmetic operations, Fuzzy Sets and Systems, Vol. 90, pp. 307-316.
[16] Trutsching W, Gonzalez- Rodriguez G, Colubi A. & Angeles Gil A, (2009) A new family of metrics
for compact, convex(fuzzy) sets based on a generalized concept of mid and spread, Inf. Sci. Vol.
179, pp. 3964-3972.
[17] Diamond P. & Kloeden P, (1994) Metric spaces of fuzzy sets, World scientific, Singapore.
[18] Emerson J.D. & Strenio J, (1983) Boxplots and batch comparisons, in: D.C. Hoaglin, F. Mosteller,
J.W. Tukey (Eds.), Understanding Robust and Exploratory Data Analysis, Wiley, New York, pp.
5896.
[19] Tukey J.W, (1977) Exploratory Data Analysis, Addison-Wesley.
[20] Hung W.L. & Yang M.S, (2006) An omission approach for detecting outliers in fuzzy regressions
models, Fuzzy Sets and Systems Vol. 157, pp. 31093122.
[21] Hsieh C.H. & Chen. S.H, (1999) Similarity of generalized fuzzy numbers with graded mean
integration representation, in Proceedings of Eighth international fuzzy system association world
congress, vol2, Taipei, Taiwan, Republic of China, pp. 551-555.
[22] Xu R. & Li C, (2001) Multidimensional least- squares fitting with a fuzzy model, Fuzzy sets and
systems, Vol. 119, pp. 215-223.
[23] Chen L. H. & Hseuh C.C, (2009) Fuzzy regression models using the least squares method based on
the concept of distance, IEEE Transactions on fuzzy systems, Vol. 17, pp. 1259-1272.
[24] Hassanpour H, Maleki H. R. & Yaghoobi M.A, (2009) A goal programmingto fuzzy linear
regression with non-fuzzy input and fuzzy output data, Asia Pacific Journal of Operational
research, Vol. 26, pp. 587-604.
[25] Draper N. R. & Smith H, (1980) Applied Regression analysis, Wiley, Newyork.
[26] Xu R. & Li C, (2001) Multidimensional least squares fitting with a fuzzy model, Fuzzy sets and
systems, Vol. 119, pp. 215-223.
[27] Tanaka H, Hayashi I. & Watada J, (1980) Possibilistic linear regression analysis for fuzzy data,
Eur. J. Opera. Res. Vol. 40, pp. 389-396.
[28] Kim B. & Bishu R. R, (1998) Evalaution of fuzzy linear regresssion models by comparing
membership functions, Fuzzy Sets and Systems, Vol. 100, pp. 343-352.
[29] Diamond P, (1988) Fuzzy least squares, Information Sciences, Vol. 46, pp. 141157.
[30] Wu B. & Tseng N. F, (2002) A new approach to fuzzy regression models with application to
business cycle analysis, Fuzzy Sets Syst, Vol. 130, pp. 3342.
[31] Hojati M, Bector C. R. & Smimou K, (2005) A simple method for computation of fuzzy linear
regression, Eur J Oper Res. Vol. 166, pp. 172184.
[32] Nasrabadi M. M, & Nasrabadi E, (2004) A mathematical-programming approach to fuzzy linear
regression analysis, Appl Math Comput. 155:873881.
[33] Chen L. H, Hsueh C. C, (2009) Fuzzy regression models using the least squares method based on
the concept of distance, IEEE Trans Fuzzy Syst. Vol. 17, pp. 12591272.