See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/313456816

An interval type-2 fuzzy regression model with crisp inputs and type-2 fuzzy
outputs for TAIEX forecasting

Conference Paper · August 2016

DOI: 10.1109/ICInfA.2016.7831906

CITATIONS READS

6 629

3 authors, including:

Ali Vahidian Kamyad A. Zare

Ferdowsi University of Mashhad Islamic Azad University, Gonabad branch
149 PUBLICATIONS 1,879 CITATIONS 91 PUBLICATIONS 2,421 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by A. Zare on 24 July 2018.

The user has requested enhancement of the downloaded file.

 Proceedings of the IEEE
International Conference on Information and Automation
Ningbo, China, August 2016

An Interval Type-2 Fuzzy Regression Model with

Crisp Inputs and Type-2 Fuzzy Outputs for TAIEX
Forecasting
Narges Shafaei Bajestani Ali Vahidian Kamyad
Department of electrical engineering Department of electrical engineering
Science and Research Branch, Islamic Azad University Science and Research Branch, Islamic Azad University
Tehran, Iran Tehran, Iran
[email protected] [email protected]

Assef Zare
Department of electrical engineering
Gonabad Branch, Islamic Azad University,
Gonabad, Iran
[email protected]

Abstract— The type-2 fuzzy logic system permits us to model while the type-2 fuzzy logic systems (FLS) can help deal with
uncertainties existing in membership functions. Accordingly, this even higher levels of uncertainty in real world applications.
study aims to propose a linear framework for an interval type-2 Type-2 fuzzy rules are more complicated than type-1 fuzzy
fuzzy regression model based on the existing possibilistic models. rules, because they use of type-2 fuzzy sets in antecedent or
In this model, vagueness is minimized, under the circumstances
where the h-cut of observed value is included in predicted value.
consequent parts. Therefore, most of the type-2 FLS studies
In this model both primary and secondary membership function only concern with interval type-2 FLSs (IT2FLS).
of predicted value fit the observed value. This model, without the Some researchers have used type-2 fuzzy regressions as the
additional complexities, demonstrates its ability compared to modeling structure. Wei and Watada have been build a type-2
previous type-2 fuzzy models. The Taiwan stock index Fuzzy Qualitative Regression Model [13], which substantially
forecasting is used to evaluate model efficiency. is a simpler version of our model when H=0. In their article,
Wei and Watada imply that they have used a general type-2
Keywords— Interval type-2 fuzzy regression; Interval type-2 fuzzy number while it seems that they have used interval type-
fuzzy number; Quadratic programming; Forecasting.
2. O. Poleshchuk and E. Komarov present a regression model
I. INTRODUCTION for interval type-2 fuzzy sets based on the least squares
estimation technique, but there isn’t any example to guarantee
Statistical linear regression has long been used in almost the good performance of their proposed approach [14].
every field of science. Developed by Tanaka et al. [1]–[3], the Hosseinzadeh et al. present a weighted goal programming
fuzzy regression models have been applied to modeling approach to fuzzy linear regression with crisp inputs and type-
systems involving vague phenomena. Fuzzy regression 2 fuzzy outputs [15]. But this model only has tried to close the
techniques can generally be classified into two distinct membership functions of observed and estimated responses by
categories. The first category, suggested by Diamond [4], is an closing some of their parameters. It seems that none of the
adaptation of the classical least squares method. The second studies mentioned above could well model the type-2 fuzzy
category, known as possibilistic regression [1]–[3], reduces regression; instead they have reduced their model to only
the problem of obtaining fuzzy coefficients of a regression some points of type-2 fuzzy numbers. This study aims to
model to one of mathematical programming. propose a framework for an interval type-2 possibilistic fuzzy
The fuzzy regression analysis can be applied to many real- regression model. Since interval type-2 fuzzy membership
life issues [5]–[8]. In which the strict supposition of classical functions could be represented only by its Footprint of
regression models cannot be satisfied. For example, people Uncertainty (FOU). This model has been built based on the
infrequently use numerical expressions to judge things in their possibility model for upper membership function (UMF) and
life. The identification, the judgment and evaluation that necessity model for lower membership function (LMF).
peoples actually perform are usually expressed in qualitative In this paper, The Taiwan stock index (TAIEX) is used as
linguistic terms that can be modeled with fuzzy sets. Many the forecasting target. TAIEX is used for forecasting and
researchers have worked on the field of fuzzy linear regression evaluating model efficiency in several research by different
(FLR) [9]–[12]. However, to this date, most models of fuzzy models [17]–[21].
regression analysis have focused only on type-1 fuzzy sets,

681
978-1-5090-4102-2/16/$31.00 ©2016 IEEE
 AαV ( A∝ , A∝ ) . Where μ AV ( x) =  μ A ( x ) , μ A ( x )  (Fig. ∝ ∝
α

1) [16].
We assume that a fuzzy phenomenon can be presented as a
fuzzy system of equations which, in turn, can be described by
the fuzzy function Yˆi = A1 xi1 + A2 xi 2 + …+ Aq xiq in which A j
is the fuzzy parameters with triangular membership functions
( a j , b j , c j ) , i=1,2,…,n. Therefore the membership function
of the fuzzy linear function Yˆi is obtained as the triangular
membership function (Ci1 , Ci 3 , Ci 2 ) in which:

 (1 (Ci 3 − Ci1 ))( y − Ci 3 ), Ci1 ≤ y ≤ Ci 3


U Yˆ ( y ) = (1 (Ci 3 − Ci 2 ))( y − Ci 2 ), Ci 3 ≤ y ≤ Ci 2 (2)
 0 otherwise
Fig. 1. α − cuts of IVFS [16] 
The rest of this paper is organized as follows: Section 2 q q q
reviews the relevant literature mentioned in this study; Section
3 describes the proposed type-2 fuzzy regression; Section 4
Ci1 = a x
j =1
j ij
, Ci 2 = c x ,
j =1
j ij
Ci 3 = b x
j =1
j ij
(3)
demonstrates and compares the ability of the proposed models
in TAIEX forecasting; Finally, Section 5 is the conclusion and
The membership function of the observed output data Yi also
implications.
should have the same special form ( pi , qi , ri ) .
II. REVIEW
According to the possibility and nessesity models of tanaka,
the type-1 fuzzy regression model can be rewiten as :
Definition 1: (Normal interval type-2 fuzzy set (IT2FS)): An
Yi ⊆ h Yˆi = A1 xi1 + A2 xi 2 +…+ Aq xiq
IT2FS A , is said to be normal if its upper membership (4)
function(MF) is normal, i.e. sup μ A ( x) = 1 Yi ⊇h Yˆi = A1 xi1 + A2 xi 2 +…+ Aq xiq
Wich that: Aj = (ai , b j , c j ) , A j = ( ai , b j , c j ) .
Definition 2: (Perfectly Normal IT2FS) An IT2FS say A , is The type-1 fuzzy regression problem can be reduced to the
said to be perfectly normal if both its upper and lower MFs are following two quadratic programming[2], [22].
normal, i.e. sup μ A ( x) = sup μ A ( x) = 1 Possibility model:
n q
min   ((c j − a j ) xij ) 2
A = c;α1 , β1 ;α 2 , β 2
i =1 j =1
Definition 3: Let , (5)
s.t
B = c′; α1′, β1′; α 2′ , β 2′ are IT2FN and wi > 0 for i=1,2,…,5, Ci 3 − (1 − H )(Ci 3 − Ci1 ) ≤ qi − (1 − H )(qi − pi )
define dw as follow[15]: Ci 3 + (1 − H )(Ci 2 − Ci 3 ) ≥ qi + (1 − H )(ri − qi )
d w ( A , B ) = w1 c − c′ + w2 α1 − α1′
Necessity model:
+ w3 β1 − β1′ + w4 α 2 − α 2′ + w5 β 2 − β 2′ n q
(1) max   ((c − a j ) xij ) 2
( A , A ) be a short notation for taking
j
Proposition 1: Let ∝ ∝
i =1 j =1

s.t
the α − cuts of the LMF, A , UMF, A , of an IT2FS, AV . Ci 3 − (1 − H )(Ci 3 − Ci1 ) ≥ qi − (1 − H )( qi − pi )
Then  μ A ( x ) , μ A ( x )  and [ μ∝ A ( x ) , μ ∝ A ( x )] are Ci 3 + (1 − H )(Ci 2 − Ci 3 ) ≤ qi + (1 − H )(ri − qi )
∝ ∝ ∝ ∝
(6)
closed intervals across all domain values x .
III. PROPOSED TYPE-2 FUZZY REGRESSION
Definition 4: (IT2FS α − cuts): The α − cuts of an IT2FS, Type-1 FLSs cannot fully handle high levels of
AV is a non-fuzzy set defined as follows: uncertainties existing in most of the real world applications.
The reason lies behind the fact that type-1 FLSs employs crisp
and “precise” type-1 fuzzy sets. A type-2 FLS can handle

682
higher uncertainty levels to produce improved performance. q q q q

Type-2 FLSs have a variety of real world applications, in Ci1U = a

j =1
uj xij Ci1L = a x
j =1
lj ij Ci 3 = b x
j =1
j ij Ci 2 L = c xj =1
lj ij

business and finance domains, in electrical energy domain, q

real-world automatic control, healthcare and medical domains
[23], [19], [24].
Ci 2U = c j =1
uj xij

The membership functions of type-2 fuzzy sets are three q q q

dimensional and include a footprint of uncertainty. It is the
third dimension of type-2 fuzzy sets, i.e. the footprint of
Ci 3 =  b j xij Ci 2 L =  clj xij Ci 2U = c uj xij (8)
j =1 j =1 j =1
uncertainty that additional degrees of freedom are provided,
which in turn makes it possible to model and handle i (
If Y = Y , Y i
i ) = ([C i 1U
, Ci 1 L ] , Ci 3 ,[ Ci 2 L , Ci 2 U ]) and y i = y , y i ( i )
uncertainties [25].
Interval type-2 fuzzy sets are the most widely used type-2 ( )
=  piu , pil , qi ,ril , riu  , the degree of the fitting of Yi to yi
fuzzy sets, since on the one hand they are simple to use and, is measured by the h-level sets of the observed and the
on the other hand it is very difficult to justify the use of any predicted fuzzy numbers respectively, where h-level have
kind of type-2 fuzzy set. In this case, the MF μ A (a) is an defined based on definition1. The fundamental idea is to
 , parameters auj , alj , b j , clj , cuj such that:
obtain A
Interval type-2 fuzzy set which can be represented only by its j
Footprint of Uncertainty (FOU). The FOU is bounded by two
Y i ⊂ h y , y i ⊂ h Y i (9)
membership functions, a lower μ  ( x (LMF) and an upper i
A
It means that we use the possibility model for UMFs and the
μ A ( x) (UMF). (I.e.by two type-1 MFs). necessity model for LMFs. Because it is enough to close the
Concerning Arianna Mencattini et al. [26], we will denote membership functions of the observed and predicted values as
much as possible and an h-cut of the observed value is
this kind of type-2 interval as 
A j = [  auj alj  , b j , clj cuj  Such
included in the predicted value (Fig. 2).
that ( )
LMFA = alj , b j , clj , UMFA = auj , b j , cuj ( ) and
2. The objective function is defined by
auj ≤ alj ≤ b j ≤ clj ≤ auj . J = J1 + J 2 − J 3 + J 4 .
The observed output data is the interval type-2 fuzzy
number yi = ([ p pil ] , qi ,[ ril riu ]) . Yi = (Y i , Y i ) is the n q n
iu
J1 =  ((a uj − alj ) xij ) 2 +  ((c uj − clj ) xij )2
predicted output and y i = ( y i , y ) is the observed output. The i =1 j =1 i =1
i
n q
basic idea of our approach based on Tanaka’s approach is to
minimize the fuzziness of the model by minimizing the total J2 = ((c
i =1 j =1
uj − auj ) xij )2
spread of the fuzzy coefficients, subject to including all the (10)
n q
given data. In type-2 fuzzy regression these should be satisfied
for primary membership functions and secondary membership J3 = ((c
i =1 j =1
lj − alj ) xij )2
functions.
This model has been built based on the possibility model n n

for UMF and the necessity model for LMF. These two models, J4 =  (q − b x ) i j ij
2

besides minimization the vagueness in secondary membership i =1 j =1

function, help to find an appropriate fit for the predicted value
into the observed value.
To formulate a fuzzy linear regression model, the followings
are assumed to hold.
1) The data can be represented by an interval type-2
fuzzy linear model, that Y , A j are Perfectly Normal IT2FS
and x j is positive:

Yi = A1 xi1 + A 2 xi 2 + ... + A q xiq (7)

Thus, the supports of LMF and UMF are obtained as the

following:
([Ci1U , Ci1L ], Ci3 ,[Ci 2 L , Ci 2U ]) = [aiu1 ail1 , bi1 , cil1 aiu1  ]xi1 + ...
+ [ aiun ai lq  , bq , cilq aiuq  ]xiq Fig. 2. Degree of fitting of Y to y

683
That: root mean square error (RMSE) has been used that was
n q n defined as follow:
J1 =  ((a
i =1
uj − alj ) xij )2 +  ((c
i =1
uj − clj ) xij ) 2 RMSE =
j =1
n
n q
 (actual value(t ) − estimated value(t )) 2

J2 =  ((cuj − auj ) xij ) 2

i =1 (15)
i =1 j =1 n
n q
J3 = ((c
i =1 j =1
lj − alj ) xij ) 2 Fig. 3, Fig. 4 and table 1 show results.
The results indicate that the proposed model in addition to
n n its simplicity and intelligible steps has been able to perform
J4 = i =1
(q i − b x )
j =1
j ij
2 predictions relatively accurate. RMSE of proposed model is
minimum comparing with previous type-2 fuzzy models. This
(11) model is a primary kind of our proposed model. With
That minimizing of J1 minimizes the vagueness of secondary optimizing it, surely this model can also present better results.
membership function; and J 2 minimizes the vagueness of
LMF of Y that is a type-1 fuzzy set. J minimizes the 4
distance point with the highest membership value in a fuzzy
estimated response and the point with the highest membership
value in the corresponding observed response. But J 3 is used
for the necessity problem, and hence it should be maximized.
Our goal is to find out the fuzzy parameters auj , alj , b j , clj , cuj ,
which are the solution of the following two QP problems and
all objective function (Fig. 2):
min J2
s.t
Ci 3 − (1 − H )(Ci 3 − Ci1U ) ≤ qi − (1 − H )( qi − piu )
Ci 3 + (1 − H )(Ci 2U − Ci 3 ) ≥ qi + (1 − H )( riu − qi )
(12)
max J3
s.t Fig 3. Interval type-2 fuzzy regression model

Ci 3 − (1 − H )(Ci 3 − Ci1 L ) ≥ qi − (1 − H )( qi − pil )

Ci 3 + (1 − H )(Ci 2 L − Ci 3 ) ≤ qi − (1 − H )( ril − qi )
(13)

More specifically, we can rewrite our problem as

following:
Min J = J1 + J 2 − J 3 + J 4
s .t
C i 3 − (1 − H )( C i 3 − C i 1U ) ≤ q i − (1 − H )( q i − p iu )
C i 3 + (1 − H )( C i 2 U − C i 3 ) ≥ q i + (1 − H )( riu − q i ) (14)

C i 3 − (1 − H )( C i 3 − C i 1 L ) ≥ q i − (1 − H )( q i − p il )
C i 3 + (1 − H )( C i 2 L − C i 3 ) ≤ q i − (1 − H )( ril − q i )
a uj ≤ a lj ≤ b j ≤ clj ≤ cuj

IV. TAIEX FORECASTING

Taiwan Stock Index of year 2000 has forecasted to
evaluate the performance of the proposed model [11,16]. Data Fig 4. TAIEX forecasting using proposed model
from 11/2 to 12/15 are training data, 12/16 to 12/29 are testing
data. To evaluate the performance of the proposed model the

684
 [14] O. Poleshchuk and E. Komarov, “A fuzzy linear regression model for
TABLE I interval type-2 fuzzy sets,” 2012 Annu. Meet. North Am. Fuzzy Inf.
RMSE CAMPARISON Process. Soc., pp. 1–5, Aug. 2012.
Type-2 fuzzy Poleshchuk WGP Proposed [15] E. Hosseinzadeh, H. Hassanpour, and M. Arefi, “A weighted goal
models model model model programming approach to fuzzy linear regression with crisp inputs and
[14] [15] type-2 fuzzy outputs,” Soft Comput., vol. 19, no. 5, pp. 1143–1151, Jun.
2014.
RMSE in training 113 107 66 [16] H. Hamrawi, S. Coupland, and R. John, “A novel alpha-cut
data representation for type-2 fuzzy sets,” in 2010 IEEE International
RMSE in testing 182 179 109 Conference on Fuzzy Systems (FUZZ), 2010, vol. 3, pp. 1–8.
data [17] Z. Wang and S. Lee, “A Neuro-Fuzzy Based Method for TAIEX
Forecasting,” in 2014 International Conference on Machine Learning and
Cybernetics, 2014, no. x, pp. 13–16.
V. CONCLITION [18] T. H. Yu, “A bivariate fuzzy time series model to forecast the TAIEX,”
This study presents a type-2 fuzzy regression model vol. 34, pp. 2945–2952, 2008.
[19] N. Shafaei Bajestani and A. Zare, “Forecasting TAIEX using improved
according to the possibility and necessity model. In this type 2 fuzzy time series,” Expert Syst. Appl., vol. 38, no. 5, pp. 5816–
model, vagueness is minimized, under the circumstances 5821, 2011.
where the h-cut of observed value is included in predicted [20] S.-M. Chen and P.-Y. Kao, “TAIEX forecasting based on fuzzy time
value. In this case, both observed values and predicted values series, particle swarm optimization techniques and support vector
machines,” Inf. Sci. (Ny)., vol. 247, pp. 62–71, 2013.
are interval type-2 fuzzy numbers. In this model both primary [21] S. M. Chen and S. W. Chen, “Fuzzy forecasting based on two-factors
and secondary membership function of predicted value fit the second-order fuzzy-trend logical relationship groups and the probabilities
observed value. of trends of fuzzy logical relationships,” IEEE Trans. Cybern., vol. 45,
TAIEX forecasting shows that this model could be used in no. 3, pp. 405–417, 2015.
[22] H. Tanaka and H. Lee, “Interval Regression Analysis by Quadratic
real application with simple concepts and reliable results. Programming Approach,” IEEE Trans. FUZZY Syst., vol. 6, no. 4, pp.
Only one type of type-2 fuzzy sets is discussed here, but 473–481, 1998.
many types of type-2 fuzzy sets can be treated through [23] H. Hagras and C. Wagner, “Towards the Wide Spread Use of Type-2
developing same concept as in this paper. Yet, most of high Fuzzy Logic Systems in Real World Applications,” Comput. Intell. Mag.
IEEE, vol. 7, no. 3, pp. 14–24, 2012.
vague phenomenon might be well identified by this model. [24 M. H. Zarinbal, M. , Fazel Zarandi, “Type-2 fuzzy image enhancement:
Fuzzy rule based approach,” J. Intell. Fuzzy Syst., vol. 26, no. 5, pp.
REFERENCES 2291–2301, 2014.
[1] H. Tanaka, S. Uejima, and K. Asai, “Linear Regression Analysis with [25] J.Mendel, Uncertain Rule-Based Fuzzy Logic Systems: Introduction and
Fuzzy Model,” IEEE Trans. Syst. MAN, Cybern., vol. 75, no. 6, pp. 903– New Directions. Prentice Hall, 2001.
907, 1982. [26] Mencattini, M. Salmeri, and R. Lojacono, “Type-2 Fuzzy Sets for
[2] H. Tanaka, “Fuzzy data analysis by possibilistic linear models,” Fuzzy Modeling Uncertainty in Measurement,” in InternationalWorkshop on
Sets Syst., vol. 24, no. 3, pp. 363–375, 1987. Advanced Methods for Uncertainty Estimation in Measurement, 2006, no.
[3] H. Tanaka, “Possibillistic linear systems and their application to the linear April, pp. 8–13.
regression model,” Fuzzy Sets Syst., vol. 27, no. 3, pp. 275–289, 1988.
[4] Phil Diamond, “fuzzy least squares,” Inf. Sci. (Ny)., vol. 46, no. 3, pp.
141–157, 1988.
[5] H. T. Lee and S. H. Chen, “Fuzzy regression model with fuzzy input and
output data for manpower forecasting,” Fuzzy Sets Syst., vol. 119, no. 2,
pp. 205–213, 2001.
[6] C. Huang and G. Tzeng, “Multiple generation product life cycle
predictions using a novel two-stage fuzzy piecewise regression analysis
method,” Technol. Forecast. Soc. Change, vol. 75, no. 1, pp. 12–31, Jan.
2008.
[7] N.-F. Pan, C.-H. Ko, M.-D. Yang, and K.-C. Hsu, “Pavement
performance prediction through fuzzy regression,” Expert Syst. Appl., vol.
38, no. 8, pp. 10010–10017, Aug. 2011.
[8] A. Kazemi, H. G. Shakouri, M. B. Menhaj, M. R. Mehregan, and N.
Neshat, “A Hierarchical Fuzzy Linear Regression Model for Forecasting
Agriculture Energy Demand : A Case Study of Iran,” in 2011 3rd
International Conference on Information and Financial Engineering,
2011, vol. 12, pp. 19–24.
[9] G. Peters, “Fuzzy linear regression with fuzzy intervals,” Fuzzy Sets Syst.,
vol. 119, no. 2, pp. 247–257, Apr. 1994.
[10] E. C. Özelkan and L. Duckstein, “Multi-objective fuzzy regression: a
general framework,” Comput. Oper. Res., vol. 27, no. 7–8, pp. 635–652,
Jun. 2000.
[11] M. Hojati, C. R. Bector, and K. Smimou, “A simple method for
computation of fuzzy linear regression,” Eur. J. Oper. Res., vol. 166, no.
1, pp. 172–184, Oct. 2005.
[12] J.-R. Yu and C.-W. Lee, “Piecewise regression for fuzzy input–output
data with automatic change-point detection by quadratic programming,”
Appl. Soft Comput., vol. 10, no. 1, pp. 111–118, Jan. 2010.
[13] Y. Wei and J. Watada, “Building a Type II Fuzzy Qualitative
Regression,” Intell. Decis. Technol., vol. 1, no. 15, pp. 145–154, 2012.

685

View publication stats