A Novel Approach to Generating an Interval Type-2
Fuzzy Neural Network Based on a Well-Behaving
Type-1 Fuzzy TSK System
Junlong Gao, Ruyi Yuan, and Jianqiang Yi, Senior Member, IEEE
Institute of Automation,
Chinese Academy of Sciences
Beijing 10090, China
{junlong.gao, ruyi.yuan, jianqiang.yi}@ia.ac.cn
Hao Ying, Fellow, IEEE
Dept. of Electrical & Computer Engineering
Wayne State University
Detroit 48202, USA
hao.ying@wayne.edu
Chengdong Li, Member, IEEE
School of Information & Electrical Engineering
Shandong Jianzhu University
Jinan 250101, China
lichengdong@sdjzu.edu.cn
Abstract—This paper presents a novel approach to
automatically creating an interval type-2 fuzzy neural network
(IT2-FNN) from a type-1 fuzzy TSK system (T1-TSK). The IT2-
FNN is constructed in such a way that it takes advantage of the
well-behaving T1-TSK. Our approach makes designing the IT2-
FNN more efficient and the resulting system is expected to
perform better than the T1-TSK due to the footprint of
uncertainty of the IT2 fuzzy sets, especially when the system is
subject to heavy external or internal uncertainties. There are two
automated procedures in the IT2-FNN formation: (1) antecedent
structure construction, and (2) learning of the parameters in both
the antecedent and consequent. The structure construction is
based on antecedent structure of the T1-TSK and consists of
three steps – IT2 fuzzy set creation, similarity categorization, and
mergence. The IT2 fuzzy sets are directly initialized from the
fuzzy sets of the T1-TSK. Then, the IT2 fuzzy sets are classified
into different groups based on their similarities. Finally, the IT2
fuzzy sets in each group are merged to create a representative
IT2 fuzzy set for each group. The parameter learning procedure
uses a hybrid learning algorithm to attain the optimal values for
all the parameters. The learning algorithm adopts a new adaptive
steepest descent algorithm and a linear least-squares method to
adjust the antecedent parameters and consequent parameters,
respectively. One benchmark modelling problem is utilized to
compare our approach with the T1-TSK systems in the literature
under various scenarios. The comparison results show our IT2-
FNN performs better than the T1-TSK systems, especially when
there are strong uncertainties. In summary, the IT2-FNN can not
only achieve better performance but its structure is simpler than
that of the similar type-2 fuzzy neural networks in the literature.
Keywords—fuzzy logic system; type transition; fuzzy set
mergence; interval type-2 fuzzy neural network; adaptive steepest
decent algorithm
I. I
NTRODUCTION
In recent decades numerous achievements which use type-1
fuzzy logic, in data-driven modeling and prediction have been
made as an important application branch with the property of
universal approximation in fuzzy logic systems, e.g. [1-3] and
so on. Since the type-2 fuzzy logic system (T2-FLS) was
brought into practical applications[4, 5], it draws researchers’
attention and becomes the hotspot of fuzzy society very
rapidly for the advantages compared to type-1 fuzzy logic
systems (T1-FLSs), i.e. footprint of uncertainty (FOU) which
can bring additional design degree of freedom to make T2-
FLS with more outstanding potential to overcome
disturbances and to reduce the rule numbers. To date, simpler
T2-FLS, i.e. interval type-2 fuzzy logic system (IT2-FLS), has
already been applied into signal processing, control, pattern
recognition, stock prediction and so forth [6]. However,
between two existing IT2-FLSs which are IT2-Mamdani
system (whose consequent part is Mamdani interval type-2
fuzzy sets (IT2-FSs)) and IT2-TSK system (whose consequent
part is polynomial functions combined with input variables)
respectively, the IT2-Mamdani system rules are hard to design
in some complex systems which do not have figurative
physical meanings whereas the IT2-TSK system rules have
specific mathematical expressions which can be understood as
special cases of piecewise approximations. Thus the IT2-TSK
system provides an easier way to reduce rule-design
difficulties at a certain degree. And either in study depth or
breadth, the IT2-TSK system is in the trend to substitute the
IT2-Mamdani system in recent years.
At present, most approaches in automate learning (i.e. self-
organizing) IT2-TSK systems for system modelling or time
series’ prediction problems adopt structure of interval type-2
fuzzy neural networks (IT2-FNNs) or interval type-2 neural
fuzzy systems (IT2-NFSs). In addition, there are two ways to
construct an IT2-FLS, the first and the most common used
This work is supported by NNSFC under grant No. 61273149, No.
61403381, No. 61421004, No. 61473176, by the NDBSR under grant No.
B1320133020 and the Natural Science Foundation of Shandong Province fo
Outstanding Young Talents in Provincial Universities under gran
ZR2015JL021 .
2016 IEEE International Conference on Systems, Man, and Cybernetics • SMC 2016 | October 9-12, 2016 • Budapest, Hungary
978-1-5090-1897-0/16/$31.00 ©2016 IEEE