100% found this document useful (1 vote)
86 views11 pages

An Interval Type-3 Fuzzy System and A New Online Fractional-Order Learning Algorithm Theory and Practice

Uploaded by

david teacher
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
100% found this document useful (1 vote)
86 views11 pages

An Interval Type-3 Fuzzy System and A New Online Fractional-Order Learning Algorithm Theory and Practice

Uploaded by

david teacher
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 11

1940 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 28, NO.

9, SEPTEMBER 2020

An Interval Type-3 Fuzzy System and a New Online


Fractional-Order Learning Algorithm: Theory
and Practice
Ardashir Mohammadzadeh , Mohammad Hosein Sabzalian, and Weidong Zhang

Abstract—The main reason of the extensive usage of the fuzzy The motivation behind this approach is that more level of
systems in many branches of science is their approximation ability. uncertainty can be modeled by GT2FSs in contrast to IT2FSs.
In this paper, an interval type-3 fuzzy system (IT3FS) is proposed. Because the secondary membership in IT2FS is a fixed number
The uncertainty modeling capability of the proposed IT3FS is
improved in contrast to type-1 and type-2 fuzzy systems (T1FS and but in GT2FS counterpart, it is obtained from type-1 fuzzy set.
T2FS). Because in the proposed IT3FS, the membership is defined Recently, some aspect of the GT2FSs has been studied. For
as an interval type-2 fuzzy set, whereas in T1FS and T2FS, the instance, the centroid type-reduction of GT2FSs is investigated
membership is crisp value and type-1 fuzzy set, respectively. An in [13]. In [14], the optimality of fuzzifier for GT2FS is studied.
online fractional-order learning algorithm is given to optimize the
In [15], the general type-2 fuzzy membership functions (MFs)
consequent parameters of the IT3FS. The stability of the learning
algorithm is proved by utilizing the Lyapunov method. The validity are parameterized. In [16], the Jaccard similarity measure is
of the proposed fuzzy system is illustrated by both simulation and studied for GT2FS. In [9], the performance potential of the
the experimental studies. It is shown that the proposed fuzzy system GT2FS and the IT2FS is compared. In [17], the high-order
and associated learning algorithm result in better approximation α-plane integration is proposed to decrease the computational
performance in comparison with the other well-known approaches.
cost of the GT2FSs.
Index Terms—Approximation performance, fractional order, Recently, the GT2FSs have attracted more attention in engi-
interval type-3 fuzzy system (IT3FS), stable learning algorithm, neering applications. For instance, in [18], a GT2FS is designed
type-2 fuzzy systems (T2FS). based on zSlice concept and it is used in a bearing fault detection
I. INTRODUCTION problem. In [19], the GT2FSs are used in the navigation of a
bipedal humanoid robot. In [20], a GT2FS is used for modeling
N THE past decades, the application of the fuzzy systems
I has been extended to many branches of science, such as con-
trol systems [1], fault detection problems [2], decision-making
the human performance by features identification. The problem
of the density-based odor classification by GT2FSs is studied
in [21]. In [22], a GT2FS is designed based on the concept
problems [3], energy management systems [4], forecasting prob- of zSlice and it is used in a forecasting problem. In [23], the
lems [5], disease diagnosis systems [6], pattern recognition GT2FSs are used for designing the H∞ -based controllers for a
problems [7], etc. The main capability of the fuzzy systems class of nonlinear systems. In [24], the GT2FSs are used in a
is their approximation ability. In most of the aforementioned problem of risk analysis and the effectiveness of the GT2FSs
applications, the fuzzy systems are used to estimate a nonlinear is demonstrated. In the aforementioned papers, it is shown that
function or to derive a dynamic model from a dataset. the use of GT2FSs result in better performance in contrast to
The type-1 and interval type-2 fuzzy systems (IT2FS) have the IT2FSs, because the GT2FSs have more ability to cope
been widely studied [8]–[12]. In recent years, some attempts with the uncertainties. Then, based on this motivation, an IT3FS
have been made to improve the approximation capability of is proposed to handle more level of uncertainties. The secondary
the fuzzy systems. One of the main approaches is to make the membership in the proposed IT3FS is an interval type-2 fuzzy
general type-2 fuzzy systems (GT2FSs) to be more applicable. set. As the authors know, the type-3 fuzzy systems have been
quite rarely studied. In [25], the definition of the fuzzy sub-
Manuscript received January 27, 2019; revised April 11, 2019, May 25, 2019, sethood is developed for the type-n fuzzy sets. In [26] and
and June 20, 2019; accepted July 5, 2019. Date of publication July 15, 2019; date [27], the concepts and applications of type-1 and type-2 fuzzy
of current version September 1, 2020. This work was supported in part by the systems (T2FS) are reviewed and a short discussion about the
National Science Foundation of China under Grant U1509211 and in part by the
National Key R&D Program of China (SQ2017YFGH001005). (Corresponding complexity of the type-3 and higher level fuzzy sets is presented.
author: Ardashir Mohammadzadeh.) In [28]–[30], a short discussion is presented about type-3 and
A. Mohammadzadeh is with the Department of Electrical Engineering, Uni- type-n MFs. In some papers [31], [32], the Takagi–Sugeno
versity of Bonab, Bonab 5551761167, Iran (e-mail: [email protected]).
M. H. Sabzalian and W. Zhang are with the Department of Automation, Shang- fuzzy systems are also known as the type-3 adaptive neurofuzzy
hai Jiao Tong University, Shanghai 200240, China (e-mail: mohammadhosein inference systems, which is quite different with the concept
@sjtu.edu.cn; [email protected]). presented in this paper.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. Another important aspect of the fuzzy systems is the learning
Digital Object Identifier 10.1109/TFUZZ.2019.2928509 algorithm. The offline learning algorithms have been widely
1063-6706 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
MOHAMMADZADEH et al.: IT3FS AND A NEW ONLINE FRACTIONAL-ORDER LEARNING ALGORITHM: THEORY AND PRACTICE 1941

applied to optimize the fuzzy neural networks. In many practical


applications, especially in adaptive control applications, it is
needed that the fuzzy model of the system to be online optimized.
Because in some cases, there is no offline data or it is impossible
that the learning algorithm to be repeated for more epochs. The
online learning of the fuzzy systems has been considered in some
researches. For instance, in [33], the active control method is
combined with the online learning fuzzy system. In [34], the
problem of the nonlinear system identification is investigated Fig. 1. Proposed interval type-3 Gaussian MF.
through the T2FS and the sliding mode online learning tech-
nique. In [35], the circuit implementation of the interval type-2
fuzzy neural network (IT2FNN) and online learning algorithm
based on the gradient descent algorithm is studied. In [36], the
online tuning of the proportional-integral controller based on the
gradient descent algorithm is investigated. The feature extraction
and online identification problem through IT2FNNs and Kalman
filter algorithms are studied in [37]. In [38], the dynamics of the
bilateral teleoperation system are modeled by the T2FSs and
the effectiveness of the T2FSs is shown using an experimental
platform.
However, the learning of the GT2FSs is rarely considered in
the literature. For instance, in [39], the GT2FSs are optimized
using genetic algorithm. In [21], the simulated annealing method Fig. 2. General type-2 Gaussian MF.
is used to optimize the GT2FSs. In [40], the GT2FSs are learned
through the adaptation laws, which are obtained from the stabil-
ity analysis. In [41], a hybrid learning algorithm based on least
square technique is proposed to optimize general type-2 fuzzy
logic systems in a classification problem.
In the most of the aforementioned learning algorithms, in or-
der to achieve the desired performance, the learning method is re-
peated for several times, as the best knowledge of the authors. In
online applications, the repetition of the learning algorithm im-
poses computational cost and it is impossible. Moreover, in the
most of the presented studies, the stability of the learning algo-
rithm is not investigated, which is very important in online practi-
cal applications. Based on the aforementioned discussion, in this Fig. 3. Interval type-2 Gaussian MF.
paper, a stable fractional-order learning algorithm is presented.
The contributions of this paper are summarized as follows.
1) An interval type-3 Gaussian fuzzy system is presented and and vertical slices is shown in Fig. 1. For comparison, the general
is compared with the other well-known fuzzy systems. type-2 MF (GT2MF) and the interval type-2 MF (IT2MF) are
2) A new learning algorithm based on the fractional-order shown in Figs. 2 and 3, respectively. As it can be seen, the
adaptation laws is proposed to train the proposed fuzzy secondary MF in the IT2MF and GT2MF is constant and type-1
system. The stability of the learning algorithm is investi- MF, respectively. But in the proposed IT3MF, the secondary MF
gated by the Lyaponuv theorem. is an IT2MF.
3) The effectiveness of the proposed method is verified The GT2MF is defined as follows [42]:
through the experimental and simulation studies.
  μ (x, u)
The rest of this paper is organized as follows. In à = Ã
, Jx ⊆ [0, 1] (1)
Section II, the proposed fuzzy system is described. The learning (x, u)
x∈X u∈Jx
algorithm and stability analysis are illustrated in Section III. 
The simulations and results are provided in Section IV and where represents the union over x and u. The IT2MF is
Section V concludes this paper. similar to the GT2MF with difference that the secondary MF is
constant
II. PROPOSED TYPE-3 FUZZY SYSTEM   1
à = , Jx ⊆ [0, 1] . (2)
A. Primary Definitions (x, u)
x∈X u∈Jx

In this section, the proposed type-3 fuzzy system is illustrated. As it can be seen from (2), in the IT2MF for all x and u, one has
The proposed interval type-3 MF (IT3MF) with some horizontal μà (x, u) = 1.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
1942 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 28, NO. 9, SEPTEMBER 2020

Fig. 4. Example of a vertical slice for the interval type-2 Gaussian MF. Fig. 5. Example of a vertical slice for the general type-2 Gaussian MF.

The proposed IT3MF is defined as follows:


 
1
  μà (x,u)
à = , Jx ⊆ [0, 1] (3)
(x, u)
x∈X u∈Jx

where 1/μà (x, u) represents the interval type-2 fuzzy MF.

B. Vertical Slice Representation


Consider the vertical slices of the IT2MF, the IT2MF can be
written as the union of the some MFs as follows.
Discrete the domain of x into N points (x1 , x2 , . . . , xN ) and
at each xi , i = 1, . . . , N , discrete the corresponding ui into Mi Fig. 6. Example of a vertical slice for the proposed interval type-3
points, then the IT2MF is represented as follows [42]: Gaussian MF.
 M1 1 MN 1
k=1 u1k k=1 uN k
à = + ··· + (4)
x1 xN
where + denotes the union.
Now, consider the GT2MF, the secondary grade correspond-
ing to each ui is not equal to one, but it is αi . Then, the GT2MF
is rewritten as follows:
 M 1 αk  M N αk
k=1 u1k k=1 uN k
à = + ··· + . (5)
x1 xN
Similar to (5), the vertical slice representation of the proposed
IT3MF is defined as follows. The domain of x is discretized
to N points (x1 , x2 , . . . , xN ) and at each xi , the domain of Fig. 7. Example of a horizontal slice for the general type-2 Gaussian MF.
primary membership ui is discretized into Mi points (uik , k =
1, . . . , Mi ), and the domain of corresponding secondary grade
to each uik is discretized into Pk points αikp , p = 1, . . . , Pk ,
then the proposed IT3MF is represented as follows: x2 , at each xi , i = 1, 2, the domain of ui is discretized into
M1 P1 αkp MN PN αkp two points ui1 and ui2 and the secondary grade corresponding
k=1 p=1 u1k k=1 p=1 uN k to each uik , k = 1, 2 is discretized into three points αik1 , αik2 ,
à = + ··· + . (6) and αik3 .
x1 xN
The aforementioned definitions are graphically illustrated by an
example. Fig. 4 depicts an IT2MF, in which the domain of x C. Horizontal Slice Representation
is discretized into two points x1 and x2 , at each xi , i = 1, 2, Consider a general Gaussian type-2 MF Ã, the horizontal slice
the domain of ui is discretized into six points and the secondary representation of à is as follows (see Fig. 7):
grade is equal to one for each uik . Fig. 5 depicts a GT2MF,
in which the domain of x is discretized into two points x1 and α1 αK
à =  + ··· +  , αk ⊆ [0, 1] (7)
x2 , at each xi , i = 1, 2, the domain of ui is discretized into six Ãα1 ÃαK
x∈X x x∈X x
points uik , k = 1, . . . , 6 and the secondary grade corresponding
to each uik is αik , k = 1, . . . , 6. Fig. 6 depicts an IT3MF, in where αk , k = 1, . . . , K represent the secondary grade (the
which the domain of x is discretized into two points x1 and level of horizontal slice). At each level αk , Ãαk can be written

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
MOHAMMADZADEH et al.: IT3FS AND A NEW ONLINE FRACTIONAL-ORDER LEARNING ALGORITHM: THEORY AND PRACTICE 1943

Fig. 10. Three domains of the proposed MF.


Fig. 8. Example of a horizontal slice for the interval type-2 Gaussian MF.

the secondary domain for all elements in the primary domain is


not a crisp value but it is a fuzzy number. In other words, the
secondary MF is an IT2MF. The third domain is equal to one
for all elements in the secondary domain (see Fig. 10).

D. Interval Type-3 Fuzzy System


Consider an IT2MF as shown in Fig. 8, when x = x , the
upper and the lower primary memberships (u) at all horizontal
slice level are fixed and they are certain values. But in GT2MF as
shown in Fig. 7, the upper and the lower primary memberships
Fig. 9. Example of a horizontal slice for the proposed interval type-3 are changed at different horizontal slice levels and also they are
Gaussian MF.
certain values. Now, consider the proposed IT3MF as shown in
Fig. 9, in addition to the upper and the lower primary member-
as its upper and lower MFs. Then, (7) is rewritten as follows: ships are changed at different horizontal slice levels, they are
also uncertain value. In other words, the upper and the lower
α1 αK
à =  +· · ·+  , α ⊆ [0, 1]. memberships are interval type-1 MFs.
[Aα1 ,Āα1 ] [AαK ,ĀαK ] k As shown in Fig. 9, a horizontal slice of an IT3MF is equal
x∈X x x∈X x
(8) to the horizontal slice of a GT2MF at two levels αk and ᾱk .
It should be noted that if à is an IT2MF, then as shown in Fig. 8, Then, the IT3MF can be created by union of all horizontal slices
the secondary grade at all slices is equal to one (αk = 1, k = α̃1 , α̃2 , . . . , α̃K . A horizontal slice of an IT3MF at level α̃k
1, . . . , K) and also the upper and the lower MFs at all slice are represents a type-2 MF in which its secondary grade is between
equal to A αk = A and Āαk = Ā, k = 1, . . . , K. [αk , ᾱk ]. The overall view of the IT3FS is shown in Fig. 11. The
Now, consider an IT3MF Ã, in contrast to the GT2MF, the output of IT3MF is obtained step by step as follows.
secondary grade is not equal to αk , but it is an interval type-1 1) Get inputs x1 , . . . , xn . The complexity of this step in the
MF. The horizontal representation of à is as follows: big-O notation is O(1).
2) Consider MFs Ãji , j = 1, . . . , M for each input. Ãji is the
1/α̃1 1/α̃K
à =  + ··· +  , αk ⊆ [αk , ᾱk ] (9) jth MF for the ith input. The complexity of this step is
Ãα̃1 Ãα̃K
x∈X x x∈X x
O(1).
3) Compute memberships of Ãji , j = 1, . . . , nAi for differ-
where αk ⊆ [αk , ᾱk ] and corresponding to αk and ᾱk , one has
ent horizontal slice levels μs = αk , k = 1, . . . , K. As-
Ãα̃k ⊆ [Aαk , Āαk ] and Ãα̃k ⊆ [Aᾱk , Āᾱk ], respectively. Then,
sume the value of xi is x i and the horizontal slice level is
(9) can be rewritten as follows (see Fig. 9):
⎡ ⎤ μs = α̃k , then (see Fig. 12)
1/α1
à = ⎣  , 1/ᾱ1 ⎦ (10) μ̄Ãj |μs =ᾱk (x i ) = mÃj |μs =α0 (x i ) + ln (1/ᾱk ) σV2 i
 Aα ,Āα
1 1
 [Aᾱ1 ,Āᾱ1 ] i i
x∈X x x∈X x

⎡ ⎤ (12)

+··· + ⎣ 1/αK , 1/ᾱK ⎦. (11) μÃj |μs =ᾱk (x i ) = mÃj |μs =α0 (x i ) − ln (1/ᾱk ) σV2 i
 Aα ,Āα
K K
 [AᾱK ,ĀᾱK ] i i
x∈X
x∈X x x (13)
Remark 1: Unlike the GT2FS, in which the secondary do-
main is a crisp value between [0,1] for all elements in the primary μ̄ Ãj |μs =α (x i ) = mÃj |μs =α0 (x i ) + ln (1/αk ) σV2 i
i k i
domain (the secondary MF is a type-1 MF), in the proposed MF, (14)

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
1944 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 28, NO. 9, SEPTEMBER 2020

2   
1

− mÃj |μs =α0 (x i ) ln (17)
i ε
where
mÃj |μs =α0 (x i ) = μÃj |μ (x i )
i i s =α0

+ μ̄Ãj |μs =α0 (x i ) 2 (18)
i
⎛ 2 ⎞

⎜ x i − C Ã j

μ̄Ãj |μs =α0 (x i ) = exp ⎝− 2
i
⎠ (19)
i σ̄Ã j
i

⎛ 2 ⎞

⎜ x i − CÃi
j

μÃj |μ (x i ) = exp ⎝− 2 ⎠ (20)
i s =α0 σ Ãj
i
Fig. 11. Overall view of the Mamdani-type IT3FS.
where CÃj , σ̄Ãj , and σ Ãj are center, upper width, and
i i i

lower width of MF Ãji at horizontal slice level μs = α0 .


The complexity of this step is O(n × M × K).
4) Compute the rule firings. Consider the lth rule as follows:
if x1 is Ãl1 and · · · and xn is Ãln
then y ⊂ [w l , w̄l ] , l = 1, . . . , M (21)
where M is the number of rules and w l and w̄l are the con-
sequent parameters corresponding to the lth rule. Since the
consequent parameters are online adjusted then the form
of rules are considered to be fixed. The ranges of input
variables are divided into M sections. For each section,
a proposed Gaussian MF is considered. The number of
Fig. 12. Example of a horizontal slice for the proposed interval type-3
Gaussian MF. rules is equal with the number of MFs. The lth rule can
be interpreted as follows. If the input vector [x1 , . . . , xn ]T
belong to the cluster [CÃl , . . . , CÃln ]T then y ⊂ [wl , w̄l ].
1
The rule firing degrees are computed as
μ Ãj |μ (x i ) = mÃj |μs =α0 (x i ) − ln (1/αk ) σV2 i
i s =αk i z̄μl s =ᾱk = Tni μ̄Ãl |μs =ᾱk (22)
i
(15)
z lμs =ᾱk = Tni μÃl |μs =ᾱk (23)
  i
where μ̄Ãj |μs =ᾱk (x i )/μ̄Ãj |μs =α (x i ) and μÃj |μ
i i k s =ᾱk
(x i )/μÃj |μ (x i ) are the upper and the lower mem-
i
z̄μl s =αk = Tni μ̄Ãl |μs =αk (24)
i
i s =αk

berships of Ãji , when the horizontal slice level is ᾱk /αk . z lμs =αk = Tni μÃl |μs =αk (25)
i
The [αk , ᾱk ] represent the uncertainty of secondary grades
where T represents the T-norm. The complexity of this
1/Δ step is O(M ).
ᾱk = αk
5) Compute the output of the proposed fuzzy system as
αk = αkΔ (16) follows:
K 
where Δ ≥ 1 is a constant and represents the uncertainty ᾱ k
¯k + αk f
f
k=1 k
of the standard division of Gaussian vertical slice at xi = y= K (26)
k=1 (ᾱk + αk )
x i . In other words, if Ãji is the GT2MF, the vertical slice
at xi = x i is a type-1 MF with standard division σV2 i (x i ) where
M  l 
(see Fig. 7), but when Ãji is the IT3MF, the vertical slice z̄μs =ᾱk w̄l + z lμs =ᾱk wl
at xi = x i is an IT2MF with standard division [1/Δ · f¯k = M 
l=1
 (27)
l=1 z̄μs =ᾱk + z μs =ᾱk
l l
σV2 i (x i ) , Δ · σV2 i (x i )] (see Fig. 9). σV2 i (x i ) is obtained
M 
as follows: z̄ l
w̄ + z l
w
l=1 μs =αk l μs =αk l
fk =   . (28)
σV2 i (x i ) = μÃj |μ (x i ) M
z̄ l + z l
i s =α0 l=1 μs =αk μs =αk

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
MOHAMMADZADEH et al.: IT3FS AND A NEW ONLINE FRACTIONAL-ORDER LEARNING ALGORITHM: THEORY AND PRACTICE 1945

The complexity of this step is O(M × K). where


K z̄ l
III. STABLE FRACTIONAL-ORDER LEARNING ALGORITHM ᾱk M z̄l μs =ᾱ+z k
l=1 ( μs =ᾱk μs =ᾱk )
k=1 l
∂y
= K
In this section, a novel algorithm is presented to optimize w̄l (ᾱ k + α )
k=1 k
the parameters of the proposed fuzzy system. The adaptation
laws are derived form a stability analysis such that the following K l
z̄μ s =αk
α k M  
sliding surface is converged to zero: k=1 z̄μs =α +z lμs =α
l
l=1
+ K
k k

s= Dt1−q e + λDt−q e (29) k=1 (ᾱk + αk )

where 0 < q < 1, λ is a positive constant, and e is the estimation l = 1, . . . , M (39)


error, which is defined as follows:
K zl
ᾱk M z̄l μs =ᾱ+z k
l=1 ( μs =ᾱk μs =ᾱk )
l
e = y (θ|u) − yd (30) ∂y
=
k=1
K
k=1 (ᾱk + αk )
wl
where y is the output of the proposed fuzzy system, yd is the
desired output, u is the input vector of fuzzy system, and θ is the K z lμs =α
α k M  k 
vector of consequent parameters, which is defined as follows: k=1 z̄μs =α +z lμs =α
l
l=1

T
+ K
k k

θ = [w1 , . . . , wM , w̄1 , . . . , w̄M ] (31) k=1 (ᾱk + αk )

where M is the number of rules. Dtq g(t) is the Caputo derivative l = 1, . . . , M. (40)
of g in order q, which is defined as follows:
 t ∂y
1 d The term in (37) is obtained as follows:
(t − τ )−q g(τ )dτ
q ∂ui
Dt g(t) = (32)
Γ(1 − q) 0 dt
K ∂ f¯k ∂f
∂y k=1 ᾱk ∂xi + αk ∂xki
where Γ(·) is Gamma function. = K (41)
k=1 (ᾱk + αk )
Theorem 1: If the adaptation laws for the proposed fuzzy ∂ui
system are chosen as follows, then the estimation error e (t) is
converged to zero when t → ∞: where

∂ f¯k
     
¯ sign(s) ξ
θ̇ = −λe − |ẏ¯d | + ς T  |u̇| (33) ∂xi
ξT ξ ⎛     ⎞
M − xi −CÃl − xi −CÃl
⎝ i
z̄μl s =ᾱk w̄l + i
z lμs =ᾱk wl ⎠
where ξ and ς are defined as ξ = ∂y(θ|u) ∂θ and ς = ∂y(θ|u)
∂u ,
l=1 σ̄ 2 l

σ2 l

¯ ¯
respectively. |ẏ d | and |u̇| are the upper bounds of ẏd and u̇, i i
= M  l 
l=1 z̄μs =ᾱk + z μs =ᾱk
l
respectively.
Proof: Consider the following Lyapunov function: ⎛     ⎞
 − xi −CÃl − xi −CÃl
1 2 f¯k l=1 ⎝
M i
z̄μl s =ᾱk + i
z lμs =ᾱk ⎠
V = s (34) σ̄ 2
Ãl
σ2 l

2 i i
− M  
where s is defined in (29). The fractional time derivative of V l=1 z̄μl s =ᾱk + z lμs =ᾱk
yields (42)

Dtq V ≤ sDtq s. (35) ∂f k


∂xi
From (29), Dtq s is obtained as follows: ⎛     ⎞
M − xi −CÃl − xi −CÃl
Dtq s = ė + λe. (36) ⎝ i
z̄μl s =αk w̄l + i
z lμs =αk wl ⎠
l=1 σ̄ 2 l σ2 l
à Ã
i i

From (30), ė is obtained as follows: = M 


l=1 z̄μl s =αk + z lμs =αk
∂y (θ|u) T ∂y (θ|u) T
ė = θ̇ + u̇ − ẏd . (37) ⎛     ⎞
∂θ ∂u M − xi −CÃl − xi −CÃl
fk ⎝ i
z̄μl s =αk + i
z lμs =αk⎠
The term ∂y
is obtained as follows: l=1 σ̄ 2 l σ2 l
∂θ Ã Ã
i i

 T − M  .
∂y (θ|u) ∂y ∂y ∂y ∂y z̄μl s =αk + z lμs =αk
= ,..., , ,..., (38) l=1
∂θ w1 wM w̄1 w̄M (43)

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
1946 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 28, NO. 9, SEPTEMBER 2020

TABLE I
EXAMPLE 1: THE VALUES OF SIMULATION PARAMETERS

From (36) and (37), we have


∂y (θ|u) T ∂y (θ|u) T
Dtq V ≤ s θ̇ + s u̇ − sẏd + sλe. (44)
∂θ ∂u
Considering ξ = ∂y(θ|u)
∂θ and ς = ∂y(θ|u)
∂u , the inequality (44) is
rewritten as follows:
Dtq V ≤ sξ T θ̇ + sς T u̇ − sẏd + sλe. (45)
By replacing the adaptation law (33) into (44), we have Fig. 13. Example 1: The prediction performance and free noise case.
    
¯ sign (s) ξ
Dtq V ≤ s ξ T −λe − |ẏ¯d | + ς T  |u̇|
ξT ξ
+ sς T u̇ − sẏd + sλe. (46)
Simplification of (46) yields
 
Dtq V ≤ − |s| |ẏ¯d | − |s| ς T  |u̇|
¯ + sς T u̇ − sẏd < 0. (47)


IV. SIMULATIONS AND EXPERIMENTAL VERIFICATION Fig. 14. Example 1: The prediction performance, noisy case, and zoomed
In this section, the effectiveness of the proposed fuzzy system view.
and the associated learning algorithm is examined.
Example 1: In this example, the proposed IT3FS is used for
IT2FNN (SEIT2FNN) [47]. The values of root-mean-square
online prediction of the Mackey–Glass time series. The dynamic
error (RMSE) are given in Table II. As it can be seen, the
of the Mackey–Glass time series is given as follows:
proposed method results in significantly better performance than
0.2x (t − τ ) the other methods, especially in noisy case.
ẋ = − 0.1x(t) (48)
0.2x10 (t − τ ) Remark 2: It must be noted that the proposed leaning algo-
where τ = 30 and x(0) = 1.2. The inputs of fuzzy system are rithm is applied as online. In other words, unlike to the other
x (t − 6), x (t − 12), x (t − 18), and x (t − 24). The objective is methods, the simulation is run for just one epoch.
to estimate x(t) by the proposed fuzzy system. The first 700/300 Example 2: In this example, the proposed method is used for
data samples from 1000 data samples are used for training and online dynamic estimation of the Chuas chaotic system
testing [43]. The simulation parameters are given in Table I.
The trajectories of the actual and estimated outputs with the es- ẏ1 = y2
timation error in noisy and free noise cases are shown in Figs. 13 ẏ2 = y3
and 14, respectively. In noisy case, a white noise with zero mean ẏ3 = f (y1 , y2 , y3 )
and 0.1 variance is considered. As it can be seen, the proposed
method results in good performance. The performance of the 14 168 1
f (y1 , y2 , y3 ) = y1 − y2 + y3
proposed method is compared with the other fuzzy models, such 1805 9025 38
as: the hierarchical T2FS (HT2FS) [43], IT2FNN [44], evolving  3
2 28 7
type-2 neural fuzzy system (eT2FS) [45], self-constructing neu- − y1 + y2 + y3
ral fuzzy inference network (SONFIN) [46], and self-evolving 45 361 95 (49)

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
MOHAMMADZADEH et al.: IT3FS AND A NEW ONLINE FRACTIONAL-ORDER LEARNING ALGORITHM: THEORY AND PRACTICE 1947

TABLE II
EXAMPLE 1: THE COMPARISON OF RMSE FOR DIFFERENT MODELS

TABLE III
EXAMPLE 2: THE COMPARISON OF RMSE FOR DIFFERENT MODELS

Fig. 16. Example 4: The experimental setup.

Fig. 15. Example 2: The trajectories of the estimated and actual states.

self-constructing neural fuzzy system (SNFS) [46], interval


where initial conditions are y1 (0) = 0, y2 (0) = 0, type-2 neural fuzzy system for system identification and feature
and y3 (0) = 0.3. The trajectories of the actual and the estimated elimination (SEIT2FNN ) [37], SEIT2FNN [47], radial-basis
states with the estimation errors are given in Fig. 15. As it can function neural-network based on supervised fuzzy cluster-
be seen, the estimated trajectories are well converged to the ing (RBFNN-SFC) [48], GT2FS with sampling defuzzification
actual trajectories. The performance of the proposed method (GT2FS-SD), GT2FS based on vertical slice centroid type-
is compared with the HT2FS [43]. In [43], all parameters of reducer (GT2FS-VSCTR), and interval type-2 fuzzy logic sys-
HT2FS are trained with the square-root cubature Kalman filter. tem (IT2FLS) [21]. The comparison of RMSE for different
The values of RMSE are compared in Table III. It must be noted models is given in Tables IV and V. As it can be seen, the
that in [43], the results are obtained after 50 epochs. However, proposed fuzzy system in less training iterations and less training
the results of the proposed method in this paper are given after parameters, results in significantly better performance in con-
just one epoch. As it can be seen from Table III, the proposed trast to the other models. It must be noted that in this example,
method results in significantly better performance than other the consequent parameters of the proposed IT3FS are learnt by
methods. the extended Kalman filter and the simulation parameters of
Example 3: To verify the effectiveness of the proposed fuzzy IT3FS are same as Example 1. Also, the inputs of IT3FS are
system, the modeling problem of the two real-world datasets is divided into 200/5000 in the MCV and MPG problems.
considered. The first one is the prediction of fuel consumption Example 4: In this example, the effectiveness of the pro-
of automobile in miles per gallon (MPG) [37], and the second posed fuzzy system and online learning algorithm is verified
one is the estimation of maintenance costs of medium-voltage on an experimental setup. As shown in Fig. 16, the temperature
electrical line (MCV) [21]. In MPG and MCV problems, 320/72 is controlled by an ON–OFF control method and the dynamic
and 396/99 samples of total datasets are used for training/testing, of the system is online estimated by the proposed IT3FS. The
respectively. The performance of the proposed IT3FS is com- parameters of IT3FS are same as Example 1. The estimation
pared with other well-known fuzzy neural systems, such as: performance is shown in Figs. 17 and 18. As it can be seen, the

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
1948 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 28, NO. 9, SEPTEMBER 2020

TABLE IV
EXAMPLE 3: THE COMPARISON OF RMSE FOR DIFFERENT MODELS, MPG PROBLEM

TABLE V
EXAMPLE 3: THE COMPARISON OF RMSE FOR DIFFERENT MODELS, MAINTENANCE COST PROBLEM

TABLE VI
EXAMPLE 5: THE COMPARISON OF RMSE AND EXECUTION TIME FOR DIFFERENT MODELS

Fig. 19. Example 5: The foot print of uncertainty at a horizontal slice for the
interval type-3 fuzzy set.
Fig. 17. Example 4: The estimation performance.

zero mean and different variances as the measurement error is


added to the state y3 . The values of the RMSE are given in
Table VI. The simulation condition is same as Example 2 and
the number of rules for all models is 3. As it can be seen, the
proposed IT3FS results in better performance, especially in the
high noise environment. The proposed structure has more ability
to handle the uncertainties because of the nature of its MFs. The
Fig. 18. Example 4: The estimation performance and zoomed view. upper and the lower of foot print of uncertainty in the GT2FS
are crisp value but in the proposed structure are type-1 fuzzy set
(see Figs. 19 and 20).
trajectory of the estimated output is well converged to the actual Remark 3: In this paper, only the consequent parameters are
output. This implementation on a real-world application shows online tuned. The choosing of other parameters is not chal-
that the proposed method is effective and can be used for online lengeable, so that the value of these parameters are same in
dynamic approximation problems. five different examples. However, the parameters of the MFs
Example 5: In this example, the estimation performance and can be determined based on the variation range of the input
the execution time of the proposed IT3FS are compared with variables and the level of the uncertainties. It must be noted
the GT2FS and IT2FS. The online dynamic estimation of the that the antecedent parameters can also be trained based on the
Chuas chaotic system is considered and the white noise with adaptation law (33). To do this, (31) and (38) must be considered

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
MOHAMMADZADEH et al.: IT3FS AND A NEW ONLINE FRACTIONAL-ORDER LEARNING ALGORITHM: THEORY AND PRACTICE 1949

[3] C.-B. Cheng, H.-S. Shih, and E. S. Lee, “Fuzzy interactive multi-level deci-
sion making,” in Fuzzy and Multi-Level Decision Making: Soft Computing
Approaches. Berlin, Germany: Springer, 2019, pp. 89–111.
[4] D. Arcos-Aviles, J. Pascual, L. Marroyo, P. Sanchis, and F. Guinjoan,
“Fuzzy logic-based energy management system design for residential grid-
connected microgrids,” IEEE Trans. Smart Grid, vol. 9, no. 2, pp. 530–543,
Mar. 2018.
[5] L. Tang, X. Wang, X. Wang, C. Shao, S. Liu, and S. Tian, “Long-term
electricity consumption forecasting based on expert prediction and fuzzy
Bayesian theory,” Energy, vol. 167, pp. 1144–1154, 2019.
[6] G. Manogaran, R. Varatharajan, and M. Priyan, “Hybrid recommendation
system for heart disease diagnosis based on multiple kernel learning with
adaptive neuro-fuzzy inference system,” Multimedia Tools Appl., vol. 77,
no. 4, pp. 4379–4399, 2018.
Fig. 20. Example 5: The foot print of uncertainty at a horizontal slice for the [7] Q. Jiang, X. Jin, S.-J. Lee, and S. Yao, “A new similarity/distance measure
general type-2 fuzzy set. between intuitionistic fuzzy sets based on the transformed isosceles trian-
gles and its applications to pattern recognition,” Expert Syst. With Appl.,
vol. 116, pp. 439–453, 2019.
as follows: [8] O. Castillo and K. Atanassov, “Comments on fuzzy sets, interval type-2
 fuzzy sets, general type-2 fuzzy sets and intuitionistic fuzzy sets,” in
θ = CÃ1 , . . . CÃM , CÃ1n , . . . CÃM
n
Recent Advances in Intuitionistic Fuzzy Logic Systems. Berlin, Germany:
1 1 Springer, 2019, pp. 35–43.
σ̄Ã1 , . . . σ̄ÃM , σ̄Ã1n , . . . σ̄ÃM
n
[9] J. M. Mendel, “Comparing the performance potentials of interval and
1 1
general type-2 rule-based fuzzy systems in terms of sculpting the state
σ Ã1 , . . . σ ÃM , σ Ã1n , . . . σ ÃM
n space,” IEEE Trans. Fuzzy Syst., vol. 27, no. 1, pp. 58–71, Jan. 2019.
1 1
T [10] G. Ruiz-Garcia, H. Hagras, H. Pomares, and I. Rojas, “Towards a fuzzy
w1 , . . . , wM , w̄1 , . . . , w̄M (50) logic system based on general forms of interval type-2 fuzzy sets,” IEEE
Trans. Fuzzy Syst., to be published, doi: 10.1109/TFUZZ.2019.2898582.
! [11] F. Olivas, F. Valdez, P. Melin, A. Sombra, and O. Castillo, “Interval type-2
∂y (θ|u) ∂y ∂y ∂y ∂y fuzzy logic for dynamic parameter adaptation in a modified gravitational
= ,..., ,..., ,..., search algorithm,” Inf. Sci., vol. 476, pp. 159–175, 2019.
∂θ CÃ1 CÃM CÃ1n CÃM
1 1 n [12] X. Yang, H.-K. Lam, and L. Wu, “Membership-dependent stability condi-
tions for type-1 and interval type-2 T–S fuzzy systems,” Fuzzy Sets Syst.,
∂y ∂y ∂y ∂y vol. 356, pp. 44–62, 2019.
,..., ,..., ,...,
σ̄Ã1 σ̄ÃM σ̄Ã1n σ̄ÃM
n
[13] Y. Chen and D. Wang, “Study on centroid type-reduction of general type-2
1 1
fuzzy logic systems with weighted Nie–Tan algorithms,” Soft Comput.,
∂y ∂y ∂y ∂y vol. 22, no. 22, pp. 7659–7678, 2018.
,..., ,..., ,..., [14] S. Kulkarni and R. Agrawal, “Determining the optimal fuzzifier range for
σ Ã1 σ ÃM σ Ã1n σ ÃM
n alpha-planes of general type-2 fuzzy sets,” in Proc. IEEE Int. Conf. Fuzzy
1 1
"T Syst., 2018, pp. 1–8.
∂y ∂y ∂y ∂y [15] J. R. Castro, M. A. Sanchez, C. I. Gonzalez, P. Melin, and O. Castillo, “A
,..., , ,..., . new method for parameterization of general type-2 fuzzy sets,” Fuzzy Inf.
w1 wM w̄1 w̄M (51) Eng., vol. 10, no. 1, pp. 31–57, 2018.
[16] D. Wu and J. M. Mendel, “Similarity measures for closed general type-
2 fuzzy sets: Overview, comparisons, and a geometric approach,” IEEE
V. CONCLUSION Trans. Fuzzy Syst., vol. 27, no. 3, pp. 515–526, Mar. 2019.
[17] E. Ontiveros, P. Melin, and O. Castillo, “High order α-planes integration:
In this paper, a new interval type-3 fuzzy system is presented. A new approach to computational cost reduction of general type-2 fuzzy
The proposed fuzzy system has more approximation capabil- systems,” Eng. Appl. Artif. Intell., vol. 74, pp. 186–197, 2018.
[18] H. Hassani, J. Zarei, M. M. Arefi, and R. Razavi-Far, “zSlices-based
ity than conventional T2FSs. Because the membership in the general type-2 fuzzy fusion of support vector machines with application
proposed fuzzy set is an interval type-2 fuzzy set not a type-1 to bearing fault detection,” IEEE Trans. Ind. Electron., vol. 64, no. 9,
fuzzy set. Furthermore, an online stable fractional-order learning pp. 7210–7217, Sep. 2017.
[19] J. Andreu-Perez, F. Cao, H. Hagras, and G.-Z. Yang, “A self-adaptive
algorithm is presented for tuning of the parameters of the pro- online brain–machine interface of a humanoid robot through a general
posed fuzzy system. The simulation and the experimental studies type-2 fuzzy inference system,” IEEE Trans. Fuzzy Syst., vol. 26, no. 1,
show that the proposed fuzzy system and learning algorithm pp. 101–116, Feb. 2018.
[20] L. A. T. Salomao, M. Mahfouf, E. El-Samahy, and C.-H. Ting, “Psy-
with less rule and leaning iteration result in significantly better chophysiologically based real-time adaptive general type 2 fuzzy mod-
performance in contrast to the other well-known fuzzy neural eling and self-organizing control of operator’s performance undertaking
networks and learning algorithms. In this paper, to decrease a cognitive task,” IEEE Trans. Fuzzy Syst., vol. 25, no. 1, pp. 43–57,
Feb. 2017.
the computational cost, only the consequent parameters are ad- [21] M. Almaraashi, R. John, A. Hopgood, and S. Ahmadi, “Learning of interval
justed. For future studies, the optimality of the rule database and and general type-2 fuzzy logic systems using simulated annealing: Theory
antecedent parameters and also decreasing of the computational and practice,” Inf. Sci., vol. 360, pp. 21–42, 2016.
[22] S. S. Pal and S. Kar, “A hybridized forecasting method based on
cost can be investigated. weight adjustment of neural network using generalized type-2 fuzzy
set,” Int. J. Fuzzy Syst., vol. 21, pp. 1–13, 2018. [Online]. Available:
REFERENCES https://doi.org/10.1007/s40815-018-0534-z
[23] F. Baghbani, M.-R. Akbarzadeh-T, and A. Akbarzadeh, “Indirect adaptive
[1] C. W. De Silva, Intelligent Control: Fuzzy Logic Applications. Boca Raton, robust mixed h2/H∞ general type-2 fuzzy control of uncertain nonlinear
FL, USA: CRC Press, 2018. systems,” Appl. Soft Comput., vol. 72, pp. 392–418, 2018.
[2] F. Li, P. Shi, C.-C. Lim, and L. Wu, “Fault detection filtering for nonho- [24] S. Bahrebar et al., “A novel type-2 fuzzy logic for improved risk anal-
mogeneous Markovian jump systems via a fuzzy approach,” IEEE Trans. ysis of proton exchange membrane fuel cells in marine power systems
Fuzzy Syst., vol. 26, no. 1, pp. 131–141, Feb. 2018. application,” Energies, vol. 11, no. 4, 2018, Art. no. 721.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.
1950 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 28, NO. 9, SEPTEMBER 2020

[25] J. T. Rickard, J. Aisbett, and G. Gibbon, “Fuzzy subsethood for fuzzy sets [46] C.-F. Juang and C.-T. Lin, “An online self-constructing neural fuzzy
of type-2 and generalized type-n,” IEEE Trans. Fuzzy Syst., vol. 17, no. 1, inference network and its applications,” IEEE Trans. Fuzzy Syst., vol. 6,
pp. 50–60, Feb. 2009. no. 1, pp. 12–32, Feb. 1998.
[26] M. F. Zarandi, R. Gamasaee, and O. Castillo, “Type-1 to type-n fuzzy logic [47] C.-F. Juang and Y.-W. Tsao, “A self-evolving interval type-2 fuzzy neural
and systems,” in Fuzzy Logic in Its 50th Year. Berlin, Germany: Springer, network with online structure and parameter learning,” IEEE Trans. Fuzzy
2016, pp. 129–157. Syst., vol. 16, no. 6, pp. 1411–1424, Dec. 2008.
[27] O. Castillo, “Towards finding the optimal n in designing type-n fuzzy [48] A. Staiano, R. Tagliaferri, and W. Pedrycz, “Improving RBF networks per-
systems for particular classes of problems: A review,” Appl. Comput. formance in regression tasks by means of a supervised fuzzy clustering,”
Math., vol. 17, no. 1, pp. 3–9, 2018. Neurocomputing, vol. 69, no. 13–15, pp. 1570–1581, 2006.
[28] I. B. Türkşen, “Type 2 representation and reasoning for CWW,” Fuzzy Sets
Syst., vol. 127, no. 1, pp. 17–36, 2002.
[29] V. Torra and Y. Narukawa, “On hesitant fuzzy sets and decision,” in Proc.
IEEE Int. Conf. Fuzzy Syst., 2009, pp. 1378–1382.
[30] B. Zhu, Z. Xu, and M. Xia, “Dual hesitant fuzzy sets,” J. Appl. Math.,
vol. 2012, 2012, Art. no. 879629.
Ardashir Mohammadzadeh was born in Azerbai-
[31] E. A. Sezer, B. Pradhan, and C. Gokceoglu, “Manifestation of an adaptive
jan, Iran. He received the B.S. degree from the Sahand
neuro-fuzzy model on landslide susceptibility mapping: Klang valley,
Malaysia,” Expert Syst. With Appl., vol. 38, no. 7, pp. 8208–8219, 2011. University of Technology, Tabriz, Iran, in 2011, the
M.S. degree from the K. N. Toosi University of Tech-
[32] J.-S. Jang, “ANFIS: Adaptive-network-based fuzzy inference system,”
nology, Tehran, Iran, in 2013, and the Ph.D. degree
IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685,
from the University of Tabriz, Tabriz, in 2016, all in
May/Jun. 1993.
control engineering.
[33] R. Xu, D. Li, and J. Jiang, “Online learning fuzzy vibration control of
He is currently an Assistant Professor with the Uni-
smart truss structure,” Proc. Inst. Mech. Engineers, Part G: J. Aerosp.
versity of Bonab, Bonab, Iran. His research interests
Eng., vol. 231, no. 3, pp. 548–557, 2017.
[34] E. Kayacan, E. Kayacan, and M. A. Khanesar, “Identification of nonlinear include soft computing, intelligent control systems,
chaotic systems, and fuzzy systems.
dynamic systems using type-2 fuzzy neural networks—A novel learning
algorithm and a comparative study,” IEEE Trans. Ind. Electron., vol. 62,
no. 3, pp. 1716–1724, Mar. 2015.
[35] C.-F. Juang and K.-J. Juang, “Circuit implementation of data-driven
TSK-type interval type-2 neural fuzzy system with online parameter
tuning ability,” IEEE Trans. Ind. Electron., vol. 64, no. 5, pp. 4266–4275,
May 2017.
Mohammad Hosein Sabzalian was born in Iran.
[36] T. Kumbasar and H. Hagras, “A gradient descent based online tuning
He is currently working toward the Ph.D. degree
mechanism for PI type single input interval type-2 fuzzy logic controllers,” in adaptive fuzzy control of nonlinear systems with
in Proc. IEEE Int. Conf. Fuzzy Syst., 2015, pp. 1–6.
Shanghai Jiaotong University, Shanghai, China. His
[37] C.-T. Lin, N. R. Pal, S.-L. Wu, Y.-T. Liu, and Y.-Y. Lin, “An interval
research is supervised by Professor Weidong Zhang.
type-2 neural fuzzy system for online system identification and feature
His main research interests are in intelligent con-
elimination,” IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 7,
trol, adaptive control, smart grids and its automation.
pp. 1442–1455, Jul. 2015.
[38] D. Sun, Q. Liao, and H. Ren, “Type-2 fuzzy modeling and control for
bilateral teleoperation system with dynamic uncertainties and time-varying
delays,” IEEE Trans. Ind. Electron., vol. 65, no. 1, pp. 447–459, Jan. 2018.
[39] P. Melin and D. Sánchez, “Optimization of type-1, interval type-2 and
general type-2 fuzzy inference systems using a hierarchical genetic algo-
rithm for modular granular neural networks,” Granular Comput., vol. 4,
pp. 211–236, 2019.
[40] A. Mohammadzadeh, S. Ghaemi, O. Kaynak, and S. Khanmohammadi, Weidong Zhang received the B.S., M.S., and Ph.D.
“Observer-based method for synchronization of uncertain fractional order degrees in control science and engineering from Zhe-
chaotic systems by the use of a general type-2 fuzzy system,” Appl. Soft jiang University, Hangzhou, China, in 1990, 1993,
Comput., vol. 49, pp. 544–560, 2016. and 1996, respectively.
[41] M. A. Sanchez, J. R. Castro, V. Ocegueda-Miramontes, and L. Cervantes, He joined Shanghai Jiao Tong University, Shang-
“Hybrid learning for general type-2 TSK fuzzy logic systems,” Algorithms, hai, China, in 1998, as an Associate Professor, where
vol. 10, no. 3, p. 99, 2017. he has been a Full Professor since 1999. From 2003
[42] J. M. Mendel, R. I. John, and F. Liu, “Interval type-2 fuzzy logic to 2004, he was with the University of Stuttgart,
systems made simple,” IEEE Trans. Fuzzy Syst., vol. 14, no. 6, Stuttgart, Germany, as an Alexander von Humboldt
pp. 808–821, Dec. 2006. [Online]. Available: https://ieeexplore.ieee.org/ Fellow. In 2011, he was appointed as a Chair Professor
abstract/document/4016089 with Shanghai Jiao Tong University. He is currently
[43] A. Mohammadzadeh and S. Ghaemi, “Synchronization of chaotic systems the Director of the Engineering Research Center of Marine Automation, Shang-
and identification of nonlinear systems by using recurrent hierarchical hai Municipal Education Commission and the Deputy Dean with the Department
type-2 fuzzy neural networks,” ISA Trans., vol. 58, pp. 318–329, 2015. of Automation, Shanghai Jiao Tong University. He has authored one book and
[44] C.-F. Juang, R.-B. Huang, and W.-Y. Cheng, “An interval type-2 fuzzy- more than 300 refereed papers, and holds 51 patents. His research interests
neural network with support-vector regression for noisy regression prob- include control theory and information processing theory and their applications
lems,” IEEE Trans. Fuzzy Syst., vol. 18, no. 4, pp. 686–699, Aug. 2010. in several fields, including power/chemical processes, USV/ROV, and aerocraft.
[45] S. W. Tung, C. Quek, and C. Guan, “eT2FIS: An evolving type-2 neural Prof. Zhang was a recipient of the National Science Fund for Distinguished
fuzzy inference system,” Inf. Sci., vol. 220, pp. 124–148, 2013. Young Scholars of China and Shanghai Subject Chief Scientist.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on March 20,2024 at 08:22:53 UTC from IEEE Xplore. Restrictions apply.

You might also like