542 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 41, NO.

2, APRIL 2011

Robust Adaptive Controller Design for a Class of

Uncertain Nonlinear Systems Using Online
T–S Fuzzy-Neural Modeling Approach
Yi-Hsing Chien, Wei-Yen Wang, Senior Member, IEEE, Yih-Guang Leu, and Tsu-Tian Lee, Fellow, IEEE

Abstract—This paper proposes a novel method of online model- Moreover, Takagi–Sugeno (T–S) fuzzy systems are nonlinear
ing and control via the Takagi–Sugeno (T–S) fuzzy-neural model systems described by a set of if–then rules. Such a model can
for a class of uncertain nonlinear systems with some kinds of approximate a wide class of nonlinear systems. In [3] and [4],
outputs. Although studies about adaptive T–S fuzzy-neural con-
trollers have been made on some nonaffine nonlinear systems, little the authors proved that the T–S fuzzy system can approximate
is known about the more complicated uncertain nonlinear systems. any continuous function to any precision.
Because the nonlinear functions of the systems are uncertain, By using well-known offline tuning algorithms for unknown
traditional T–S fuzzy control methods can model and control them nonlinear systems, an initial fuzzy-neural model with adjustable
only with great difficulty, if at all. Instead of modeling these uncer- parameters can be constructed. However, the derived fuzzy-
tain functions directly, we propose that a T–S fuzzy-neural model
approximates a so-called virtual linearized system (VLS) of the neural model with offline-tuned parameters cannot cope with
system, which includes modeling errors and external disturbances. parameter changes arising from external disturbances [5]. Thus,
We also propose an online identification algorithm for the VLS and offline algorithms cannot be applied to situations where real-
put significant emphasis on robust tracking controller design using time processing is required, such as adaptive control and sig-
an adaptive scheme for the uncertain systems. Moreover, the sta- nal processing. In these situations, the adjustable parameters
bility of the closed-loop systems is proven by using strictly positive
real Lyapunov theory. The proposed overall scheme guarantees must be tuned online during operation to compensate for
that the outputs of the closed-loop systems asymptotically track undesirable effects. The objective of adaptive control is to
the desired output trajectories. To illustrate the effectiveness and maintain consistent performance of a system in the presence
applicability of the proposed method, simulation results are given of uncertainties. Further issues are the stability analysis and
in this paper. controller design of T–S fuzzy-neural controlled systems [6],
Index Terms—Fuzzy-neural model, online modeling, robust [7]. These have been extensively investigated in the literature.
adaptive control, uncertain nonlinear systems. The existence of a common positive definite matrix for a set
of Lyapunov inequalities is a sufficient condition for stabi-
I. I NTRODUCTION lization [8]–[10]. However, this is very difficult to achieve
using an online approach, even when using the well-known
T HE STUDY of controller design for some physical sys-
tems has a long-standing history. Most physical systems
are described by a set of differential equations. Research has
linear matrix inequality (LMI) method [8]–[10]. Therefore, in
this paper, adaptive schemes are used for simultaneous online
focused on the development of various design techniques for modeling and controller design, instead of offline modeling. An
controllers of these systems. The existence of a mathematical additional benefit is that the stability analysis of the adaptive
model of the system is assumed for model-based control. T–S fuzzy-neural controlled systems is easier than that of the
Controllers are designed to modify the behavior of the system LMI method.
and achieve some desired performance [1]. For this purpose, a Tracking control designs for unknown nonlinear systems are
systematic way to construct a model mapping the inputs to the important issues for practical applications. In [11]–[13], the au-
outputs is needed. Fuzzy models are usually used in the case thors only consider the stabilization problem for affine systems.
where the model structure and parameters are unknown [2]. Moreover, theoretical justification development presented in
[14]–[17] is valid only for single-input–single-output nonlinear
Manuscript received August 7, 2009; revised February 9, 2010 and
systems and therefore is hardly practical in real applications
July 13, 2010; accepted July 18, 2010. Date of publication September 20, such as the trajectory control of robot manipulators and space
2010; date of current version March 16, 2011. This work was supported by vehicles. Although Hwang and Hu [18] proposed a robust
the National Science Council, Taiwan, under Grant NSC 96-2221-E-027-116.
This paper was recommended by Associate Editor E. Santos, Jr.
neural learning controller for multiple-input–multiple-output
Y.-H. Chien and T.-T. Lee are with the Department of Electrical Engineer- (MIMO) manipulators, the state feedback control scheme does
ing, National Taipei University of Technology, Taipei 106, Taiwan (e-mail: not always hold in practical applications, because models of
[email protected]).
W.-Y. Wang and Y.-G. Leu are with the Department of Applied Electronics those systems are not always known. Furthermore, an adaptive
Technology, National Taiwan Normal University, Taipei 10610, Taiwan (e-mail: controller design [19], [20] for a class of MIMO uncertain non-
[email protected]). linear systems was proposed only for a simple case of system
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. outputs. It is therefore the objective of this paper to develop
Digital Object Identifier 10.1109/TSMCB.2010.2065801 an online adaptive T–S fuzzy-neural modeling approach for the

1083-4419/$26.00 © 2010 IEEE

CHIEN et al.: CONTROLLER DESIGN FOR A CLASS OF UNCERTAIN NONLINEAR SYSTEMS 543

Fig. 1. Configuration of a typical fuzzy-neural network.

design of robust tracking controllers for the more complicated the subspace corresponding to the premise part [5]. The T–S
unknown systems. fuzzy-neural model is defined as
On the whole, this paper deals with the T–S fuzzy-neural
model because of its ability to approximate dynamic nonlin- R(i) : If z1 is F1i , . . . , zn is Fni , . . . , zn+m is Fn+m
i

ear systems [14], [21], [22]. Although studies about adaptive

Then ȳl = pil1 z1 + pil2 z2 + · · · + pil(n+m) zn+m (1)
T–S fuzzy-neural controllers have been made on some MIMO
nonlinear systems, little is known about more complicated
where z = [z1 , z2 , . . . , zn+m ]T ∈ n+m is a vector of lin-
nonlinear systems. Compared with the previous approaches
guistic variables, y l represents the output of the fuzzy-
[22]–[24], the main contribution of this paper is an investigation
neural network, Fji (j = 1, 2, . . . , n + m) are fuzzy sets, and
of the more complicated unknown systems using online adap-
tive T–S fuzzy-neural controllers. Because the nonlinear func- pilk (i = 1, 2, . . . , h, l = 1, 2, . . . , n, k = 1, 2, . . . , n + m) are
tions of the systems are unknown, traditional T–S fuzzy control adjustable parameters which are called the weighting factors.
methods can model and control them with great difficulty, if Fig. 1 shows the configuration of the T–S fuzzy-neural
at all. Instead of modeling these unknown functions directly, model [14], which is a typical T–S fuzzy inference system [21]
we propose that a T–S fuzzy-neural model approximates a so- constructed from a neural network structure. It has a total of six
called virtual linearized system (VLS) of the system, which layers. The T–S fuzzy-neural model is essentially a multimodel
includes modeling errors and external disturbances. Moreover, approach in which a set of linear models is combined to
we can prove that the closed-loop system controlled by the describe the global behavior of the system [8], [14]. Based
proposed controller is robust stable and that the effect of all on this idea, the T–S fuzzy-neural model is appropriate for
the modeling errors and external disturbances on the tracking developing fuzzy-neural controllers because many systems can
error can be attenuated. be expressed locally in some form of mathematical model.
The rest of this paper is organized as follows. Section II The T–S fuzzy-neural model can approximate a wide class of
introduces the T–S fuzzy-neural model for the VLS. Section III nonlinear systems. In [3] and [4], it was proved that the T–S
presents a controller design for online modeling and robust fuzzy-neural system can approximate any continuous function
tracking. In Section IV, simulation results are presented to to any precision.
confirm the effectiveness and applicability of the proposed Suppose that the MIMO unknown nonaffine nonlinear
method. Finally, conclusions are given in Section V. system is

ẋ1 = f1 (x, u) + de1

II. T–S F UZZY-N EURAL M ODEL FOR THE VLS
ẋ2 = f2 (x, u) + de2
The T–S fuzzy-neural model approximates a nonlinear sys- ..
tem with a combination of several linear systems. It is formed .
by fuzzy partitioning of the input space. The premise of a fuzzy
ẋn = fn (x, u) + den
implication indicates a fuzzy subspace of the input space, and
each consequent part expresses a local input–output relation in y = Cx (2)
544 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 41, NO. 2, APRIL 2011

.
where x = [x1 , x2 , . . . , xn ]T = [xu ..xs ]T ∈ n is a vector of different mean value points x∗i (i = 1, 2, . . . , n) and u∗i (i =
states, xu ∈ p is an uncertain state vector, xs ∈ n−p 1, 2, . . . , n) for different functions fi (i = 1, 2, . . . , n). Thus,
is a bounded state vector, where xs  < ∞, and u = we can transform the real system (2) into (3) by a VLS as
[u1 , u2 , . . . , um ]T ∈ m and y = [y1 , y2 , . . . , yg ]T ∈ g (g ≤ follows:
m ≤ p) are the vectors of the control inputs and system outputs,
respectively. C ∈ g×n is a constant matrix demand that CCT ⎡ ⎤
a11 a12 ... a1n
is nonsingular and can be represented as ⎢ a21 a22 ... a2n ⎥
⎡ ⎤ ẋ = ⎢
⎣ ... .. .. ⎥ xξ
. ⎦
..
. . .
C1 .. C2
C = ⎣ · · · · · · · · · ⎦ ∈ g×n an1 an2 . . . ann
⎡ ⎤
C3 b11 b12 . . . b1m
⎢ b21 b22 . . . b2m ⎥
where C1 ∈ p×p ; we define cij ∈ C1 ; then, cij = 0 if i = +⎢
⎣ ... .. .. ⎥ uξ
. ⎦
..
j, and cij = 0 if i = j. Moreover, C2 ∈ p×(n−p) , C3 ∈ . .
(g−p)×n , and de = [de1 , de2 , . . . , den ]T represents external bn1 bn2 . . . bnm
disturbances. fi : n+m → 1 (i = 1, 2, . . . , n) are unknown ⎡ f (x̄, ū) + d ⎤
1 e1
functions whose first derivatives with respect to x and u exist. ⎢ f2 (x̄, ū) + de2 ⎥
Without loss of generality, we assume that solutions for (2) +⎢
⎣ .. ⎥
⎦
exist. .
Remark 1: In (2), we consider the uncertain state vector xu , fn (x̄, ū) + den
in which states may be bounded or unbounded, and the bounded = Axξ + Buξ + dd (3)
state vector xs , in which states are bounded. C1 is defined such
that the ith output yi of the system is composed of only the ith
where dd = [dd1 , dd2 , . . . , ddn ]T with ddi = fi (x, u) +
uncertain state xi ∈ xu , with the rest of yi ’s value coming from
dei (i = 1, 2, . . . , n), x = [x1 , x2 , . . . , xn ]T = t1 x and
the bounded state vector xs . In other words, xi ∈ xu only con- u = [u1 , u2 , . . . , um ]T = t2 u with 0 < t1 , t2 < 1 are the
tributes to yi ’s value. For example, yi = cii xi + nj=p+1 cij xj ,
vectors of critical points, xξ = [xξ1 , xξ2 , . . . , xξn ]T = x − x,
where 1 ≤ i ≤ g, and xi ∈ xu . System (2), which we consider
uξ = [uξ1 , uξ2 , . . . , uξm ]T = u − u, aij = ∂fi (x∗i , u∗i )/∂xj ,
in this paper, covers systems which have been published in
and bik = ∂fi (x∗i , u∗i )/∂uk , i = 1, 2, . . . , n, j = 1, 2, . . . , n,
many related papers. Therefore, system (2) is a class of more
k = 1, 2, . . . , m. We can choose the parameters of t1 and t2 to
complicated uncertain nonlinear systems.
find the values of x and u.
Definition 1 [25]: Suppose that a function f is continuous
Remark 2: The VLS models the unknown nonlinear system
in the closed interval [x, x] and differentiable in the interval’s
(2). Because the nonlinear functions of the systems (2) are
interior (x, x), where x = t1 x, 0 < t1 < 1. Then, for some x∗
unknown, traditional T–S fuzzy control methods can rarely
between (x, x), we have f (x∗ ) = (f (x) − f (x))/(x − x). We
model and control them. Instead of modeling the unknown
call x a critical point and x∗ a differential mean point of f on
systems (2) directly, the T–S fuzzy-neural model in (1) (or
(x, x).
Fig. 1) is used to approximate the VLS in (3) which is used
By using the mean value theorem [25], there are
to model the unknown nonlinear system (2).
points x∗ij (i = 1, 2, . . . , n, j = 1, 2, . . . , n) and u∗ik (i =
From (1) and Fig. 1, the coefficient plk (l = 1, 2, . . . , n, k =
1, 2, . . . , n, k = 1, 2, . . . , m) on the linear segments joining
1, 2, . . . , n + m) of the T–S fuzzy-neural model is
xj to xj (j = 1, 2, . . . , n) and uk to uk (k = 1, 2, . . . , m)
for every function fi (i = 1, 2, . . . , n). Therefore, the un-
known nonaffine nonlinear functions fi (i = 1, 2, . . . , n) can h n+m
i=1 pilk μFji (zj )
be formed as follows: j=1
plk = (4)
∂fi (x∗i , u∗i ) 
h
n+m
fi (x, u) = fi (x̄, ū) + (x1 − x̄1 ) j=1 μFji (zj )
∂x1 i=1

∂fi (x∗i , u∗i )

+ (x2 − x̄2 ) + · · · where μFji (zj ) is the value of the membership function. For the
∂x2
tuning of the weighting factors pilk , we define
∂fi (x∗i , u∗i )
+ (xn − x̄n )
∂xn n+m
∂fi (x∗i , u∗i ) ∂fi (x∗i , u∗i ) μFji (zj )
+ (u1 − ū1 ) + (u2 − ū2 ) w ≡
i j=1
, i = 1, 2, . . . , h. (5)
∂u1 ∂u2

h n+m
∂fi (x∗i , u∗i ) μFji (zj )
+ ··· + (um − ūm ) i=1 j=1
∂um

where x∗i = [x∗i1 , x∗i2 , . . . , x∗in ]T and u∗i = [u∗i1 , u∗i2 , . . . , u∗im ]T The antecedent part of the fuzzy implication describes
are the mean points of fi (i = 1, 2, . . . , n). Note that we can get the conditions of the state deviations and input deviations
CHIEN et al.: CONTROLLER DESIGN FOR A CLASS OF UNCERTAIN NONLINEAR SYSTEMS 545

[xTξ , uTξ ]T . The consequent part of the fuzzy implication rep- Assumption 1 [14], [22]: Let xξ and uξ belong to the
resents the VLS in (3). For the purpose of approximating the compact sets Ux and Uu , respectively, where
VLS in (3), the ith fuzzy implication (1) can be described as
Ux = {x ∈ n : x ≤ mx < ∞}

R(i) : If xξ1 is F1i , . . . , xξn is Fni and uξ1 is Fn+1

i
, Uu = {u ∈ m : u ≤ mu < ∞}
i
. . . , uξm is Fn+m and mx and mu are design parameters. We define φlj =
[p1lj , p2lj , . . . , phlj ], l = 1, 2, . . . , n, j = 1, 2, . . . , n + m. It is
Then x̂˙ = Âi xξ + B̂i uξ (6) known that the optimal adjustable parameters φ∗lj lie in some
convex regions
where  
Mφlj = φlj ∈ h : φlj  ≤ mφlj ,
⎡ pi pi12 ··· pi1n ⎤ l = 1, 2, . . . , n; j = 1, 2, . . . , n + m
11
⎢ pi21 pi22 ... pi2n ⎥
Âi = ⎢
⎣ .. .. .. .. ⎥
⎦ where the radii mφlj are constant and
. . . .
pin1 pin2 ... pinn φ∗lj = arg min
φlj ∈Mφlj
⎡ pi pi1(n+2) ··· pi1(n+m) ⎤  
1(n+1)  i 
⎢ pi2(n+1) pi2(n+2) ... pi2(n+m) ⎥ × sup plj (xξ , uξ ) − p̂ilj (xξ , uξ |φlj ) ,
⎢ ⎥
B̂i = ⎢ .. .. .. .. ⎥. (7) xξ ∈Ux ,uξ ∈Uu
⎣ . . . . ⎦
pin(n+1) pin(n+2) ... i
pn(n+m) l = 1, 2, . . . , n; j = 1, 2, . . . , n + m.

According to Assumption 1, we define the optimal adjustable

After applying (4), (5), and some commonly used defuzzifi- matrices as
cation strategies, the real system (2) or the VLS (3) becomes ⎡ p∗i p∗i · · · p∗i ⎤
11 12 1n
⎢ p∗i p∗i ... p∗i
2n ⎥

h
Â∗i = ⎢
21 22
.. ⎥
ẋ = x̂˙ + dd + df = wi {Âi xξ + B̂i uξ } + dd + df ⎣ .. .. ..
. ⎦
. . .
i=1 p∗i
n1 p∗i
n2 ... ∗i
pnn
⎡ ⎤ ⎡ p∗i ⎤
p11 p12 ··· p1n p∗i ··· p∗i
1(n+1) 1(n+2) 1(n+m)
⎢ p21 p22 ... p2n ⎥ ⎢ p∗i p∗i p∗i ⎥
=⎢
⎣ .. .. .. .. ⎥
⎦ xξ ⎢ 2(n+1) 2(n+2) ... 2(n+m) ⎥
. . . . B̂∗i = ⎢ .. .. .. .. ⎥.
⎣ . . . . ⎦
pn1 pn2 . . . pnn
⎡ ⎤ p∗i
n(n+1) p∗i
n(n+2) ... ∗i
pn(n+m)
p1(n+1) p1(n+2) ··· p1(n+m)
⎢ p2(n+1) p2(n+2) ... p2(n+m) ⎥ Lemma 1 [26]: Suppose that a matrix Λ ∈ g×g is given.
⎢ ⎥
+⎢ .. .. .. .. ⎥ uξ + dd + df For every symmetric positive definite matrix Q ∈ g×g , the
⎣ . . . . ⎦
Lyapunov matrix equation ΛT Γ + ΓΛ = −Q has a unique
pn(n+1) pn(n+2) ... pn(n+m) solution for Γ = ΓT > 0 if and only if Λ ∈ g×g is a Hurwitz
y = Cx (8) matrix.
We define r = [r1 , r2 , . . . , rg ]T , where rj (j = 1, 2, . . . , g)
  are bounded reference signals. The tracking error vector is e =
where df = (A − hi=1 wi Âi )xξ + (B − hi=1 wi B̂i )uξ
y − r = [e1 , e2 , . . . , eg ]T . A coefficient matrix is defined as
and pij (i = 1, 2, . . . , n, j = 1, 2, . . . , n + m) is used to
⎡ ⎤
approximate aij (i = 1, 2, . . . , n, j = 1, 2, . . . , n) and −λ1 0 ··· 0
bij (i = 1, 2, . . . , n, j = 1, 2, . . . , m) of the VLS in (3) ⎢ 0 −λ2 · · · 0 ⎥
Λ=⎢ ⎣ ... .. .. ⎥ (9)
. ⎦
which models the unknown nonlinear system (2). In this ..
. .
paper, (3) is a state equation, and (8) is an approximate state 0 0 ··· −λg
equation. x̂ is the estimation of the state vector x. Considering
approximation error df , (3), and (6), we can obtain (8) where the coefficients λ1 , λ2 , . . . , λg are selected such that the
ẋ = x̂˙ + dd + df . matrix Λ is a Hurwitz matrix. From (3), based on the certainty
equivalence approach, a control input can be chosen as
 −1
III. C ONTROLLER D ESIGN FOR O NLINE M ODELING AND uξ = (CB)T (CB)(CB)T (−CAxξ + ω − Cus ) (10)
ROBUST T RACKING
where ω = [ω1 , ω2 , . . . , ωg ]T = ṙ + Λe and us is an error
To design a robust controller for (2), the following assump- compensator which is designed to compensate for dd . However,
tions are required. the right side of (3) is unknown, so we replace A and B by
546 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 41, NO. 2, APRIL 2011

h   
wi Âi and hi=1 wi B̂i in (8), respectively. Moreover, we 
N
 k
i=1  =Ω T
α −1 −1
α (−ΩΩ + αI) T
define Ω = C hi=1 wi B̂i . The inverse of the square matrix
k=0
ΩΩT can be derived as
 −1 
h

(ΩΩ ) T −1
= αI − (−ΩΩ + αI) T × −C wi Âi xξ + ṙ + Λe − Cus + uΔ + ū
i=1
= α−1 I + α−2 (−ΩΩT + αI) + α−3 (−ΩΩT + αI)2 (13)
+ α−4 (−ΩΩT + αI)3 + · · ·
where uΔ is designed to estimate dΔ .

N
 k Remark 3: We consider two cases based on the determinant
= α−1 α−1 (−ΩΩT + αI) + dΩ (11)
of (ΩΩT ). If |ΩΩT | > ε, (12) can be calculated. However, if
k=0
|ΩΩT | ≤ ε, we use uΔ to approximate dΔ . Then, the con-
where α is required to be sufficiently large so that all eigenval- troller u becomes (13). In the second case, there are many
ues of α−1 (−ΩΩ T
∞ + αI) have magnitudes less than one [27] projection algorithms [28], [29] which can be applied to solve
−1 −1 T the parameter drift problem. The principal idea behind such
and dΩ = α k=N +1 (α (−ΩΩ + αI))k . From (10) and
(11), we redesign the fuzzy-neural control input as approaches is to project the directions of adaptations (i.e.,
˙ ˙
  Âi , B̂i , q̇k , q̇Δj ), whenever they have the tendency to move
N
 −1 k into or stay at |ΩΩT | < ε, so that the determinant of (ΩΩT )
T −1 T
u =Ω α α (−ΩΩ + αI) can become larger than or equal to ε.
k=0
From ė = ẏ − ṙ and substituting (12) and (13) for (8), the

h error dynamic equation of the VLS becomes
× −C wi Âi xξ + ṙ + Λe − Cus
i=1

+ (1 − β)dΔ + βuΔ + ū ė = Cẋ − ṙ

= Cx̂˙ + Cdd + Cdf − ṙ

where

h 
h

h
=C wi Âi xξ + C wi B̂i uξ
dΔ = ΩT dΩ −C wi Âi xξ + ṙ + Λe − Cus i=1 i=1
i=1

h
= [dΔ1 , dΔ2 , . . . , dΔm ]T . + Cdd + CA − C wi Â∗i xξ
i=1
If |ΩΩT | > ε, we set β = 0. The fuzzy-neural control input
becomes 
h 
h
+ CB − C wi B̂∗i uξ + C wi (Â∗i − Âi )xξ
 

N
 −1 k i=1 i=1
T −1 T
u =Ω α α (−ΩΩ + αI)

h
k=0
+C wi (B̂∗i − B̂i )uξ − ṙ

h i=1
× −C wi Âi xξ + ṙ + Λe − Cus + dΔ + ū
i=1 
h 
h
=C wi Âi xξ + −C wi Âi xξ + ṙ + Λe − Cus

h
i=1 i=1
= ΩT (ΩΩT )−1 −C wi Âi xξ + ṙ + Λe − Cus + ū
i=1 
h 
h
+C wi Â∗i − Âi xξ + C wi (B̂∗i − B̂i )uξ
(12)
i=1 i=1

where ε is a small positive constant. 

h

If |ΩΩT | ≤ ε, we choose β = 1. The fuzzy-neural control + Cdd + CA − C wi Â∗i xξ

input becomes i=1


h

N
 k + CB − C wi B̂∗i uξ − ṙ − βΩdΔ + βΩuΔ
u= ΩT α−1 α−1 (−ΩΩT + αI)
i=1
k=0


h 
h 
h

× −C w  xξ + ṙ + Λe − Cus
i i
+ dΔ = Λe + C wi Ãi xξ + C wi B̃i uξ
i=1 i=1
i=1

+ (uΔ − dΔ ) + ū + Cd̃ − Cus − βΩdΔ + βΩuΔ (14)

CHIEN et al.: CONTROLLER DESIGN FOR A CLASS OF UNCERTAIN NONLINEAR SYSTEMS 547

where Ãi = Â∗i − Âi , B̃i = B̂∗i − B̂i , and d̃ = dd +(A− Moreover, we define uΔ and eΔ as
h h
i ∗i
i=1 w  )xξ +(B−
i ∗i ˜ ˜ ˜ T
i=1 w B̂ )uξ = [d1 , d2 , . . . , dn ] .
⎡ sign(ē ) 0 ··· 0 ⎤
Δ1
We define us (the error compensator) and eΔ as ⎢ 0 sign(ē ) · ·· 0 ⎥
Δ2
⎡ sign(e ) 0 ··· 0 ⎤ uΔ = ⎢ ⎣ . . . .. ⎥ d̂Δ
⎦
Δ1 .. .. .. .
⎢ 0 sign(eΔ2 ) ··· 0 ⎥ ···
us = ⎢
⎣ .. .. .. .. ⎥ d̂
⎦
0 0 sign(ēΔm )
. . . .
= Diag [sign(ēΔ )] d̂Δ (20)
0 0 ··· sign(eΔn )
ēΔ = eT ΓΩ = [ēΔ1 ēΔ2 ··· ēΔm ] (21)
= Diag [sign(eΔ )] d̂ (15)

eΔ = eT ΓC = [eΔ1 eΔ2 ··· eΔn ] (16) where d̂Δ = [dˆΔ1 , dˆΔ2 , . . . , dˆΔm ]T and Γ > 0 is a Lyapunov
matrix. The fuzzy implications are defined to obtain d̂Δ , which
where d̂ = [dˆ1 , dˆ2 , . . . , dˆn ]T and Γ > 0 is a Lyapunov matrix. is the estimate of dΔ , as follows:
The fuzzy implications are defined to obtain d̂, which is the
estimate of d̃, as follows: If ēΔ1 is FēiΔ1 and x is Fx
i
Then dˆΔ1 = qΔ1
i

If eΔ1 is FeiΔ1 and x is Fx

i
Then dˆ1 = q1i If ēΔ2 is FēiΔ2 and x is Fx
i
Then dˆΔ2 = qΔ2
i

If eΔ2 is FeiΔ2 and x is Fx

i
Then dˆ2 = q2i ..
.
.. If ēΔm is FēiΔm and x is Fx
i
Then dˆΔm = qΔm
i
. (22)
.

If eΔn is FeiΔn and x is Fx

i
Then dˆn = qni . (17) After applying some commonly used defuzzification strategies,
we have
After applying some commonly used defuzzification strategies,
hd̂
we have Δ i
qΔj μFēi (ēΔj )μF i (x)
Δj x

hd̂
dˆΔj = i=1
qki μFei (eΔk )μF i (x) hd̂
Δ
Δk x
dˆk = i=1 μFēi (ēΔj )μF i (x)
x

hd̂ Δj
i=1
μFei (eΔk )μF i (x)
Δk x hd̂
i=1 Δ
i i
= τΔj qΔj , j = 1, 2, . . . , m (23)

hd̂
i=1
= τki qki , k = 1, 2, . . . , n (18)
i=1 where
where  hd̂
T
 T
1
qΔj = qΔj 2
qΔj ··· qΔjΔ , j = 1, 2, . . . , m
h
qk = qk1 qk2 ··· qk d̂ , k = 1, 2, . . . , n  T
hd̂
 T
1
τ Δj = τΔj 2
τΔj ··· τΔjΔ
h
τ k = τk1 τk2 ··· τk d̂ ⎡
⎡
⎢
⎢ μFē1 (ēΔj )μF 1 (x)
⎢ μFe1 (eΔk )μF 1 (x) =⎢ Δj x
⎢ x ⎢ hd̂
=⎢h Δk
⎣ Δ
⎣d̂ μFēi (ēΔj )μF i (x)
μFei (eΔk )μF i (x) i=1 Δj x
Δk x
i=1
μFē2 (ēΔj )μF 2 (x)
μ 2 (eΔk )μ 2 (x) x
Fe F x
× Δj

× Δk hd̂
Δ

hd̂
μFēi (ēΔj )μF i (x)
μFei (eΔk )μF i (x) i=1 Δj x
Δk x
i=1
⎤T
⎤T
μ h (ēΔj )μ h (x) ⎥
μ h (eΔk )μ h (x) ⎥ Fē
d̂Δ d̂
FxΔ ⎥
d̂
FeΔk d̂
Fx ⎥ ··· h Δj ⎥ ,
··· ⎥ , ⎥

hd̂ ⎦ d̂Δ
⎦
μFei (eΔk )μF i (x) μFēi (ēΔj )μF i (x)
Δk x Δj x
i=1 i=1

k = 1, 2, . . . , n. (19) j = 1, 2, . . . , m. (24)
548 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 41, NO. 2, APRIL 2011

 
Assumption 2 [14], [22]: |d˜k | ≤ τ Tk q∗k (k = 1, 2, . . . , n), 1 
h
˙ i ) + eT ΓCd̃ − eT ΓCu
and |dΔj | ≤ τ TΔj q∗Δj (j = 1, 2, . . . , m), where q∗k and q∗Δj are + tr (CB̃i )T Γ(CB̃ s
η2 i=1
the optimal adjustable vectors and qk and qΔj represent the
1  T
n
estimates of q∗k and q∗Δj , respectively.
− βe ΓΩdΔ + βe ΓΩuΔ −
T T
q̃ q̇i
On the basis of the aforementioned discussion, the following γ1 i=1 i
theorem can be obtained.
β  T
m
Theorem 1: Consider the MIMO unknown nonaffine nonlin- − q̃ q̇Δj . (31)
ear system (2), which is approximated as (8). If the controllers γ2 j=1 Δj
are designed as (12) and (13) with update laws
From Lemma 1, substituting ΛT Γ + ΓΛ = −Q into (31), we
˙i
 = η1 wi CT (CCT )−1 exTξ , i = 1, 2, . . . , h (25) have
i
˙  
B̂ = η2 wi CT (CCT )−1 euTξ ,
h h
i = 1, 2, . . . , h (26) 1
v̇ = − eT Qe + eT ΓC wi Ãi xξ + eT ΓC wi B̃i uξ
q̇k = γ1 τ k eΔk , k = 1, 2, . . . , n (27) 2 i=1 i=1
q̇Δj = − γ2 τ Δj ēΔj , j = 1, 2, . . . , m (28)
 
1 
h
+ ˙ i)
tr (CÃi )T Γ(CÃ
where η1 , η2 , γ1 , and γ2 are positive constants, then the closed- η1 i=1
loop system is robust stable, and limt→∞ e(t) = 0.
Consider the Lyapunov-like function candidate  
1 
h
+ ˙ i)
tr (CB̃i )T Γ(CB̃
1   
h η2 i=1
1
v = eT Γe + tr (CÃi )T Γ(CÃi )
2 2η1 i=1
+ eT ΓCd̃ − eT ΓCus − βeT ΓΩdΔ + βeT ΓΩuΔ
1 
h  
1  T β  T
n m
+ tr (CB̃i )T Γ(CB̃i )
2η2 − q̃i q̇i − q̃ q̇Δj . (32)
i=1 γ1 i=1 γ2 j=1 Δj
1 n
+ (q∗i − qi )T (q∗i − qi )
2γ1 i=1 When the determinant of the matrix ΩΩT is bigger than ε, we
β m
 T   set β = 0, and (32) becomes
+ q∗Δj − qΔj q∗Δj − qΔj . (29)
2γ2 j=1 v̇ = Δ + eT ΓCd̃ − eT ΓCDiag [sign(eΔ )] d̂

1  T
n
We define q̃i = q∗i − qi (i = 1, 2, . . . , n) and q̃Δj = q∗Δj −
− q̃ q̇i
qΔj (j = 1, 2, . . . , m). Differentiating (29) with respect to γ1 i=1 i
time, we get
  
n 
n
1  T
n

1 T 1 T 1 
h
i =Δ + eΔi d˜i − |eΔi |dˆi − q̃ q̇i
v̇ = ė Γe + e Γė + ˙ T i
tr (CÃ ) Γ(CÃ ) i=1 i=1
γ1 i=1 i
2 2 2η1 i=1
  
n 
n
1  T
n
1 
h
i T ˙ i ≤Δ + |eΔi ||d˜i | − |eΔi |dˆi − q̃ q̇i
+ tr (CÃ ) Γ(CÃ ) γ1 i=1 i
2η1 i=1 i=1 i=1
   
1 
h n n
i
+ ˙ T
tr (CB̃ ) Γ(CB̃ ) i ≤Δ + |eΔi |τ Ti q∗i − |eΔi |τ Ti qi
2η2 i=1 i=1 i=1
 
1 
h
1  T
n
+ tr (CB̃i )T Γ(CB̃˙ i)
2η2 i=1 − q̃ q̇i
γ1 i=1 i
1  T β  T
n m
− q̃i q̇i − q̃ q̇Δj . (30) 
n
1  T
n
γ1 i=1 γ2 j=1 Δj =Δ + |eΔi |τ Ti q̃i − q̃ q̇i
i=1
γ1 i=1 i
Inserting (14), (15), and (20) into the aforementioned equation  
1 T
yields = Δ + |eΔ1 |τ T1 − q̇1 q̃1
γ1
1 h  
v̇ = eT (ΛT Γ + ΓΛ)e + eT ΓC wi Ãi xξ 1 T
2 + |eΔ2 |τ 2 − q̇2 q̃2
T
i=1 γ1
    
1 
h h
i
T
+ e ΓC i i
w B̃ uξ + i T ˙
tr (CÃ ) Γ(CÃ ) 1 T
η1 i=1 + · · · + |eΔn |τ n − q̇n q̃n
T
(33)
i=1 γ1
CHIEN et al.: CONTROLLER DESIGN FOR A CLASS OF UNCERTAIN NONLINEAR SYSTEMS 549

   
where 1 1
=Δ + |eΔ1 |τ T1 − q̇T1 q̃1 + |eΔ2 |τ T2 − q̇T2 q̃2
γ1 γ1
1 
h 
h
 
Δ = − eT Qe + eT ΓC wi Ãi xξ + eT ΓC wi B̃i uξ 1
2 + · · · + |eΔn |τ Tn − q̇Tn q̃n
i=1 i=1 γ1
   
1 
h
˙ i) 1 T
+ tr (CÃi )T Γ(CÃ − |ēΔ1 |τ Δ1 + q̇Δ1 q̃Δ1
T
η1 i=1 γ2
 
  1 T
1  − |ēΔ2 |τ Δ2 + q̇Δ2 q̃Δ2
h T
i
+ i T ˙
tr (CB̃ ) Γ(CB̃ ) γ2
η2 i=1  
1
1 − · · · − |ēΔm |τ TΔm + q̇TΔm q̃Δm (35)
= − eT Qe γ2
2
⎛ ⎞ where Δ is the same as (34). If we select Âi , B̂i , q̇k , and q̇Δk
 h h i T ˙i
(CÃ ) Γ(CÂ )
+ tr ⎝ wi ÃiT CT ΓexTξ − ⎠ as (25)–(28), (33) and (35) become
i=1 i=1
η1
1
⎛ ⎞ v̇ = − eT Qe ≤ 0. (36)
 h h i T ˙i 2
(CB̃ ) Γ(CB̂ ) ⎠
+ tr ⎝ wi B̃iT CT ΓeuTξ − .
i=1 i=1
η2 Equations (29) and (36) only guarantee that e(t) ∈ L∞ but
not that it converges. The boundedness of e(t) implies the
(34)
boundedness of x(t). Since the operating states are finite, xξ
is bounded. Based on Assumption 1 and the boundedness of
When the determinant of the matrix ΩΩT is smaller than or xξ , uξ is bounded. Therefore, ė(t) is bounded, i.e., ė(t) ∈ L∞ .
equal to ε, we set β = 1, and (32) becomes Integrating both sides of (36) yields

v̇ = Δ + eT ΓCd̃ − eT ΓCDiag [sign(eΔ )] d̂ − eT ΓΩdΔ "t

1
1 
n
1 
m v(t) − v(0) ≤ − λmin (Q) e(τ )2 dτ (37)
+ eT ΓΩuΔ − q̃Ti q̇i − q̃TΔj q̇Δj 2
γ1 γ2 0
i=1 j=1


n 
n 
m where λmin (Q) > 0 is the minimum eigenvalue of Q. When t
=Δ + eΔi d˜i − |eΔi |dˆi − ēΔj dΔj tends to approach infinity, (37) becomes
i=1 i=1 j=1
"∞

m
1 
n
1  T
m
v(0) − v(∞)
+ |ēΔj |dˆΔj − q̃Ti q̇i − q̃ q̇Δj e(τ )2 dτ ≤ 1 . (38)
j=1
γ1 i=1
γ2 j=1 Δj 2 λmin (Q)
0


n 
n 
m
≤Δ + |eΔi ||d¯i | − |eΔi |dˆi − |ēΔj ||dΔj | Since the right side of (38) is bounded, e ∈ L2 . Therefore, by
i=1 i=1 j=1 using Barbalat’s lemma [30], we have e(t) → 0 as t → ∞.
This completes the proof.

m
1 
n
1  T
m
+ |ēΔj |dˆΔj − q̃Ti q̇i − q̃ q̇Δj
j=1
γ1 i=1
γ2 j=1 Δj
IV. S IMULATION R ESULTS

n
≤Δ + |eΔi |τ Ti q∗i This section presents the simulation results of the proposed
i=1
controller, showing that the tracking error of the closed-loop
system can be made arbitrarily small. In addition, the simula-

n 
m
tion results confirm that the effect of modeling errors and exter-
− |eΔi |τ Ti qi − |ēΔj |τ TΔj q∗Δj
i=1 j=1
nal disturbances on the tracking error is attenuated efficiently
by the proposed controller. The experiments use a personal

m
1  T
n
1  T
m
computer with the programming language MATLAB 7.0.4 to
+ |ēΔj |τ TΔj qΔj − q̃i q̇i − q̃ q̇Δj
j=1
γ1 i=1 γ2 j=1 Δj obtain the output responses of the closed-loop systems.
Example 1: Consider the Duffing forced oscillation system

n 
m
[31] with disturbance
=Δ + |eΔi |τ Ti q̃i − |ēΔj |τ TΔj q̃Δj
i=1 j=1 ẋ1 = x2
1 
n
1 
m
ẋ2 = − 0.1x2 − x31 + 12 cos t + u + d
− q̇Ti q̃i − q̃TΔj q̇Δj
γ1 i=1
γ2 j=1 y = x1
550 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 41, NO. 2, APRIL 2011

Fig. 4. Time response of the control input. (a) Using the previous method in
[31]. (b) Using the proposed method.

Fig. 2. Steady trajectories (0–15 s) of the state x1 with the proposed method
and the previous method in [31].

Fig. 5. Trajectories of the outputs y1 , y2 and reference signals r1 , r2 in

example 2. (a) Case 1. (b) Case 2.

control input, the easier the implementation of the controller for

a real system.
Example 2: Consider a more complicated MIMO nonaffine
nonlinear system

ẋ1 = f1 (x, u)+dd1 = (2+sin(x1 x2 )) (u1 +eu1 −1)+dd1

Fig. 3. Transient trajectories (0–1.5 s) of the state x1 with the proposed  
method and the previous method in [31]. ẋ2 = f2 (x, u)+dd2 = x1 x2 +20 log 2+x21 u2 +2u1 +dd2
    
y1 −2 0 x1
where u is the control input, y is the system output, and d is the y= =
y2 0 3 x2
external disturbance. For comparison purposes, it is assumed
that the initial states x1 (0) = 1 and x2 (0) = 0, and the external where x1 , x2 ∈ xu . fi (x, u), i = 1, 2, are unknown nonlinear
disturbance is a random value in the interval [−0.1, 0.1]. functions, u1 and u2 are the control inputs, and both dd1
Therefore, the system response of the overall system using the and dd2 are external disturbances which are assumed to be
proposed adaptive T–S fuzzy-neural controller can be simulated random values in the interval [−0.1, 0.1]. In this example, three
and compared with that of a previous method reported in [24]. different cases for critical points x, u and reference signals are
The design parameters for the proposed approach are chosen as simulated.
η = 0.002, λ = 5, and Q = 2. The control objective is to derive Five fuzzy sets over the interval [−6, 6] are defined for
the control law so that the state x1 of the nonlinear system xξ = [xξ1 , xξ2 ]T with the term sets (PB, PS, Z, NS, NB) and
tracks the reference signal r = sin(t). three fuzzy sets over the interval [−1400, 1400] for uξ =
Figs. 2 and 3 show the steady trajectories (0–15 s) and the [uξ1 , uξ2 ]T . The design parameters are selected as η = 0.003,
transient trajectories (0–1.5 s), respectively, of the state x1 of λ1 = 5, λ2 = 6, and Q = [1 0; 0 2]. The initial states of
the proposed method, the reference signal r, and the state x1 the system are assumed to be x(0) = [−1, 0]T (cases 1) and
of the previous method. As shown in Figs. 2 and 3, the rise x(0) = [−0.2, −0.2]T (case 2). We use the proposed control
time and the settling time of the closed-loop system using the laws in (12) and (13) to control the output y1 of the system
proposed controller are much smaller compared to that of the to track the reference signals r1 = sin(0.5t) + cos(t) (cases
previous controller. Furthermore, the magnitude of the derived 1) and 1/(1 + et−8 ) − 1.5/(1 + et−15 ) + 1/(1 + et−30 ) + 0.5
control input using the proposed approach is also much smaller, (case 2) and the output y2 of the system to track the reference
as shown in Fig. 4(a) and (b). The significantly reduced magni- signals r2 = cos(0.5t) + sin(t) (cases 1) and 1/(1 + et−8 ) −
tude of the control input of the proposed approach demonstrates 1.5/(1 + et−15 ) + 1/(1 + et−30 ) + 0.2 (case 2). Fig. 5 shows
its superiority in practical applications, because the smaller the the curves of the outputs y1 and y2 of the closed-loop system.
CHIEN et al.: CONTROLLER DESIGN FOR A CLASS OF UNCERTAIN NONLINEAR SYSTEMS 551

[8] K. Tanaka and M. Sugeno, “Stability analysis and design of fuzzy

control systems,” Fuzzy Sets Syst., vol. 45, no. 2, pp. 135–156,
Jan. 1992.
[9] H. O. Wang, K. Tanaka, and M. F. Griffin, “An approach to fuzzy control
of nonlinear systems: Stability and design issues,” IEEE Trans. Fuzzy
Syst., vol. 4, no. 1, pp. 14–23, Feb. 1996.
[10] H. Zhang, C. Li, and X. Liao, “Stability analysis and H-infinity con-
troller design of fuzzy large-scale systems based on piecewise Lyapunov
functions,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 36, no. 3,
pp. 685–698, Jun. 2006.
[11] C. Lin, Q.-G. Wang, and T. H. Lee, “H∞ output tracking con-
trol for nonlinear systems via T–S fuzzy model approach,” IEEE
Trans. Syst., Man, Cybern. B, Cybern., vol. 36, no. 2, pp. 450–457,
Fig. 6. Responses of control inputs u1 and u2 in example 2. Apr. 2006.
[12] H. K. Lam and E. W. S. Chan, “Stability analysis of sampled-data
fuzzy-model-based control systems,” Int. J. Fuzzy Syst., vol. 10, no. 2,
pp. 129–135, 2008.
The responses of control inputs u1 and u2 are shown in Fig. 6. [13] Y.-B. Zhao, G.-P. Liu, and D. Rees, “Modeling and stabilization of
The simulation results indicate that the effect of modeling errors continuous-time packet-based networked control systems,” IEEE Trans.
Syst., Man, Cybern. B, Cybern., vol. 39, no. 6, pp. 1646–1652,
and external disturbances on the tracking error is attenuated Dec. 2009.
efficiently by the proposed controller. [14] Y.-G. Leu, W.-Y. Wang, and T.-T. Lee, “Robust adaptive fuzzy-neural
controllers for uncertain nonlinear systems,” IEEE Trans. Robot. Autom.,
vol. 15, no. 5, pp. 805–817, Oct. 1999.
[15] W.-Y. Wang, Y.-H. Chien, and I.-H. Li, “An on-line robust and adaptive
V. C ONCLUSION T–S fuzzy-neural controller for more general unknown systems,” Int. J.
Fuzzy Syst., vol. 10, no. 1, pp. 33–43, 2008.
We have proposed a novel approach by using an online T–S [16] Y.-C. Hsueh and S.-F. Su, “Fuzzy sliding controller design with adaptive
fuzzy-neural modeling and a robust adaptive control for a class approximate error feedback,” Int. J. Fuzzy Syst., vol. 11, no. 1, pp. 36–43,
of unknown systems to achieve the efficient attenuation of 2009.
[17] Y.-J. Huang, T.-C. Kuo, and S.-H. Chang, “Adaptive sliding-mode control
modeling errors and external disturbances. Instead of modeling for nonlinear systems with uncertain parameters,” IEEE Trans. Syst., Man,
unknown systems directly, the T–S fuzzy-neural model approx- Cybern. B, Cybern., vol. 38, no. 2, pp. 534–539, Apr. 2008.
imates the VLS of a real system, with modeling errors and [18] M. C. Hwang and X. Hu, “A robust position/force learning controller
of manipulators via nonlinear H∞ control and neural networks,” IEEE
external disturbances. Compared with previous approaches, the Trans. Syst., Man, Cybern. B, Cybern., vol. 30, no. 2, pp. 310–321,
main contribution of this paper is an investigation of more com- Apr. 2000.
plicated MIMO unknown systems using online adaptive T–S [19] W.-Y. Wang, Y.-H. Chien, Y.-G. Leu, and T.-T. Lee, “Adaptive T–S fuzzy-
neural modeling and control for general MIMO unknown nonaffine non-
fuzzy-neural controllers. The initial values of the parameters of linear systems using projection update laws,” Automatica, vol. 46, no. 5,
the fuzzy-neural model are tuned to their true values through pp. 852–863, May 2010.
online update laws. In Theorem 1, it is proved that, although [20] X. Jia, D. Zhang, X. Hao, and N. Zheng, “Fuzzy H∞ tracking control
for nonlinear networked control systems in T–S fuzzy model,” IEEE
the bounds of modeling errors and external disturbances are Trans. Syst., Man, Cybern. B, Cybern., vol. 39, no. 4, pp. 1073–1079,
unknown, the tracking errors of the closed-loop system can Mar. 2009.
be made arbitrarily small. The simulation results support the [21] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap-
plication to modeling and control,” IEEE Trans. Syst., Man, Cybern.,
theoretical arguments about the T–S fuzzy-neural modeling and vol. SMC-15, no. 1, pp. 116–132, Jan. 1985.
the tracking performance of the design algorithms under the [22] L.-H. Chien, W.-Y. Wang, I.-H. Li, and S.-F. Su, “T–S fuzzy
adaptive tuning methods. control for uncertain nonlinear systems using adaptive fuzzy approach,”
in Proc. IEEE Int. Conf. Fuzzy Syst., Vancouver, BC, Canada, Jul. 2006,
pp. 800–805.
[23] Y.-G. Leu, W.-Y. Wang, and T.-T. Lee, “Observer-based direct adap-
tive fuzzy-neural control for nonaffine nonlinear systems,” IEEE Trans.
R EFERENCES
Neural Netw., vol. 16, no. 4, pp. 853–861, Jul. 2005.
[1] H. J. Lee, H. Kim, Y. H. Joo, W. Chang, and J. B. Park, “A new intelligent [24] C.-A. Chen, H.-K. Chiang, and J.-C. Shen, “Fuzzy sliding mode control
digital redesign for T–S fuzzy system: Global approach,” IEEE Trans. of a magnetic ball suspension system,” Int. J. Fuzzy Syst., vol. 11, no. 2,
Fuzzy Syst., vol. 12, no. 2, pp. 274–284, Apr. 2004. pp. 97–106, 2009.
[2] X. Liu and Q. Zhang, “Approaches to quadratic stability conditions and [25] S. I. Grossman and W. R. Derrick, Advanced Engineering Mathematics.
H∞ control designs for T–S fuzzy systems,” IEEE Trans. Fuzzy Syst., New York: Harper & Row, 1998.
vol. 11, no. 6, pp. 830–839, Dec. 2003. [26] M. Vidyasagar, Nonlinear Systems Analysis. Englewood Cliffs, NJ:
[3] G. Feng, S. G. Cao, N. W. Rees, and C. K. Chak, “Design of fuzzy Prentice-Hall, 1993.
control systems with guaranteed stability,” Fuzzy Sets Syst., vol. 85, no. 1, [27] C. T. Chen, Linear Systems Theory and Design. New York: 1Holt,
pp. 1–10, Jan. 1997. Rinehart and Winston, 1999.
[4] S. G. Cao, N. W. Rees, and G. Feng, “Stability analysis and design for a [28] W.-Y. Wang, Y.-G. Leu, and C.-C. Hsu, “Robust adaptive fuzzy-neural
class of continuous-time fuzzy control systems,” Int. J. Control, vol. 64, control of nonlinear dynamical systems using generalized projection
no. 6, pp. 1069–1087, Aug. 1996. update law and variable structure controller,” IEEE Trans. Syst., Man,
[5] C. W. Park and Y. W. Cho, “T–S model based indirect adaptive fuzzy Cybern. B, Cybern., vol. 31, no. 1, pp. 140–147, Feb. 2001.
control using online parameter estimation,” IEEE Trans. Syst., Man, [29] P. A. Ioannou and A. Datta, “Robust adaptive control: A unified ap-
Cybern. B, Cybern., vol. 34, no. 6, pp. 2293–2302, Dec. 2004. proach,” Proc. IEEE, vol. 79, no. 12, pp. 1736–1768, Dec. 1991.
[6] N. Manamanni, B. Mansouri, A. Hamzaoui, and J. Zaytoon, “Relaxed [30] K. Hornick, M. Stinchcombe, and H. White, “Multilayer feedforward
conditions in tracking control design for a TS fuzzy model,” J. Intell. networks are universal approximators,” Neural Netw., vol. 2, no. 5,
Fuzzy Syst., vol. 18, no. 2, pp. 185–210, Apr. 2007. pp. 359–366, 1989.
[7] R.-J. Wai and Z.-W. Yang, “Adaptive fuzzy neural network control de- [31] W.-Y. Wang, I.-H. Li, L.-C. Chien, and S.-F. Su, “On-line modeling and
sign via a T–S fuzzy model for a robot manipulator including actuator control via T–S fuzzy models for nonaffine nonlinear systems using a
dynamics,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 38, no. 5, second type adaptive fuzzy approach,” Int. J. Fuzzy Syst., vol. 9, no. 4,
pp. 1326–1346, Oct. 2008. pp. 152–161, Sep. 2007.
552 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 41, NO. 2, APRIL 2011

Yi-Hsing Chien was born in Taipei, Taiwan, in 1978. Yih-Guang Leu received the M.S. and Ph.D. de-
He received the M.S. degree in electrical engineering grees in electrical engineering from National Taiwan
from Fu Jen Catholic University, Xinzhuang City, University of Science and Technology, Taipei,
Taiwan, in 2007. He is currently working toward the Taiwan, in 1995 and 1999, respectively.
Ph.D. degree in electrical engineering from National From 1999 to 2006, he was an Associate Professor
Taipei University of Technology, Taipei. with the Department of Electronics Engineering,
His research interests include fuzzy logic systems Hwa Hsia College of Technology and Commerce,
and adaptive control. Taipei. In 2005, he was an Associate Professor and
the Chairman of the Department of Electronics Engi-
neering, Hwa Hsia Institute of Technology, Zhonghe,
Taiwan. He is currently an Associate Professor with
the Department of Applied Electronics Technology, National Taiwan Normal
University, Taipei. His current research interests and publications are in the
areas of fuzzy logic control, robust adaptive control, machine learning, and
neural networks.
Dr. Leu is a member of the IEEE Systems, Man, and Cybernetics Society.

Tsu-Tian Lee (M’87–SM’89–F’97) was born in

Taipei, Taiwan, in 1949. He received the B.S. degree
in control engineering from National Chiao Tung
University (NCTU), Hsinchu, Taiwan, in 1970 and
the M.S. and Ph.D. degrees in electrical engineering
from The University of Oklahoma, Norman, in 1972
and 1975, respectively.
In 1975, he was an Associate Professor with
NCTU, where he was a Professor and the Chairman
of the Department of Control Engineering in 1978
and was a Professor and the Director of the Institute
of Control Engineering in 1981. In 1986, he was a Visiting Professor with
the University of Kentucky, Lexington, where he was a Full Professor of
Wei-Yen Wang (M’00–SM’04) received the M.S. electrical engineering in 1987. In 1990, he was a Professor and the Chairman
and Ph.D. degrees in electrical engineering from Na- of the Department of Electrical Engineering, National Taiwan University of
tional Taiwan University of Science and Technology, Science and Technology, Taipei, where he was a Professor and the Dean of
Taipei, Taiwan, in 1990 and 1994, respectively. the Office of Research and Development in 1998. Since 2000, he has been
From 1990 to 2006, he worked concurrently as a with the Department of Electrical and Control Engineering, NCTU, where he
Patent Screening Member with the National Intellec- is currently a Chair Professor. Since 2004, he has been a President of National
tual Property Office, Ministry of Economic Affairs, Taipei University of Technology (NTUT), Taiwan. His professional activities
Taiwan. In 1994, he was appointed as Associate include serving on the Advisory Board of the Division of Engineering and
Professor with the Department of Electronic Engi- Applied Science, National Science Council, serving as the Program Director
neering, St. John’s and St. Mary’s Institute of Tech- of the Automatic Control Research Program and the National Science Council,
nology, Tamsui, Taiwan. From 1998 to 2000, he was and serving as an Advisor of the Ministry of Education, Taiwan. He also held
with the Department of Business Mathematics, Soochow University, Taipei. numerous consulting positions. He is the author or coauthor of more than
From 2000 to 2004, he was with the Department of Electronic Engineering, 200 refereed journal and conference papers in the areas of automatic control,
Fu Jen Catholic University, Xinzhuang City, Taiwan, where he became a robotics, fuzzy systems, and neural networks. His current research interests
Full Professor in 2004. In 2006, he was a Professor and the Director of the include motion planning, fuzzy and neural control, optimal control theory and
Computer Center, National Taipei University of Technology, Taipei. He is application, and walking machines.
currently a Professor with the Department of Applied Electronics Technology, Prof. Lee was a Fellow of the New York Academy of Sciences in 2002. He
National Taiwan Normal University, Taipei. His current research interests and has been actively involved in many IEEE activities. He has served as a member
publications are in the areas of fuzzy logic control, robust adaptive control, of the Technical Program Committee and of the Advisory Committee of many
neural networks, computer-aided design, digital control, and charge-coupled IEEE-sponsored international conferences. He is currently the Vice President
device camera-based sensors. He has authored or coauthored over 100 refereed for Membership of the IEEE Systems, Man, and Cybernetics Society. He was
conference and journal papers in the aforementioned areas. the recipient of the Distinguished Research Award from the National Science
Dr. Wang is currently serving as an Associate Editor of the IEEE T RANSAC - Council, Taiwan, during 1991–1992, 1993–1994, 1995–1996, and 1997–1998,
TIONS ON S YSTEMS , M AN , AND C YBERNETICS —PART B: C YBERNETICS , a the Academic Achievement Award in engineering and applied science from the
Managing Editor of the International Journal of Fuzzy Systems, an Area Editor Ministry of Education, Taiwan, in 1997, the National Endow Chair from the
of the International Journal of Intelligent Systems Science and Technology, and Ministry of Education, Taiwan, and the TECO Science and Technology Award
a member of the editorial board of the Open Cybernetics and Systemics Journal. from the TECO Technology Foundation in 2003.