Fuzzy CMAC Control for MIMO Nonlinear Systems

C. M. Lin, Senior Member, IEEE, T. Y. Chen, C. H. Chen, and F. S. Ding

Abstract— A fuzzy cerebellar model articulation adaptive laws. Moreover, the robust controller is employed to
control system is presented for multi-input multi-output efficiently suppress the influence of residual approximation
(MIMO) nonlinear systems. The proposed control system error between the ideal controller and FCMAC, so that the
contains of a fuzzy CMAC (FCMAC) and a robust controller. robust tracking performance of the system can be guaranteed.
FCMAC is a main tracking controller utilized to approximate Finally, two uncertain nonlinear systems are demonstrated
an ideal controller and the parameters of FCMAC are on-line to verify the effectiveness of the proposed control system.
tuned by the derived adaptive laws from the Lyapunov function.
The robust controller is designed to suppress the influence of II. PROBLEM STATEMENT
approximate error between the ideal controller and adaptive
FCMAC, so that the robust tracking performance of the system Consider the nth-order multivariable nonlinear systems
can be guaranteed. Finally, two MIMO uncertain nonlinear expressed in the following form:
systems, a chua’s chaotic circuit and a two-inverted pendulum x ( n ) (t ) == f n ( x(t )) + Gn u(t ) + l ( x (t ), t )
system, are performed to illustrate the effectiveness of the
proposed control scheme. y = x (t ) (1)
where u(t ) = [u1 (t ), u 2 (t ), ", u m (t )]T ∈ ℜ m is the control
I. INTRODUCTION input vector of the system,

V ARIOUS control efforts have been utilized to control the y = x (t ) = [ x1 (t ), x2 (t ), " , xm (t )] ∈ ℜ is the system output
T m

uncertain nonlinear systems. Sliding mode control

vector, x (t ) = [ x T (t ), x T (t ), ", x ( n-1)T (t )]T ∈ ℜ mn is the state
(SMC) has been confirmed as a powerful robust scheme for
controlling the nonlinear systems with uncertainties [1], [2]. vector of the system, f n ( x (t )) is a nominal function
The most outstanding features of SMC are insensitive to and Gn is a nominal constant gain. Without loss of
parameter variations, fast dynamic response and external generality, Gn is assumed to be positive definite and is
disturbance rejection [1].
Cerebellar model articulation controller (CMAC) is invertible, and l ( x (t )) is referred to as the lumped
classified as a non-fully connected perceptron-like uncertainty. Equation (1) can be expressed as state and output
associative memory network with overlapping equations as follows:
receptive-fields [3]; and it intends to resolve the fast x (t ) = Am x (t ) + Bm [ f n ( x (t )) + G n u(t ) + l ( x(t ), t )]
size-growing problem and the learning difficult in currently y = C mT x (t ) (2)
available types of neural networks (NNs). Comparing to
neural networks, CMACs possess good generalization
capability, fast learning ability and simple computation [3].
This network has already been shown to be able to Where
approximate a nonlinear function over a domain of interest to
any desired accuracy [4]-[7]. For the reasons, CMACs have
been adopted widely for the closed-loop control of complex The control objective is to design a suitable controller such
dynamical systems in recent literatures [7]-[10]. This study that the system state vector x (t ) can track a desired trajectory
proposes a FCAMC-based control system for multivariable x d (t ) . Define the tracking error e ∆ x d − x ∈ ℜ m , and the
nonlinear systems. The proposed control system is comprised tracking error vector of the system is defined as
of an FCMAC and a robust controller. FCMAC is a main
tracking controller utilized to mimic an ideal controller, and e ∆ [e T , e T , ", e ( n −1)T ]T ∈ ℜ mn . The reference trajectory
the parameters of FCMAC are on-line tuned by the derived dynamic equation can be expressed as
x d (t ) = Am x d (t ) + Bm xd( n ) (3)
Manuscript received March 1, 2007. This work was supported in part The Subtracting (2) from (3), gives
National Science Council of Republic of China through grant NSC
94-2213-E-155-010. e = Am e + Bm [ xd( n ) − f n ( x (t )) − Gn u(t ) − l ( x (t ), t )] (4)
Chih-Min Lin is with the Department of Electrical Engineering,
Yuan-Ze University Chung-Li, Tao-Yuan, 320, Taiwan, Republic of China
(phone: +886-4638800-2418; fax: +886-4639355; e-mail: III. SLIDING MODE CONTROL
[email protected]). In order to implement SMC, the first step is to select a
C. H. Chen is with the Department of Computer Sciences and Information
Engineering, China University of Technology, Hukou Township sliding surface. Then, design the control law such that the
303,Taiwan, R.O.C.( e-mail: [email protected]) system state trajectories are forced toward the sliding surface
F. S. Ding is with the Department of Electrical Engineering, Yuan-Ze and stay on it. The sliding surface can be defined as:
University, Chung-Li, Tao-Yuan, 32026, Taiwan, Republic of China (e-mail:
[email protected]).

1-4244-0991-8/07/$25.00/©2007 IEEE 3107

 n −1
d 
s (t )∆ + λ  e (t ) = K T e (5) Ξw , Ξc , Ξv
s (e ,t )
Wˆ , cˆ , vˆ

 dt 
Adaptive Laws

Weight Memory

where K = [ λ n−1 I , (n − 1) λ n− 2 I , ", I ]T ∈ ℜ mn×m satisfies that Input Space I

Space W

Output Space O
µφ1k

all roots of the equation

p1 1k ∑ o1
φk wkpko
xd + e (t ) Sliding pn a
µφnk ∑ onO u FCMAC + u(t ) Uncertain Nonlinear x
z n−1 I + (n − 1) λz n −2 I + " + (n − 1) λ n−2 zI + λ n −1 I = 0
Uncertain Systems
(6) Hyperplane ank

− + Multivariable Systems
x u RC
are in the open left half-plane, in which z is the Laplace
Receptive --Field
Space R
Association Memory
Space A

operator. A sufficient condition for the existence and RCMAC Network

FCMAC Network Architecture
Architecture
reachable of the sliding hyperplane in the system state space s (e , t ) −
(2 R 2 ) 1 ( R 2 + I )
is to choose the control law such that the following reaching Robust Controller

condition is satisfied:
1 d T m m
Fig. 1. The configuration of the FAMAC robust control system.
2 dt
∑
s (t ) s ( t ) = s T (t ) s(t ) =
i =1
∑
si (t )si (t ) ≤ − σ i si (t ) (7)
i =1 1) Input space I : For a given p = [ p1 , p2 , " , pn ]T ∈ ℜ n , a

where σ i is a small positive constant. Taking the time where na is the number of input state variables, each input
derivative of both sides of (5) and using (4), yields
state variable pi must be quantized into discrete regions
s(t ) = K T Am e + xd( n ) − f n ( x(t )) − Gn u(t ) − l ( x(t ), t ) (8)
(called elements) according to given control space. The
Therefore, an ideal controller uid which guarantees the number of elements, ne , is termed as a resolution.
reaching condition must satisfy the following condition: 2) Association memory space A : Several elements can be
s T (t ) s(t ) = s T (t )[ K T Am e + xd( n ) − f n ( x (t )) − Gn u(t ) − l ( x (t ), t )
accumulated as a block, the number of blocks, nb , is usually
m

≤− ∑σ i s i (t ) (9) greater than or equal to two. A denotes an association

i =1
memory space with nc ( nc = na × nb ) components. In this
If the system dynamics and the lumped uncertainty are
space, each block performs a receptive-field basis function,
exactly known, an ideal controller can be designed as follows
the Gaussian membership function is adopted here as the
to satisfy inequality (9).
receptive-field basis function, which can be represented as
uid = Gn−1 [− f n ( x) − l ( x, t ) + xd( n ) + K T Am e + σ sgn ( s(t )) ] (10)
 − ( pi − cik ) 2 
where σ = diag (σ1 , .... , σi , .... , σ m ) . However, in practical µik = exp   , for k = 1, 2," nb (12)
 vik2 
applications, the ideal controller in (10) is unobtainable.
Thus, a fuzzy CMAC control system is proposed in the where µik represents the output of the k-th receptive-field
following section to achieve robust tracking performance. basis function for the i-th input with the mean cik and
variance vik . Figure 3 depicts the schematic diagram of a
IV. FUZZY CMAC CONTROL SYSTEM DESIGN
two-dimensional FCMAC with ne = 5 and n f = 4 ( n f is the
The configuration of the FCMAC robust control
algorithm, which consists of an FCMAC and a robust number of elements in a complete block); in which p1 is
controller, is depicted in Fig. 1. The control system is divided into blocks Ba1 and Bb1 , and p2 is divided into
assumed to take the following form: blocks Ba 2 and Bb 2 .
u(t ) = uFCMAC + uRC (11)
3) Receptive-field space R : Areas formed by blocks, referred
where uFCMAC is an FCMAC and uRC is a robust controller.
to as Ba1 Ba 2 and Bb1 Bb 2 are called receptive-fields. The
The FCMAC uFCMAC is a main tracking controller utilized to
number of receptive-fields, nd , is equal to nb in this study.
approximate the ideal controller. The robust controller uRC is
The k-th multi-dimensional receptive-field function is defined
employed to efficiently suppress the influence of residual as
approximation error, so that the robust tracking performance n
 n − ( pi − cik ) 2 
φk ( p, c k , v k ) = ∏ µik = exp ∑
a a

of the system can be guaranteed.  (13)

A. Description of FCMAC i =1  i =1 vik2 
An FCMAC is proposed and shown in Fig. 2. This where c k = [c1k , " , c n k ]T and c k = [c1k , " , c n k ]T . The
a a

FCMAC is composed of input space, association memory multi-dimensional receptive-field functions can be expressed
space, receptive-field space, weight memory space and output in a vector form as
space. The signal propagation and the basic function in each Φ( p, c , v ) = [φ1 , " , φk , " , φn ]T (14)
space are described as follows. d

where c = [c1T , " , c kT , " , c nT ]T and v = [v1T , " , v kT , " , v nT ]T . d d

4) Weight memory space W : Each location of R to a

particular adjustable value in the weight memory space can be
expressed as

3108
 W = [w1 , " , w p , " , wn ] o
(15) where Ŵ , Φ̂, cˆ and vˆ are the estimated matrix and vectors
where w p = [w1 p , " wkp , " wn p ] and wkp denotes the d
T
of W * , Φ * , c * and v * respectively. Thus, the dynamic
connecting weight value of the p-th output associated with the equation (18) can be expressed via (19) and (20) as
~ ~
k-th receptive-field. s(t ) = Gn [W T Φ * + Wˆ T Φ + ε − uRC ] − σ sgn [ s (t )] (21)
5) Output space O : The output of FCMAC is the algebraic ~ ~
where W = W * − Wˆ and Φ = Φ * − Φˆ . Moreover, the
sum of the activated weights in the weight memory, and is ~
expansion of Φ in Taylor series can be obtained as
expressed as   ∂φ1     ∂φ1  T 
T
nd
      
op = w Φ =T
p ∑w
k =1
kp φk , for p = 1, 2," no (16)  φ~1

 #
   ∂c
 
  ∂φ
#
 

  ∂v  
 # 
    ∂φ  T 
T
~ ~  =   k   | c = cˆ (c * − cˆ) +   k   | v = vˆ (v * − vˆ) + β
The outputs of FCMAC can be expressed in a vector notation Φ = φk
    ∂c     ∂v  
as  #
φ~
 
 
#   # 
T
  ∂φ  T 
o = [o1 , " o p , " on ]T = W T Φ (17)  n d   ∂φ n d

   n   d

O
 ∂c  
   ∂v  
In the two-dimensional case shown in Fig. 3, the output of
≡ Φc c~ + Φv v~ + β (22)
FCMAC is the sum of the value in receptive-fields Bb1 Bb 2 , T

Bd 1 Bd 2 , B f 1 B f 2 and Bg 1 Bg 2 , where the input state is (0.8,0.8).  ∂φ ∂φ ∂φn  n ×n n

where Φc =  1 , " , k , " ,  |c =cˆ ∈ ℜ ; d d a d

Weight Memory
Space W
 ∂c ∂c ∂c 
T
Input Space I
 ∂φ ∂φ ∂φ n 
; c~ = c * − cˆ ;
Output Space O n ×n n
µ
φ1k Φv =  1 , ", k , ",  |v =vˆ ∈ ℜ
d d a d

 ∂v ∂v ∂v 
p1 1k ∑ o1
φk
v~ = v * − vˆ and β ∈ ℜ n is a vector of higher-order terms.
wkp
ko
d
pna
µφn k ∑ onO
a
nk

Rewriting (22), it can be obtained that

Receptive -Field
- Φ * = Φˆ + Φc c~ + Φv v~ + β (23)
Space R
Association Memory Substituting (22) and (23)into (21), yields
~
s(t ) = Gn [W T Φˆ + Wˆ T (Φc c~ + Φv v~ ) + ω − uRC ]
Space A

Fig. 2. Architecture of an FCMAC. (24)

− σ sgn [ s(t )]
Variable p2 ~
where ω ≡ W *T β + W T (Φ c~ + Φ v~ ) + ε. c v
1.5
B Bf 2
h2
µ 28
Bd 2 0.9
B f 1B f 2 In case of the existence of ω, consider a
µ 24 µ27 ∗ ∗ State (0.8, 0.8)
µ 23 µ 26
Bb 2 0.3
Bd 1 Bd 2 specified H ∞ tracking performance [11]
Bg 2 -0.3 m
~ ~
∑∫ si2 (t ) dt ≤ sT (0)Gn−1 s (0) + tr [W T (0) Ξ w−1W (0)] + ~
c T (0) Ξ c−1c~ ( 0)
µ22 Bb1 Bb 2 T
µ25
Be 2 µ 21 -0.9
Bc 2 Ba 2 B g1 Bg 2 Variable p1 0
-1.5 i =1
-1.5 -0.9 -0.3 0.3 0.9 1.5
Layer 4 µ11 m

+ v~ T (0) Ξ v−1v~ (0) + ∑r ∫

T
Layer 3 Ba1 µ15
Layer 2
Layer 1
Bb1 Layer 1
i
2
ω i2 (t ) dt (25)
Bc1 µ12 µ16 Bd 1 Layer 2 i =1
0

Be1 µ13 µ17 Bf1 Layer 3 where Ξ w , Ξ c and Ξ v are the learning-rate matrices, and
µ14 µBh1
B g1 Layer 4
ri is a prescribed attenuation constant. If the system starts
18

Fig. 3. A two-dimensional FCMAC with n f = 4 and ne = 5 . ~

with initial conditions s (0) = 0, W (0) = 0, c~ (0) = 0,
B. Robust controller design v~ (0) = 0, then the H ∞ tracking performance in (25) can be
Subtracting (10) from (8), yields
rewritten as
s(t ) = Gn [uid − u] − σ sgn[ s (t )] (18) m  si 
Assume there exists an optimal FCMAC u*FCMAC to estimate sup ∑  ω ≤ ri 

(26)
ωi ∈L2 [0,T ] i =1
 i 
the ideal sliding mode controller uid such that T T

∫ ∫
2 2

uid = u *
( p, W , c , v ) + ε ≡ W Φ + ε
* * * *T *
(19) where si = si2 (t ) dt and ωi = ω i2 (t ) dt . This shows
FCMAC 0 0

where ε = [ε 1 , ...., ε i , .... , ε m ]T is a minimum reconstructed that ri is an attenuation level between the approximation
* *
error vector; W , Φ , c and v are the optimal parameter
* * error ωi (t ) and system output function si (t ). If ri = ∞ , this
matrix and vectors of W , Φ, c * and v * respectively. is the case of minimum error tracking control without
approximation attenuation [11]. Therefore, the following
However, the optimal FCMAC can not be obtained; thus, an
theorem can be stated and proved.
estimating FCMAC is used to estimate the optimal FCMAC.
Theorem 1: Consider the nth-order MIMO nonlinear systems
From (17), the control law (11) can be rewritten as follows:
represented by (1). FCMAC is given as in (17), in which the
u(t ) = uFCMAC ( p, Wˆ , cˆ, vˆ) + uRC ≡ Wˆ T Φˆ + uRC (20) adaptive laws of FCMAC are designed as in (27)-(29) and the
robust controller is designed as in (30). Then, the robust

3109
tracking performance in (25) can be achieved for the iL R

prescribed attenuation level ri , i = 1, 2, ..., m .

 +
Wˆ = Ξ w Φˆ s T (t )
+
(27) L
C2 vC 2
C1 vC 1
g
− −
cˆ = Ξ Φ T Wˆ s (t ) (28) u3 ± u2 u1
c c

vˆ = Ξ Φ T Wˆ s(t )
v v (29) Fig. 4. Chua’s chaotic circuit.
uRC = (2 R ) ( R + I ) s (t )2 −1 2
(30) The dynamic equations of the Chua’s circuit are written as
[12]
Proof: The Lyapunov function candidate is given by
1 1 
~ 1
[ ~ ~
V(s(t),W, c~, v~) = sT (t)Gn−1s(t) + tr (WT Ξw−1 W) + c~T Ξc−1c~ + v~T Ξv−1v~ (31)
2
] vC =  (vC − vC ) − g (vC ) + u1 (t )  + d1 (t )
1
C1  R 2

1 1

Taking the derivative of the Lyapunov function and using 1 1 

(30), yields vC =  (vC − vC ) + iL + u2 (t )  + d 2 (t )
~ ~
2
C2  R 1 2

V ( s (t ), W , ~
c , v~ ) = s T (t ) [W T Φˆ + Wˆ T (Φ c c~ + Φ v ~
v ) + ω − u RC ]
1
~  iL = (−vC + u3 (t )) + d 3 (t ) (38)
− sT (t )Gn−1σ sgn [ s(t )]v−1 − tr (W T Ξ w−1 Wˆ ) − c~T Ξ c−1cˆ − v~T Ξ v−1vˆ (32) L 2

~ ~
It can be noted that s T (t ) W T Φˆ = tr (W T Φˆ s T (t )) and where u(t )∆[u1 (t ), u2 (t ), u3 (t )]T denotes the control input
s T (t )Gn−1 σ sgn [ s (t )] ≥ 0 , equation (32) can be written as and d (t )∆[d1 (t ), d 2 (t ), d3 (t )]T denotes the external
~
V (s(e , t),W , c~, ~

~ ˆ T
v ) ≤ trW T [Φ


(
s (e , t ) − Ξw−1Wˆ ] + c~T ΦcTWˆ s(e , t) − Ξc−1cˆ ) disturbance. The voltages vC (t ),
1
vC (t ) 2
and the
current iL (t ) are the state variables. Thus, the state vector of
( v v )
+ v~ T Φ T Wˆ s(e , t ) − Ξ −1vˆ + [ s T (e , t )ω − s T (e , t ) u ] (33) RC chaotic system is defined
From (27)-(30), equation (33) can be rewritten as as x (t )∆[vC (t ), vC (t ), iL (t )]T = [ x1 (t ), x2 (t ), x3 (t )]T . The
1 2

~ m
r2 +1
V ( s (t ), W , c~ , v~ ) ≤ ∑ [ si (t )ω i (t ) − si2 (t ) i 2 ] dynamic equation (31) can be rewritten as
i =1 2 ri x (t ) = f ( x ) + G ( x ) u(t ) + d (t ) (39)
m
s 2 (t ) ri 2ω i2 (t ) where
∑
≤ [− i
2
+
2
] (34)
 1 1  1 1  1 
T

i =1 f ( x) =   ( v C − v C ) − g ( v C ) ,  ( v C − v C ) + i L , ( − v C )
 C1  R  C2  R  L 
2 1 1 1 2 2

Assuming ωi ∈ L2 [0, T ], ∀T ∈ [0, ∞), integrating the above

equation from t = 0 to t = T , yields and G ( x ) = diag ( 1 , 1 , 1 ) .
C1 C 2 L
m
1 T 2 r2 T 2
V (T ) − V (0) ≤ [−
2 0
∑ ∫
si (t ) dt + i
2 0
ω i (t ) dt ] (35) ∫ The external disturbance is given as
i =1
 d1 (t )   sin(2t ) exp(−0.2t ) + 0.3 
Since V (T ) ≥ 0 , the above inequality implies the following
d (t ) = d 2 (t ) =  cos(2t )exp(−0.2t ) − 0.5 .
inequality  d 3 (t )  sin(3t )exp(−0.2t ) + 0.2 
1 m T 2 1 m 2 T 2
∑∫
2 i=1 0
si (t ) dt ≤ V (0) +
2 i=1 0
∑ ∫
ri ω i (t ) dt (36) The desired trajectories come from the reference model
outputs that are chosen from x di (t ) = −4 xdi (t ) + 4 γ i ,
Using (36), the above inequality is equivalent to the following
m
~ ~
where γ i is the input signal to the reference model. The initial
∑∫ si2 (t ) dt ≤ s T (0)G n−1 s ( 0) + tr [W T ( 0) Ξ w−1W ( 0)] + c~ T (0) Ξ c−1~
T
c (0 ) conditions of the Chua’s chaotic circuit and the reference
0
i =1

m models are given as

+ v~ T (0) Ξ v−1v~(0) + ∑r ∫
T

i
2
ω i2 (t ) dt (37) x1 (0) = −1, x2 (0) = 0 and x3 (0) = 0, xd 1 (0) = 0, xd 2 (0) = −1,
0
i =1
and xd 3 (0) = 1. The reference inputs are unit periodic
Thus the proof is completed.
rectangular signals. For the proposed control scheme, the
V. SIMULATION STUDIES sliding hyperplane is designed as s (t )∆ e (t ). The proposed
In this section, two MIMO nonlinear systems, a chua’s FCMAC is characterized as:
chaotic circuit and a two-inverter pendulum system, are y number of input state variables: na = 3
performed to verify the effectiveness of the proposed control y number of elements for each state variable: ne = 5
method. y generalization: n f = 4
A. Chua’s chaotic circuit
A typical Chua’s chaotic circuit consists of one linear y number of blocks for each state variable: nb = 8 (blocks)
resistor ( R ), two capacitors ( C1 , C2 ), one inductor ( L ) and y number of receptive-fields: nd = 8 (receptive-fields)
one nonlinear resistor ( g (vC ) ) as shown in Fig. 4.
1
y receptive-field basis functions: µik = exp[−( pi − cik ) 2 / vik2 ]

3110
The learning-rate matrices of FCMAC are selected as u1
mb
u2
mb
Ξ w = 30 I 8×8 , Ξ c = Ξ v = 0.5 I 24×24 and the specified θ1 θ2
attenuation constant diagonal matrix is R = 0.2 I 3×3 . l l
k
The simulation results for Chua’s chaotic circuit are α ((tt )
shown in Fig 5. The trajectories of the system states are
α ((tt )
plotted in Figs. 5(a)-(c) for vC (t ) , vC (t ) and iL (t ) ,
1 2 ma ma
respectively. The associated control efforts
δ1
u1 (t ), u2 (t ), u3 (t ) are depicted in Figs. 5(d)-(f). From the δ2
simulation results, it can be seen that the proposed control Fig. 6. Inverted double pendulums on carts system.
scheme can provide a good control performance with small
g 1
tracking error. θ1 = r θ1 + u1 + N 21 (θ1 , θ1 )
Capacitor voltage (volt) vC
1
cm l mb cm l 2
x1
 k ( α (t ) − c m l ) 
xd1 + 2
(−α (t )θ1 + α (t )θ 2 − δ1 + δ2 ) (40)
 m c
b m l 
g 1
θ2 = r
θ2 + u 2 + N 22 (θ 2 , θ2 )
cm l mb cm l 2
(a) Time (sec)
Capacitor voltage (volt) vC 2
 k ( α (t ) − c m l ) 
+ 2
(−α (t )θ 2 + α (t )θ1 − δ1 + δ2 ) (41)
 mb cml 
xd 2
x2
where θ i (t ), θi (t ) are the angle and angular velocity of
(b) Time (sec) pendulums. u(t )∆[u1 (t ), u2 (t )]T are the control torques of the
Inductor current (ampere) iL
x3 pendulums. N 21 (θ1 , θ1 ), N 22 (θ 2 , θ2 ) are the system
xd 3 uncertainties. k and g r denote the spring and the gravity
constant, respectively. ma and mb are the masses of the carts
and the pendulums, respectively. l is the length of the
(c) Time (sec) pendulums and L is the natural length of the spring. δ1 and
Control input (ampere) u1
δ2 are the displacements of the moving carts. α (t ) = sin(5t ),
cm = mb /(ma + mb ), Define
x (t )∆[θ1 (t ), θ 2 (t )]T = [ x1 (t ), x2 (t )]T , then the dynamic
equations (40)-(41) can be expressed as
x(t ) = f ( x (t )) + G ( x (t )) u(t ) + d (t ) (42)
(d) Time (sec)
Control input (ampere) u2 where
 g r   k (α(t ) − cm l ) 
 c l θ1   m c l 2 ( −α (t )θ1 + α(t )θ 2 − δ1 + δ2 )
f ( x(t )) =  m  +  b m

 g r θ   k (α(t ) − cm l ) ( −α (t )θ + α(t )θ − δ + δ )
 c m l   mb c m l
2 2 2 1 1 2


(e) 1 1
Time (sec)
and G ( x(t )) = diag ( , ) . The external
Control input (volt) u3 mb cm l 2 mb cm l 2
 N (θ , θ ) 
disturbance is d (t ) =  21 1 1  .
 N 22 (θ 2 , θ 2 )
The parameters of inverted double pendulums are
ma = mb = 50 kg , k = 1 , gr = 1 , l = 1 , L = 2,
(f) Time (sec)
Fig. 5. Numerical simulations for Chua’s chaotic circuit. mb  2 m 2
N 21 (θ1 , θ1 ) = − θ1 sin(θ1 ), N 22 (θ 2 , θ2 ) = − b θ2 sin (θ 2 ),
ma ma
B. Inverted double pendulum system δ1 = sin(2t ), δ2 = sin(3t ) + L. The tracking reference signals
The system of inverted double pendulums on carts is
shown in Fig. 6 and its dynamic equations are given as π
are xd 1 (t ) = [ sin (t ) + 0.3sin(3t )] and
follows [13]: 6.28

3111
 Control input (N-m) u2
π
xd 2 (t ) = − [ sin(t ) + 0.3sin(3t )] . The initial conditions of
6.28
the pendulum system are given
as x1 (0) = 0.5, x2 (0) = −0.5 , x1 (0) = 0 and x2 (0) = 0. The
sliding hyperplane is designed as s (t )∆ e(t ) + 3e (t ). The (f) Time (sec)
proposed FCMAC is characterized as: Fig. 7. Numerical simulations for inverted double
y number of input state variables: na = 2 pendulum system.
y number of elements for each state variable: ne = 5
VI. CONCLUSION
y generalization: n f = 4 (elements/ block)
This study has successfully applied a fuzzy CMAC control
y number of blocks for each state variable: nb = 8 (blocks) system for MIMO nonlinear systems. The effectiveness of the
y number of receptive-fields: nd = 8 (receptive-fields) proposed control system is verified by controlling two MIMO
y receptive-field basis functions: µik = exp[−( pi − cik ) 2 / vik2 ] nonlinear systems, a Chua’s chaotic circuit and a
two-inverted pendulum. The simulation results confirm that
The learning-rate matrices of FCMAC are chosen as
the proposed control system not only guarantee the system
Ξ w = 35I 8×8 , Ξ c = Ξ v = 0.85I 16×16 and the specified stability but also maintains a good tracking performance.
attenuation constant diagonal matrix R = 0.1I 3×3 . The
simulation results for inverted pendulums system are shown REFERENCES
in Fig. 7. The trajectories of system states are plotted in Figs. [1] J. J. E. Slotine and W. P. Li, Applied Nonlinear Control. Englewood
7(a)-(d) for θ1 (t ), θ 2 (t ), θ1 (t ) and θ2 (t ), respectively. The [2]
Cliffs, NJ: Prentice-Hall, 1991.
J. Y. Hung, W. Gao, and J. C. Hung, “Variable structure control: a
associated control efforts u1 (t ) and u2 (t ) are depicted in Figs. survey,” IEEE Trans. Ind. Electron., vol. 40, no. 1, pp. 2-22, 1993.
[3] J. S. Albus, “A new approach to manipulator control: the cerebellar
7(e) and (f). The simulation results show that the states can model articulation controller (CMAC),” J. Dyn. Syst. Meas. Contr.,
track the reference trajectories as the sliding surface is vol. 97, no. 3, pp. 220-227, 1975.
controlled to approach to zero. [4] S. H. Lane, D. A. Handelman, and J. J. Gelfand, “Theory and
First pendulum angle (rad) θ1 development of higher-order CMAC neural networks,” IEEE Contr.
Syst. Mag., vol. 12, no. 2, pp. 23–30, 1992.
[5] F. J. Gonzalez-Serrano, A. R. Figueiras-Vidal, and A.
x1 Artes-Rodriguez, “Generalizing CMAC architecture and training,”
xd1 IEEE Trans. Neural Networks, vol. 9, no. 6, pp. 1509-1514, 1998.
[6] J. C. Jan and S. L. Hung, “High-order MS_CMAC neural network,”
IEEE Trans. Neural Networks, vol. 12, no. 3, pp. 598-603, 2001.
(a) Time (sec) [7] C. M. Lin and Y. F. Peng, “Adaptive CMAC-based supervisory
Second pendulum angle (rad) θ 2 control for uncertain nonlinear systems,” IEEE Trans. Syst., Man,
Cybern. B, vol. 34, no. 2, pp. 1248-1260, 2004.
[8] C. M. Lin and Y. F. Peng, “Missile guidance law design using
xd 2
adaptive cerebellar model articulation controller,” IEEE Trans.
x2 Neural Networks, vol. 16, no. 3, pp. 636-644, 2005.
[9] J. Y. Chen, P. S. Tsai, and C. C. Wong, “Adaptive design of a fuzzy
cerebellar model arithmetic controller neural network,” IEE
(b) Time (sec) Proc.,Control Theory Appl., vol. 152, no. 2, pp. 133-137, 2005.
First pendulum angular velocity (rad/sec) θ1 [10] T. F. Wu, P. S. Tsai, F. R. Chang, and L. S. Wang, “Adaptive fuzzy
xd1
CMAC control for a class of nonlinear systems with smooth
compensation,” IEE Proc.,-Control Theory Appl., vol. 153, no. 6, pp.
647-657, 2006.
[11] B. S. Chen, C. H. Lee, Y. C. Chang, “ H ∞ tracking design of
x1
uncertain nonlinear SISO systems: adaptive fuzzy approach,” IEEE
Trans. Fuzzy Syst., vol. 4, no. 1, pp. 32-43, 1996.
(c) Time (sec)
Second pendulum angular velocity (rad/sec) θ 2 [12] Y. C. Chang, “Robust H ∞ control for a class of uncertain nonlinear
time-varying systems and its application,” IEE Proc.,-Control
Theory Appl., vol. 151, no. 5, pp. 601-609, 2004.
[13] H. K. Lam, F. H. F. Leung, and Y. S. Lee, “Design of a switching
x2 controller for nonlinear systems with unknown parameters based on
xd 2 a fuzzy logic approach,” IEEE Trans. Syst., Man, Cybern. B, vol. 34,
no. 2, pp. 1068-1074, 2004.
(d) Time (sec)
Control input (N-m) u1

(e) Time (sec)

3112