Proceedings of the 2007 American Control Conference ThB17.

3
Marriott Marquis Hotel at Times Square
New York City, USA, July 11-13, 2007

Multivariable Adaptive Fuzzy CMAC Control

for a Class of Nonlinear Systems

Ter-Feng Wu, Li-Sheng Wang, Senior Member, IEEE, Fan-Ren Chang, and Pu-Sheng Tsai

(CMAC) has attracted much attention because of the faster

Abstract— In this paper, a modified multivariable adaptive learning, better generalization and simpler computation. In
fuzzy cerebellar model articulation controller (CMAC) scheme particular, a trained CMAC can approximate nonlinear func-
is proposed to solve the tracking problem for a class of nonlinear tions in a generalized lookup table manner over a domain to
systems. A FCMAC (Fuzzy CMAC) module is used to ap-
proximate a nonlinear multivariable (multi-input multi-output
any desired accuracy. Applications include robot manipula-
(MIMO)) system involving uncertainty to create the desired tors [10] and spacecrafts [11], etc., in which some parameters
ideal control inputs. Suitable control and adaptive laws with are unknown. Lately, various modifications of CMAC have
output feedback based on sliding surface concept are incorpo- been proposed to enhance its performance. The CMAC with a
rated with FCMAC to yield a MIMO adaptive FCMAC (MIMO robust compensation can achieve H ∞ tracking performance
AFCMAC) control system without prior off-line learning phase as shown in [12]. The merging of CMAC and the Hamil-
required. To overcome the chattering problem associated with ton-Jacobin-Bellman optimization theory yields an optimal
discontinuity derived from conventional switching robust com-
control design [13]. Combining a fuzzy reasoning mechanism,
pensation, a smooth compensation is proposed. By Lyapunov
stability analysis, it is shown that all of the closed-loop signals the resulting fuzzy CMAC brings about a simple control
are bounded, and the tracking errors converge exponentially to architecture [11, 14-15].
a residual set, whose size can be controlled. While the tracking In this paper, to solve the chattering problem on the control
precision may be sacrificed slightly, the quality of the control signals associated with the switching function used in many
signal can be improved significantly. Simulation results are adaptive CMAC design, we propose a modified multivariable
presented to demonstrate the validity and the applicability of adaptive fuzzy CMAC (MIMO AFCMAC) control scheme.
the proposed methodology.
The method is based on the previous work [16] for SISO
systems to design a tracking controller for a class of nonlinear
I. INTRODUCTION
multivariable systems. The AFCMAC approximation is

T O deal with nonlinearities in a dynamical system, various

control methodologies have been developed from the
perspective of system theory and traditional feedback control
adopted as a rough tuning, and the smooth compensation is
developed as a fine tuning such that: (i) the design method-
ology is easy to realize, (ii) all of the control gains, including
theory [1-2]. Notably, these methods mostly depend on a the CMAC weights, can be updated on-line without a prior
sufficient knowledge of the system dynamics, so that their learning phase, (iii) chattering can be reduced, and (iv) the
application is limited for uncertain systems. Therefore, con- tracking performance is guaranteed. Comparing to the SISO
trol systems are increasingly required to exhibit high dy- case, the development of AFCMAC for MIMO case needs to
namical performance and robustness in complex and uncer- deal with more complex systems with higher dimension, and
tain situations [3]. To deal with this problem, some tech- there are more design parameters to be assigned.
niques in intelligent control [4] have found their applications.
Examples include the neural networks using appropriated II. FUZZY CMAC DESIGN
learning phases [5-6] and fuzzy control by capturing human
In this section, the concepts of basic CMAC and fuzzy
experiences [7-8].
CMAC (FCMAC) are introduced.
After the introduction by [9], one subclass of the neural
networks, called the Cerebellar Model Articulation Controller A. Basic CMAC Design
Fig. 1 shows the mapping and the signal flow in a basic
Manuscript received September 15, 2006. CMAC. The architecture of the basic CMAC includes input
T.-F. Wu is with the Department of Electrical Engineering, National Ilan space S ⊂ R L , association memory space, receptive-field
University, Ilan, Taiwan, ROC (corresponding author, phone: 886 -
03-9357400ext666; fax: 886-03-9369135; e-mail: [email protected]).
space, weight memory space and output space Z ⊂ R m . The
L.-S. Wang is with the Institute of Applied Mechanics, National Taiwan characterizations are described in [17]. The output of the
University, Taipei, Taiwan, ROC (e-mail: [email protected]). CMAC enters the physical system to be controlled through
F.-R. Chang is with the Department of Electrical Engineering, National Taiwan
University, Taipei, Taiwan, ROC (e-mail: [email protected]).
the m input variables of the plant. A basic CMAC is designed
P.-S. Tsai is with the Department of Electronic Engineering, China Institute of [14, 16] such that the CMAC output variable is given by
Technology, Taipei, Taiwan, ROC (e-mail: [email protected]).

1-4244-0989-6/07/$25.00 ©2007 IEEE. 3365

 ThB17.3

µF ( s1 )
Association Weight 1

Receptive-field N 1 P
memory space , A memory
space , T
space, W PA: → R m Output
: → space, Z Output
GSA Aa ww
w1
1 " w1m zCMAC1
Input
space, S
Ab w2 w2  −2 −1 0
s2
1 2 s1
s2
Ba w3 w3
2 P 2
s1 Bb w4
(1,1)
w4
 b ( −2, 2) ( −1, 2) ( 1, 2) ( 2, 2)

#
(-2,2)
Cc
Cd
w5
w6
w5
w6 
# f
d
1
( −2, 1) ( −1, 1) ( 1, 1) ( 2, 1)
µ F ( s2 )
2
1
1

sL 0 0
Dc w7 w7 s1
a ( −2, −1) ( −1, −1) (1, −1) (2, −1)
Dd w8 w8
Ee w9 w9
 zCMACm
e
c
−1
( −2, −2) ( −1, −2) (1, −2) (2, −2) N
−1

Ef w10 w10 −2 −2
for L=2 −2 −1 0 1 2
case Fe w11 w11
Ff w12 " w12 A B
for M=12
case w1 " wm C D

Fig. 1. Mapping and signal propagation for a basic CMAC. E F

Fig. 2. Sketch of CMAC integrated with fuzzy sets.

zCMAC = F (s) ∈ Z (1)
L m
where F : R → R is a nonlinear function of CMAC input ª1 1 1 1 0 0 0 1 1 1 1 1º
«1 1 0 0 0 1 0 1 1 1 1 1»
variable s = [ s1 ," , sL ]T ∈ S . Mathematically, the relation (1) A=« ». (9)
can be represented by a pair of mappings [9] , «1 0 0 0 1 1 1 1 1 1 1 1»
« »
G : S → A; s 6 G (s) = a(s) ∈ A ∈ R M (2) ¬1 0 1 1 0 0 1 1 1 1 1 1¼
P : A → Z ; a 6 P (a). (3)
III. MULTIVARIABLE ADAPTIVE FUZZY CMAC
Particularly, we may choose the function P which generates
CONTROLLER DESIGN
the output zCMAC ∈ R m as follows
In this section, an adaptive scheme is included to achieve
zCMAC = P(a) = (aT W )T (4) basic performance for a MIMO system. Moreover, to sup-
where W≡[ w1,", wm] ∈RM×m denotes the CMAC weight matrix. press the approximated errors of the FCMAC, a smooth
compensation is evolved to overcome the chattering issue.
B. Fuzzy CMAC Design
A. Problem Description
To simplify the input partition, cf. Fig. 2, the structure of a
single output FCMAC was proposed in [16]. For a 2-input Consider a multivariable nonlinear system [18]:
m-output problem, a fuzzy system with N fuzzy rules may be y ( n ) = f ( x ) + G ( x )u (10)
designed, each of which being given by ( nm ) T
where y = [ y1,", ym ]T ∈ Rm and y ( n ) = ª¬ y1( n1 ) ,", y º¼ ∈ R m m
R ( i ) : IF s1 is F1i and s2 is F2i , THEN z (fi ) is (aTi W )T (5)
denote the system output and its derivative, respectively;
where the THEN-part is extracted from CMAC. Given the m

membership function µ F i of fuzzy set Fki , k = 1,2, the fol-

k
n = {n1 ," , nm } with ¦n
i =1
i = n being the system degree;
lowing defuzzification process is adopted to calculate the ( n1 −1) ( nm −1)
x = [ x1 , x2 ,", xn ]T = [ y1 , y1 ," , y1 , y2 ,", ym ]T ∈ R n is
output z FCMAC :
T m
N N
the state vector; f ( x ) = [ f1 ( x ),", f m ( x )] ∈ R is an un-
zFCMAC = ¦ vi (aTi W)T / ¦vi (6) known system dynamics; G(x) = [g1 (x),", gm (x)] ∈ Rm×m is a T
i=1 i=1
2
where vi = ∏ µ F i ( sk ). Equation (6) may be re-written com- bounded matrix with g i ( x ) = [ gi1 ( x ),", gim ( x )]T ∈ R m for
k =1 k
i = 1,", m ; and u = [u1 ,", um ]T ∈ R m is the system input. To
pactly as
steer the state vector x in a certain controllability region
z FCMAC = (hT AW )T (7)
c ∈ R n , it is necessary that G( x ) is invertible. Additionally,
where
T
N
T
in implementing the adaptive law, the G( x ) is assumed to be
h = [ h1 ," , hN ] , hi = vi / ¦ vi , and A = [ a1 ," , a N ] . (8) known in this paper. The objective is, given the desired
i =1

In (7), the matrix A (determined by CMAC) and the vector output trajectory y d = [ yd 1 , yd 2 ,", ydm ]T ∈ R m , to construct a
h (determined by fuzzy rules) are typically fixed, but the multivariable adaptive fuzzy CMAC controller u such that
weight matrix W is adjustable. the tracking error defined as x = [y 1T ,", y Tm ]T ∈Rn where
For the 2-input problem, a set of membership functions
may be chosen as shown in Fig. 2. The matrix A in (7) is y i = [( yi − ydi ),( yi − ydi ),",( yi( ni −1) − ydi( ni −1) )]T , i =1,",m , can
designed according to [16], given by exponentially converge to an arbitrarily small residual set

3366
 ThB17.3

without control chattering problem. Hurwitz. Accordingly, for any ni × ni symmetric posi-
If the system dynamics f is known, the ideal control law tive-definite matrix Qi , there exists an ni × ni symmetric
u* = G −1 ª¬ −f ( x ) + y (dn ) − Cx º¼ (11) positive definite matrix Pi such that the following Lyapunov
matrix equations hold;
may be applied, where
Pi Fi + FiT Pi = −Qi , for i = 1,", m (20)
ªc1n " c11 0 0 º
« 1 » m×n The solution Pi of (20) can then be used to construct a
C≡ « 0 % 0 » ∈R (12) Lyapunov function candidate in designing a modified MIMO
« 0 0 c " c » AFCMAC control scheme later. Additionally, the optimal
¬ mnm m1 ¼
weight W * and the bound D are unknown in most applica-
is a constant matrix with cini ," , ci1 being the coefficients of a tions. In particular, the optimal weight W * is conventionally
Hurwitz polynomials ∆i (λ ) = λ ni + ci1λ ni −1 + " + cini , for trained by some off-line learning phase, which may take a
longer time. Alternatively, an adaptive mechanisms can be
i = 1,", m . Moreover, the control law (12) can be separated
adopted to estimate W * and D on-line.
into two parts: the approximated part u*A and the known part
B. Modified MIMO AFCMAC with Smooth Compensation
u k defined respectively as
To deal with the problems mentioned above, the following
u*A ≡ G −1 [ −f (x) ] and u k ≡ G −1 ª¬ y (dn ) − Cx º¼ . (13) adaptive laws are respectively used to obtain the estimates
Using (11), the closed-loop error dynamics becomes Wˆ ≡ [w ˆ m ]∈ RM ×m and D
ˆ 1 ,", w ˆ ≡ diag[Dˆ ]∈ Rm×m :
i

y i( ni ) + ci1 y i( ni −1) + " + cini y i = 0 , for i = 1," , m (14) 

ˆ ˆ T T
W = − W11 − ( 1φh A ) (21)
and the tracking error x exponentially approaches zero. ˆ
D = − D1ˆ +  -, D ˆ (0) > 0 (22)
However, the control law (11) can not be utilized if the 2 2
m
function f is unknown. A scheme to approximate f must be
where φ= [φ1,",φm] ≡ ¦ei ∈Rm with ei ≡ ( y Ti Pi Gi )T ∈ R m and
T

obtained such that the approximated control law u*A can be i=1

adopted. The AFCMAC developed herein provides such a - ≡ diag ¬ª φ1 ,", φm ¼º ∈ R m×m for i = 1,", m ;
scheme. m×m m×m
In the 2-input m-output FCMAC built in Subsection II.B, 1 ≡ diag[γ 1i ]∈ R , 2 ≡ diag[γ 2i ]∈ R ,
m×m m×m
T
the input variable is s = [ s1 , s1 ] ∈ R 2 where s1 = dT x with 11 ≡ diag[σ1i ]∈ R and 12 ≡ diag[σ 2i ]∈ R are the positive
T design parameter matrices, to assure the convergence rates of
the coefficient vector d =[ d1,",dn ] ∈Rn . Therefore, the errors the estimation. The first terms on the right-hand side of (21)
in various orders are combined as a single variable, which is -(22) represents the σ-modification which can increase the
similar to the concept of the sliding surface [1]. The optimal robustness [19]. Since G( x ) is known, the variable ei is
weight matrix W * ∈ R M ×m , where M = 12 here, is assumed available, and hence the above laws are feasible. Based on
to exist; the output of FCMAC, Eq. (7), rewritten as these estimates, the control law is designed as
u∗FCMAC = (hT AW * )T ∈ R m , would be close to u*A . Let u = u AFCMAC + u AR + u K (23)
T
0 = [ε1 " ε m ] ∈ R m be defined as where
T u AFCMAC = (hT AW ˆ )T (24)
0 = u∗FCMAC − u*A = ¬ª(uFCMAC
∗ * ∗ *
º (15)
1 − uA1 ),",(uFCMACm − uAm ) ¼
u = −D ˆ ⋅ sat ( φ ) (25)
AR b
which is assumed to be bounded with a positive bound Di [5],
with
i.e. ε i ≤ Di , for i = 1," , m . Substituting the ideal control T
sat b (φ) ≡ [sat b (φ1 ),",sat b (φm )] (26)
(11) into the system (10) yields,
where the constant b > 0 specifies the boundary layer and
y ( n ) = y (n) 
d − Cx + G[ u − u ]
*
(16)
sat b (⋅) denotes the saturation function [1], given by
which implies that
y i = Fi y i + Gi [u - u* ] , for i = 1,", m (17) ­sgn( z ), if z > b
sat b ( z ) ≡ ® (27)
where ¯ z / b, otherwise
ª0(ni −1)×1 Ini −1 º ª0(ni −1)×m º for any argument z ∈ R . Here the signum function is
Fi = «!!!!! » ∈Rni ×ni , Gi = «!!!» ∈Rni ×m. (18) °­ z / z , z ≠ 0
«¬ − cTi »¼ «¬ gi »¼ sgn( z ) ≡ ® . (28)
°̄0, z=0
Replacing u* by (11) and (15), the error equations become
Figure 3 shows the architecture of the modified MIMO
y i = Fi y i + Gi ª¬u − u∗FCMAC + 0 − u K º¼ , for i = 1,", m (19) AFCMAC control scheme. The following theorem summa-
Since ∆ i ( λ ) is Hurwitz, the ni × ni matrix Fi is also rizes the main result of this study.

3367
 ThB17.3

Theorem 1: Consider the nonlinear system (10) controlled by xd − + x u

the modified AFCMAC scheme (21)-(25) and (13). If the y ( n ) = f ( x ) + G ( x )u
uK
x
design parameters ( 1 ,  2 , 11 , 1 2 , b) are positive, u K ≡ G −1 ª¬ y (dn ) − Cx º¼ + + uA
and the parameters (11 , 1 2 ) are chosen such that T
e = ( x PG ) T u AFCMAC
m e
§ 3 · λ ( Q) φ = ¦ ei u AFCMAC ˆ )T
T
= (h AW
max ¨ σ 1i , σ 2i , for i = 1,", m ¸ < min , where λmin (< ) i =1
+
+ u AR
© 4 ¹ λmax ( P) - ≡ diag ¬ª φ1 ! φm ¼º
ˆ = − W1
W ˆ 1 − ( 1φhT A )T
and λmax ( <) are the minimum and maximum eigenvalues of
the associated matrix, respectively, then
1. all signals in the closed-loop system are bounded. ˆ ⋅ sat b (φ)
u AR = − D
2. the tracking error x converges exponentially to a re-
sidual set, which can be made small by adjusting the pa- ˆ = − D1
D ˆ 2 +  2-, D
ˆ (0) > 0
rameters ( 1 ,  2 , 11 , 1 2 , b) .
Proof: Let w  ≡w i
ˆ − w∗ and D ≡ Dˆ − D for i = 1," , m .
i i i i i
Fig. 3. Architecture of the modified MIMO AFCMAC control scheme.
m
Consider the Lyapunov function candidate V = ¦ Vi , where ª 1
m
b σ
V ≤ ¦ « − y Ti Qi y i + ( D i + Di ) − 1i w  Ti w
i
i =1
i =1 ¬ 2 4 2γ 1i
1 1 T 1 2
Vi ≡ ( y Ti Pi y i +  w
w  + D ) , for i = 1,", m , (29) σ 2i  2 σ 1i * 2 σ 2i 2 º
2 γ 1i i i γ 2i i − Di + wi + Di »
2γ 2i 2γ 1i 2γ 2i ¼ (34)
whose rate along the trajectory of (19), is computed as m
§ 1  w
w T
i 3 D 2
m
ª 1 ≤ ¦ ¨ − y Ti Qi y i − σ 1i i
− ( σ 2i ) i

V = ¦ «− y Ti Qi y i + eTi (u AFCMAC − uFCMAC + uAR + 0)

∗
i =1 © 2 2 γ 1i 4 2γ 2i
i =1 ¬ 2
(30) σ 1i * 2 5σ 2i 2 γ 2i 2 ·
+ wi + Di + b ¸.
1 T 1 º 2γ 1i 8γ 2i 4σ 2i ¹
+ w  + ( Dˆ − Di )Dˆ i »
ˆ w
γ 1i i i γ 2i i ¼ For any β 2 > 0 , the inequality (34) may be re-described as
m
using (23) and V ≡ ¦Vi . The laws (21)-(25), yield m ª 2
y
i =1 V ≤ ¦«−β2Vi + ( β2λmax (Pi ) − λmin (Qi )) i
i=1 «
¬ 2
m
ª 1 (35)
V = ¦ « − y Ti Qi y i − φi Dˆ i sat b (φi ) + φiε i
 Tw · D º
2
i =1 ¬ 2
w  § 3
(31) + ( β2 −σ1i ) i i + ¨ β2 − ( σ2i ) ¸ i » +α2
σ 1i T σ º 2γ1i © 4 ¹ 2γ 2i ¼
+ ( Dˆ i − Di ) φi − w ˆ − 2i D Dˆ
 w
γ 1i i i γ 2i i i »¼ where the constant α 2 is given by
Depending on the values of φi for i = 1,", m , the following m
§ σ 1i 5σ γ ·
α2 ≡ ¦ ¨
2
w*i + 2i Di2 + 2i b2 ¸ > 0 . (36)
three cases are considered separately; i =1 © 2γ 1i 8γ 2i 4σ 2i ¹
Case 1. φi > b , ∀i = 1," , m : Since sat b (φi ) = sgn(φi ) , Case 3. φi > b for i = 1," , k (< m) and φi ≤ b for
the inequality (31) yields i = k + 1,", m : Applying similar techniques in Case 1 and 2
m
ª 1 σ σ σ σ º
V ≤ ¦ «− y Ti Qi y i − 1i w
2
 i − 2i D i2 + 1i w*i + 2i Di2 »
 Ti w for i ≤ k and i > k , respectively, for any β3 > 0 , we find
i =1 ¬ 2 2γ 1i 2γ 2i 2γ 1i 2γ 2i ¼
m ª y i
2
ª V ≤ ¦
2
m
y (32) « − β 3Vi + ( β 3λmax ( Pi ) − λmin (Qi ) )
≤ ¦ «−β1Vi + ( β1λmax (Pi ) − λmin (Qi )) i i =1 « 2
i =1 «
¬ 2 ¬ (37)
 Tw · D º
2
T
 w  D º
2 w  § 3
w
+ ( β1 − σ1i ) i i + ( β1 − σ 2i ) i » + α1 + ( β 3 − σ 1i ) i i + ¨ β 3 − ( σ 2i ) ¸ i » + α 3
2γ 1i 2γ 2i ¼ 2γ 1i © 4 ¹ 2γ 2i ¼
where the constant α1 is defined by where the constants α 3 is defined by
m
§ σ 1i σ ·
k
§ σ 1i σ ·
α3 ≡ ¦ ¨
2
α1 ≡ ¦ ¨
2
w*i + 2i Di2 ¸ > 0 (33) w*i + 2i Di2 ¸ +
i =1 © 2γ 1i 2γ 2i ¹
i =1 © 2γ 1i 2γ 2i ¹ (38)
m
§σ 5σ γ ·
¦ ¨ 1i w*i + 8γ 2i Di2 + 4σ2i b2 ¸
2
Case 2. φi ≤ b , ∀i = 1," , m : Since sat b (φi ) = φi / b , the
i = k +1 © 2γ 1i 2i 2i ¹
inequality (31) leads to
implying α1 ≤ α 3 ≤ α 2 . Let

3368
 ThB17.3

ªQ1 0 0 º ª P1 0 0º ­ yd( 2)1 = −5 y d 1 − 5 yd 1 + 5r1

« » n×n , « » ° (3)
Q ≡ « 0 % 0 »∈R P≡ «0 % 0 » ∈ R n×n . yd2 2 − 2.75 y d 2 − 0.75 yd 2 + 0.75r2
® yd 2 = −3  (45)
«0 0 Q » «0 0 Pm »¼ ° (1)
¬ m¼ ¬
¯ yd 3 = −3 yd 3 + 3r3
λ
It follows that min ( Q ) λ ( Q ) i = 1,", m . Comparing
≤ min i for where r1 , r2 , and r3 are piecewise constant functions. Sup-
λmax ( P ) λmax ( Pi )
pose that the system dynamics in (44) is uncertain so that the
(37) to (32) and (35), one may choose
proposed control scheme is applicable. The design parame-
­ λ (Q) 3 ½
β = β1 = β2 = β3 = min ® min , σ 1i , σ 2i , for i = 1,", m¾ (39) ters, ( 1 ,  2 , 11 , 1 2 , b) are chosen as small as possible ac-
λ
¯ max ( P ) 4 ¿
cording to Theorem 1 as
such that V ≤ − βV + α , which infers that all the signals are
2 1 = diag([10 15 10]),  2 = diag([1 1 1]),
bounded according to [19]. Moreover, if the parameters 11 (46)
11 = 1 2 = diag([0 0 0]), b = 0
and 1 2 are chosen such that With the other parameters being
­ 3 ½ λ ( Q) C1 = [2 3]T , C2 = [60 47 12]T , C3 = 6,
max ®σ 1i , σ 2i , for i = 1,", m ¾ < min (40)
¯ 4 ¿ λmax ( P) Q1 = diag([10 10]), Q 2 = diag([20 20 20]), Q 2 = 3, (47)
and those constants are selected as ˆ (0) = diag([1 1 1]),
d = [1 0.1 1 0.1 0.1 1] , D
­ 3 ½
β = β1 = β 2 = β3 = min ® σ 1i , σ 2i , for i = 1,", m ¾ (41) that the corresponding F , P , and λmin (Q) / λmax (P ) are,
¯ 4 ¿
respectively,
the inequailties (32), (35) and (37) imply that
2 ª 0 1 0 º
x ª0 1º « 0
V ≤ − βV + [ βλmax ( P) − λmin (Q)] + α2 . (42) F1 = « , F
» 2 « = 0 1 » , F3 = 6 (48)
¬−2 −3¼ »
2 «¬−60 −47 −12»¼
Now define the residual set as
ª96.4048 59.381 0.1667º
2 2α j ª12.5 2.5º « »
Γ j ≡ {x : x < } , for j = 1, 2, 3 . (43) P1 = « » , P2 = « 59.381 62.496 1.4762 » , (49)
λmin (Q) − βλmax ( P) ¬ 2.5 2.5¼ «¬ 0.1667 1.4762 0.9563»¼
It is seen that Γ1 ⊆ Γ3 ⊆ Γ 2 , and outside the residual set, we P3 = −6, λmin (Q) / λmax ( P) = 0.0708.
have V ≤ − βV . As a result, the tracking error x converges Three cases are then considered in the simulation.
exponentially to an adjustable residual set Γ 2 , cf. [19]. Case A: Sign function without σ-modification is used. All
The initial tracking errors are normally large, so the initial of the design parameters are set to be those in (47)-(49), and
regime is typically located in Case 1. The error x converges the numerical results are as follows. Figures 4(a) and 4(b) plot
exponentially to Γ1 . As the errors are reduced, the status of the tracking responses and tracking errors (yt ≡ y-yd), re-
spectively. Figure 4(c) shows the control input which is
the system should become Case 3 and then to Case 2. bounded but suffers from severe chattering, because this
Therefore, the steady-state tracking error can be steered to the scheme is designed with switching compensation. Although a
residual set Γ 2 , which can be made arbitrarily small by ad- favorable tracking response is obtained, the chattering asso-
justing the parameters.  ciated with the control efforts reduces tracking accuracy.
Figures 4(d) plot the upper bound of the approximation error.
IV. SIMULATION RESULTS Even if all of the gains are bounded, the drift of the upper
To demonstrate the performance of the proposed modified bounds caused by the drop in the tracking accuracy and the
MIMO AFCMAC scheme, the tracking of the following numerical quantization error cannot be eliminated.
multivariable highly nonlinear system is considered, Case B: The design parameters are chosen the same as
those in Case A, except b =1, and 11 = 12 = diag([0.070.070.07]) .
ª y1(2) º ª−5y1 y1 + 4y2 + 3y2 + 2y2 + y3 º ª1 2 3º ªu º
2

« (3) » « 2 − y2 3
» « 1
» «u » (44) These parameters meet the requirements in (40) and the per-
y
« 2 » = «−
y2 − 2 e − 3sin( y2 ) + 4y1 + 5y1 + 6y3 » +
«
4 5 6
»« 2 » formances are shown in Fig. 4(e)-4(g). Although the tracking
« y3(1) » «−6ln( y3 +1) +5y + 4y + 3y + 2y + y » ¬«7 8 8¼¬ » «u3 »¼
¬ ¼ «¬ 3 1 1 2 2 2»
¼ errors are slightly larger than that of Case 1, the control signal
and the drift phenomena of approximation upper bounds has
where y1 , y2 , y3 and u1 , u2 , u3 are the system outputs and
improved greatly.
inputs, respectively. Let the desired output be generated by Case C: The design parameters are set the same as those
the following equations:

3369
 ThB17.3

1.5
y
1 0.06
ponentially to a residual set, whose size can be adjusted.
1
0.04
yt3 yt1 Simulation results demonstrate the effectiveness of the pro-
tracking responses (y)

tracking errors (yt)

3
0.5 0.02
posed methodology in applying to a multivariable highly
0
0
-0.02 nonlinear system. Application of the method to real nonlinear
y
-0.5 2
-0.04 yt2 system is currently studied, and we hope to report the result in
-0.06
-1

-0.08
the near future.
-1.5
0 1 2 3 4 5 6 0 1 2 3 4 5 6
time, sec time, sec

(a) Tracking responses of Case A. (e) Tracking errors of Case B. REFERENCES

40
0.06 u1 [1] J.J.E. Slotine, and W. Li, Applied Nonlinear Control, Prentice-Hall, 1991.
30
0.04
yt3 yt1

20
[2] P. Ioannou, and J. Sun, Robust Adaptive Control, Prentice Hall, 1996.
tracking errors (yt)

[3] C.H. Wang, T.C. Lin, T.T Lee, and H.L. Liu, “Adaptive Hybrid Intel-
control inputs (u)

0.02
10
0
0 ligent Control for Uncertain Nonlinear Dynamical Systems,” IEEE
-0.02

-0.04
-10
u3
Trans. Systems, Man and Cybernetics, Part B, vol. 32, no. 5, pp.
-0.06
-20
583-597, 2002.
yt2 -30 u2
-0.08 [4] C.T. Lin, and C.S. George Lee, Neural Fuzzy Systems: a Neuro-Fuzzy
-40
0 1 2 3
time, sec
4 5 6 0 1 2 3
time, sec
4 5 6 Synergism to Intelligent Systems, Prentice-Hall, 1996.
(b) Tracking errors of Case A. (f) Control inputs of Case B. [5] R.-J. Wai, C.-M. Lin, and Y.-F. Peng, “Robust CMAC Neural Network
40 1.5
Control for LLCC Resonant Driving Linear Piezoelectric Ceramic
u1 1.4 Motor,” IEE Proc.-Control Theory and Applications, vol. 150, no. 3, pp.
approximation error bounds (Dh)

30

20
1.3
221-232, 2003.
1.2
control inputs (u)

10
1.1 [6] M.-F. Yeh, and K.-C. Chang, “A Self-Organizing CMAC Network
0
u3
1 With Gray Credit Assignment,” IEEE Trans. Systems, Man and Cy-
0.9
-10
0.8
bernetics, Part B, vol. 36, no. 3, pp. 623-635, 2006.
-20
u2 0.7 [7] L.-X. Wang, A Course in Fuzzy Systems and Control, Prentice-Hall, 1997.
-30
0.6
[8] C.-C. Wong, and J.-Y. Chen, “Fuzzy Control of Nonlinear Systems via
-40 0.5
0 1 2 3
time, sec
4 5 6 0 1 2 3
time, sec
4 5 6
Rule Adjust-ment,” IEE Proc.- Control Theory and Applications, vol.
(c) Control inputs of Case A. (g) Updated approximation 146, no. 6, pp. 578-584, 1999.
error bound of Case B. [9] J.S. Albus, “A New Approach to Manipulator Control: The Cerebellar
1.5 Model Articulation Controller (CMAC),” Trans. ASME, J. Dyn. Syst.
0.06
1.4
yt1 Meas. Contr., vol. 97, pp. 220-227, 1975.
approximation error bound (Dh)

1.3 0.04
yt3
[10] K.-S. Hwang, and C.-S. Lin, “Smooth Trajectory Tracking of Three-Link
tracking errors (yt)

1.2 0.02

1.1 0
Robot: a self-organizing CMAC approach,” IEEE Trans. Systems, Man, and
1
-0.02 Cybernetics, vol. 28, no. 5, pp. 680-692, 1998.
0.9

0.8 -0.04 [11] C.M. Kwan, F.L. Lewis, L. Haynes, and J.D. Pryor, “Robust Spacecraft
yt2
0.7 -0.06 Attitude Control Using Fuzzy CMAC,” in Proc. IEEE Int Conf. on In-
0.6

0.5
-0.08
telligent Control, 1996, pp. 43-48.
0 1 2 3 4 5 6 0 1 2 3 4 5 6
time, sec time, sec [12] C.-M. Lin, Y.-F. Peng, and C.-F. Hsu, “Robust Cerebellar Model Articula-
(d) Updated approximation (h) Tracking errors of Case C. tion Controller Design for Unknown Nonlinear Systems,” IEEE Trans.
error bounds of Case A. Circuits and Systems, vol. 51, no. 7, pp. 354-358, 2004.
[13] Y.H. Kim, and F.L. Lewis, “Optimal Design of CMAC Neural-Network
Fig. 4. Simulation results of the modified MIMO AFCMAC.
Controller for Robot Manipulators,” IEEE Trans. System, Man and Cyber-
in Case A, except b = 0.1 , and netics, Part C, vol. 30, no. 1, pp. 22-31, 2000.
[14] J.-Y. Chen, P.-S. Tsai, and C.-C. Wong, “Adaptive Design of a Fuzzy
11 = diag([0.05 0.05 0.05]), 11 = diag([0.03 0.03 0.03]) . (50) Cerebellar Model Arithmetic Controller Neural Network,” IEE
The performance is the same as that in Case B, except that the Proc.-Control Theory and Applications, vol. 152, no. 2, pp. 133-137,
residual set of the tracking error is smaller. This comparison 2005.
[15] T.-F. Wu, P.-S. Tsai, and F.-R. Chang, “Robust Adaptive Fuzzy CMAC
justifies that the residual set can be made smaller by using Control for Unknown Systems,” in the 16th IFAC World Congress,
smaller 11 , 1 2 and b. Prague, Czech Republic, 2005.
[16] T.-F. Wu, P.-S. Tsai, F.-R. Chang, and L.-S. Wang, “Adaptive Fuzzy
CMAC Control for a Class of Nonlinear Systems with Smooth Com-
V. CONCLUSION pensation,” IEE Proc.-Control Theory and Applications, vol. 153, no. 6,
In this study, a modified multivariable adaptive fuzzy pp. 647-657, 2006.
cerebellar model articulation controller (modified MIMO [17] C.-M. Lin, and Y.-F. Peng, “Adaptive CMAC-Based Supervisory
Control for Uncertain Nonlinear Systems,” IEEE Trans. Systems, Man,
AFCMAC) design methodology was developed to solve the and Cybernetics, vol. 34, no. 2, pp. 1248-1260, 2004.
tracking problem for a class of multivariable nonlinear sys- [18] W.-S. Lin, and C.-S. Chen, “Robust Adaptive Sliding Mode Control
tems. The proposed method is based on the CMAC technique, Using Fuzzy Modelling for a Class of Uncertain MIMO Nonlinear
which is incorporated into the THEN-part of a fuzzy rea- Systems,” IEE Proc.-Control Theory and Applications, vol. 149, no. 3,
pp. 193-201, 2002.
soning mechanism. The fuzzy CMAC method was then [19] P.A. Ioannou, and P.V. Kokotovic, “Instability Analysis and Im-
combined with the adaptive law, so the entire control gains or provement Robustness of Adaptive Control,” Automatica, vol. 20, no. 5,
weights can be adjusted on-line, without preliminary off-line pp. 583-594, 1984.
learning. The scheme with saturation compensation prevents
the chattering and steers the tracking error to converge ex-

3370