Automatica 37 (2001) 1149}1160

Neural network adaptive robust control of nonlinear systems

in semi-strict feedback form夽
J. Q. Gong, Bin Yao*
School of Mechanical Engineering, Purdue University, West Lafayette, IN 47907, USA
Received 14 June 2000; revised 29 January 2001; received in "nal form 3 March 2001

Neural network adaptive robust control (ARC) design is generalized to synthesize performance
oriented control laws for a class of nonlinear systems in semi-strict feedback forms through the
incorporation of backstepping design techniques.

Abstract

In this paper, the recently proposed neural network adaptive robust control (NNARC) design is generalized to synthesize
performance oriented control laws for a class of nonlinear systems transformable to the semi-strict feedback forms through the
incorporation of backstepping design techniques. Both repeatable (or state dependent) unknown nonlinearities and nonrepeatable
unknown nonlinearities such as external disturbances are considered. In addition, unknown nonlinearities can exist in the virtual
control input channel of each intermediate step as well. All unknown but repeatable nonlinear functions are approximated by outputs
of multi-layer neural networks to achieve a better model compensation for an improved performance. All NN weights are tuned
on-line with no prior training needed. In order to avoid the possible divergence of the on-line tuning of neural networks,
discontinuous projections with "ctitious bounds are used in the NN weight adjusting law to make sure that all NN weights are tuned
within a prescribed range. By doing so, even in the presence of approximation error and nonrepeatable nonlinearities such as
disturbances, a controlled learning is achieved and the possible destabilizing e!ect of on-line tuning of NN weights could be avoided.
Certain robust control terms can then be constructed to attenuate various model uncertainties e!ectively for a guaranteed output
tracking transient performance and a guaranteed "nal tracking accuracy in general. The precision motion control of a linear motor
drive system with slow electrical dynamics is used as a case study to illustrate the proposed NNARC methodology. Experimental
results with a simulated slow electrical dynamics are presented.  2001 Elsevier Science Ltd. All rights reserved.

Keywords: Neural networks; Adaptive control; Robust control; Function approximation; Uncertainty; Nonlinear systems

1. Introduction nonlinearities that can be used to capture those physical

phenomena. These factors make it di$cult to design
It is in general di$cult to treat various nonlinearities a high performance and yet universal enough controller
under a uni"ed framework. In some situations, due to the for general uncertain nonlinear systems. The appearance
limited knowledge about certain nonlinear physical of neural networks (NNs) helps one considerably in the
phenomena, it is also impossible to precisely describe the design of such a needed controller. Theoretically, as long
as a su$cient number of neurons are employed, a neural
network can approximate a continuous function to an
夽
This paper was not presented at any IFAC meeting. This paper was arbitrary accuracy on any compact set (Funahashi, 1989;
recommended for publication in revised form by Guest Editors Hornik, 1991; Cybenko, 1989; Park & Sandberg, 1991).
F.L. Lewis and K.S. Narendra. This universal approximation capability of neural net-
 Supported by the National Science Foundation under the CA- works has enabled researchers to model certain complex
REER grant CMS-9734345. nonlinear physical phenomena e!ectively through vari-
* Corresponding author. Tel.: #1-765-494-7746; fax: #1-765-494-
0539. ous judicious use of NNs. It is thus natural and of
E-mail addresses: [email protected] (J. Q. Gong), practical signi"cance to investigate the proper way
[email protected] (B. Yao). of introducing neural networks in nonlinear controller

0005-1098/01/$ - see front matter  2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 0 1 ) 0 0 0 6 9 - 3
1150 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160

designs for an improved control performance, which is design a robust adaptive controller (RAC) for multi-link
the focus of the paper. rigid robots. Radial basis function (RBF) networks were
In the research "eld of neural networks, focus has been used in Sanner and Slotine (1992). Systems in normal
on the investigation of various NN characteristics, such form were considered in Tzirkel-Hancock and Fallside
as network structure, stability, convergence, and unique- (1992), and cerebellar model articulation controller
ness of weights, etc. Many works have been done on the (CMAC) neural network was studied in Commuri and
stability of a variety of neural networks (Kelly, 1990; Lewis (1997) for MIMO systems in Brunowskii canonical
Liang & Yamaguchi, 1997) and the evolution of the form. In Polycarpou (1996) and Zhang et al. (1996), with
weights of neural networks (Sussmann, 1992; Hirsch, the
-modi"cation type weight-tuning laws, backstep-
1989). However, in these papers, in order to guarantee the ping method was used to derive adaptive neural network
stability of neural networks and/or the uniqueness of the control schemes for nonlinear systems with un-matched
weights, the NN weights have to satisfy some restrictive model uncertainties. In all these works, the input-hidden
conditions, which may limit the approximation capabil- weights of neural networks are assumed to be known. It
ity of neural networks in practice since weights can only may be bene"cial if this assumption can be relaxed so
be tuned in a relatively small region. Fortunately, when that one can fully explore the generality and #exibility of
neural networks are used for control design purposes, the neural networks. Furthermore, since the
-modi"cation
main focus is on the performance of the closed-loop type of weight tuning method was used (Lewis, Yesidirek,
system in terms of output tracking as long as all signals & Liu, 1995; Polycarpou, 1996; Commuri & Lewis, 1997),
are bounded. Whether or not the NN weights converge asymptotic output tracking cannot be achieved even
to their ideal values may not be the key issue. As such, the when the unknown nonlinear function is in the functional
NN weights can be tuned in a relatively large region. range of the neural network. In other words, the perfect
Consequently, the approximation range of a neural net- learning capability of neural networks in the ideal case is
work becomes large, and a better approximation capabil- not guaranteed. In addition, transient tracking perfor-
ity can be expected, which is helpful in the control of mance is in general not known. Transient period may be
nonlinear systems when little is known about the nonlin- long and large transient tracking errors may exhibit,
earities in the system. Thus, in this paper, not much which may not be suitable for practical applications.
attention will be paid to the convergence of weights of Recently, the adaptive robust control (ARC) approach
neural networks, as long as the boundednesses of all the has been proposed in Yao (1997) and Yao and Tomizuka
signals in neural networks are guaranteed. (2001) for nonlinear systems in the presence of both
Neural networks have been used in the control "eld parametric uncertainties and uncertain nonlinearities
recently (Hunt, Sbarbaro, Zbikowski, & Gawthrop, such as disturbances. The resulting ARC controllers
1992) and various results have been obtained (Lewis, achieve a guaranteed output tracking transient perfor-
Yesidirek, & Liu, 1995; Polycarpou, 1996; Zhang et al., mance and "nal tracking accuracy in general. In addi-
1996; Ioannou, & Chien, 1996; Sanner & Slotine, 1992; tion, in the presence of parametric uncertainties only,
Tzirkel-Hancock & Fallside, 1992; Commuri & Lewis, asymptotic output tracking is achieved. These strong
1997). A survey of the application of neural networks to performance results achieved by ARC controllers moti-
control "eld was given in Hunt, Sbarbaro, Zbikowski, vate us to investigate whether the essential idea of ARC
and Gawthrop (1992), where modeling, identi"cation, approach can be extended to the NN-based controller
and control of nonlinear systems via neural networks designs to further improve the achievable performance of
were discussed. Two main issues have to be dealt with in NN-based controllers. At the same time, since only
the use of neural networks for nonlinear control design. a special class of unknown nonlinear functions*a linear
Firstly, since the ideal synaptic weights of a neural net- combination of known basis functions with unknown
work for approximating an unknown nonlinear function weights*have been considered in Yao (1997) and Yao
are usually unknown, certain algorithms need to be de- and Tomizuka (1997), such an extension is also of
rived to tune these weights. In terms of control terminol- signi"cant theoretical values since a more general class
ogy, adaptation laws are needed. Secondly, the ideal NN of unknown functions can be dealt with via neural
weights for the neutral network to reconstruct an un- networks.
known nonlinear function exactly may not exist, i.e., the In Gong and Yao (2000b), neural networks and the
unknown nonlinear function to be approximated may ARC design philosophy are integrated to design perfor-
not be in the functional range of the neural network. The mance oriented control laws for single-input}single-
approximation error between the ideal output of a neural output (SISO) nonlinear systems with matched model
network and the true nonlinear function cannot be as- uncertainties in a normal form. The experimental results
sumed to be zero in general although it may be very small on the precision motion control of an epoxy core
within a compact set. Thus, the issue of robustness to the linear motor, Gong and Yao (2001) have demonstrated
approximation errors needs to be considered. In Lewis et the high performance nature of such an integrated neural
al. (1995), backpropagation neural networks were used to network adaptive robust control (NNARC) design. In
 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160 1151

this paper, through the incorporation of backstepping system input, and
(x, t) represents the lumped non-
G
design (Krstic, Kanellakopoulos, & Kokotovic, 1995), repeatable nonlinearities of the ith subsystem, which
the NNARC design will be generalized to synthesize could be state-dependent disturbances.
performance oriented control laws for nonlinear systems Since f (x ) is not assumed to possess any special form,
G G
with unmatched model uncertainties*uncertain nonlin- a three-layer neural network will be employed to approx-
ear systems transformable to the semi-strict feedback imate it for a better performance. Thus, the following
forms (Polycarpou & Ioannou, 1993; Yao & Tomizuka, assumption is made.
2001). The form allows unknown nonlinearities existing
in both system model and the virtual input channel of Assumption 1. The NN approximation error associated
each intermediate step, and the unknown nonlinearities with the nonlinear function f is assumed to be boun-
G
could include nonrepeatable nonlinearities such as ex- ded by  f (x )!w2G u G (V G x G )) G (x )d G , ∀x 3RG,
G G D D D ? D G D G
ternal disturbances as well. All unknown but repeatable i"1,2, n, where  G (x ) is a nonnegative known shape
D G
nonlinear functions will be approximated by the outputs function, d G is an unknown positive constant,
D
of multi-layer neural networks to achieve a better model x G "[x 2 !1]2 is the augmented input vector to the
? G
compensation for an improved performance. All NN neural network (!1 term denotes the input bias),
weights are tuned on-line with no prior training needed. w G "[w G ,2, w G DG ]2 is the hidden-output weight
D D  D P
Discontinuous projection method with "ctitious bounds vector, V G "[* G ,2,* G DG ]23RPDG "G> is the input-
D D  D P
(Gong & Yao, 1999) will be used to make sure that all hidden weight matrix with * G 3RG>", r G is the
D H D
NN weights are tuned within a prescribed range. By number of neurons of the hidden layer, and
doing so, even in the presence of approximation error u G (V G x G )"[g G (*2G x G ),2, g G DG (*2G DG x G )]2 is the
D D ? D  D  ? D P D P ?
and nonrepeatable nonlinearities such as disturbances, activation function vector (Hornik, 1991).
a controlled learning can be achieved to avoid the pos-
sible destabilizing e!ect of on-line tuning of NN weights. Remark 1. According to theorems in Funahashi (1989)
Certain robust control terms can then be constructed to and Hornik (1991), f (x )!w2G u G (V G x G )) G ,
G G D D D ? D
attenuate various model uncertainties e!ectively for ∀x 3A G , i"1,2, n where f is a continuous function,
G D G
a guaranteed output tracking transient performance and A G is a compact set and A G LRG, and  G is an arbit-
D D D
a guaranteed "nal tracking accuracy in general*a track- rarily small positive number. Correspondingly, in As-
ing performance result that existing NN-based robust sumption 1,  G (x) can be chosen as 1 and d G can be
D D
adaptive controllers (Lewis et al., 1995; Polycarpou, arbitrarily small when x 3A G .
G D
1996; Zhang et al., 1996; Sanner & Slotine, 1992) could
not achieve. Furthermore, if the unknown nonlinear Assumption 2. The input gain b (x ) is nonzero with
G G
functions are in the functional range of the neural net- known sign. Thus, without loss of generality, assume
works and the ideal weights fall within the prescribed b (x )*b '0, ∀x 3RG, i"1,2, n, where b , i"
G G JG G JG
ranges, asymptotic output tracking is also achieved to 1,2, n, are known positive constants.
retain the perfect learning capability of neural networks
in the ideal situation*a steady-state tracking perfor- Assumption 3. Nonlinear input gains b (x ) can be
G G
mance that existing NN-based robust adaptive control- approximated by the output of multi-layer neural net-
lers (Lewis et al., 1995; Polycarpou, 1996; Zhang et al., works with b (x )!w2G u G (V G x G )) G (x )d G , ∀x 3RG,
G G @ @ @ ? @ G @ G
1996; Sanner & Slotine, 1992) cannot have. i"1,2, n, where w G , u G , V G ,  G , and d G are de"ned in
@ @ @ @ @
similar ways as in Assumption 1, respectively. It is as-
sumed that the number of neurons in the hidden-layer is
2. Problem formulation r G . The dimensions of w G , u G , and V G are de"ned accord-
@ @ @ @
ingly. Similar to Remark 1, d G can be made arbitrarily
@
The system to be considered in this paper has the small on some compact set A G .
@
following form (Polycarpou, 1996; Yao, 1997; Yao &
Tomizuka, 1997): Assumption 4. The nonrepeatable nonlinearity
is
G
bounded by
x "f (x )#b (x )x #
(x, t), i"1,2, n!1, (1)
G G G G G G> G

(x, t))
G (x , t)d
G (t), i"1,2, n, (4)
x "f (x)#b (x)u#
(x, t), (2) G G
L L L L
where 
G (x , t) is a nonnegative known function of x and
y"x , (3) G G
 time t, and d
G (t) is an unknown but bounded positive
where x "[x ,2, x ]2 is the state vector, x"x , y is time-varying function.
G  G L
the system output, f (x ) and b (x ) are unknown continu-
G G G G
ous functions representing the repeatable state-depen- In the following derivations, w( G "[w( G ,2,w( G DG ]2
D D  D P
dent unknown nonlinearity and the unknown nonlinear represents the estimate of the hidden-output weight vec-
input gain of the ith subsystem respectively, u is the tor of the NN approximating f , w "w( !w is the
G D D D
1152 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160

corresponding estimation error of the hidden-output Tomizuka (1994) where bounds are assumed known. Let
weight vector, VK G "[*( G ,2,*( G DG ]2 is the estimate of ( 夹KI and ( 夹KI be the "ctitious lower and upper bound
D D  D P J S
the input-hidden weight matrix, VI G "VK G !V G is the for 夹 , where 夹 could be any of the unknown weight
D D D KI
corresponding estimation error matrix, and u G is the vector or matrix. Based on these "ctitious lower and
D
shorthand notation for u G (V G x G ). Notations w( G , w G , upper bounds, same as in Sastry and Bodson (1989) and
D D ? @ @
VK G , VI G , and u G are de"ned in the same way. For simpli- Goodwin and Mayne (1989), a discontinuous projection
@ @ @
city of notations, let f *"w2G u G (V G x G ), fK "w( 2G u G mapping Proj( ) can be de"ned as Proj夹K ( )"Proj夹K ( )
G D D D ? G D D KI
䢇 䢇 䢇

(VK G x G ), bH"w2G u G (V G x G ), and bK "w( 2G u G (VK G x G ). with its mkth element being
D ? G @ @ @ ? G @ @ @ ?
Proj夹K ( )"
KI
䢇


3. Preliminaries

夹K "( 夹KI and '0,
KI S KI
䢇
0 if
In this section, the approximation properties of NNs 夹K "( 夹KI and (0, (5)
KI J KI
䢇

and the projection mapping that are needed for the

controller designs are reviewed. otherwise.
KI
䢇

3.1. Approximation properties of neural networks For simplicity of notations, de"ne ( 夹KI "max( 夹KI ,
J
( 夹KI  , and denote ( 夹 "( 夹KI and 夹 "夹KI .
For a neural network with x 3RN> being its input S
 Throughout the paper, all parameter estimates will be
vector, V"[* ,2,* L ]23RPL "N> being its input-
 P updated by the projection type of adaptation laws given
hidden weight matrix, u being the activation function by
vector, w3RK"PL being its hidden-output weight matrix,


the following theorem and lemma hold (Fu, Chang, Proj夹) ( ), if 夹 is a vector
夹) " (6)
Yang, & Kuo, 1997). Proj夹) ( )2], if 夹 is a vector
where  is any diagonal positive-de"nite adaptation rate
Theorem 2. The NN output w2u(Vx ) can be approxi-
 matrix with proper dimension, and is any adaptation
mated by its estimate w( 2u(VK x ) with w2u(Vx )"
  function. Same as in Yao and Tomizuka (1994), it can be
w( 2u( !w 2(u( !u(
VK x )!w( 2u(
VI x #d , where u( "
  ,, shown that the above projection type of adaptation law
u(VK x ), u(
"diagg(
,2, g(
with g(
"g
(*( 2x )"dg (z)/
  P G G G  G has the following nice properties:
dz *G 2( x  , i"1,2, r , and residual term d "!
X L ,,
w 2u(
Vx #wO(VI x ) with O(VI x ) being the sum of the P1. The parameter estimates are always within the
  
higher-order terms. known prescribed bound used in de"ning the pro-
jection mapping, i.e.,
Lemma 3. The residual term d can be bounded by a lin-
,, ( 夹KI )夹K )( 夹KI . (7)
ear-in-parameter function (Fu, Chang, Yang, & Kuo, 1997), J KI S
i.e., d  )
2Y, where the known function vector P2. In addition, if the true parameter 夹 is actually
,,  within the prescribed range, i.e., ( 夹KI )夹 )
Y"[1, x  , w(  x  , VK  x  ]2,
is an un- J KI
  $   $   ( 夹KI , then,
known vector constituting of positive elements,   is the S

䢇

2-norm of a vector , and   denotes the Frobenius norm 夹 2(\Proj夹) ( )! ))0, ∀ , if 夹 is a vector
$
䢇 䢇

of a matrix . 䢇
(8)
Trace 夹 (\ Proj夹) ( )! ) )0
3.2. Projection mapping O
∀ , if 夹 is a matrix
O
Although the weights of the ideal NNs approximating
various unknown nonlinearities are unknown, they
are constants and bounded. Thus, it is assumed that 4. NNARC of nonlinear systems in semi-strict feedback
each element of w G , w G , V G , and V G is bounded, form
D @ D @
i.e.,  DGK )w G ) DGK ,  DGKI )v G ) DGKI , m" Although it is usually assumed that the input-hidden
JU D K SU JT D KI ST
1,2, r G , k"1,2, i#1, i"1,2, n,  @GK )w G ) weights are known (Lewis et al., 1995) and it can be
D JU @ K
 @GK ,  @GKI )v G ) @GKI , m"1,2, r G , k"1,2, achieved by the o!-line training of neural networks, it
SU JT @ KI ST @
i#1, i"1,2, n, where the lower and upper bounds might be more practical and bene"cial if this assumption
 DGK ,  DGK ,  DGKI ,  DGKI ,  @GK ,  @GK ,  @GKI , and can be relaxed and input-hidden weights can be tuned
JU SU JT ST JU SU JT
 @GKI may not be known. It is natural to require that the on-line. In the following, for simplicity, the sigmoid func-
ST
estimates of the weights should be within the correspond- tion will be used as activation functions. Other type of
ing bounds. However, due to the fact that these bounds activation functions (e.g., RBF (Sanner & Slotine, 1992),
may not be known a prior, certain "ctitious bounds have bipolar sigmoid function (Lewis et al., 1995)) can be
to be used (Gong & Yao, 1999). The design is thus worked out in the same way as long as the activation
di!erent from the conventional ARC designs in Yao and functions and their derivatives are bounded functions.
 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160 1153

4.1. Step 1 For notation consistence with later discussions, de"ne

Adp "0. Choose < " z . By using the third equality
 X  
In principle, if all bounds are known, the ARC scheme in (14), the time derivative of < can be written as
X
in Yao (1997) can be applied to the "rst subsystem in (1)
to synthesize a virtual control law  for x so that <Q "b z z #b  z !Adp
  X     Q  
x can track its desired trajectory x (t) with the desired
 B #z Res #b  . (16)
properties stated in Yao (1997). However, in the pro-    Q
posed backstepping NNARC design, since the bounds 4.2. Step i
are unknown, the following NNARC control function is
suggested (Gong & Yao, 2000) At step i, the NNARC design used in the above step
 " # # , will be generalized to construct a virtual control law
 ? Q Q  for x so that x can track its desired virtual control
1 k G G> G
 " [!f K #x ],  "! Q z , law  synthesized in Step i-1 with a desired transient
? bK  B Q 
(9) G\
b
J performance (for consistency, denote  (t)"x (t)).
  B
Let z "x ! . Recursively de"ne the following
where  is the model compensation part, z "x ! G G G\
?   functions for step j (j*2) from those in the previous
x is the tracking error, k "k #c with k '0 and
B Q    steps:
c being a positive function to be speci"ed later, and the

second robust term  is chosen to satisfy the following JH\ 
Q
I (x, t)"
(x, t)!
H\
,
robust control conditions as in Gong and Yao (2000): H H x J
J J
(i) z Res #b (x ) ) , (10)
    Q X



(ii) z  )0, (11) !(u( J !u(
J VK J x J ) H\ z # H\, l)j!1,
 Q D D D ? x H UDJ
H J " J
where UD
(u( J !u(
J VK J x J )z , l"j,
Res "[ f (x )!f K ]#[b (x )!bK ] #
, (12) D D D ? H
       ? 

 
(  !(   ) 


 " 1# UD  UD   
X (   X !x J H\ z w( 2J u(
J # H\, l)j!1,
UD  ? x H D D TDJ
H J " J
TD
 
(  !(   ) 
# 1# U@  U@   x J z w( 2J u(
J , l"j,
(   X ? H D D
U@ 
# d    #d    #d
   (13)
D  X @  X  X
H J "
and  , i"1,2,5, are the positive design constants. U@


XG 
Let z denote the discrepancy between  and x , i.e.,
   !(u( J !u(
J VK J x J ) H\ x z # H\ , l)j!2,
z "x ! . Under the proposed virtual control law @ @ @ ? x J> H U@J
   J
(9), by applying Theorem 2, the closed-loop equation of 
the "rst subsystem is derived as follows: !(u( J !u(
J VK J x J H\ x z # H\
@ @ @ ? x J> H U@J

z "x !x "f (x )#b (x )[z # ]#
!x # (u( J !u(
J VK J x J )z z ,
  B        B @ @ @ ? H\ H
l"j!1,
"b z #b  #Res #b  "b z
   Q   Q  
(u( J !u(
J VK J x J ) z , l"j,
# b  #b  #
@ @ @ ? J? H
 Q  Q 


# [ f !f *#d  ]#[b !bH#d  ] 
  D ,,   @ ,, ? H\ x z w( 2 u(
# H\, l)j!2,
!x
# [!w 2 (u(  !u(
 VK  x  ) ? x J> H @J @J
J T@J
D D D D ? J
! w( 2 u(
 VI  x  ]#[!w 2 (u(  ! u(
 VK  x  ) 
D D D ? @ @ @ @ ? !x J H\ x z w( 2J u(
J
? x J> H @ @
!w( 2 u(
 VI  x  ]  . (14) H J " J
@ @ @ ? ? T@
# H\ #x J z z w( 2J u(
J , l"j!1,
From the last equality of (14), the associated adaptation T@J ? H\ H @ @
functions are chosen as x J  z w( 2J u(
J , l"j.
? J? H @ @
  "(u(  !u(
 VK G x  )z ,   "x  z w( 2 u(
 ,
UD @ @ @ ?  TD ?  D D (17)
  "(u(  !u(  VK  x  )z  ,  "x   z w( 2 u(
 .

U@ @ @ @ ?  ? T@ ? ?  @ @ From (4),
I (x, t) is bounded by 
I (x, t))I
H (x , t)dI
H (t),
H H H
(15) where dI
H (t)"maxd
J (t), l"1,2, j , and I
H (x , t) is any
H
1154 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160

smooth function satisfying is given by

   
 H
I
H *j max 
H (x , t),max H\ 
(x , t), <Q "b z z #
b  z ! z Adp
H x J J XH H H H> J JQ J J J
J J


#z [b  #Res ] . (26)
l"1,2, j!1 . (18) J J JQ J
Proof. Lemma 4 can be proved via direct substitu-
Lemma 4. At step j)i, choose the following virtual con- tions. 䊐
trol function:
Remark 5. The robust term  in (19) can be chosenas
HQ
 " # # ,
H H? HQ HQ  "!k z (27)
HQ HQ H

1 H\  H\ 
 " !f K #
H\ f K #
H\ bK x where k is a gain large enough such that k *
H? bK H x J x J J> HQ HQ
H J J J J
 h /4 min  with h *(1/(b ) u( 
G HG J2H XJG H JH DH


# H\ !bK z , (   w( H   #   ( H   ) ] #
H\  ( 1 / ( b )   
D UD JH H\
/x 
K
t H\ H\ K
u( K (w( K #( K ) , h *[(1/(b )u( H (w( H #
k D D UD H JH @ @
 "! HQ z , (19) ( H ) ]#
H\ [(1/(b ) u( K  (w( K #( K )
HQ b H U@ H? K JH @ @ U@ H\
JH /x  x ]#[(1/(b )u H\  (w( H\ #( H\ )
where k "k #c with k '0 and c being a positive K K> JH @ @ U@
HQ H H H H (( /x )x #z )], h *[(1/(b ) H ]#
function to be specixed later, and the second robust term H\ H\ H H\ H JH D
 is chosen to satisfy the following conditions
H\ [(1/(b ) /x  K ], h *[(1/(b )
HQ K JH H\ K D H JH
(i) z Res #b  ) , (20)  H  ]#
H\ [(1/(b )  K  /x x ]#[(1/
H H H HQ XH @ H? K JH @ H\ K K>
(ii) z  )0, (b ) H\ (( /x )x #z )], h *[(1/(b )
H HQ
(21) JH @ H\ H\ H H\ H JH
I
H ]. By doing so, using similar techniques as in Gong
where and Yao (1999), it can be shown that (20) and (21) are
Res "[ f !f K ]#[b !bK ] #[b !bK ]z satis"ed.
H H H H H H? H\ H\ H\
H\ 
#
H\ [ f K !f ] 4.3. Step n
x J J
J J
H\  Theorem 6. Consider the control law u" with the fol-
#
H\ x [bK !b ]#
I L
x J> J J H lowing adaptation laws
J J
w( H "Projw( DH ( H L H ), VQK H "ProjVK DH [( H L H )2],
 
JH (  !( J  )  JH D UD UD D TD TD
 "
1# UDJ  UD   #

XH ( J  XJ w( H "Projw( @H ( H L H ), VQK H "ProjVK @H [( H L H )2],
J UD  J @ U@ U@ @ T@ T@

 
( J  !( J  )  (28)
; 1# U@  U@  
( J  XJ
U@  where j"1,2, n, and 's are diagonal positive-dexnite
JH JH JH matrices. By choosing c large enough such that
#
d J  
d J   #
dI
J   . H
D  XJ @  XJ  XJ
J J J L bc L
!
H H z!
z Adp )0 (29)
(22) b H H H
Then, the jth error subsystem is H JH H
then, all closed-loop signals are bounded and < is
z "b z #b  !b z !Adp XL
H H H> H HQ H\ H\ H bounded above by
#b  #Res (23) 
H HQ H < (t))exp(!2k t)< (0)# T [1!exp(!2k t)] (30)
with XL T XL 2k T
T
H\  H\ PDJ  where k "mink , j"1,2, n , and  "
HL  .
Adp "
H\ w( #

H\ *( T H T H XH
H w( J DJ *( J DJ K
J D J K D K Proof. See Appendix A. 䊐
H\  H\ P@J 
#
H\ w( #

H\ *( (24)
w( J @J *( J @J K Remark 7. Theorem 6 shows that the proposed NNARC
J @ J K @ K can achieve a guaranteed transient performance and "nal
and the derivative of the augmented positive-dexnite func- output tracking accuracy in general*in the sense that
tion the exponentially converging rate 2k and the bounds of
T
< "< #z, (25) the "nal tracking error index can be adjusted freely by
XH XH\  H
 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160 1155

suitably choosing the design parameters. These results Theorem 9. When the input-hidden weights are known, in
are thus much stronger than those in Lewis et al. (1995), addition to the results in Theorem 6, the proposed NNARC
Polycarpou (1996) and Zhang et al. (1996). In all those law with the output-weights being updated using the adapta-
schemes of Lewis et al. (1995), Polycarpou (1996) and tion functions (36) also guarantees that
Zhang et al. (1996), transient performance is not guaran- B. In the ideal case that f (x )"f H, and b (x )"bH, (i.e.,
G G G G G G
teed. the nonlinear functions being approximated are in the func-
tional range of the corresponding neural networks), asymp-
Remark 8. Noting the recursive de"nition of adaptation totic output tracking is achieved provided that there is
functions in (17), in implementation, the adaptation func- no external disturbance (i.e.,
"0, i"1,2, n), and all
G
tions at the nth step can be written as follows: ideal weights of NNs lie within the xctitious bounds
(i.e., ( DGK )w G )( DGK , ∀m"1,2, r G , ∀i"1,2, n,

 
L\  JU D K SU D
L H "(u( H !u(
H VK H x H ) z !
Jz (31) and ( @GK )w G )( @GK , ∀m"1,2, r G , ∀i"1,2, n).
UD D D D ? H x J> JU @ K SU @
JH H
Proof. See Appendix B. 䊐
 
L\ 
L H "x H z !
Jz w( 2 u(
, (32)
TD ? H x J> DH DH
JH H The conditions in Theorem 9 represent the ideal situ-
L H "(u( H !u(
H VK H x H ) ation where NNs are to be used, and results of Theorem
U@ @ @ @ ?
9 indicate that perfect learning is realized by the pro-
 
L\ 
; z  #z z !
Jx z , (33) posed NNARC for such an ideal situation. It is noted
H H? H H> x H> J>
JH H that such an result cannot be attained by the existing

 
L\  robust adaptive NN controls (Lewis et al., 1995; Polycar-
L H "x H z  #z z !
Jx z w( 2 u(
,
T@ ? H H? H H> x H> J> @H @H pou, 1996; Zhang et al., 1996; Commuri & Lewis, 1997)
JH H
(34)
Remark 10. Same as in Yao and Tomizuka (1997) and
From the de"nition of projection mapping, it is known Krstic et al. (1995), trajectory initialization can be used to
that Proj( ) )) ) . Furthermore, in viewing of Eq. (24), render zero initial state tracking errors (i.e., z (0)"0,
adaptation laws in (28), and adaptation func- H
j"1,2, n, or < (0)"0) to further reduce the transient
tions (31)}(34), it is clear that
L z Adp )z2 XL
H H H tracking error.
Az, where z"[z ,2,z ]2, and A is a known
 L
matrix whose elements are positive and can be
calculated. Thus 5. Experimental case study
Left-hand side of (29))!z2Cz#z2Az
The proposed NNARC design has been applied to the
"!z2(C!A)z, (35) precision motion control of linear motor drive systems
where C"diag[c ,2, c ]. Since all the entries in A are with negligible electrical dynamics in Gong and Yao
 L (2001), where dynamics is of matched model uncertain-
calculable, c 's in C can be chosen accordingly to make
H ties*a special case of the systems studied in this paper.
sure that C!A is a positive-de"nite matrix. Condition
(29) is thus satis"ed. The experimental results in Gong and Yao (2001) have
demonstrated the high performance nature of the
Now consider the special case that the input- proposed NNARC; for example, with an encoder resolu-
hidden weights are known as assumed in Lewis et al. tion of 1 m, for a low-speed back-and-forth move-
(1995), Polycarpou (1996), Zhang et al. (1996), Sanner ment, the position tracking error is kept within $2 m
and Slotine (1992), Tzirkel-Hancock and Fallside (1992) during most of the execution time while the maximum
and Commuri and Lewis (1997). For this case, the ad- tracking error during the entire run is kept within
aptation laws for the input-hidden weights are not $5.6 m. In this section, linear motor drive systems with
needed, and, from (31) and (34), the adaptation functions nonnegligible electrical dynamics will be considered. The
for the output weights are simpli"ed to dynamics of such systems are of un-matched model un-
certainties, which makes the NNARC design in Gong

 
L\ 
L H "u H z !
Jz , and Yao (2001) inapplicable and complicates the devel-
UD D H x J> opment and implementation of NNARC controller sig-
JH H
ni"cantly.
 
L\ 
L H "u H z  #z z !
Jx z . (36)
U@ @ H H? H H> x H> J>
JH H 5.1. Linear motor with nonnegligible electrical dynamics
In such a case, in addition to the results in Theorem 6, the
proposed NNARC also achieve the following stronger The state space model of the linear motor drive
performance results: systems with nonnegligible electrical dynamics can be
1156 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160

presented as follows (Yao & Xu, 1999): 5.3. NN Structures

x "x
  Based on the particular properties of f , b , f , and
1   
b for the linear motor (37), the following NNs are used
x " !Bx #f (x )#f (x ) 
 M       to estimate them:
# [K #K (x )]x #f f K "w(  x #w(  tanh(v( x )
$ $V      D   D  D 
R # w(  sin ( L x )#w(  ,
x "! x #u (37) D     D 
 ¸  *
bK "w(  #w(  sin( L x ),
 @  @    
where x"[x , x , x ]2"[x, x , i]2, x represents the dis-
   f K "w(  x #w(  , bK "w(  , (40)
placement of the inertia load, i is the motor current  D   D   @
amplitude, M is the mass of the inertia load plus the iron where w(  and w(  represent the bias terms of NNs for
core translator, B is the combined coe$cient of the D  D 
f K and f K , respectively. It is noted that bias terms for
damping and viscous friction on the load, f represents  
 bK and bK are absorbed into w(  and w(  , respectively.
the nonlinear friction, f is the cogging force, K is   @  @
  $ The term w(  tanh(v( x ) in f K is used to mimic the
the average force constant, K (x )x represents the ef- D  D  
$V   e!ects of friction (36), and w(  sin((2/0.03)x ) tries to
fects of force ripple, f represents any external distur- D  
  compensate for the e!ects of periodical cogging forces.
bances such as cutting force in machining, R and ¸ are
Similarly, the term w(  sin((2/0.03)x ) is included in
the armature resistance and inductance, respectively. @  
bK to capture the periodic dependency of K on the
It has been shown in Yao and Xu (1999) that both  $V
position. Following the construction of adaptation laws
f and K are periodical function of x. For the
  $V in Section 4, the adaptation laws in Table 1 are obtained
experimental system used, the period is 0.03 m. As such,
for all NN weight estimates.
linear combinations of a set of harmonic functions of
position will be used to approximate the e!ects of
5.4. Parameters
f and K . A simple mathematical model of friction
  $V
f can be written as (Canudas de Wit, Olsson, Astron,
 5.4.1. Parameters of physical model
& Lischinsky, 1995)
The experiments are conducted on an epoxy core lin-


  
x K ear motor. The detailed experimental set-up is given in
f "! f #( f !f ) exp ! sgn(x ), (38)
    x Gong and Yao (2001). It should be noted that the actual

linear motor has a rather fast electrical dynamics (band-
in which f represents the level of static friction, f is the
  width above 1000 Hz), which can be neglected as done in
Coulomb friction, and x and  are empirical parameters
 Gong and Yao (2001). In order to investigate the e!ects
used to describe the Stribeck e!ect. of slow electrical dynamics, which may occur in some
It is seen that system (37) is in the semi-strict feedback types of linear motors, a rather slow electrical dynamics
form (1)}(3) studied previously with f ,0, b ,1,
  having a break frequency of 47 rad/s is added to the

,0; f "1/M[!Bx #f (x )#f (x )], b "
         experimental set-up via software simulation. With this
( 1 / M ) [ K # K ( x ) ] ,
" ( 1 /M ) f ( t , x , x ) ;
$ $V       modi"cations, the nominal values and common
f "!(R/¸)x , b "1/¸, and
"0.
    parameter variation ranges of the system are given in
Table 2 (All parameters are normalized with respect to
5.2. NNARC of linear motor drive systems the unit of input voltage).

With the backstepping NNARC design in Section 4,

the following control functions are obtained:

 "x !(k #k )z
 B Q Q 



L

 
1  
 " !f K #  x #  !z !(k #k )z
 bK  x  t  Q Q 
 
GFFFFFHFFFF! I

?

 
1    
u" " !f K #  f K #  x #  bK x #  !bK z !(k #k )z . (39)
 bK  x  x  x   t   Q Q 
   
GFFFFFFFFFFHFFFFFFFFF }FF}! I

?
 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160 1157

Table 1 NNARC. The trajectory initialization similar to Yao and

Adaptation laws for NN weights
Xu (1999) is also used to minimize the initial transient
fK w(  "Proj ( D [ D x z ], error, in which the "lter is of third order with coe$cients
 D  U U   given by a "120, a "4800, and a "64,000. The in-
  
D  U  
w(  "Proj ( D  D tanh(v(  x )
U D 
itial values of the "ltered reference trajectory are chosen
as x (0)"x (0), x (0)"x (0), and xK (0)"bK (0)x (0)#
B  B  B  
 f K (0).
1
! v( x z ,
cosh(v(  x ) D   
D  The experimental results are shown in Figs. 1}3. It can

 
2 be seen from Fig. 1 that the position tracking error
w(  "Proj ( D  D sin( x )z ,
D  U U 0.03   converges to approximate $20 m very quickly. In fact,
w(  "Proj ( D [ D z ], v(  during most of the execution period, the tracking error is
D  U U  D
within $5 m, which is indicated by the two dashed

 
"Proj ( D  D x z w( 
1
,
lines in Fig. 1. It should be noted that such a tracking
T T   D  cosh(v( x ) accuracy is in the order of encoder resolution of $1 m.
D 
The spikes in Fig. 1 are caused by Coulomb friction since
bK w(  "Proj ( @ [ @ z  ],
 @  U U  ? they occur at the time instances when the velocity cha-
w(  "Proj ( @ [ @ sin ( L x )z  ]
@  U U     ? nges directions. It is not surprise to see these spikes since
fK w(  "Proj ( D [ D x z ],
NNs can only approximate continuous functions to ar-
 D  U U   bitrary accuracy and may not be able to handle discon-
w(  "Proj ( D [ D z ]
D  U U  tinuous nonlinearity like Coulomb friction well. The
bK w(  "Proj ( @ [ @ z  ] actual control input sent to the motor driver is shown in
 @ U U  ?
Fig. 2. It is seen that the control input is quite noisy
during the transient period when the NNs have not
learned well. After the short initial learning period, the
Table 2
input signal becomes less noisy. The estimate of f K shown
Physical parameters of an epoxy core linear motor 
in Fig. 3 reveals the shape of sharp changes when velocity
Names M B K
$
R ¸ changes directions, which demonstrates the good learn-
Units kg N/m/s N/A  mH ing capability of the proposed NNARC in dealing with
Values 0.027 0.273 1.0 1.4 30 hard nonlinearities such as Coulomb friction.
Variations 0.025}0.1 0}0.275 0.5}2 1.25}3 20}50

5.4.2. Controller parameters 6. Conclusion

In experiments, constant feedback gains are used to
simplify the control law for real-time implementation. Neural networks and adaptive robust control design
Speci"cally, k #k "400, k #k "10, and philosophy have been integrated to design performance
Q Q Q Q
k #k "8. The parameters used for adaptation oriented control laws for a class of nonlinear systems in
Q Q
laws are summarized in Table 3, in which bounds of the semi-strict feedback form. Both repeatable (or state
weight estimates are calculated or estimated based on the dependent) unknown nonlinearities and nonrepeatable
variation ranges of physical parameters given in Table 2. unknown nonlinearities have been taken into account.
Within the given bounds, the initial values of all NN Unknown nonlinearities can also exist in the virtual
weight estimates are set randomly. control input channel of intermediate design steps as
well. All unknown but repeatable nonlinear functions are
5.5. Experimental results approximated by outputs of multi-layer neural networks
to achieve a better model compensation. All NN weights
A sinusoidal desired trajectory of x "0.01 sin(2t) is are tuned on-line with no prior training needed. Discon-
*B
used to illustrate the tracing capability of the proposed tinuous projections with "ctitious bounds are used in the

Table 3
Parameters for weight estimates

Estimates w(  w(  v(  w(  w(  w(  w(  w(  w(  w( 
D  D  D D  D  @  @  D  D  @
Lower bounds !11 !1.12 20 !14 !3 5 !2.88 !150 !10 25
Upper bounds 0 !0.168 1000 14 3 80 2.88 !25 10 50
Initial values !5 !0.5 100 0 0 6 0 !100 0 30
Adaptation rates 3.1 1500 600 10 7000 5 200 200 40 200
1158 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160

Fig. 3. Estimate of f .
Fig. 1. Tracking error history. 

L L bc L
)!
k z!
H H z!
z Adp
H H b H H H
H H JH H
L
#
z [Res #b  ]. (A.1)
H H H HQ
H
If (29) is satis"ed, noting (i ) of (20),
L L
<Q )!
k z#
 )!2k < # (A.2)
XL H H XH T XL T
H H
which leads to Theorem 6. 䊐

Appendix B
To prove B of the Theorem 9, choose the following
positive-de"nite function:

 
1 L
< "< #
[w 2H \ H w H #w 2H \H w H ] . (B.1)
Fig. 2. Control input history. ?L XL 2 D UD D @ U@ @
H
When f "f H, b "bH, and
"0, noting (23), (36), the
G G G G G
NN weight adjusting laws to make sure that all NN time derivative of < is derived as follows
?L
weights are tuned within a prescribed range. Conse- L L
quently, a controlled learning may be achieved and the <Q "
z z #
[w 2H \ H w H #w 2H \H w H ]
?L H H D UD D @ U@ @
possible destabilizing e!ect of on-line tuning of NN H H
weights can be avoided. Cer tain robust control terms L L
)!
k z#
z Res
can then be constructed to achieve a guaranteed output H H H H
tracking transient performance and a guaranteed "nal H H
L
tracking accuracy in general. The proposed NNARC #
[w 2H \ H w H #w 2H \H w H ]
strategy is then applied to the precision motion control of D UD D @ U@ @
H
linear motor drive systems with nonnegligible electrical L L
dynamics. Experimental results reveal the high perfor- "!
k z#
[!w 2H L H !w 2H L H ]
H H D UD @ U@
mance nature of the proposed NNARC method. H H
L
#
[w 2H \ H Projw( DH ( H L H )
Appendix A D UD UD UD
H
Since z "0, (26) leads to #w 2H \H Projw( @H ( H L H )]
L> @ U@ U@ U@
L L
<Q "
z b  !z Adp #z [Res #b  ] )!
k z)0 (B.2)
XL H H HQ H H H H H HQ H H
H H
 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160 1159

in which (29) and (ii) in (21) are used in the "rst inequality Sanner, R. M., & Slotine, E. J. -J. (1992). Gaussian networks for direct
in (B.2), and property (8) has been used in the second adaptive control. IEEE Transactions on Neural Networks, 3(6),
inequality of (B.2). It is thus easy to show that z P0 as 837}863.
H Sastry, S., & Bodson, M. (1989). Adaptive control: Stability, convergence
tPR by Barbalat's lemma (Khalil, 1996). 䊐 and robustness. Englewood Cli!s, NJ: Prentice Hall.
Sussmann, H. J. (1992). Uniqueness of the weights for minimal feed-
forward nets with a given input-output map. Neural Networks,
References 5, 589}593.
Tzirkel-Hancock, E., & Fallside, F. (1992). Stable control of nonlinear
Canudas de Wit, C., Olsson, H., Astrom, K. J., & Lischinsky, P. (1995). systems using neural networks. International Journal of Robust and
A new model for control of systems with friction. IEEE Transactions Nonlinear Control, 2, 63}86.
on Automatic Control, 40(3), 419}425. Yao, B. (1997). High performance adaptive robust control of non-
Commuri, S., & Lewis, F. L. (1997). Cmac neural networks for control linear systems: a general framework and new schemes. Pro-
of nonlinear dynamical systems: structure, stability and passivity. ceedings of the IEEE conference on decision and control
Automatica, 33(4), 635}641. (pp. 2489}2494).
Cybenko, G. (1989). Approximation by superpositions of sigmoidal Yao, B., & Tomizuka, M. (1994). Smooth robust adaptive sliding mode
function. Mathematics of Control, Signals and Systems, 2, 303}314. control of robot manipulators with guaranteed transient perfor-
Fu, L.-C., Chang, W.-D., Yang, J.-H., & Kuo, T.-S. (1997). Adaptive mance. Proceedings of the American control conference
robust bank-to-turn missile autopilot design using neural networks. (pp. 1176}1180). (The full paper appeared in ASME Journal of
Journal of Guidance, Control, and Dynamics, 20(2), 346}354. Dynamic Systems, Measurement and Control, 118(4) 764}775, 1996.)
Funahashi, K. -I. (1989). On the approximate realization of continuous Yao, B., & Tomizuka, M. (1997). Adaptive robust control of
mappings by neural networks. Neural Networks, 2, 183}192. SISO nonlinear systems in a semi-strict feedback form Automatica,
Gong, J. Q., & Yao, B. (1999). Adaptive robust control without know- 33(5), 893}900. (Part of the paper appeared in Proceedings of the
ing bounds of parameter variations. Proceedings of the 38th IEEE 1995 American control conference (pp. 2500-2505).)
conference on decision and control, Phoenix, AZ, USA, December Yao, B., & Tomizuka, M. (2001). Adaptive robust control of MIMO
7}10 (pp. 3334}3339). nonlinear systems in semi-strict feedback forms. Automatica, in
Gong, J. Q., & Yao, B. (2001). Neural network adaptive robust control press. (Parts of the paper were presented in the 1995 IEEE confer-
with application to precision motion control of linear motors. ence on decision and control (pp. 2346}2351), and the 1996 IFAC
International Journal of Adaptive Control and Signal Processing world congress, vol. F (pp. 335}340).)
(Accepted for the special issue on Developments in Intelligent Yao, B., & Xu, L. (1999). Adaptive robust control of linear motors for
Control for Industrial Applications). precision manufacturing. Proceedings of the 14th IFAC world con-
Gong, J. Q., & Yao, B. (2000b). Neural network-based adaptive robust gress, Beijing, vol. A (pp. 25}30). (The revised full version will appear
control of a class of nonlinear systems in normal form. Proceedings in the International Journal of Mechatronics.)
of the American control conference, Chicago, IL, USA Zhang, Y., Ioannou, P. A., & Chien, C. C. (1996). Parameter conver-
(pp. 1491}1523). (The revised full paper will appear in the Asian gence of a new class of adaptive controllers. IEEE Transactions on
Journal of Control.) Automatic Control, 41(10), 1489}1493.
Goodwin, G. C., & Mayne, D. Q. (1989). A parameter estimation
perspective of continuous time model reference adaptive control.
Automatica, 23(1), 57}70. Jian Qin Gong received his B.E. and M.E.
Hirsch, M. W. (1989). Convergent activation dynamics in continuous degrees from Xi'an Jiaotong University,
time networks. Neural Networks, 2, 331}349. China, in 1993 and 1996, in Electrical En-
Hornik, K. (1991). Approximation capabilities of multilayer feed- gineering and Systems Engineering, re-
forward networks. Neural Networks, 4, 251}257. spectively. He received another M.E.
degree from National University of Sin-
Hunt, K. J., Sbarbaro, D., Zbikowski, R., & Gawthrop, P. J. (1992).
gapore, Singapore in 1999 in Electrical En-
Neural networks for control systems-a survey. Automatica, 28(6), gineering. He is currently a Ph.D. student
1083}1112. in the School of Mechanical Engineering
Kelly, D. G. (1990). Stability in contractive nonlinear neural networks. at Purdue University, West Lafayette. His
IEEE Transactions on Biomedical Engineering, 37(3), 231}242. research interests include intelligent ve-
Khalil, H. K. (1996). Nonlinear systems (2nd ed.). Englewood Cli!s, NJ: hicle highway system (IVHS), control of
Prentice Hall. rigid/#exible smart material robots, control of linear motor drive sys-
Krstic, M., Kanellakopoulos, I., & Kokotovic, P. V. (1995). Nonlinear tem, control of hard disk drive (HDD), adaptive control, robust control,
and adaptive control design. New York: Wiley. and neural network control.
Lewis, F. L., Yesidirek, A., & Liu, K. (1995). Neural net robot controller
with guaranteed tracking performance. IEEE Transcations on Neural Bin Yao received the Ph.D. degree in
Networks, 6, 703}715. Mechanical Engineering from the Univer-
Liang, X. B., & Yamaguchi, T. (1997). On the analysis of global and sity of California at Berkeley in February
absolute stability of nonlinear continuous neural networks. IEICE 1996. Since 1996, he has been an Assistant
Transactions on Fundamentals of Electronics Communications and Professor in the school of Mechanical En-
Computer Sciences, E80-A(1), 223}229. gineering at Purdue University. Dr. Yao's
Park, J., & Sandberg, I. W. (1991). Universal approximation using research focuses on the development of
radial-basis-function networks. Neural Computation, 3, 246}257. a general framework for the design of high
Polycarpou, M. M. (1996). Stable adaptive neural control scheme for performance adaptive robust control
algorithms. Integration of the approach
nonlinear systems. IEEE Transcations on Automatic Control, 41, with actual sensor properties, physical
447}451. properties, and control-oriented model-
Polycarpou, M. M., & Ioannou, P. A. (1993). A robust adaptive nonlin- ing has been carried out for di!erent applications such as the control
ear control design. Proceedings of the American control conference of electro-hydraulic systems, linear motors, machine tools, and
(pp. 1365}1369). robot manipulators. Dr. Yao is the recipient of the 1997 Caterpillar
1160 J. Q. Gong, B. Yao / Automatica 37 (2001) 1149}1160

Engineering Young Faculty Development Fund for his work on design and coordinated control of intelligent high performance electro-
the electro-hydraulic control, and the 1998 National Science mechanical/electro-hydraulic systems, optimal adaptive and robust
Foundation CAREER award for his work on the engineering synthe- control, nonlinear observer design and neural networks for virtual
sis of high performance adaptive robust controllers for mechanical sensing, modeling, fault detection, diagnostics and adaptive fault-toler-
systems and manufacturing processes. His research interests include ant control, robotics, and data fusion.