You are on page 1of 12

388 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. I , NO.

2, MARCH 1996

Frank L. Lewis, Fellow, IEEE, Aydin Yegildirek, and Kai Liu, Senior Member, IEEE

Abstract-A multilayer neural-net(NN) controller for a general not need such initial estimates; in that work, the NN is linear
serial-link rigid robot arm is developed. The structure of the NN in the tunable weights so that suitable basis functions must be
controller is derived using a filtered error/passivityapproach. No first selected (e.g., RBF’s).
off-line learning phase is needed for the proposed NN controller
and the weights are easily initialized. The nonlinear nature of In this paper we extend work in [19]-1211 to confront these
the NN, plus NN functional reconstruction inaccuracies and deficienciesfor the full nonlinear three-layer NN with arbitrary
robot disturbances, mean that the standard delta rule using activation functions (as long as the function satisfies an
backpropagation tuning does not suffice for closed-loop dynamic approximation property and it and its derivatives are bounded).
control. Novel on-line weight tuning algorithms, including cor- The approximation accuracy is generally better in nonlinear
rection terms to the delta rule plus an added robustifying signal,
guarantee bounded tracking errors as well as bounded NN multilayer NN than in linear two-layer NN. Moreover, in
weights. Specific bounds are determined, and the tracking errar comparison to the linear-in-the-parameter NN used in [38],
bound can be made arbitrarily small by increasing a certain [351, 1371, [321, [331, [181, and [21], in the multilayer NN
feedback gain. The correction terms involve a second-order no basis functions are needed; it is only necessary-that an
forward-propagated wave in the backprop network. New NN approximation property be satisfied. This seems to be because
properties including the notions of a passive NN, a dissipative
NN, and a robust NN are introduced. the first layers effectively compute (i.e., “learn”) the basis
functions for the specific application while the last layer
combines them appropriately.
I. INTRODUCTION
In this paper, some notions in robot control [17] are tied to
UCH has been written about neural net (NN) for system some notions in NN theory. The NN weights are tuned on-line,
identification (e.g., [5], [ 121, and [27]) or identification- with no “off-line learning phase” needed. Fast convergence of
based (“indirect”) control, little about the use of NN in direct the tracking error is comparable to that of adaptive controllers.
closed-loop controllers that yield guaranteed performance. A novel controller structure that includes an outer tracking
Some results showing the relations between NN and direct loop ensures good performance during the initial period if the
adaptive control [lo], [16], as well as some notions on NN NN weights are initialized at zero. Tracking performance is
for robot control, are given in [2], [14], [151, 1251, [291, and guaranteed using a Lyapunov approach even though there do
[45], see also [24]. not exist “ideal” weights such that the NN perfectly recon-
NN used in the dynamic control loop pose problems not structs a required nonlinear robot function. By “guaranteed”
present when they are used for classification or system iden- we mean that both the tracking errors and the neural net
tification. Persistent problems that remain to be adequately weights are bounded.
addressed include ad hoc controller structures and the inability The controller is composed of a neural net incorporated
to guarantee satisfactory performance of the system in terms of into a dynamic system, where the structure comes from
small tracking error and bounded. NN weights (which ensures some filtered error notions standard in robot control. Unlike
bounded control inputs). Uncertainty on how to initialize the adaptive robot control, where a “regression matrix of robot
NN weights leads to the necessity for “preliminary off-line functions” must be tediously computed from the dynamics
tuning” [25], [7]. Some of these problems have been addressed of each specific arm [6], [17], the NN approach allows the
for the two-layer NN, where linearity in the parameters holds use of activation functions standard in NN theory for any
(e.g., using radial basis function (RBF) nets) [38], [351, 1371, serial-link robot arm. This means that the same NN controller
[32], [33], [18]. Multilayer nets using a projection algorithm works for any rigid robol: with no preliminary analysis. The
to tune the weights have been used for idenufication purposes NN controller affords the design freedom of trading off the
(only) in [32] and [33]. A technique based on deadzone weight complexity (i.e., number of hidden-layer neurons) octhe NN
tuning is given in 143 and 1221. Generally, in all these papers with the magnitude of a certain robust control term added for
initial estimates for the NN weights are needed that stabilize guaranteed stability.
the closed-loop system; such stabilizing weights may be very It is shown that standard delta rule tuning technique using
hard to determine. In [21] a two-layer NN is given that does
backpropagation [431, [44], [36] generally yields unbounded
Manuscnpt received March 28, 1994, revised February 27, 1995 This work NN weights if: 1) the net cannot exactly reconstruct a certain
was supported by NSF Grant IRI-9216545. required nonlinear robot function; 2) there are bounded un-
The authors are with the Automation and Robotics Research Institute, The
University of Texas at Arlington, Ft. Worth, TX 76118 USA known disturbances or modeling errors in the robot dynamics;
Publisher Item Identifier S 1045-9227(96)00162-2. or 3) the robot arm has more than one link (i.e., nonlinear
1045-9227/96$051.00 0 1996 IEEE
LEWIS et al.: MULTILAYER NEURAL-NET ROBOT CCINTROLLER 389

case). Modified weight tuning rules baseld on the delta rule


with backpropagation guarantee tracking and bounded weights
for the general case. The modifications consist of 1) the
e-modification in [26]; 2) a novel term corresponding to a
second-order forward propagating wave in the backprop tuning
network [27]; and 3) a robustifying control signal.
New passivity properties of NN as introduced for linear
nets in [21] are extended to the three-layer nonlinear net. It
is shown that the backpropagation tuning algorithm yields a
passive NN. This, coupled with the dissipativity of the robot Fig. 1. Three-layer NN structure.
dynamics, guarantees that all signals in the closed-loop system
are bounded under additional observability or persistency time to provide suitable performance of the net. That is, the
of excitation (PE) conditions. The modified weight tuning NN should exhibit “leaming-while controlling” behavior.
algorithms given herein avoid the need folr PE by making the Typical selections for o(.)include, with z E R
NN robust, that is, strictly passive in a sense defined herein. 1
.(z) = ___ sigmoid
+
1 ea2 ’
11. BACKGROUND 1 - e--cUZ
.(z) = ___ hyperbolic tangent (tanh)
Let R denote the real numbers, Rn denote the real n- 1+ e-aZ ’
vectors, and R-” the real m x n matrices. Let S be a = e-(z-mjI2 /s,, radial basis functions (RBF).
compact simply connected set of Rn. With map f S -+ Rm,
The NN equation may be conveniently expressed in ma-
define Cm(S) as the space such that J’ is continuous. We
trix format by defining z = [ z o x l z 2 . . . z ~ l ] ~ ,=y [ y ~
denote by I( . (1 any suitable vector norm. When it is required
y2 . . . y N 3 I T , and weight matrices W T = [wi3],V T = [‘ujk].
to be specific we denote the p-norm by I( ( I p . The supremum
Including xo = 1 in z allows one to include the threshold
1

norm of f(x) (over S) is defined as [3]


vector [B,1B,2 . . . B , N ~ ] ~as the first column of V T ,so that
VT contains both the weights and thresholds of the first- to
second-layer connections. Then
Given A = [ a i j ], B E Rmn the Frobenius norm is defined
y = WTc(VTx) (2)
by
where, if z = [ z l z 2 e..]’ a vector we define o ( z ) = [o(z1)
o ( z 2 ). . .IT. Including one as a first term in the vector o ( V T z )
allows one to incorporate the thresholds Ow, as the first column
with tr( ) the trace. The associated inner product is ( A ,B ) F =
of W T .Any tuning of W and V then includes tuning of the
tr(ATB).The Frobenius norm is nothing but the vector two-
thresholds as well.
norm over the space defined by stacking the matrix columns
Although, to account for nonzero thresholds, x may be
into a vector. As such, it cannot be defined as the induced
augmented by xo = 1 and o b the constant first entry of
matrix norm for any vector norm, but is compatible with the
two-norm so that I(Axllz 5 [[A((~((z112, with A E Rmn and one, we loosely say that x E Rdl and g: RN2 4 R N 2 .
A general function f(x) E Cm(S) can be written as
x E R”.
When x ( t ) E Rn is a function of time we may use the +
f(z) = W T e ( V T z ) .(x) (3)
standard Lp norms [17]. We say z ( t ) is bounded if its L ,
with N I = n, N3 = m, and E(.) a NN functional reconstruc-
norm is bounded. We say A(t) E Rmx” is bounded if its
tion error vector. If there exist N2 and constant “ideal” weights
induced matrix cm-norm is bounded.
W and V so that E = 0 for all x E S, we say f(x) is in the
functional range of the NN. In general, given a constant real
A. Neural Networks
number EN > 0, we say f(x) is within EN of the NN range if
Given x E R N 1 ,a three-layer NN (Fig. 1) has a net output there exist N2 and constant weights so that for all x E Rn,
given by ( 3 ) holds with 1 1 ~ l <l E N .
Various well-known results for various activation functions
a(.),based, e.g., on the Stone-Weierstrass theorem, say that
any sufficiently smooth function can be approximated by a
2 =l,...,N3 (1) suitably large net [8], [13], [31], [38]. The functional range of
NN (2) is said to be dense in C”(S) if for any f E C m ( S )
with e(.) the activation function, ‘ujk the first-to-second layer and E N > 0 there exist finite N z , and W and V such that ( 3 )
interconnection weights, and w i j the second-to-third layer holds with ((&I[ < E N , N= ~ n,N3 = m. Typical results are
interconnection weights. The O,, Ownz, m = 1,2, . . . , are like the follovhg, for the case of e the “squashing functions”
threshold offsets and the number of neurons in layer l is Ne, (a bounded, measurable, nondecreasing function from the real
with N2 the number of hidden-layer neurons. In the NN we numbers onto [0, l]), which include for instance the step, the
should like to adapt the weights and thrlesholds on-line in real ramp, and the sigmoid.
390 EEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 7, NO. 2, MARCH 1996

Theorem 2.1: Set Nl = n, N3 = m and let be any gravity vector, and F ( 4 ) the friction. Bounded unknown dis-
squashing function. Then the functional range of NN (2) is turbances (including, e.g., unstructured unmodeled dynamics)
dense in C” ( S ). are denoted by ‘ r d , and the control input torque is r(t).
In this result, the metric defining denseness is the supremum Given a desired arm trajectory q d ( t ) E Rn the tracking
norm. Moreover, the last layer thresholds w are not needed error is
for this result. The issues of selecting 0, and of choosing N2
for a specified S c Rn and E N are current topics of research 44 = q d ( 9 - Q(t). (8)
(see, e.g., [28] and [31]). In standard use in robotics is the filtered traclung error

B. Stability and Passive Systems r=&+he (9)


Some stability notions are needed to proceed. Consider the where A = AT > 0 is a design parameter matrix,
nonlinear system selected diagonal. Differentiating r ( t ) and using (7), the arm
dynamics may be written in terms of the filtered tracking error
U , t ) , y = h ( z ,t )
k = f(z, as
with state z ( t ) E Rn.We say the solution is uniformly M;. = -V,r -T +f +7d (10)
ultimately bounded (UUB) if there exists a compact set U C
Rn such that for all to) = ~0 E U ,there exists an E > 0 where the nonlinear robot function is
and a number T ( Ezo), such that (lrc(t)I)< E for all t 2 t o +T.
As we shall see in the proof of the theorems, the compact set f(z)= M(q)(id + A & ) f vm(q, g)(gd -tne) G(q) + + F(4)
U is related to the compact set on which NN approximation (1 1)
property (3) holds. Note that U can be made larger by selecting
more hidden-layer neurons. and, for instance, we may select
Some aspects of passivity will subsequently be important T .T T .T ..T T
[Ill, [16], [17], [41]. A system with input u(t)and output y ( t )
z = [e e q d Qd Qd 1 * (12)
is said to be passive if it verifies an equality of the so-called Define now a control input torque as
“power form”
r, = f”+
K,r (13)
L ( t ) = yTu - g ( t ) (4)
gain matrix K, = KT>O and f(z)an estimate of f ( z )
with L ( t ) lower bounded and g ( t ) 2 0. That is provided by some means not yet disclosed. The closed-loop

1’ y T ( r ) u ( r d) r 2
LT g ( r ) d r - y2 (5)
system becomes
M+=-(K,+K,)rt-jf~d 3 -(K,+V,)r+& (14)
for all T 2 0 and some y 2 0. where the functional estimation error is given by
We say the system is dissipative if it is passive and in
addition f = f -f. (15)

I” y T ( r ) u ( r d) r # 0 implies (6)
This is an error system wherein the filtered tracking error
driven by the functional estimation error.
The control T~ incorporates a proportional-plus-derivative
iq

A special sort of dissipativity occurs if g ( t ) is a monic


(PD) term in K,r = K,(1+ he).
quadratic function of ( ( z(1 with bounded coefficients, where
In the remainder of the paper we shall use (14) to focus on
~ ( tis) the intemal state of the system. We call this state-
selecting NN tuning algorithms that guarantee the stability of
strict passivity, and are not aware of its use previously in the
the filtered tracking enor r ( t ) . Then, since (9), with the input
literature (although cf. [ 111). Then the Lz norm of the state is
considered as ~ ( and
t ) the output as e(t) describes a stable sys-
overbounded in terms of the La inner product of output and
tem, standard techniques [23], [41] guarantee that e ( t ) exhibits
input (Le,, the power delivered to the system). This we use to
stable behavior. In fact, llellz 5 llrl12/~mln(A),]]&I12 I l1r1I2,
advantage to conclude some intemal boundedness properties
with umln(h)the minimum singular value of A. Generally A
of the system without the usual assumptions of observability
is diagonal, so that umm(h)is the smallest element of A.
(e.g., persistence of excitation), stability, etc.
The following standard properties of the robot dynamics are
required [17] and hold for any revolute rigid serial robot arm.
C. Robot Arm Dynamics Property I : M ( q ) is a positive definite symmetric matrix
The dynamics of an n-link robot manipulator may be bounded by
expressed in the Lagrange form [17]
mll I M(q)5 m d
+
M ( Q ) i V,(q, 4)4 + G(q) 4- F ( Q )+ Td = 7 (7)
with m l , mz known positive constants.
with q ( t ) E Rn the joint variable vector, M ( q ) the iner- Property 2: V”(q, 4) is bounded by wb(q)11411, with Wb(9) E
tia matrix, Vm(q,4) the corioliskentripetal matrix, G(q) the C’(S).
LEWIS et al.: MULTILAYER NEURAL-NET ROBOT CONTROLLER 391

Property 3: The matrix ~- ZV, is skew-symmetric. Fact 3: For each time t , x ( t ) in (12) is bounded by
Property 4: The unknown disturbance satisfies 11Td < b d ,
with bd a known positive constant. 11x11 5 ‘ c2llrll
ciQd r (19)
Property5: The dynamics (14) from &(t)to r ( t ) are a for computable positive constants c, (c2 decreases as A
stat- strict passive system. increases).
ProofofProperty 5: See [21]. The next discussion is of major importance in this paper;
it is the key to extending linear NN results to nonlinear
111. NN CONTROLLER NN’s. Proper use of these Taylor series-based results gives
a requirement for new terms in the weight tuning algorithms
In this section we derive a NN controller for the robot for nonlinear NN’s that do not occur in linear NN’s.
dynamics in Section 11. We propose various weight-tuning Let V ,I@ be some estimates of the ideal weight values, as
algorithms, including standard backpropagation. It is shown provided for instance by the weight tuning algorithms to be
that with backpropagation tuning the NN can only be guaran- introduced. Define the weight deviations or weight estimation
teed to perform suitably in closed loop under unrealistic ideal errors as
conditions (which require, e.g., f(x) linear). A modified tuning
algorithm is subsequently proposed so that the NN controller v=v-V,
A -

W=W-k, z=z-z (20)

.
performs under realistic conditions.
and the hidden-layer output error for a given x as
Thus, assume that the nonlinear robot function (11) is given
by an NN as in (3) for some constant “ideal” NN weights W 6= - 6 G o ( V T x )- .(V%). (21)
and V, where the net reconstruction error ~ ( xis) bounded by
a known constant E N . Unless the net is “minimal,” suitable The Taylor series expansion for a given II: may be written as
“ideal” weights may not be unique [1], [42]. The “best”
weights may then be defined as those which minimize the
.(VTx) = . ( V T x ) + a’(VT2)VTz+ O ( V T x ) 2 (22)
supremum norm over S of E ( z ) . This issue is not of major with d ( 2 ) and O ( Z ) denoting
do(z)/dzIZ=~, ~ terms of
concern here, as we only need to know that such ideal weights order two. (Compare to [33] where a different Taylor series
exist; their actual values are not required. was used for identification purposes only.) Denoting 3’ =
According to Theorem 2.1, this mild approximation assump- a ’ ( c T x ) ,we have
tion always holds for continuous functions. This is in stark
contrast to the case for adaptive control, where approximation
assumptions such as the Erzberger or linear-in-the-parameters Different bounds may be put on the Taylor series higher-
assumptions may not hold. The mildness of this assumption order terms depending on the choice for .(.). Noting that
is the main advantage to using multilayer nonlinear nets over
linear two-layer nets. O(G’T~)2= [.(V’x) - o ( V T z ) ]- a ’ ( Q T ~ ) V * x (24)
For notational convenience define the matrix of all the
we take the following.
weights as
Fact 4: For sigmoid, RBF, and tanh activation functions,

z=[y ;]. the higher-order terms in the Taylor series are bounded by

A. Some Bounding Assumptions and Facts where ci are computable positive constants.
Fact 4 is direct to show using (19), some standard norm
Some required mild bounding assumptions are now stated.
inequalities, and the fact that .(.) and its derivative are
The two assumptions will be true in every practical situation,
bounded by constants for RBF, sigmoid, and tanh.
and are standard in the existing literature. The facts are easy
The extension of these ideas to nets with greater than three
to prove given the assumptions.
layers is not difficult, and leads to composite function terms in
Assumption 1: The ideal weights are bounded by known
the Taylor series (giving rise to backpropagation filtered error
positive values so that IJVJJF5 V M IJW)JF
, 5 W M ,or terms for the multilayer net case-see Theorem 3.1).

B. Controller Structure and Error System Dynamics


with ZM known. Define the NN functional estimate of (11) by
Assumption 2: The desired trajectory is bounded in the
sense, for instance, that
j(x) = I@T.(VTx) (25)
with ?,I@ the current (estimated) values of the ideal NN
weights V,W as provided by the tuning algorithms subse-
(18) quently to be discussed. With T~ defined in (13), select the
control input
where Qd E R is a known constant. 7- = To - II = I@T.(PTx) + K,r - ‘U (26)
392 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 7, NO. 2, MARCH 1996

Fact 5: The disturbance term (31) is bounded according to

IIflxl with C, computable known positive constants.

N- Robust
-V
RoboticSywcm --c
9 C. Weight Updatesfor Guaranteed Tracking Performance
We give here some NN weight-tuning algorithms that guar-
antee the tracking stability of the closed-loop system under
Fig. 2. NN control structure. various assumptions. It is required to demonstrate that the
tracking error r ( t ) is suitably small and that the NN weights
with w ( t ) a function to be detailed subsequently that provides
p,I@ remain bounded, for then the control ~ ( tis)bounded.
The key features of all our algorithms are that stability is
robustness in the face of higher-order terms in the Taylor
guaranteed, there is no off-line learning phase so that NN
series. The proposed NN control structure is shown in Fig. 2,
control begins immediately, and the NN weights are very easy
where q s [qTiTIT,g = [eTkTIT.
to initialize without the requirement for “initial stabilizing
Using this controller, the closed-loop filtered error dynamics
weights.”
become
Ideal Case-Backpropagation Tuning of Weights: The next
+ +
M+ = - ( K , Vm)r w T o ( v T x )- I @ ~ ~ ( V ’ ~ C result ) details the closed-loop behavior in a certain idealized
+ + +(E 7d) 21. case that demands: 1) no net functional reconstruction error;
2) no unmodeled disturbances in the robot arm dynamics; and
Adding and subtracting W T &yields 3) no higher-order Taylor series terms. The last amounts to the
+ + +
M + = - ( K , Vm)r WTe WTG ( E T d ) v + + + assumption that f(x) in (10) is linear. In this case the tuning
rules are straightforward and familiar. Our contribution lies in
with 6 and 0 defined in (21). Adding and subtracting now
the proof and the conditions thus determined showing when
W T 8 yields the algorithm works, and when it cannot be relied on.
+ +
M+ = -(K, Vm)T IvTe + + WT6 Theorem 3.1: Let the desired trajectory be bounded and
+ + +(E 7d) 21. (27)
suppose the disturbance term w l ( t ) in (28) is equal to zero.
Let the control input for (7) be given by (26) with w ( t ) = 0
The key step is the use now of the Taylor series approxi- and weight tuning provided by
mation (23) for 6, according to which the closed-loop error
system is W =FerT (33)
M+ = - ( K , +Vm)r+WT6+fiT6’VTx+ w1+U (28) V = Gz(6-/Tl@r)T (34)
where the disturbance terms are and any constant positive definite (design) matrices F, G.Then
+ W T O ( V T ~+
w l ( t ) = T,1/T6-/V’T~ ) 2( E +~ d ) . (29) ) to zero with t and the weight
the tracking error ~ ( tgoes

Unfortunately, using this error system does not yield a com-


estimates e,& are bounded.
Proofi Define the Lyapunov function candidate
pact set outside which a certain Lyapunov function derivative
is negative. Therefore, write finally the error system L = ;T’MT+ i t r + i t r (V’G-lV).
(WTF1r/t.) (35)
M + = -(K, + Vm)r+ W T ( e- &-’GTz) Differentiating yields
+
*Ti’VTn; w 21 + +
L -(K, +
Vm)r 51 + (30) L = r T M t + + r T h r + tr ( w T F - ~ & ) + tr ( v T G - ~ V )
where the disturbance terms are whence substitution from (28) (with w1 = 0, ‘U = 0) yields
~ ( t=)WT6’VTn;+ WTO(VTz)’+ ( E + ~ d ) . (31)
L = --rTK,r + +rT(l$f- 2Vm)r + tr @*(P-’& + 6 r T )
It is important to note that the NN reconstruction error
~ ( z ) the
, robot disturbances T d , and the higher-order terms
+ tr V T ( G - l V + zrTfiT3’).
in the Taylor series expansion of f ( x ) all have exactly the The skey symmetry-property makes the second t e p zero,
same influence as disturbances in the error system. The next and since W = W - W with W constant, so that d W / d t =
key bound is required. Its importance is in allowing one to - d W / d t (and similarly for V ) ,the tuning rules yield
overbound w ( t ) at each time by a known computable function;
it follows from Fact 4 and some standard norm inequalities. L = -rTK,r.
LEWIS et al.: MULTILAYER NEURAL-NET ROBOT CONTROLLER 393

Since L > 0 and L 5 0 this shows stability in the sense of or when f ( x ) is nonlinear, cannot be guaranteed to yield
v,
Lyapunov so that T , and I&’(and hence p,I@)are bounded. bounded weights in the closed-loop system.
Moreover General Case: To confront the stability and tracking per-

I” -L d t < 00. (36)


formance of a NN robot arm controller in the thorny general
case, we require: 1) the modification of the weight tuning rules
and 2) the addition of a robustifying term w(t). The problem in
LaSalle’s extension [17], [41] is now used to show that this case is that, though it is not difficult to conclude that T ( t ) is
T(t) in fact goes to zero. Boundedness of T guarantees the bounded, it is impossible without these modifications to show
boundedness of e and 6, whence boundediness of the desired that the NN weights are bounded in general. Boundedness of
trajectory shows q , g , x are bounded. Property 2 then shows the weights is needed to verify that the control input T ( t )
boundedness of V,(q,g). Now, i = --2rTK,i, and the remains bounded.
boundedness of M - l ( q ) and of all signals on the right-hand The next theorem relies on an extension to Lyapunov theory.
side of (28) verify the boundedness of k , and hence the The disturbance T d , the NN reconstruction error E , and the
uniform continuity of E. This allows one to invoke Barbalat’s nonlinearity of f(x) make it impossible to show that the
Lemma [17], [41] in connection with (36) to conclude that L Lyapunov derivative L is nonpositive for all r ( t ) and weight
goes to zero with t , and hence that T ( t ) vanishes. values. In fact, it is only possible to show that L is negative
Note that the problem of net weight initialization occurring outside a compact set in the state space. This, however, allows
in other approaches in the literatUte does not arise. In fact, one to conclude boundedness of the tracking error and the
selecting the initial weights W ( 0 ) V
, ( 0 )as zero takes the NN neural net weights. In fact, explicit bounds are discovered
out of the circuit and leaves only the outer tracking loop in during the proof. The required Lyapunov extension is [17,
Fig. 2. It is well known that the PD term K,T in (26) can then Theorem 15-61, the last portion of our proof being similar to
stabilize the plant on an interim basis. A formal proof reveals the proof used in [26].
that K , should be large enough and the initial filtered error Theorem 3.2: Let the desired trajectory be bounded by (18).
~ ( 0 small
) enough. The exact value of K , needed for initial Take the control input for (7) as (26) with robustifying term
stabilization is given in [9], though for practical purposes it is
only necessary to select K,, large.
Note next that (33) and (34) are nothing but the continuous-
time version of the backpropagation algorithm. In the scalar and gain
sigmoid case, for instance

a’(z) = a ( z ) ( l - .(z)) (37) K, C2 (40)

so that with C2 the known constant in (32). Let NN weight tuning


be provided by
P I @ T = .(VTx)[l- a(vi‘x)]I@T (38)

which is the filtered error weighted by the current estimate I@


and multiplied by the usual product involving the hidden-layer
outputs.
Theorem 3.1 indicates when backprop alone should suffice;
with any constant matrices F = FT > 0 , G = GT > 0 ,
namely, when the disturbance w l ( t ) is equal to zero. Ob-
and scalar design parameter K > > . Then, for large enough
serving the first term in (29) reveals that this is a stronger
control gain K,,Athe filtered tracking error r ( t ) and NN
assumption than simply linearity of the robot function f(x) in
weight estimates V ,I@ are UUB, with practical bounds given
(10). That is, even when E(.) = 0 , T d = 0, and f(x) is linear,
specifically by the right-hand sides of (43) and (44). Moreover,
backprop tuning is not guaranteed to afford successful tracking
the tracking error may be kept as small as desired by increasing
of the desired trajectory. Note that f(.) is linear only in the
the gains K, in (26).
one-link robot arm case. The assumption w l ( t ) = 0 means,
Proof: Let the approximation property (3) hold with
moreover, that the NN can exactly appr’oximatethe required a given accuracy EN for all x in the compact set U, =
function over all of Rn.This is a strong assumption.
(~11x115 b,} with b, > c l Q d in (19). Define U, = {rllrll 5
Theorem 3.2 further reveals the failure of simple back-
(b, - C l Q d ) / C z } . Let ~ ( 0 )E U,. Then the approximation
propagation in the general case. In fact, in the two-layer NN
property holds.
case V = 1 (i.e., linear in the parameters), it is easy to
Selecting now the Lyapunov function (35), differentiating,
show that, using update rule (33), the weights W are not
and substituting now from the error system (30) yields
generally bounded unless the hidden-layer output ~ ( xobeys )
a stringent PE condition [18]. In the three-layer (nonlinear)
case, PE conditions are not easy to derive as one is faced with
the observability properties of a certain bilinear system. Thus,
backpropagation used in a net that cannlot exactly reconstruct
f(x),or on a robot arm with bounded uninodeled disturbances,
394 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. I, NO. 2, MARCH 1996

Fig. 3. Neural net closed-loop error system.

weight errors are fundamentally bounded by ZM (through C3).


The parameter K, offers a design trade-off between the relative
eventual magpitudes of I1~11and l1Zll~.
The first terms of {41) and (42) are nothing but the standard
backpropagation algorithm. The last terms correspond to the
e-modification [26] in standard use in adaptive control to
guarantee bounded parameter estimates; they form a special
sort of forgetting term in the weight updates. The second term
in (41) is novel and bears discussion. The standard backprop
terms can be thought of as backward propagating signals in a
nonlinear "backprop" network [27] that contains multipliers.
The second term in (41) seems to correspond to a forward
which is guaranteed positive as long as either travelling wave in the backprop net that provides a second-
order correction to the weight tuning for W .
(43) Note that there is design freedom in the degree of complex-
ity (e.g., size) of the NN. For a more complex NN (e.g., more
or hidden units), the bounding constants will decrease, resulting

llZ(lF> c 3 / 2 + Jm 5 bz
in smaller tracking errors. On the other hand, a simplified NN
(444) with fewer hidden units will result in larger error bounds; this
degradation can be compensated for, as long as bound E N is
where
known, by selecting a larger value for K, in the robustifying
c3= z, + Cl/'. (45) signal v(t), or for A in (9).
An alternative to guaranteeing the boundedness of the NN
Thus, L is negative outside a compact set. The form of the weights for the two-layer case (i.e., linear in the parameters) is
right-hand side of (43) shows that the control gain K, can presented in [32], [33], and [35], where projection algorithms
be selected large enough so that b, < ( b , - q Q d ) / c Z . Then, are used for tuning W . In the nonlinear case a deadzone tuning
any trajectory ~ ( tbeginning
) in U, evolves completely within algorithm is given in [22].
U,.. According to a standard Lyapunov theorem extension [17],
[26], this demonstrates the UUB of both I1rI1 and 11211~. H Iv. PASSIVITY PROPERTIES OF THE NN
Some remarks are in order. First, since any excursions of
I(rl( or ((21 1~ the bounds given in (43) and (U), The closed-loop ersor system appears in Fig. 3, with the
beyond
respectively, lead to a decrease in the Lyapunov function L , signal c 2 defined as
it follows that the right-hand sides of (f3.3) and (44) can be <z(t) = for error system (28)
taken as practical bounds on llrll and / ( Z / ( Frespectively,
, in
c 2 ( t ) = -wT(3- 3'PTz), for error system (30). (46)
the sense that the norms will be restricted to these values plus
E for any arbitrarily small E > 0. (In the former case, signal w ( t ) should be replaced by z u l ( t ) . )
A comparison with the results of [26] for adaptive control Note the role of the NN, which is decomposed into two
shows that the NN reconstruction error E N , the bounded effective blocks appearing in a typical feedback configuration,
disturbances b d , and the higher-order Taylor series terms, all in contrast to the role of the NN in the controller in Fig. 2.
embodied in the constants CO,C3, increase the bounds on Passivity is important in a closed-loop system as it guar-
Ilrll and llZll~in a very interesting way. Note from (43), antees the boundedness of signals, and hence suitable per-
however, that arbitrarily small tracking error bounds may formance, even in the presence of additional unforseen dis-
be achieved by selecting large control gains K,. (If K, is turbances as long as they are bounded. In general, an NN
taken as a diagonal matrix, KO,,, is simply the smallest cannot be guaranteed to be passive. The next results show,
gain element.) On the other hand, (44) reveals that the NN however, that the weight tuning algorithms given here do in
LEWIS et al.: MULTILAYER NEURAL-NET ROBOT CONTROLLER 395

fact guarantee desirable passivity properties of the NN, and


hence of the closed-loop system.
The first result is with regard to error system (28).
Theorem 4.1: The backprop weight tuning algorithms (33),
(34) make the map from ~ ( tto) -fiT6-, and the map from
r ( t ) to -kT6.’VTz,both passive maps. -
Proof: The dynamics with respect to J@, V are

W = -F3rT (47)
V =- G x ( i F k ~ ) ~ . (48)

1) Selecting the nonnegative function


L = i t r WT~-l$jl
and evaluating L along the trajectories of (47) yields
L = t r r,?rT~-lW= -tr I?.‘T&r?’ = yT(-r;t”&)
which is in power form (4). and evaluating L yields
2) Selecting the nonnegative function
L = t r V T G - ’ ~= TT(-@T6’VTx)
L = i t r VTG-~P
- K1lrll(llVll$ - (v,V ) F )
and evaluating L along the trajectories of (48) yields 5 ‘rT(-%’T8’VTx)- ~ [ ~ ‘ r ~- ~V M(l l~V I ~l F )V ~ ~ $
L = t r VTG-lV==-tr VTz(+.’T@r)T
= rT(-@T$VTx). which is in power form with the last function quadratic
which is in power form. in IlVIlF.

Thus, the robot error system in Fig. 3 is state strict passive It is exactly the special forms of L that allow one to show
(SSP) and the weight error blocks are passive; this guarantees the boundedness of I@ and V when the first terms (power
the dissipativity of the closed-loop system [41]. Using the input) are bounded, as in the proof of Theorem 3.2.
passivity theorem one may now conclude ithat the input-output It should be noted that SSP of both the robot dynamics and
signals of each block are bounded as long iis the external inputs the weight-tuning blocks does guarantee SSP of the closed-
are bounded. loop system, so that the norms of the internal states are
Unfortunately, though dissipative, the closed-loop system bounded in terms of the power delivered to each block. Then,
is not SSP so, when disturbance wl(t) is nonzero, this does boundedness of input-output signals assures state boundedness
not yield boundedness of the internal states of the weight without any sort of observability requirement.
blocks (i.e., I@, V ) unless those blocks are observable, that is We define an NN as passive if, in the error formulation,
persistently exciting (PE). Unfortunately.,this does not yield it guarantees the passivity of the weight-tuning subsystems.
a convenient method for defining PE in ,a three-layer NN, as Then, an extra PE condition is needed to guarantee bound-
the two weight-tuning blocks are coupled, forming in fact a edness of the weights [18]. We define an NN as robust if,
bilinear system. By contrast, PE conditions for the two-layer in the error formulation, it guarantees the SSP of the weight-
case V = I (i.e., linear NN) are easy to deduce [18]. tuning subsystem. Then, no extra PE condition is needed for
The next result shows why a PE condition is not needed boundedness of the weights. Note that 1) SSP of the open-loop
with the modified weight update algorithm of Theorem 3.2; it plant error system is needed in addition for tracking stability
is in the context of error system (30). and 2) the NN passivity properties are dependent on the weight
Theorem 4.2: The modified weight tuning algorithms (41), tuning algorithm used.
(42) make the map from r ( t ) to -WT(6 - 6’VTz),and the
map from r ( t ) to -l@T6’VTx, both SSP maps. v. ILLUSTRATIVE DESIGN
AND SIMULATION
Pro08 The revised dynamics relative to @, V are given A planar two-link arm used extensively in literature for
by illustration purposes appears in Fig. 4. The dynamics are
W = -FerT + F6’VTzrT + K F ~ ~ T ~ ~ %(49)
’ given, for instance in [17]; no friction term was used in this
example. This arm is simple enough to simulate conveniently,
V =- G x ( ~ ? . ’ ~ k +
r ) KGIITIIP.
~ (50) yet contains all the nonlinear terms arising in general n-link
manipulators. The joint variable is q = [q142IT.We should
like to illustrate the NN control scheme derived herein, which
1) Selecting the nonnegative function
will require no knowledge of the dynamics, not even their
L = gtr WTF-’I,?r structure which is needed for adaptive control.
396 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 7, NO. 2, MARCH 1996

Adaptive Contmllu with Unmodclcd D p d u


2, 1

0'

H .

-1.5 ; I I I I
0 2 4 6 8 10
Time (Sec)

Fig. 4. Two-link planar elbow arm. (a)

Adaptive Controller with Unmodeled D p d c a

-1.5 I
0
I
2 4
I
6
I I
8 10 0 2 4 6 8 10
T i e (Sed Time (Sec)

Fig. 6 . Response of adaptive controller with unmodeled dynamics. (a) Actual


Adaptive Controller: Mass Estimates and desired joint angles. (b) Parameter estimates.

v-l
35 I I

The (1, 1) entry of the robot function matrix Yis t : ( i j d l +


15 AI&) +Clg cos q1 (with A = diag{Al, A2)). To demonstrate
1
M- the deleterious effects of unmodeled dynamics in adaptive
[kd control, the term e,g COS q1 was now dropped in the con-
troller. The result appears in Fig. 6 and is unsatisfactory. It
is emphasized that in the NN controller all the dynamics are
-20
0
I
2
I
4
I
6
I
8 10 unmodeled.
T i e (Sec) NN Controller with Backprop Weight Tuning: Some pre-
(b) processing of signals yields a more advantageous choice for
Fig. 5. Response of adaptive controller. (a) Actual and desired joint angles. x ( t ) than (12) that already contains some of the nonlinearities
(b) Parameter estimates. inherent to robot arm dynamics. Since the only occurrences of
the revolute joint variables are as sines and cosines, the vector
Adaptive Controller-Baseline Design: For comparison, a x can be taken as
standard adaptive controller is given by [40]
T=Y$+K,r (51)
where c1 = q d + he,c2 = q d + h e and the signum function
\ir = F Y T r (52)
is needed in the friction terms. The NN controller appears in
with F = FT > 0 a design parameter matrix, Y ( e ,e, q d , &,&) Fig. 2.
a fairly complicated matrix of robot functions that must be The response of the controller (26) (with v ( t ) = 0) with
explicitly derived from the dynamics for each arm,and 9 the backprop weight tuning (e.g., Theorem 3.1) appears in Fig. 7.
vector of unknown parameters, in this case simply the link The sigmoid activation functions were used, and 10 hidden-
masses ml , m2. layer neurons. The values for q d ( t ) , A,F, K , were the same
We took the arm parameters as t 1 = 2! = lm,ml = 0.8 as before, and we selected G = diag(l0,lO). In this case the
kg, m 2 = 2.3 kg, and selected q 1 d ( t ) = sin t , q 2 d ( t ) = cos t , NN weights appear to remain bounded, though this cannot in
Kv= diag {20,20}, F = diag (10, lo}, = A = dfag {5,5). general be guaranteed.
The response with this controller when q(0) = O,Q(O) = The choice of 10 hidden-layer neurons was made as follows.
O,riZ1(0) = O , f i z ( O ) = 0 is shown in Fig. 5. Three simulations were performed, using five, 10, then 15
Note the good behavior, which obtains since there are only hidden-layer neurons. It was observed that going from five-10
two unknown parameters, so that the single mode (e.g., two neurons significantly improved the performance, but going
poles) of q d ( t ) guarantees PE [ l l ] . from 10-15 neurons made no perceptible improvement. It is
LEWIS et al.: MULTILAYER NEURAL-NET ROBOT CONTROLLER 397

- 1 . I 4 -1.5 / I I I I
0 2 4 6 8 10 0 2 4 6 8 10
Time (sac) Time (Sec)

(4

6
’“------

Weight8

...............
-1 I I I I I
0 2 4 6 8 10 0 2 4 6 8 I
Time (Sec) Time (Sec)

(b) (3)
Fig. 7. Response of NN controller with backprop weight tuning. (a) Actual Fig. 8. Response of NN controller with improved weight tuning. (a) Actual
and desired joint angles. (b) Representative weight estimates. and desired joint angles. (b) Representative weight estimates.

Controller without NN: Desired and Actual Joint An&


easy in our software to specify a new number of hidden-layer 1.5 , 1
neurons, and each simulation run took about one minute of
computer time. Similar simulation iterations may be made to
determine effective numbers for the various design gains.
NN Controllerwith Improved Weight Tuning: The response
of the controller (26) with the improved weight tuning in
Theorem 3.2 appears in Fig. 8, where we took K = 0.1. The
tracking response is better than that using straight backprop,
and the weights are guaranteed to remain bounded even though -1.5 ! I I I 1 I
0 2 4 6 8 10
PE may not hold. The comparison with the performance TLne (Sec)
of the standard adaptive controller in Fig. 5 is impressive, Fig. 9. Response of controIler without NN. Actual and desired joint angles.
even though the dynamics of the arm were not required to
implement the NN controller.
No initial NN training or learning phase was needed. The An improved weight tuning algorithm was derived to correct
NN weights were simply initialized at zero in this figure. these deficiencies. The improved algorithm consists of a
To study the contribution of the NN, Fig. 9 shows the backprop term, plus the e-modification term from adaptive
response with the controller T = K,r, t h a t is, with no neural control, plus a novel second-order forward propagating wave
net. Standard results in the robotics literature indicate that from the backprop network. A robustifying control term is
such a PD controller should give bounded errors if K, is also needed to overcome higher-order modeling error terms.
large enough. This is observed in the figure. It is very clear, The improved tuning algorithm makes the NN strictly state
however, that the addition of the NN makes a very significant passive, so that bounded weights are guaranteed in practical
improvement in the tracking performance. nonideal situations.
No NN off-line learning or training phase was needed;
simply initializing the NN weights at zero made for fast
VI. CONCLUSION convergence and bounded errors. Structured or partitioned
A multilayer (nonlinear) NN controller for a serial-link NN’s can be used to simplify the controller design as well
robot arm was developed. The NN controller has a structure as make for faster weight updates [21].
derived from robot control theory passivity notions and offers
guaranteed tracking behavior. Backpropagation tuning was
shown to yield a passive NN, so that it only performs well REFERENCES
under ideal conditions that require linear robot dynamics, no [l] F. Albertini and E. D. Sontag, “For neural nets, function determines
NN reconstructionerror, and no robot unmodeled disturbances. form,” in Proc. IEEE Con$ Decision Contr., Dec. 1992, pp. 26-31.
IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 7, NO. 2, MARCH 1996

E. Alcorta and E. N. SBnchez, “On-line function approximation via [33] -, “Neural networks as on-line approximators of nonlinear sys-
neural networks,” preprint, 1992. tems,’’ in Proc IEEE Con5 Decision Contr., Tucson, AZ, Dec. 1992,
R. G. Bartle, The Elements of Real Analysis. New York Wiley, 1964. pp. 7-12.
F.-C. Chen and H. K. Khalil, “Adaptive control of nonlinear systems [34] T. Poggio and F. Girosi, “A theory of networks for approximation and
using neural networks,” Int. J. Contr., vol. 55, no. 6, pp. 1299-1317, learning,” Artificial Intell. Lab, Massachusettes Inst. Technol., Tech.
1992. Rep. 1140, 1989.
S. R. Chu and R. Shoureshi, “Neural-based adaptive nonlinear system [35] G. A. Rovithakis and M. A. Christodoulou, “Adaptive control of
identification,”Intelligent Control Systems, DSC-vol. 45, ASME Winter unknown plants using dynamical neural networks,” IEEE Trans. Syst.,
Annual Meeting, 1992. Man, Cybem., to appear, 1994.
J. J. Craig, Adaptive Control of Robot Manipulators. Reading, MA: [36] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning internal
Addison-Wesley, 1988. representations by error propagation,” in Parallel Distributed Process-
X. Cui and K. G. Shin, “Direct control and coordination using neural ing, D. E. Rumelhart and J. L. McClelland, Eds. Cambridge, MA:
networks,” IEEE Trans. Syst., Man, Cybem., vol. 23, no. 3, May/June MI’T Press, 1986.
1993. [37] N. Sadegh, “Nonlinear identification and control via neural networks,”
G. Cybenko, “Approximation by superpositions of a sigmoidal func- Control Systems with Inexact Dynamics Models, DSC-vol. 33, ASME
tion,” Math. Contr. Signals Syst., vol. 2, no. 4, pp. 303-314, 1989. Winter Annual Meeting, 1991.
D. M. Dawson, Z . Qu, F. L. Lewis, and J. F. Dorsey, “Robust control [38] R. M. Sanner and J.-J. E Slotme, “Stable adaptive control and recursive
for the tracking of robot motion,” Int. J. Contr., vol. 52, no. 3, pp. identification using radial Gaussian networks,” in Proc. IEEE Con$
581-595, 1990. Decision Contr., Brighton, U.K., 1991.
G. C. Goodwin, “Can we identify adaptive control?’ in Proc. European [39] S. S. Sastry and M. Bodson, Adaptive Control. Englewood Cliffs, NJ:
Contr. Con$, July 1991, pp. 1714-1725. Prentice-Hall, 1989.
G. C. Goodwin and K. S. Sin, Adaptive Filtering, Prediction, and [40] J.-J. E. Slotine and W. Li, “Adaptive manipulator control: A case study,”
Control. Englewood Cliffs, NJ: Prentice-Hall, 1984. IEEE Trans. Automat. Contr., vol. AC-33, no. 11, pp. 995-1003, Nov.
B. Horn, D. Hush, and C. Abdallah, “The state space recurrent neural 1988
network for robot identification,” Advanced Control Issues for Robot [41] -, Applied Nonlinear Control. Englewood Cliffs, NJ: Prentice-
Manipulators, DSC-vol. 39, ASME Winter Annual Meeting, 1992. Hall, 1991.
K. Hornik, M. Stinchombe, and H. White, “Multilayer feedforward [42] H. J. Sussman, “Uniqueness of the weights for minimal feedforward nets
networks are universal approximators,” Neural Networks, vol. 2, pp. with a given inpnt-output map,” Neural Networks, vol. 5, pp. 589-593,
359-366, 1989. 1992.
Y. Iiguni, H. Sakai, and H. Tokumaru, “A nonlinear regulator design in [43] P. J. Werbos, “Beyond regression: New tools for prediction and analysis
the presence of system uncertainties using multilayer neural networks,” in the behavior sciences,” Ph D. dissertation, Committee Appl. Math.,
IEEE Trans. Neural Networks, vol. 2, no. 4, pp. 4 1 M 1 7 , July 1991. Harvard Umv., 1974.
L. G. Kraft and D. P. Campagna, “A summary comparison of CMAC [44] -, “Backpropagation: Past and future,” in Proc. I988 Int. Conf
neural network and traditional adaptive control systems,” in Neural Neural Nets, vol. 1, pp. 1343-1353, 1989.
Networks for Control, W. T. Miller, R. S. Sutton, and P. J. Werbos, [45] T. Yabuta and T. Yamada, “Neural network controller characteristics
Eds. Cambridge, MA: MIT Press, 1991, pp. 143-169. with regard to adaptive control,” IEEE Trans. Syst., Man, Cybem., vol.
Y. D. Landau, Adaptive Control: The Model Reference Approach. New 22, no. 1, pp. 170-176, Jan. 1992.
York: Marcel Dekker, 1979.
F. L. Lewis, C. T. Abdallah, and D. M. Dawson, Control of Robot
Manipulators. New York Macmillan, 1993.
F. L. Lewis, K. Liu, and A. Yegildirek, “Neural net robot controller
with guaranteed tracking performance,” in Proc. IEEE Int. Symp. Intell.
Contr., Chicago, IL, Aug. 1993, pp. 225-231.
F. L. Lewis, A. Yegildirek, and K. Liu, “Neural net robot controller:
Structure and stability proofs,” in Proc. IEEE Con$ Decision Contr.,
San Antonio, TX, Dec. 1993, pp. 2785-2791.
-, “Neural net robot controller: Structure and stability proofs,” J.
Intell. Robotic Syst., to appear.
F. L. Lewis, K. Liu, and A. Yegildirek, “Neural net robot controller
with guaranteed tracking performance,” IEEE Trans. Neural Networks,
to appear. Frank L. Lewis (S’78-M’Sl-SM’86-F’94) was
1 C.-C. Liu and F.-C. Chen, “Adaptive control of nonlinear continuous- born in Wurzbnrg, Germany, subsequently studying
time systems using neural networks general relative degree and MIMO in Chile and Scotland. He received the bache-
cases,” Int. J. Contr., vol. 58; no. 2, pp. 317-335, 1993. lor’s degree in physics/electrical engineering and
1231 K. Liu and F. L. Lewis, “Robust control techniques €or general dynamic
the master’s degree in electrical engineering at Rice
systems,” J. Intell. Robotic Syst., vol. 6, pp. 3 3 4 9 , 1992.
University, Houston, TX, in 1971. In 1977 he re-
[24] W. T. Miller, R. S. Sutton, and P. J. Werbos, Eds., Neural Networks for
Control. Cambridge: MIT Press, 1991.
ceived the M.S. degree in aeronautical engineering
[25] K. S. Narendra, “Adaptive control using neural networks,” Neural from the University of West Florida, Pensacola. In
Networks for Control, W. T. Miller, R. S. Sutton, and P. J. Werbos, 1981 he received the Ph.D. degree at The Georgia
Eds. Cambridge: MIT Press, 1991, pp. 115-142. Institute of Technology in Atlanta, where he was
[26] K. S. Narendra and A. M. Annaswamy, “A new adaptive law for robust employed from 1981 to 1990
adaptation without persistent excitation,” IEEE Trans. Automat. Contr., He spent six years in the U.S Navy, serving as Navigator aboard the frlgate
vol. AC-32, no. 2, pp. 134-145, Feb. 1987. USS Tnppe (FF-1075), and Executive Officer and Acting Commanding Offi-
[27] K. S. Narendra and K. Parthasarathy, “Identificationand control of dy- cer aboard USS Sahnan (ATF-161) He has studied the geometric propertm of
namical systems using neural networks,” IEEE Trans. Neural Networks, the Riccati equatlon and implicit systems, his current interests include robotics,
vol. 1, pp. 4-27, Mar. 1990. intelligent control, nonlinear systems, and manufacturing process control
[28] J. Olvera, X. Guan, and M. T. Manry, “Theory of monomial networks,” Dr Lewis was awarded the Moncrief-O’Donnell Endowed Chair in 1990
in Proc. Symp. Implicit Nonlinear Syst., Dec. 1992, pp. 96-101. at the Automation and Robotics Research Institute of The University of
[29] T. Ozaki, T. Suzuki, T. Furuhashi, S. Okuma, and Y. Uchikawa, Texas at Arlington He is the authorha-author of 80 journal papers and
“Trajectorycontrol of robotic manipulators,”IEEE Trans. Ind. Electron., the books Optimal Control, Optimal Esitmation, Applied Optimal Control and
vol. 38, pp. 195-202, June 1991. Estimatzon, Aircraft Control and Simulation, Control of Robot Manipulators,
[30] Y. H. Pao. Adaptive Pattern Recognition and Neural Networks. Read- and the IEEE repnnt volume Robot Control He is a registered Professional
ing, MA: Addison-Wesley, 1989. Engineer in the State of Texas, Associate Editor of Circuits, Systems, and
[31] J. Park and I.W. Sandberg, “Universal approximation using radial-basis- Slgnal Processing, and the recipient of an NSF Research Initiation Grant, a
function networks,” Neural Computu., vol. 3, pp. 246-257, 1991. Fulbright Research Award, the Amencan Society of Engineering Education
[32] M. M. Polycarpou and P. A. Ioannu, “Identification and control using F E Terman Award, three Sigma Xi Research Awards, the UTA Hallibnrton
neural network models: design and stability analysis,” Dep. Elect. Eng. Engineenng Research Award, and the UTA University-Wide Distinguished
Syst., Univ. S. Cal., Tech. Rep. 91-09-01, Sept. 1991. Research Award.
LEWIS et al.: MULTILAYER NEURAL-NET ROBOT CONTROLLER 399

Aydin Yeyildirek was born in Istanbul, Turkey, in Kai Liu (S’89-M’90-SM’95) received the B.S.
1965. He received the B.S. degree in electronics degree in physics in 1980 and the M.S. degree
and telecommunication engineering from Istanbul in electrical engineering in 1982, both from the
Technical University, Turkey, in 1986, the M.S. Beijing Polytechnic University, Beijing, P.R. China,
degree in electrical engineering and applied physics and the Ph.D. degree in electrical engineering from
from Case Westem Reserve University, Cleveland, the Georgia Institute of Technology, Atlanta, in
OH in 1992, and the Ph.D. degree from The 1990.
University of Texas at Arlington, in 1994. From 1982 to 1986 he was a faculty mem-
He is currently a Research Associate at the her in the Department of Electrial Engineering,
Automation and Robotics Research Institute, Fort Beijing Polytechnic University, where his research
Worth. TX. His research iinterests include neural - techniques of micro-
interests focused on the design
networks, robotics, nonlinear systems, adaptive control, and image processing. computer-based control systems. He is the Head Research Engineer in the
Dr. Yesildirek is a recipient of the Turkish Educational Ministry Advanced Controls Group, the Automation and Robotics Research Institute
Scholarship. of The University of Texas at Arlington. His current research interests include
fuzzy logic and neural nets controls, automation, modeling and control of
dynamic systems, robust control theory, and robotics.
Dr. Liu is a member of Sigma Xi, the IEEE Control Society, and the IEEE
Robotics and Automation Society.