Professional Documents
Culture Documents
NETWORK CONTROL
The Kluwer International Series on
ASIAN STUDIES IN COMPUTER
AND INFORMATION SCIENCE
Series Editor
Kai-Yuan Cai
Beijing University ofAeronautics and Astronautics, Beijing, CHINA
Editorial Advisory Board
Han-Fu Chen, Institute of System Science, Chinese Academy of Sciences
Jun-Liang Chen, Beijing University of Post and Telecommunication
Lin Huang, Peking University
Wei Li, Beijing University of Aeronautics and Astronautics
Hui-Min Lin, Institute of Software Technology, Chinese Academy of Sciences
Zhi-Yong Liu, Institute of Computing Technology, Chinese Academy of Sciences
Ru-Qian Lu, Institute of Mathematics, Chinese Academy of Sciences
Shi-Tuan Shen, Beijing University of Aeronautics and Astronautics
Qing-Yun Shi, Peking University
You-Xian Sun, Zhejiang University
Lian-Hua Xiao, National Natural Science Foundation of China
Xiao-Hu You, Southeast University
Bo Zhang, Tsinghua University
Da-Zhong Zheng, Tsinghua University
Bing-Kun Zhou, Tsinghua University
Xing-Ming Zhou, Changsha University of Technology
Also in the Series:
by
S. S. Ge
C. C. Hang
T. H. Lee
T. Zhang
TJ217.S7362001
629.8'3--dc21 2001050337
and
Kai-Yuan Cai
Series Editor
Department ofAutomatic Control
Beijing University ofAeronautics and Astronautics
Beijing 100083
China
I do not know what I may appear to the world, bu.t to myself I seem to have
been only like a boy playing on the seashore, diverting myself in now and then
finding a smoother pebble or a prettier shell than ordinary, while the great ocean
of truth lay all u.ndiscovered before me.
-Isaac Newton
Contents
Preface xiii
1 Introduction 1
1.1 Introduction . 1
1.2 Adaptive Control . 2
1.3 Neural Network Control 3
1.4 Instability Mechanisms in Adaptive Neural Control Systems .5
1.5 Outline of the Book 9
1.6 Conclusion 10
2 Mathematical Preliminaries 11
2.1 Introduction . 11
2.2 Mathematical Preliminaries 11
2.2.1 Norms for Vectors and Signals 12
2.2.2 Properties of Matrix 15
2.3 Concepts of Stability . 16
2.4 Lyapunov Stability Theorem 17
2 ..5 Useful Theorems and Formula. 19
2.5.1 Sliding Surface 19
2..5.2 Mean Value Theorem 20
2.5.3 Integral Formula 20
2.5.4 Implicit Function Theorem 23
2.5.5 Input-Output Stability . 25
2.6 Conclusion 26
8 Conclusion 261
8.1 Conclusion 261
8.2 Design Flexibility . 262
8.3 Further Research 263
References 265
Index 281
Preface
Recent years have seen a rapid development of neural network control tech-
niques and their successful applications. Numerous simulation studies and
actual industrial implementations show that artificial neural network is a good
candidate for function approximation and control system design in solving the
control problems of complex nonlinear systems in the presence of different kinds
of uncertainties. Many control approaches/methods, reporting inventions and
control applications within the fields of adaptive control, neural control and
fuzzy systems, have been published in various books, journals and conference
proceedings. In spite of these remarkable advances in neural control field, due
to the complexity of nonlinear systems, the present research on adaptive neural
control is still focused on the development of fundamental methodologies. From
a theoretical viewpoint, there is, in general, lack of a firmly mathematical basis
in stability, robustness, and performance analysis of neural network adaptive
control systems.
This book is motivated by the need for systematic design approaches for
stable adaptive control using approximation-based techniques. The main objec-
tives of the book are to develop stable adaptive neural control strategies, and to
perform transient performance analysis of the resulted neural control systems
analytically. Other linear-in-the-parameter function approximators can replace
the linear-in-the-parameter neural networks in the controllers presented in the
book without any difficulty, which include polynomials, splines, fuzzy systems,
wavelet networks, among others.
Stability is one of the most important issues being concerned if an adaptive
neural network controller is to be used in practical applications. In this book,
Lyapunov stability techniques playa critical role in the design and stability
analysis of the adaptive systems. Under different operating conditions and
a priori knowledge, stable neural controller designs are presented for several
classes of nonlinear systems, including (i) single-input single-output (SISO)
nonlinear systems, (ii) nonlinear systems in strict-feedback form, (iii) non affine
nonlinear systems, and (iv) multi-input and multi-output (MIMO) nonlinear
in triangular form. Stability of the overall neural network systems is rigorously
xiii
xiv Preface
controller is firstly developed for a class of SISO nonlinear systems. The con-
trol singularity problem, which usually met in feedback linearization adaptive
control, is completely solved using the newly proposed control method. The
developed control schemes ensure global stability of the adaptive systems and
asymptotic convergence of output tracking error. Then, adaptive control design
is developed for strict-feedback nonlinear systems through combining multilayer
NNs with backstepping technique. It is proven that under certain conditions,
the semi-globally uniformly ultimate boundedness is achievable for the closed-
loop adaptive systems. The relationship between the transient performance
and the design parameters is also investigated to guide the tuning of the neural
controller. Finally, adaptive NN Control is presented for a class of MIMO non-
linear systems having triangular structure in control inputs using multi-layer
neural networks. Without imposing any constraints on the system interconnec-
tions, the developed controller guarantees the stability of the adaptive neural
system and the convergence of the mean square tracking errors to small residual
sets.
In Chapter 6, adaptive NN control is investigated for a class of non affine
nonlinear systems. Both state and output (using a high-gain observer for state
estimation) feedback controllers are given for linearly parameterized and mul-
tilayer neural networks, and their effectiveness are verified by numerical simu-
lation.
In Chapter 7, controller design is investigated for several classes of trian-
gular nonlinear systems using quadratic Lyapunov function for its convenience
of analysis and simplicity of the resulting controllers. Firstly, we investigate
a class of systems in strict-feedback form with gn(Xn-l) which is indepen-
dent of X n . This nice properties can be exploited for better controller design.
Secondly, we study the nonlinear strict-feedback systems which include both
parametric uncertainty and unknown nonlinear functions, and constant gi so
that the parametric certainties can be solved using model based adaptive con-
trol techniques and the unknown nonlinear functions be approximated using
NN approximation. Thirdly, we investigate the control problem a class of non-
linear pure-feedback systems with unknown nonlinear functions. This problem
is considered difficult to be dealt with in the control literature, mainly because
that the triangular structure of pure-feedback systems has no affine appear-
ance of the variables to be used as virtual controls. Finally, the extension
from SISO nonlinear systems in triangular forms to MIMO nonlinear systems
in block-triangular forms have also been considered in this chapter.
In summary, this book covers the analysis and design of neural network
based adaptive controller for different classes of nonlinear systems, which in-
clude SISO nonlinear systems, nonlinear systems in strict-feedback and pure-
feedback forms, MIMO nonlinear systems in triangular form, and nonaffine
nonlinear systems. Numerical simulation studies are used to verify the effec-
xvi Preface
tiveness and the performance of the control schemes. This book is aimed at
a wide readership and is a convenient and useful reference for research stu-
dents, academics and practicing engineers in the areas of adaptive control,
neural/fuzzy modelling and control.
For the creation of the book, we are very fortunate to have the appropri-
ate suggestions from and helpful discussions with our colleagues, friends and
co-workers. In particular, we would like to express our sincere gratitude to C.
Canudas de Wit of Laboratoire d'Automatique de Grenoble, K Y. Cai of Bei-
jing University of Aeronautics and Astronautics, C. J. Harris of University of
Southampton, F. L. Lewis of University of Texas at Arlington, 1. M. Y. Mareels
of University of Melbourne, Y. Miyasato of Institute of Statistical Mathematics,
T. Parisini of Politecnico di Milano, M. Polycarpou of University of Cincinnati,
1. Postlethwaite of University of Leicester, J. Si of Arizona State University, M.
W. Spong of University of Illinois at Urbana-Champaign, G. Tao of University
of Virginia, C. W. de Silva of University of British Columbia, H. Wang of the
University of Manchester Institute of Science and Technology, B. Yao, Purdue
University, Y. H. Tan, Guilin Institute of Electronic Technology, A. P. Loh and
J. X. Xu of the National University of Singapore for their constructive and
helpful suggestions and comments.
The first author owes a big thank you to his parents and his wife for their
love, supports and sacrifice throughout the years, to his children, Yaowei Jas-
mine, Yaolong George and Yaohong Lorraine, for their understanding of not
being able to play with them for numerous weekends and evenings.
Last but not the least, the first author would like to thank his current
and former postdoctoral fellows and graduate students, especially Z. Sun, J.
Wang, and G.Y. Li, C. Wang, Z.P. Wang, J. Zhang, and J.Q. Gong for their
help in technical analysis, and many critical discussions. In particular, he is in
great debt to J. Wang, C. Wang and J. Zhang for their unconditional help in
formatting the book without any hesitation. Special thanks go to M. Fearon
for her assistance and help in the process of publishing the book.
Introduction
1.1 Introduction
In recent years, adaptive control of nonlinear systems has received much atten-
tion and many significant advances have been made in this field. Due to the
complexity of nonlinear systems, at the present stage, research on adaptive non-
linear control is still focused on development of the fundamental methodologies.
This book addresses adaptive control design for several classes of nonlinear sys-
tems using approximation-based techniques. The main objectives of the book
are to develop stable adaptive neural control strategies, and to perform tran-
sient performance analysis of the resulted neural control systems analytically.
In this chapter, a brief background of adaptive control and neural network
control is firstly examined, including a description of the historical develop-
ment in these research areas. Then, several challenging problems in adaptive
neural control are discussed through analyzing the instability mechanisms of
the available neural control schemes, which motivates the research of the book
and highlights the importance of this work. Finally, the organization of the
book is presented with a description of the purposes, contents, and methodolo-
gies used in each chapter.
In the last few decades, a great number of control approaches/methods, in-
ventions, and control applications within the fields of adaptive control, neural
control and fuzzy systems, have been published in various books, journals and
conference proceedings. Making a complete description for all aspects of adap-
tive control techniques is difficult due to the vast amount of literature. In the
following, a brief introduction related to the research in this book, i.e., adaptive
control and neural network control, is given to obtain an outline of the historic
development, and present status in these areas.
For clarity and conciseness, this book is dedicated to adaptive neural net-
1. Parameter Drift
The characteristic of parameter drift is the tendency of NN weights to
drift to infinity with time due to the existence of function approximation
errors. This phenomenon is quite similar to that of the identification and
adaptive control for parametric systems in the presence of external distur-
bances. Such a problem has been successfully solved in many NN-based
6 Chapter 1. Introduction
where f(x) and g(x) are unknown nonlinearities, a commonly used control
structure is
(1.2)
with flex, Og) and j(x, Of) being the estimates of the unknown nonlinear-
ities. Additional precautions have to be made for avoiding the possible
singularity problem because a standard adaptive law does not guarantee
fl(x,Og) being bounded away from zero. To deal with such a problem,
projection algorithms have been applied to project the estimated para-
meter Og in a convex subset 8 g in which lfJ(x,Og)1 ~ go> 0 [149, 179j.
This approach looks very simple; however, the projection algorithm relies
on the exact knowledge of the set 8 g . Such a requirement is much re-
stricted because there is no physical meaning of neural network weights,
which leads to the construction of the convex set 8 g extremely difficult
if not impossible.
sliding control term for confining x(t) to the region n for all time [161].
This technique requires considerable large amount of a priori information
for the studied plants. In addition, the sliding control design is regarded
as a high-gain control method which is usually expensive in practical
applications, and may excite the unmodeled dynamics of the plants [169].
In this book, the problems mentioned above will be considered and ad-
dressed. The techniques used include applying projection algorithms for avoid-
ing parameter drift, introducing a novel Lyapunov function to construct sin-
gularity free controller, providing explicit conditions on initial states, design
parameters and reference signals for guaranteeing transient behaviour and sta-
bility of the adaptive systems.
The main goals of the book are to develop constructive and systematic
methods for neural adaptive control design, prove system stability, and perform
transient performance analysis of the closed-loop systems.
8 Chapter 1. Introduction
The first objective of this book focusses on the stability issues of adaptive
neural network control. In view of the facts that the operation regions of
all physical plants are bounded, and the function approximation capability
of neural networks holds only over compact sets, a regionally stable adaptive
NN control scheme is firstly developed for nonlinear systems in a Brunovsky
controller form. The stability of the closed-loop systems is guaranteed when
the initial states, design parameters, and allowed reference signals satisfy some
certain conditions. In the case that the upper bounds of system nonlinearities
are known, globally stable neural control is achievable.
The second objective of this book is to analyse the transient behaviour for
the designed adaptive system. As discussed before, transient behaviour of a
NN-based adaptive system is important for guaranteeing the system's stability.
Most of the NN control approaches developed in the book include the analysis
of the bounds on tracking error and system states which are helpful for designer
to construct neural networks and choose controller parameters. These bounds
also serve as a guide for control performance improvement.
ondly, we study the nonlinear strict-feedback systems which include both para-
metric uncertainty and unknown nonlinear functions, and constant 9i so that
the parametric certainties can be solving using model based adaptive control
techniques and the unknown nonlinear functions be approximated using NN
approximation. Thirdly, we investigate the control problem of a class of non-
linear pure-feedback systems with unknown nonlinear functions. This problem
is considered difficult to be dealt with in the control literature, mainly because
that the triangular structure of pure-feedback systems has no affine appear-
ance of the variables to be used as virtual controls. Finally, the extension
from SISO nonlinear systems in triangular forms to MIMO nonlinear systems
in block-triangular forms have also been considered in this chapter.
1.6 Conclusion
In this chapter, after a brief description of the background of adaptive control
and neural network control, including a description of the historical develop-
ment in these research areas, several challenging problems in adaptive neural
control have been discussed through analyzing the instability mechanisms of the
available neural control schemes, which motivates the research of the book and
highlights the importance of this work. Finally, the organization of the book
is presented with a description of the purposes, contents, and methodologies
used in each chapter.
Chapter 2
Mathematical Preliminaries
2.1 Introduction
In this chapter, some basic concepts concerning signals and systems are pre-
sented. From the view point of system analysis, a control system is a closed-loop
system. The behaviours of the tracking error signals, the control signals and
all the internal signals in the system are very important in control system de-
sign. It is, therefore, essential to have appropriate measures (or norms) for the
size of these signals for analysis and controller design. From these norms, we
can define deduced norms to measure the "gain" of the systems. In between,
we also present the concepts of compact sets, continuous and differentiable
functions, the Lipschitz condition, and Barbalat's Lemma. Then, some basic
matrix properties, the concepts of stability and Lyapunov Stability, are in-
troduced. Finally, we present the definitions and properties/operations of (i)
the stable sliding surface, (ii) Mean Value Theorems, (iii) Integral formula in-
cluding integration by parts, Change of variables, Comparison theorem, and
differentiation of integrals, (iv) Implicit Function Theorem, and among others
for completeness. All the above concepts and formulas are essential tools used
in the book.
(2.2)
The three most commonly used norms are IlxliI, IIxl12 (or IIxll for simplicity)
and IIxlle<>' All p-norms are equivalent in the sense that if II ·lIpt and II ·llp2 are
two different p-norms, then there exist positive constants CI and C2 such that
(2.3)
IIAxli p
IIAllp = sup -11-11- = sup IIAxli p (2.4)
x#O x P Ilxllv=1
It is easy to show that the induced norms are also equivalent in the same
sense as for the vector norms, and satisfying
(2.5)
Given a matri-x A = [aij], the Froben'ius norm is defined as the root of the
sum of the squares of all elements [113]
with trC) the matrix trace (i.e., sum of diagonal elements). Though the Frobe-
nius norm is not an induced norm, it is compatible with the vector 2-norm in
that
Compact sets
Definition 2.2. A subset S c R n is said to be open if every vector xES,
there is an E-ne'ighbo'urhood of x
Continuous Function
A function f: Rn ---4 R"m is said to be contimtotts at a point x if f (x+bx) ---4 f (x)
whenever bx ---4 O. Equivalently, f is continuous at x if, given I' > 0, there is
b > 0 such that
for all (Xa,Xb) E U. We say,L a Lipschitz constant for f(x). We say f(x) is
Locally Lipschitz if each point of U has a neighborhood Slo in U such that the
restriction f(x)ISlo is Lipschitz.
Let C k be the space of continuous functions with k continuous derivatives,
where k ::::: 0 is integer.
Lemma 2.1. Let a mapping f(x) : U ---4 R be C 1 , Then, f(x) is locally
Lipschitz. Moreover, if Sl C U is compact, then, the restriction f(x)ISl is
Lipschitz /67/,
14 Chapter 2. Mathematical Preliminaries
Differentiable Function
A function 1: R --4 R is said to be differentiable at a point x if the limit
ofdo Xl 0ldoxn]
J(x) = [~~] = [ : '. : E R mxn (2.10)
OfmiOX1 .. : Ofmiox n
Function Norms
Definition 2.4. Let f(t): R+ --4 R be a continuous function or piecewise con-
tinuous function. The p-norm of f is defined by
L1 ~ {I: R+ --4 R 1111111 = 100 111dt < 00, convolution kernel} (2.13)
L2 ~ {f: R+ --4 R IIIfll2 = 100 1112 dt < 00, finite energy} (2.14)
From a signal point of view, the I-norm, IIx1l1' of the signal x(t) is the integral
of its absolute value, the square IIxll§ of the 2-norm is often called the energy
of the signal x(t), and the oo-norm is its absolute maximum amplitude or peak
value. The definition of the norms for vector functions are not unique.
A function 1 may belong to L1 and not be bounded. Conversely, a bounded
function needs not belong to L 1 . However, if 1 E L1 n L oo , then 1 E Lp for all
p E [1, (0). It is also true that f E Lp does not imply that f --4 0 as t --4 00.
This is not even guaranteed if 1 is bounded. However, we have the following
results.
2.2 Mathematical Preliminaries 15
lim j(t)
t->oo
=0 (2.16)
lim
t->oo
it
0
f(7) d7 (2.17)
Corollary 2.2. If f(t), ht) E L oo , and f(t) E L p , for some p E [1,00), then
f(t) - t 0 as t - t 00.
Corollary 2.3. For the differentiable function f(t), if limHoo f(t) = k < 00
and f(t) exists, then f(t) - t 0 as t - t 00.
• strictly positive if a(t) > 0 or for some E> 0, a(t) 2: E for all t.
• symmetric if AT = A;
• skew-symmetric if AT = -A;
Some of the most useful facts on real matrices are summarised below [52].
f(x*) =0 (2.19)
Definition 2.7. The equilibrium point x = 0 is said to be stable if, for any
given € > 0, there exists a positive [) such that if
IIx(O)1I < [)
2.4 Lyapunov Stability Theorem 17
then
'V(x, t) ~ a(llxll)
for all t 2: 0 and in the neighbourhood N of the origin Rn;
V(x, t) s 13(lIxll)
for all t 2: 0 and in (the neighbourhood N of the origin) Rn.
Definition 2.11. Given a continuously differential function V: R n x R+ -t R,
together with a system of differential equations
x= f(x, t)
the derivative of V along the system is defined as
• uniformly stable if for x EN, there exists a scalar function V(x, t) SItch
that V(x, t) > 0 and decrescent and V(x, t) ::; 0;
;1 E R(n-l)x(n-l),bs = [~1
-A n - l 1
(2.22)
(2.23)
It has been shown in [169] that definition (2.23) has the following additional
properties:
(i) if the magnitude of e s is bounded by a constant C > 0, then the error
e(t) Ene, \ie(O) E nc with
(2.24)
20 Chapter 2. Mathematical Preliminaries
~ C and e(O) rt Oc, then the error e(t) will converge to Oc within
(ii) if les(t)1
a short time-constant (n - 1}/>..
The Mean Value Theorem 2.5 for functions from R to R is false, in general,
for vector-valued functions from R n to R m , when 111 > 1. However, a correct
equation is obtainable if ~ is suitably chosen. This gives a useful generalization
of the Mean-Value Theorem for vector-valued functions. Let L(x, y) denote the
line segment joining two points x, y E R n , i.e.,
L(x,y) = {z I z = Ox+ (1-0)y, 0 ~ 0 :51}
Theorem 2.6 (Mean Value Theorem). Assume that f(x) -+ R n is contin-
uously differentiable at each point x of an open set S C Rm. Let x and y be
two points of S such that the line segment L(x, y) C S. Then there exists a
point z of L(x, y) such that f4J
F(x) = 19 (X)
g(a)
f(t)dt, Vx E [a,b]
F(x) = l x
f(g(t»g'(t)dt
In particular, we have
1 9 (b)
g(a)
f(x)dx =
lb
a
f(g(t»g'(t)dt
2.5 Useful Theorems and Formula 21
1 9 (d)
f{t)da{t) =
lb f{g{x»da{g(x»
g(c) a
1 b
f(x)dg(x) = f(x)g(x)
Ib
a -1 b
g{x)df(x)
Lemma 2.3 (Riemann Integral Inequality). Assume f(x) and g{x) are
integrable on [a, b], then, we have the following inequalities:
1b f{x)dx :S 1b g{x)dx
m :S b _1 a lba f{x)dx:S M
3. Cauchy-Schwartz inequality
(l b
f{X)9{X)dX) 2 :S (l b
f 2{X)dX) (l g2{X)dX)
b
ib f(x)da(x) ~ lb g(x)da(x)
:t l b
(t) j(x)dx = j(b(t» d:~t)
d jb(t) jb(t) db(t) da(t)
-d f(x, t)dx = j£(x, t)dx + j(b(t), t)-d- - j(a(t), t)-d-
t aCt) aCt) t t
Lemma 2.6. If a fttnction F(z, () : R x RTil --> R is differentiable with respect
to z, then for any two points Zl, Z2 E R, the following equation holds
f(X,gi(Xl)) =0
lim gi(X) = gi(X;-)
x-).x~
24 Chapter 2. Mathematical Preliminaries
which implies
f(a.;,g(x)) = 0, '<:IxE S
Next, let us prove the continuous function being patched up over the set is
unique as well by contradiction. Suppose that there are at least two continuous
functions, say g'(x) and g"(x), satisfying f(x,g'(x)) = 0 and f(x,g"(x)) = 0,
'<:Ix E S, and there is at least one point, say X2 E S, such that g'(X2) =J g"(X2).
Again, using Implicit Function Theorem 2.10, for point X2, there exists a
neighborhood U2 of X2 and a unique continuous function g2(X) such that
which implies
f(x,g(x)) = 0,
<>
For different mathematical proof and arguments, readers can refer to [32]
for details. Interested readers can also refer to [11] where it shows that the
polynomial function being patched-up can be extended continuously to the
beyond of the neighborhood constructively.
2.5 Useful Theorems and Formula 25
8f(x, y)
I 8y (x,y)1 > d > 0, \f(x,y) ERn X R
then there exists a uniq'ue continuous (smooth) function y = g(x) such that
f(x,g(x)) = O.
Proof For clarity, let us consider the case that af~x,y) (x, y) > d > 0, \f(x, y) E
R n x R. y
First we prove that for every x ERn, there exists a unique g(x) E R such
that f(x, g(x)) = O. Because the lower positive bound of the partial derivative
af~~,Y) > d > 0, \f(x, y) E R n x R, if f(x, 0) = c I- 0, by Mean Value Theorem
(Theorem 2.5)' then we have
f(x, _I~I) = f(x, 0) + 8f~ y) Iy=y; (_I~I - 0) < f(x, 0) - Icl :::; 0
where y; E (0, I~I) and y; E (_I~I, 0). Therefore, by the fact that f(x, y) is
monotonically increasing with respect to y, there exists a unique y E (- ~, ~)
such that f(x, y) = O. If c = 0, it is obvious that y = 0 is the point such that
f(x, y) = O.
Then, we assert that y = g(x) is a continuous (smooth) function of x.
Fix Xo, apply Implicit Function Theorem 2.10 to f(x, y) around the point
(xo, y(xo)). Because the positivity of partial derivative to y, we see that
there exists a continuous (smooth) function Yl (x) around x = xo, such that
f(x, Yl(X)) = 0 around x = Xo. By the strict increasing of f(x, y) with respect
to y, we have Yl(X) = g(x) for every x, therefore g(x) is continuous (smooth)
in x around Xo. Since Xo is arbitrary, we see that g(x) is a (smooth) continuous
function in x.
The prooffor the case that ~(x, y) < -d < 0, \f(x, y) E R n x R can be
similarly obtained, and is omitted.y
<)
(2.27)
(2.29)
2.6 Conclusion
In this chapter, we have presented some basic concepts, operators and their
properties, some theorems that are essentials for the book, which include (i)
the concepts of norms, compact sets, continuous and differentiable functions,
the Lipschitz condition, and Barbalat's Lemma, (ii) some basic matrLx proper-
ties, the concepts of stability and Lyapunov Stability, and (iii) the definitions
and properties/operations of (a) the stable sliding surface, (b) Mean Value
Theorems, (b) Integral formula including integration by parts, Change of vari-
ables, Comparison theorem, and differentiation of integrals, and (d) Implicit
Function Theorem, and among others for completeness.
Chapter 3
3.1 Introduction
In this chapter, the sturctures, approximation properties as well as weights
updating algorithms of the neural networks used in this book are overviewed.
Firstly, general function approximation abilities of neural networks are briefly
introduced. Then, two classes of commonly used neural networks, linearly
parameterized networks and non-linearly parametrized networks, are discussed
in details respectively. For each class of neural network, after the introduction
of network structures, the approximation properties are analyzed. Moreover,
the neural network weights learning algorithms are given with the rigorous
convergence proof for the easy utilization in the rest part of the book.
d(f(W*,x),f(x)) $ E (3.1)
(3.2)
(3.3)
It has been justified in [131] that for a continuous positive function s(-) on
[0,00), if its first derivative is completely monotonic, then this function can
be used as a radial basis function. Commonly used RBFs are the Gaussian
functions, which have the form
where Pi = [pil, Pi2, ... ,Pin]T is the center of the receptive field and Ui is the
width of the Gaussian function. The radial basis functions can also be chosen
as Hardy's multiquadric form
(3.5)
(3.6)
30 Chapter 3. Neural Networks and Function Approximation
Lemma 3.1. For the Gaussian RBF (3.4), if z= z - E1/; where 1/; is a bounded
vector and constant E > 0, then
S(2) = S(z) + ESt (3.7)
The Taylor series expansions of exp [2«Z~,)T,~] and exp [-«!t ll )2] at 0 are
exp [2E(Z-j..ti)
O'{
T-]
'IjJ
= 1 + EOli,
where Oli and 02i denote the remaining factors of the higher-order terms of
the Taylor series expansions. Thus, (3.8) can be written as
+ EOli)(1 + E02i) = Si(Z) + ESti
si(i) = si(z)(1
where Sti = Si(Z)EOli(1 + E02i) + Si(Z)02i. Since si(i) and Si(Z) are bounded
basis functions, Sti is also bounded. It follows from (3.3) that
S(2) = S(z) + ESt
where St = [Stl, St2, ... , Stl]T is a bounded vector function. <>
Another linearly parametrized approximator is the higher order neural net-
works which are expansion of the first order Hopfield [68] and Cohen-Grossberg
[28] models that allow higher-order interactions between neurons. The strong
storage capacity and approximation capability of higher order neural networks
have been shown in [100,144]. The structure of HONNs can be expressed as
g(W,z) = WTS(z), Wand S(z) E Rl, (3.9)
S(z) = [Sl(Z),S2(Z), ""Sl(Z)]T, (3.10)
Si(Z) = II [s(Zj)]dj(k), i=I,2, ... ,l (3.11)
JEIk
Lemma 3.2. Consider the basis functions of HONN (3.10) with z and Z E Rl
being the input vectors, and define St = ~[S(z) - S(z)]. Then St has the
following properties ~
(3.12)
(3.14)
guarantees that W(t) E Loa for bonnded initial weights W(O), where 0 <m <
00,rtu = r~ > 0 and btU > 0 are constant design parameters.
Proof Choose the Lyapunov function candidate as
(3.15)
(3.16)
32 Chapter 3. Neural Networks and Function Approximation
where
(3.19)
(3.20)
(3.21)
Since e s and II 511 are bounded, we conclude that lie. and li are bo.unded.
Therefore, 6 w is a compact set. It follows from (3.18) and (3.21) that Vw :s; 0,
whenever W is outside 6 w . Therefore the NN weight errors W will be bounded
by 6 w . (;
Remark 3.1. If 0 < m :s; 1, then the proof can be treated more elegantly as
follows.
Choose the Lyapunov function candidate
V. =
w
~wTr-lw
2 w
Because land 810 are positive constants, we conclude that W(t) E Loo. However
it cannot be generalized to 1 < Tn < 00 because of the possibility that es = O.
Lemma 3.4. If e s is bounded, then the weight update law
guarantees the boundedness of W, where KO, KI, K2 and 8 are positive con-
stants.
Proof Consider the Lyapunov function candidate
V= ~WTW
2
and taking the derivative of V along (3.24), we have
Since the filtered tracking error e s is bounded, we conclude that 810 is a compact
set, and V ::; 0 as long as Wet) is outside 8 10 . <>
34 Chapter 3. Neural Networks and Function Approximation
Lemma 3.5. If the adaptive law is chosen using projection algorithm (741 as
with the adaptive gain 'Y > 0, then the boundedness of W can be guaranteed.
Proof Consider the Lyapunov function candidate
(3.27)
(3.28)
V = WTS(Z)e s :S 0 (3.29)
(3.30)
where Z = [Zl, Z2," . ,znV is the input vector, Vjk are the first-to-second
layer interconnection weights, Wj are the second-to-third layer interconnection
weights, Ow and Ovj are the threshold offsets. The activation function s(·) can
be chosen as the continuous and differentiable nonlinear sigmoidal
1
S{Zi) = 1 + e-"(Z; , 'Vzi E R (3.32)
with "I > 0 being a constant, or a hyperbolic tangent function
(3.33)
Throughout this book, the sigmoidal function (3.32) shall be used as activation
functions of multilayer neural networks with "I = 1, Le.,
1
s{z·) - 'Vzi E R (3.34)
t - 1 + e-z;'
The important result, that MNNs with one or more hidden layers are ca-
pable of approximating any continuous nonlinear function, was obtained inde-
pendently by Funahashi [44], Cybenko [30] and Hornik et. al. [69]. The main
approximation result given by [44] is summarised in the following theorem for
a three-layer network case:
36 Chapter 3. Neural Networks and Function Approximation
Hidden-layer
Zl
Z2 • (
g(z)
Theorem 3.1. Let s(·) be a non constant, bounded and monotone increasing
continuous function. Let S1 z be a compact subset of Rn and g(z) be a real
valued continuo'us function on S1 z . Then for any arbitrary J.L > 0, there exists
an integer I and real constants Wj, Vjk, ()vj, and ()tu SllCh that (44/
The above theorem reveals that for a three-layer network, if the node num-
ber of the hidden layer is large enough, it can approximate any continuous
function to arbitrary accuracy on a compact set. It should be pointed that the
above theorem is only an existence theorem, which is not constructive in the
sense that no method is presented for finding the ideal weights of NNs and no
or little information is given for the number of hidden neurons to be used for
function approximation.
For simplicity of presentation, MNNs (3.31) are re-expressed in a matrix
format as below. Define
(3.36)
(3.37)
expressed as
(3.39)
(3.41)
h S' = S(V'T-)
were Z, S', = d·lag {" "
S1,S2,··· ,sl "
"} WZ·th Si=s '('T-)
Vi Z = d[8(Za)11
dZ a za=i/[z,
i = 1,2, ... , l, and the residual term du is bounded by
where o(frT Z)2 denotes the sum of the high order terms in the Taylor series
expansion. Using (3.44), we have
WTS(VTz) - W*TS(V*T z )
= (W + w"fs - W"T [S - S'VTz + O(VTz)2]
= WT S + (W - wf S'VT Z - W*T O((iT z)2
= WTS+ WTS'VTz- WTS'(V - V*)T z - W*T O(V T z)2
= WT(S - S'yT z ) + VTzWTS' + d.u (3.45)
(3.48)
3.4 Non-linearly Parametrized Networks 39
I I
IIS'VTzll :SL IvTzs'(vTz)1 :S 0.22391, IIS'IIF :SL s'(vTz) :S 0.2.51 (3.49)
i=l i=l
From (3.44), we know that the high order term O(VT z)2 is bounded by
(3 ..51)
where A = ['\1, '\2, ... , '\n-1jT is an appropriately chosen coefficient vector such
that sn-1 + '\n_1Sn-2 + ... + '\1 is Hurwitz.
Lemma 3.8. If sigmoid activation junction (3.34) is ttsed, and z ~ [.;-T, (f, vjT,
z = [zT, l]T, then there exist positive constants O!l to 0!4 sttch that the residltal
term du is bounded by /50/
(3.56)
(3.57)
Considering (3.54) and II';-dll < c (Assumption 3.3), we can derive 11.;11 ::;
d 1 + d2 1e s l with constants d 1 , d2 > O. From (3.55), we can derive Ivl ::; d3 +
d4 iesl with constants d3 , d4 > O. Thus, from (3.57), we know that there exist
computable positive constants C3 and C4, such that IIzll ::; C3 + c4le s l·
Substituting it into (3.47) and letting O!l = Co + C1C3, 0!2 = C1C4, 0!3 = C2C3
and 0!4 = C2C4, we conclude that (3.56) holds. <)
Lemma 3.9. If sigmoid activation function (3.34) is lIS ed, and z = [xT, es , Vl]T ,
z = [zT, l]T, then there exist positive constants O!l to 0!4 sttch that the residual
term d u is bOlmded by
(3.58)
Proof The proof can be done similarly as in Lemma 3.8. The only difference
is that (3.57) should be changed to
(3.59)
If c is small, then es and es/c are varaibles of different scales. It has been
found that by feeding both of them as the input to the neural network, function
approximation accuracy can be improved.
Corollary 3.2. If sigmoid activation function (3.34) is llsed, and z ~ [xT, es ,
~, Vl]T, z = [zT,l]T, then there exist positive constants O!l to 0!4 sttch that
the residual term du is bO'unded by
(3.60)
Proof The proof can be done similarly as in Lemma 3.8. The only difference
is that (3.57) should be changed to
1
Ilzll ::; Ilzll + 1::; II xii + (1 + -c )Iesl + hi + 1 (3.61)
3.4 Non-linearly Parametrized Networks 41
g'uarantee that Wet), Vet) E Loo for bounded initial weights W(O) and V(O),
where 0 < rn < 00, r w = r~ > 0, r v = r;; > 0, 6w > 0 and 6v > 0 are
constant design parameters.
Proof Choose the Lyapunov function candidate as
V.u ="21 [w
- T -1 -
rw W +tr{V- T rv-1 V}
- ]
(3,64)
Using the properties of MNNs (see (3.49) in the proof of Lemma 3,7 and As-
sumption 3.2)
42 Chapter 3. Neural Networks and Function Approximation
we obtain
::; -
1 + lesl m [(
2
-
bwllWIl - bw (1
lle sl )2 -2
+ leslm) + bvllVllF + a(e s )
]
(3.67
)
where
(3.68)
(3.69)
which are all well defined at e s = O. For bounded e s ERe., and let
and define
Since es is bounded, 1 < Tn < 00, we conclude that ae and a are bounded.
Therefore, 8 wv is a compact set. It follows from (3.67) a~d (3.70) that Vw ::; 0,
whenever (W, V) is outside 8 wv . Therefore the NN weight errors (W, V) will
be bounded by 8 wv . <>
Remark 3.2. !f0 < Tn::; I, the proof can be treated moreelegently as follows.
Choose the Lyapunov function candidate
1
lAT
l!:w = -W
2 r-w lAW + -tr{V
2
AT
r-v lAV}
3.4 Non-linearly Parametrized Networks 43
Using the property tr{V T zWT 8'} = W T 8'VT z, the time derivative of Vw
along (3.62) and (3.63) is
Because every element of 8 is not larger than one, we know that 11811 2 :::; 1 with
1 the NN hidden-layer node number. Therefore, Vw < 0 once
Because I, 8w and 8v are positive constants, we conclude that W(t), V(t) E Leo.
However, it cannot be generalized to 1 < Tn < 00 because of the possibility
that e s = O.
V= -fv[ZWT8Ies+6vlesl",T] (3.73)
(3.74)
and ",T(O) is bounded, then W(t) E 8 w and V(t) E Leo, 't;/t ;::: O.
Proof Consider the positive function
Therefore
Vv = ~tr{VTr;lV}
and using the property that WTS'VTZ = tr{VTZWTS'}, its time derivative
along (3.73) may be written as
Both of multilayer NNs and RBF NNs shall be used for different adaptive
control designs in this book. The choice of the networks depends on the con-
ditions and prior knowledge for the studied systems. If dimensionality of the
input vector is not very high and the ranges of network input signals can be
determined/guaranteed a priori, RBF networks shall be used to simplify the
design and analysis. Otherwise, MNNs will be applied for general function
approximation.
3.6 Conclusion
In this chapter, two classes of function approximators, linearly parametrized
NNs and nonlinearly parametrized MNNs, have been discussed for function
approximation. The comparison studies show that they have quite different
properties in solving various identification and control problems. It is believed
that a better NN control design should fully utilize their advantages and avoid
their shortcomings. Though several commonly used tuning algorithms have also
been presented for both LPNNs and MNNs, they are by no means exclusive.
Other forms or variations are possible in solving individual practical problems.
Chapter 4
4.1 Introduction
The development of feedback linearization techniques provides a powerful tool
for nonlinear system control. Under some geometric conditions, the input-
output response of a class of single-input single-output (8180) nonlinear sys-
tems can be rendered to the following Brunovsky form [77J
(4.1)
where x = [Xl, X2, ... ,Xn]T ERn, u E R, y E R are the state variables, system
input and output, respectively; a(x) and b(x) are smooth functions; and d(t)
denotes the external disturbance bounded by a known constant do > 0, i.e.,
Id(t)1 '5 do. For the controller design using feedback linearization techniques,
the most commonly used control structure is
before the controller is put into operation. Other methods include applying
the projection algorithm to project W inside a feasible set where no singu-
larity problem occurs [171,179], and modifying the controllers by introducing
switching control portions to keep the control magnitudes bounded [52,99, 186J.
In [161,171]' stable direct adaptive controllers are studied for a class of
nonlinear systems using radial basis function (RBF) neural networks. Though
the restrictions mentioned above are eliminated, the absolute value of the first
derivative of b(x) is assumed to be bounded by a known positive constant,
which is usually difficult to be checked/determined in practical applications.
Past experience shows that a good design of Lyapunov function should fully
utilize the property of the studied systems. Adaptive control of nonlinear
systems has been an active research area and many good theoretical results
have been obtained in the literature [87,91,104,128, 165J and the references
therein.
In this chapter, the adaptive control problem is studied for a class of non-
linear systems in Brunovsky form (4.1) satisfying 8b(x)/8x n = 0, Vx E 0 (a
compact set) (Assumption 4.3) or Vx E R n (Assumption 4.6). Although the
assumption 8b(x)/8x n = 0 may restrict the range of applicable plants, it brings
us a nice property
where A = [/\1, /\2, ... , '\n_1]T is chosen such that the polynomial sn-1 +
'\n_1Sn-2+ ... + /\1 is Hurwitz. From the Property (i) of sliding surface given
in Chapter 2, we know that e1(t) ---> 0 as e s ---> O.
Define 1/ = _y~n) + [0 AT]e, the time derivative of es may be written as
Assumption 4.2. The sign of b(x) is known and there exist two constants bo
and bI such that b1 2: Ib( x) I 2: bo > 0 on a compact subset fl C Rn.
Assumption 4.2 implies that smooth function b(x) is strictly either positive
or negative. From now onwards, without losing generality, we shall assume
b(x) > O.
* 1 [1 1 b(x) ] (4.7)
u = - b(x) [a (x) + v]- cb(x) + cb 2 (x) - 2b 2 (x) es
where c > 0 is a design parameter, then limt--;oo Ile(t)1I = o.
50 Chapter 4. SISO Nonlinear Systems
Proof Substituting the DFC u = u* (4.7) into (4.6), under the assumption of
d(t) = 0, we have the closed-loop system
(4.8)
(4.8), we obtain
. 2 2
b(x) 2 es es
v:. -
S -
1 .
--e e - - - e - - - - - - -
b(x) s s 2b2(x) s - c:b(x) t:b2(x)
(4.9)
Remark 4.1. For system (4.6), the DFC is not unique. Different choices of
DFC will also have signaficant effect on the neural network controller structure.
In fact, since a(x) and b(x) are unknown, all the terms of u*(z) in (4.7) are
unknown, and subsequently, neural networks should be used to parameterize
it as a whole as there is no known terms to be separated from the unknown
functions.
u
*
= -
1
b(x) [a(x) + v] -
[1 + 1
-;;
b(x) ]
t:b 2 (x) - 2b2(x) es (4.10)
.=
es -b(x) ;
[1 + 1 b(x) ]
c:b2(x) - 2b 2 (x) es (4.11)
2
Choosing a positive definite function Vs 2:(X) and differentiating it along
(4.11), we obtain
(4.12)
4.2 NN Control with Regional Stability 51
Since bo ::; b(x) ::; b1 (Assumption 4.2), it follows from (4.12) that Vs is a
Lyapunov function. According to Lyapunov theorem, we have limt->oo les I = 0
which implies limt->oo Ilell = o. <:;
From (4.12), it can be seen that Vs will be more negative if a smaller c: is
chosen. Therefore, the convergence rate of the tracking error can be adjusted
by tuning the design parameter c.
Remark 4.2. Comparing with the DFC in (4.7), there is a known term -es/c
in (4.10), which may be separated out to reduce the complexity of the function
that needs to be approximated by neural networks. In fact, (4.10) can be
re-written as
U * = u *() 1 s
1 Z - -e (4.13)
c
where
(4.14)
Neural networks are then used to approximate the unknown function u*(z).
It should be noted that though the elements e s and es/c in z are dependent
52 Chapter 4. SISO Nonlinear Systems
due to constant c, they are in different scales when a very small c is chosen.
Thus es/c is also fed into NNs as an input for improving the NN approximation
accuracy.
For simplicity, let us consider the high-order neural networks (HONNs)
(3.9). Because the DFC input u*(z) (4.15) is a continuous function on 0., for
an arbitrary constant Ji-o > 0, there exist an integer 1* and an ideal constant
weight vector W*, such that for all I 2: 1*
(4.18)
Adding and subtracting b(x)u*(z) on the right hand side of (4.20) and noting
(4.17), we have
'T T
£Os = a(x) + l/ + b(x)[W S(z) - W* S(z) - Ji-d
+b(x)u*(z) + d(t), Vz EO. (4.21)
(4.22)
Since NN approximation (4.17), and Assumptions 4.2 and 4.3 are only valid on
the compact set 0, it is necessary to guarantee the system's states remaining
in 0 for all time. In the following theorem, we show that for appropriate ini-
tial condition x(O) and suitably chosen design parameters, adaptive controller
(4.18)-(4.19) guarantees x E 0, Vt 2: o.
4.2 NN Control with Regional Stability .53
Proof The proof includes two steps. We shall firstly assume that x E 0 holds for
all time, and find the upper bounds of system states. Later, for the appropriate
initial condition x(O) and controller parameters, we prove that state x indeed
remains in the compact set 0 for all t ~ O.
Step 1 Suppose that x E 0, Vt ~ 0, then NN approximation (4.17), Assump-
tions 4.2 and 4.3 are valid. Consider the Lyapunov function candidate
Vs =
1[
2
e; + W- T r
b(x}
-1 - ]
W (4.23)
(4.24)
and noting that Ip,d s:; J.to, Id(t)1 s:; do and 0 < b(x) < b1 , the following
inequality holds
(4.28)
where ao = max{c, AUa-ax} and ao = cJl5bl + ~d5 +aIIW*II2. Solving the above
inequality using Lemma 2.9, we have
(4.29)
Since \1;;(0) is bounded, inequality (4.29) shows that es and W(t) are bounded.
By (4.23) and 0 < b(x) :S bl , it follows that lesl :S J2\1;;b(x) :S J2Vsbl .
Combining with (4.29), we obtain
Next, we determine an upper bound of the error vector e. Define ( = [el, e2,
... , en_I]T. From the properties for (2.20), we know that (i) there is one state
space representation for mapping es = [AT 1]e, i.e., (; = As( + bse s with As
being a stable matrix; (ii) there are two constants ko > 0 and /\0 > 0 such that
Ile Ast II :S koe- Aot , and (iii) the solution for ( is
where kA = ko(1 + IIAII). Since k A, ao, Ao, and a are positive constants, and
«0) and es(O) depend on x(O) - Xd(O), we conclude that there exists a positive
constant R(c, ,-\max, x(O), W(O») depending on c, '-\max, x(O) and W(O) such
that
(4.34)
It is easy to see that for all x(O) E no and Xd End, we have x E n, 'rit 2: o.
Then, for the system with x(O) E 0 0 , bounded W(O) and Xd End, the following
constants ,-\" and c .. can be determined
,-\" = sup {Amax 1{xI IIx - xdll < R (0, ,-\max, x(O), W(O») } C 0,
A",,,.,ER+
Xd E nd} (4.36)
Xd E Od} (4.37)
Therefore, for the initial condition x(O) E 0 0 , bounded initial NN weight W(O),
and the desired signal Xd E Od, if the controller parameters are chosen such
that l 2: l", Amax < A" and c $ c", then system state x indeed stays in 0 for
all time. This completes the proof. <)
Remark 4.4. The result obtained in Theorem 4.1 is regionally stable because
the initial states are required to be within 0 0 . This is reasonable because the
neural network approximation is only valid on a compact set and Assumptions
4.2 and 4.3 hold on O. The stability analysis of Theorem 4.1 also provides
a method to estimate the allowed initial state region for the given controller.
For controller (4.18) with parameters c, l, and r satisfying condition (ii) in
Theorem 4.1, the following compact set can be found
Remark 4.5. Theorem 4.1 shows that the initial NN weights of the adaptive
controller can be set to any bounded values, which relaxes the strong condition
required in [24], where the initial values of the NN weights are required to be
chosen sufficiently close to the ideal values.
The above discussion only presents the boundedness of the signals in the
adaptive system, no transient performance of the system's output is provided.
In practical applications, the transient performance is also of great important.
Theorem 4.2 (Transient performance). For the closed-loop adaptive sys-
tem consisting of (4.1), (4.18) and (4.19), if x(O) E no and the design para-
meters I 2': 1*, /\max < A* and f: :S f:*, then
(i) the transient bOtmd of tracking error
(4.39)
Applying Mean Value Theory, there exists a constant /\~ E [AO, 2!o] such that
t
1 )te-,x:t. Let >'s = min{/\o, -2
e-,xot = e- 2a o - (/\0 - -2 1 }, we have >'s :S
no 0:0
minpo, /\~}. Combining with (4.41), it can be shown that
(4.42)
Because ao = max {£, A"~qx} and Amax is defined as the largest eigenvalue of
r- 1 . The above inequality indicates that the tracking error converges to a
small residual set which depends on the NN approximation error ILl, external
disturbance d(t) and controller parameters £, a and r. Decreases in £ and a
or increases in the adaptive gain rand NN node number l result in a better
tracking performance.
Remark 4.8. It should be pointed out that HONNs used in this book may
be replaced by any other linear approximator such as radial basis function
networks [52], spline functions [141] or fuzzy systems [170], while the stability
and performance properties of the adaptive system still hold.
(4.43)
1
b(x) = ml6~(x1) (4.44)
It can be checked that Assumptions 4.2 and 4.3 are satisfied and ~ ::; b(x) ::;
4, Vx E n.
Figure 4.1: Variable length pendulum with leO) = lo + II cos(O), ldlo = 0.5,
gllo = 10 and rnl5=1
The parameters, initial conditions and the neural networks of the adaptive
controller (4.18)-(4.19) are chosen as: A = 10.0; a = 0.005; the initial NN
weight W(O) = 0.0; Z = [xl,x2,e s ,e s lc,vl T ; a two-order neural network with
the elements Si(Z) in (3.11) chosen as S(Zl)' S(Z2)' S(Z3), S(Z4), S(Z5), and the
possible combinations of them (Le., s(zds(zj) with i,j = 1,2,3,4,5). The
total number of the NN nodes is 20. Figure 4.2 presents the simulation results
for the designed NN controller with r = diag{2.0} and c = 0.25. Although
no exact model of the pendulum plant is available and the initial NN weights
are set to zero, through the NN learning phase, it is shown that the tracking
error given in Figure 4.2(a) converges to a small region after 15 seconds. The
boundedness of the NN weight estimates, control input and system states are
shown in Figures 4.2(b)-(d), respectively.
To study the control performance for different design parameters, the fol-
lowing two cases have been investigated. Firstly, we increase the adaptation
gain from r = diag{2.0} to r = diag{6.0} with all other parameters fixed.
The simulation results are given in Figure 4.3. Comparing Figure 4.3(a) with
Figure 4.2(a), it can be seen that smaller tracking error in Figure 4.3(a) is ob-
tained during the initial period (from 0 to 10 seconds). This confirms that fast
adaptation may improve the transient performance of the adaptive system.
Secondly, we reduce the control parameter c from 0.25 to 0.05 and keep other
simulation parameters the same as in the first simulation. Figure 4.4 provides
4.2 NN Control with Regional Stability 59
the simulation results. Comparing the tracking errors in Figure 4.4(a), Figure
4.2(a) and Figure 4.3(a), we find that smaller final tracking error is achieved
for smaller c, while all other signals are bounded .
..(J.os
".1
-0.15
".2
10 15 ro ~ ~ $ ~ ~ ~
10
:IS 45 50
(c) HONN weight IIWII (d) States Xl ("--") and X2 ("- -")
v as follows
U '" = u "'()
1 Z - 1 s,
-e z -- txT ,8,
e v]T E 0 Z C Rn +2 (4.45)
e
where
"'( 1 [ b(x) es
ul z) = - b(x) a(x) + v] + 2b2 (x) es - eb2 (x) (4.46)
Oz = {(x, e s , v) I x E 0; Ilxdll $ c}
When nonlinear functions a(x) and b(x) are unknown, the nonlinear ui(z) is
not realizable.
According to the approximation property of neural networks in Chapter 3,
the smooth function ui(z) may be approximated by multilayer neural networks
(3.31) as below
(4.47)
Assumption 4.4. On the compact set Oz, there exist ideal constant weights
W" and V" such that J1.1 is bounded by 1J1.11 $ J1.o with constant J1.o > O.
Now, we are ready to present the direct adaptive controller
(4.48)
(4.49)
The term U r is introduced for improving the controller robustness in the pres-
ence of NN residual error du . The NN weights are updated as given in Lemma
3.11, Le.,
Adding and subtracting b(x)u! on the right hand side of (4.52) and noting
(4.46), we obtain
.
es = -b(x) ~
[1 + cb 1(x) -
2
b(x) ]
2b 2 (x) e s
As neural network approximation (4.47), and Assumptions 4.2 and 4.3 are only
valid over the compact set n, we first show that by suitably choosing design
parameters, the proposed controller guarantees state x belonging to n, Vt 2: 0
under some given initial conditions.
Theorem 4.3. For system (4.1), controller (4.48) and adaptive laws (4.50)
and (4.51), there exist a compact set no c n, and positive constants c* and €*
such that if
(i) the initial conditions x(O) E Do and W(O) E 8 w1
(ii) the desired signal satisfying "Xd" ::; c*, and control parameter € ::; €* I
then all the signals in the closed-loop system are bounded and the states x
remain in the compact set D for all time (I 91J.
Proof The proof includes two steps. We shall firstly assume that xED holds
for all the time, and find the upper bounds of the system states. Then, for the
appropriate initial condition x(O) and controller parameters, we prove that the
state x indeed remains in the compact set D for t 2: O.
Step 1 Suppose XED, then the NN approximation (4.47) holds and the upper
bounds of the system states can be found as follows. Consider the Lyapunov
function candidate,
Since WTSevTz) :::: v'i IiWII :::: 1.2241jow (see Lemma 3.11), W*TS(V*T z) ::::
v'i IIW*II, UTeS:::: 0, IJ.LLI :::: J.Lo and Itffil < ~, we obtain
(4.55)
with
Integrating this inequality over [0, t] and then multiplying it by e-!:f t , we obtain
b
Vs(t) :::: e-~tVs(O) + 102b0oo ( 1 -
2 2 b)
e-'f-t :::: Vs(O) + 102b0oo ,
2 2
'it 2:: 0 (4.57)
bl
les(t)1 :::: -[e;(O) + 102051 , (4.58)
bo
Let ( = [el, e2, ... , en-IV, then a state representation of es = [AT l]e may
be written as
(4.59)
where As and bs are defined in (2.21). Therefore, there exist constants ko > 0
and /\0 > 0 such that IleAstl1 :::: koe->-'ot. The solution for ( can be written as
(4.60)
where
- - b1
+ c:2aol + c
2
R(c, c:, x(O)) = koll(,(O)1I + kl" bo [e~(O)
c* = sup {c
cER+
I {x I IIxll < R(c, 0, x(O)) } co}
Furthermore, for x(O) E 0 0 and c:S c*, a positive constant c:* can be found
Remark 4.9. It is worth noting that Theorem 4.3 guarantees the bounded-
ness of the closed-loop system in the sense of regional stability, i.e., the initial
conditions are required to satisfy x(O) E 0 0 and W(O) E 8 w for ensuring x E O.
This is reasonable because Assumptions 4.2 and 4.3 hold on compact set 0,
and neural network approximation property (4.47) is only valid on Oz' As the
regions 8 w and 0 0 can be estimated by (3.74) and (4.63), it is not difficult to
find the initial conditions to achieve a stable closed-loop system.
Remark 4.10. In adaptive laws (4.50) and (4.51), the e-modification [137J
terms are introduced to improve the robustness in the presence of the NN
approximation error. It should be pointed out that convergence of the NN
weight estimates is not guaranteed due to the lack of persistent excitation
(PE) condition for the system signals. In fact, even for linear adaptive control
systems the PE condition is very difficult to check/guarantee. In references
[64,161]' PE conditions were studied for radial basis function networks. For
4.2 NN Control with Regional Stability 65
c:= sup {c
cER+
I {x I IIxll :S R(c, e, x(O»} c~ , 2: o}
'ilt
Then, we say that c is the maximum bound of the desired signal IIXdll for
the given controller and the initial state x(O), i.e., under the initial condition
x(O) E St o, all the reference signals satisfying IIXdll < c can be followed by using
this controller.
Theorem 4.3 only ensures the boundedness of the signals in the closed-loop
system, no transient performance is given. The following theorem reveals the
control performance of the proposed adaptive NN controller.
Theorem 4.4. j191} For the closed-loop system consisting of (4·1), (4.48),
(4·50) and (4.51), if x(O) E ~o, W(O) E 8 w , IIXdll :S c*, and c: :S c:*, then
where Cl, C2 > 0 are comp'utable constants, constant f30 :::: 0, and constant
V1(0) :::: 0 depending on system initial conditions x(O), W(O) and V(O),
and
(4.65)
66 Chapter 4. SISO Nonlinear Systems
VI =
e; + W- T r w-1 W- + tr{V- T r v-1 V}
'12 [ b(x) - ]
(4.66)
Differentiating (4.66) along (4.50), (4.51) and (4.54), and using (3.42), we have
Since
and IJ-Ld :S J-Lo, Id(t)1 :S do, inequality (4.70) can be further written as
(4.71)
4.2 NN Control with Regional Stability 67
with
with
(4.74)
Remark 4.11. Two performance criteria (4.64) and (4.65) reveal the transient
response of the closed-loop system. It can be seen that large initial errors
es(O), W(O) and V(O) may lead to a large mean square tracking error during
the initial period of adaptation. However, inequality (4.64) implies
e
m
u
M
where Xl = [Xl, X2]T with Xl the angle displacement of the pendulum from
the vertical configuration; g = 9.8mj 8 2 is the gravity acceleration coefficient;
4.2 NN Control with Regional Stability 69
M and 111 are the mass of the cart and the pole, respectively; L is the half-
length of the pole, and u is the applied force control. The true values of the
parameters are M = 1.0kg, 111 = O.lkg and L = 0.5m. The initial states are
[XI(0),X2(0)f = [:O,of, and the control objective is to make Y = Xl track a
reference signal Yd = ~ sin(t). If we require that the system states remain in
the following compact set
then it can be checked that Assumptions 4.2 and 4.3 are satisfied and bo <
b(x) ::; b1 with bo = 0.998 and bl = 1.4634. The multilayer neural networks are
chosen with the input vector
- T
Z = [Xl, X2, es , 1/, eo] ,l = 5, 'Y = 100.0, and eo = 0.1
1
Ixdt)1 ::; 11(11 + IYdl ::; 1((0)1 + 3.0 (4.76)
Considering the initial states, it is known that 1((0)1 = :0, les(O)1 = 2~ and
IYdl ::; 3~' Supposing MNNs satisfying IIW*II ::; 1.0 and J.ll ::; 0.1 over the
compact set S1 z , from (4.56) we have ao < 14.58. It follows from (4.76) that
for all c: < f;* = 0.106, the state IXII < 7[/4, 'Vt ~ O. Noticing
and inequalities (4.58) and (4.61), it can be checked that IX21 < 37[/2, 'Vt ~ 0
for c: < c:*. Hence, if we choose the design parameter c: < c:*, states (Xl, X2) do
remain in the required set S1 for all time.
Figure 4.6 shows simulation results for the proposed NN controller with
c: = 0.1 and NN node number l = 5. The initial weights W(O) = 0 and the
elements of V(O) are taken randomly in the interval [-0.2,0.2]. Although only
five neural nodes are taken and no exact model of the inverted pendulum is
available, through the NN learning phase, the system output y shown in Figure
4.6(a) tracks the desired signal Yd successfully, and the output tracking error
70 Chapter 4. SISO Nonlinear Systems
shown in Figure 4.6(c) converges to a quite small set after 30 seconds. This
confirms that the proposed MNN adaptive controller possesses the abilities
of learning and controlling nonlinear systems. Figure 4.6{b) indicates system
states Xl and X2 within the compact set Sl. The control input u, is given in
Figure 4.6{d) and the boundedness of the NN weight estimates are shown in
Figures 4.6{e) and 4.6{f).
In order to verify the methods for improving the transient performance, the
following two simulations have been done. Firstly, we increase the number of
the neural network nodes from l = 5 to l = 10 with other simulation parame-
ters being fixed. The simulation results are given in Figure 4.7. Comparing
Figure 4.7{c) with Figure 4.7{a), it shows that smaller tracking error and faster
convergence rate are obtained in Figure 4.6(a). This is due to the approxi-
mation error J.ll being reduced by choosing larger NN node number. Secondly,
we reduce the control parameter e from 0.1 to 0.05 and keep other simulation
parameters be the same as the first simulation. Figure 4.8 provides the simula-
tion results. Comparing the tracking errors in Figure 4.6{c) and Figure 4.8{a),
we can see that the transient performance is improved and the final tracking
error is also reduced in Figure 4.8{a) when decreasing the parameter e.
Finally, the contribution of neural networks is shown by comparing the
adaptive controller with the non-adaptive controller (4.48) with e = 0.1, W ==
0.0 and V == 0.0, i.e., no neural networks and adaptive learning mechanism.
The other simulation parameters are the same as before. In this case, the
controller becomes u = -~es. Since e s = A{y - Yd) + (y - Yd), this non-
adaptive controller is a PD controller. Figure 4.9 shows the control results for
this fixed gain PD controller. Comparing the tracking errors in Figures 4.6(c),
4.7(a), 4.8(a), and Figure 4.9(a), we conclude that adaptive NN controllers
performs much better than the non-adaptive controller because of the learning
ability of neural networks.
0.1,
0.05
-0.05
-0.1
-0.15
" "
0 10 20 40 50 70 80 90 100
0.06
0.05
1.5
0.04
0.03
0.02 0.5
0.01 ; ...
-0.5
-0.01
-<>.02 -.
-0.030 -1.5
1.2
0.
0.6
0.4
02
~ " 40 50 W 70 80 90 _
0.1
1 .
0.05 ...
a ..
-1
-0.05
-2
-0.1
-.4.150 10 20 30 40 so 60 10 so 90 100 ~ ID ro ~ ~ ro 00 ro M ro ~
(a) Tracking error Y - Yd (b) Control input u(t)
Figure 4.9: Responses of the fixed-gain controller( u = -~es with e = 0.1 (no
neural networks and adaptive mechanism)
for limiting the systems states inside the compact set O. In fact, other control
scheme such as variable structure control (VSC) can also be used to perform
the similar role in the controller design.
Assumption 4.7. On the compact set Oz, there exist ideal constant weights
> O.
W* and V* such that J.Ll is bounded by lJ.Lll ::; J.Lo with constant J.Lo
Considering (3.43) and Assumption 4.7, we obtain
(4.78)
74 Chapter 4. SISO Nonlinear Systems
where
(4.79)
where the variable structure control term
AT
u vsc = -1( 1f;sgn(e s ) (4.80)
with j( being the estimate of 1(* and sgn(-) the standard sign function. Fol-
lowing the similar derivations in Section 4.2.3, equations (4.52)-(4.54) for the
VSC controller (4.79) can be further expressed as
es [ 1 b(X)] -T A A,AT_ ATA,-T_
b(x) = - ks + c:b 2 (x) - 2b 2 (x) es + W (8 - 8 V z) + W 8 V z
d(t)
+u vsc + du - Itl + b(x) (4.81)
In Theorem 4.5, adaptive laws for tuning the NN weights and the control gains
of the VSC terms are presented, and the closed-loop is proven to be semi-
globally asymptotic stable.
Theorem 4.5. For system (4.1) with Assumptions 4.5-4.6 being satisfied, if
the controller is g'iven by (4.79) and the controller parameters are ttpdated by
~ -,AT_
= -rw(8 - 8 (4.82)
A
W V z)e s
T
V =
A
-rvzw A A
8'e s (4.83)
k = rk7JJies I (4.84)
where rw = r~ > 0, r v = r~ > 0, r k = rl >0 are constant design parame-
ters, then for bounded initial conditions,
(i) the tracking error converges to zero asymptotically, and
(ii) all signals in the closed-loop system are battnded, and the states x remain
in the compact set
where ( = [Xl, X2," . ,Xn_l]T and ko, '\0 and VI (0) are positive constants.
Proof (i) Consider the Lyapunov function candidate
where k = k - J{*.
Differentiating (4.86) along (4.81)-(4.84) and noticing the fact that
WTS'VTz = tr{VTzWTS'}
we have
1 . b( x ) 2 - T -1':' - T -1':' -"T -1':"
VI = b(x)eSeS - 2b 2(x)e S +W rw W+tr{V rv V}+J\ fk K
2 e~ d(t) -'T I
= -kses - cb 2 (x) + V'usees + [d u - J-tt + b(x) ]e s + R wles
Using (4.78) and U vse = _KT wsgn(es), we obtain
. ') e2 AT T - T
VI :S -kse; - s( ) -
-2
cb x
J( wlesl + J(* wlesl + J( 7ple s l
2 e; (4.87)
= -kse s - cb 2 (x)
Hence, if the initial values es(O), W(O), V(O) and K(O) are bounded, then
(4.88)
(4.89)
we have
(4.90)
with
Therefore,
(4.92)
Considering (4.91) and (4.93), we conclude that the system states x remain in
Ox for all time. This completes the proof. <>
Remark 4.14. The result of Theorem 4.5 is semi-global in the sense that, if
the node number of the neural networks is chosen large enough such that the NN
approximation holds on Ox, then the closed-loop stability can be guaranteed
for bounded initial states and NN weights.
Remark 4.15. It is worth noting that the initial conditions x(O}, Ur(O} and
V(O) affect the size of the compact set Ox significantly. From the NN approxi-
mation results [44], we know that for any given neural network (3.31) with fixed
NN node number l and parameter "I, the smaller the NN approximation region
Ox is, the smaller I W'" II, IIV'" I and /-Lo should be. In this case, the upper bound
of the parameter II J('" II decreases and the gain k of the sliding control term
U vsc might also be reduced. Eq.(4.85} indicates that if the initial values ((O)
and e s (0) are set to zeros by appropriately initializing the reference trajectory
Xd(O), a small Ox can be obtained. On the other hand, by noting (4.85) and
(4.91), we can see that increasing the gains ks, r w, r v and r k may also reduce
the size of the compact set st x .
4.3 VSC - Semi-Global Stability 77
(4.94)
with
(4.95)
where es > 0 is a small constant and tanh(·} denotes the hyperbolic tangent
function, which has a nice property [150, 184J
a
o ::; lal - a tanh( -) ::; 0.2785es, 'Va E R (4.96)
es
Since such a modification may cause the estimated gain k and the NN weights
Wand V grow unboundedly due to the loss of asymptotic tracking, the fol-
lowing modified adaptive laws can be applied by using the a-modification
scheme [74J:
~ = -rw[(S-S'VTz)es+awW] (4.97)
~ = -rv[zwTs1es+avV] (4.98)
Theorem 4.6. For system {4.1}, if controller {4.79} with modified sliding mode
control (4.94) is chosen and adaptive laws (4.97}-{4.99) are taken, then, for
bounded initial conditions,
{i} all signals in the closed-loop system are uniformly ultimately bounded,
(ii) there exists a compact fix such that x E fix, 'Vt :2: 0, and
{iii} the tracking error converges to a small neighbourhood of zero, which can
be freely adjusted by tuning the design parameters ks, es, a w , a v and ak.
78 Chapter 4. SISO Nonlinear Systems
(4.101)
with constant
f30 = 0.2785E s(kr + k2 + k;) + a; IIW*112 + ~v IIV*II} + ~k 111(*11 2
Hence, Vs < 0 when outside a compact set
4.4 Conclusion
In this chapter, adaptive control schemes have been studied for a class of non-
linear dynamical systems using both LPNN and MNN. A regionally stable
adaptive controller is firstly developed for the systems with Assumptions 4.2-
4.3 holding on a given compact set. The designed control structure avoids the
possible singularity problem, and the stability of the closed-loop adaptive sys-
tem is guaranteed. In addition, the transient behaviour of the adaptive neural
system has been investigated, and possible methods were provided for improv-
ing the system response. Then, for systems with Assumptions 4.5-4.6 satisfied
on whole space, a semi-globally stable VSC-based controller has been studied
by combining multilayer neural networks with variable structure control strat-
egy. To avoid the chattering problem, a modified adaptive controller was also
provided by introducing a smooth switch function to replace the discontinuous
sign function. The modified scheme guarantees that the tracking error con-
verges to a small neighbourhood of zero, which can be adjusted by the designer
through the tuning of the controller parameters.
80 Chapter 4. SISO Nonlinear Systems
06,---.--.---.--.---,----,
0.2
-0.2
.,"
...... 1. .. ~ I
-0.4
\ :I
\]
-O.60~--:---,':---.L----L---.L---'---"o
2.5
1.5
O.SO':-------':---":-----'-15----'-ro----'-25---".1O
5.1 Introduction
In nonlinear control design, Lyapunov stability techniques play an important
role in constructing controllers and performing the closed-loop stability analy-
sis. Nevertheless, for a given nonlinear system, there is in general no sys-
tematic procedure for choosing a suitable Lyapunov function to guarantee the
system stability. Different choices of Lyapunov functions may result in differ-
ent control structures and control performance. Past experience shows that
a good design of Lyapunov function should fully utilize the property of the
studied systems. Adaptive control of nonlinear systems has been an active
research area and many good theoretical results have been obtained in the lit-
erature [87,91,104,128,165] and the references therein. In this chapter, a novel
kind of Lyapunov functions is developed for achieving globally stable nonlinear
adaptive control.
As we have seen, most of neural adaptive control schemes in the literature
are based on feedback linearization strategy, i.e., trying to remove the effects
of the system's nonlinearities and forcing the resulting system as a linear one.
In fact, there is no necessity to cancel all the nonlinearities for a stable closed
loop [104]. In many cases, some nonlinearities of the systems may be helpful
for achieving control objective. By introducing an integral Lyapunov function
(ILF), novel control structures are developed for several classes of SISO non-
linear systems, strict-feedback nonlinear systems and MIMO nonlinear systems
in triangular form in this chapter. The proposed designs completely avoid the
singularity problem existing in feedback linearization approaches.
In this chapter, we shall first introduce the definition of the Integral Lya-
punov Function (ILF) in solving adaptive control problems of SISO nonlinear
systems with matching condition. Then, ILF is applied to solve for a class of
where x = [Xl X2 ... xnjT ERn, U E R , Y E R are the state variables, system
input and output, respectively; f(x) and g(x) are unknown functions of x.
Though the SISO system takes the same form as system (4.1) in Chapter
4, the design methodology and the obtained controller used in this section are
different from those in Chapter 4.
Assumption 5.1. The sign of g(x) is known, and there exist known con-
tinuous functions f{x) ;::: 0, g(x) > 0 and a known constant 90 such that
If(x)1 ::; f{x) and 0 < go::; Ig(x)1 ::; g(x), Vx ERn.
The above assumption implies that the continuous function g(x) is strictly
either positive or negative. Without losing generality, we shall assume g(x) ;:::
go > 0 for all x E Rn.
Assumption 5.2. The desired trajectory vector Xd = [Yd, 'IJd, ... ,yd'lT E Od C
R n +l with Od a compact subset.
The control objective is to find a controller u such that output y follows a
given reference signal Yd. Define vectors Xd, e and a filtered error e s as
. (n-l)jT
/\r-
Xd = [Yd Yd ... Yd ,
es = (:t + l
el = [AT l]e (5.2)
Ve = i 0
e,
0'[0 1 + (a
2 e~ [. 4 2 2 2
+ 1.11) ]da = 4" (e s + 31.11) + gl.ll + 20 1 .
]
Clearly, the above function is positive definite and radially unbounded with
respect to e s .
Ve = l' i
o
e, l'
ada = -e s
2
2
(5.5)
84 Chapter 5. ILF for Adaptive Control
(5.7)
Lemma 5.1. For system (5.1) with Assmnptions 5.1-5,2 satisfied, if x E Leo,
and the desired continuous Lyapunov-based controller is chosen as
where the botmded continuous function k(t) ~ k* > 0 with constant k*, the
weighting function a(x) renders the Ve in (5,4) satisfying conditions (1) and
(2) in Definition 5.1, and the smooth function h(Z) is defined in (S.7), then
the system tracking error converges to zero asymptotically.
Proof Taking the time derivative of Ve given in (5.4) along (5.3), we have
Ve = g;;1 (x)ese s
+ i es
£T
[8g,;-l(xn-l,£T+VIL
8 X n-l +
8g,;-I(Xn_l,£T+vd.
8 VI
Jd£T
i
o X n- l VI
f(X) ] e• ~ 8g,;-I(Xn_l' £T + vd
= a (x)e s [ g(x) + u + g';- (x)esv +
1
0 £T l~ 8Xi Xi+l
(5.10)
5.2 Matching SISO Nonlinear Systems 85
lo
e• 8g;;1(xn_1,{1'+V1). d i e . og;;1(xn-b{1'+v1)d
{1'
8Vl
V1 (1' = -v
0
0-
80-
0-
. [l(x)
Ve = a(x)e s g(x) + u + h(Z) ] (5.11)
with h(Z) defined in (5.7). Substituting the controller given in (5.9), (5.11)
becomes
v.e = -k(t)e 2 < -k*e 2 < 0
8 - S -
(5.12)
Integrating it, we have
Ve = l e
• {1'g-1(X n _1, {1' + IIdoo (5.13)
Since 0 < go :S g(x) :S g(x), x ERn, from Assumption 5.1, we know that Ve is
a positive definite and radially unbounded function with respect to e s .
The corresponding equations of (5.6)-(5.8) for this case become
() l es[
PI e s =
o
a
~[)g-I(Xn_l,a+vr)
~
i=1
8 Xi+1
Xi
hl(Z) = PI(e S ) , Z = [T
X ,e s , v, VI
]T (5.16)
es
(5.17)
with the bounded continuous function k(t) ~ k* > 0 with constant k*, and
the smooth function h(Z) defined in (5.16), then the system tracking error
converges to zero asymptotically.
Proof It follows directly from Lemma 5.1.
(5.21)
.5.2 Matching SISO Nonlinear Systems 87
(5.22)
(5.23)
-1 (-
vgmax Xn-l, VI )
(5.24)
g(x)
which shows that hmax(Z) is smooth and well defined.
Equation (5.23) can be further decomposed as
hmax(Z) = h 1max (Z) + h 2max (Z) Z = [xT, es , V, Vl]T (.5.25)
where
h
Imax
(Z) Plmax
= esg-(x) ' Plmax -
-ie. o
~ ag;"~x(Xn-l'
a ~
i=1
a
a + VI) . d (526)
Xi
X.+l a .
(5.27)
(.5.28)
h
max (Z)
= _1_11
_( )
g X 0
(J
es
~ ag;"~x(Xn-l' (Je s + VI) . d(J
~
i=1
a
Xi
X,+1
Similarly, we have the following corollary that provides a desired feedback con-
trol for achieving the control objective.
wher'e the smooth function hmax(Z) is given in (5.30) with Z = [xT, e s , v, Vl]T
and the b01tnded continuous function k(t) 2 k* > 0 with constant k* I then the
system tracking error converges to zero asymptotically.
Proof It follows directly from Lemma 5.1.
Y= Xl
W;
(1) if it is the properties of the system, then E RPl and W; E RP2 are the
vectors of unknown constant parameters, Stg{x) E RPl and Sg{x) E RP2
are known regressor vectors, and functions fg{x),go{x) E C 1 are known;
(2) if fg{x) and g{x) are unknown and approximators are used for parame-
trization, then Wi E RPl and W; E RP2 are the vectors of unknown
constant weights, Stg{x) E RPl and Sg{x) E RP2 are known regressor
basis functions, functions J6{x), 90{x) E C 1 denote the unknown residual
errors over a compact set n.
Clearly, the unknown parameter vector W; enters into system (5.1) non-
linearly. Many practical systems, such as pendulum plants [6, 19] and fer-
mentation processes [14], can be described by system (5.1) and possess such
a kind of nonlinear parameterizations. For convenience of analysis, define
sg{x) = Sg{x)a{x) and sgo{x) = go{x)a{x), then
(5.35)
(5.36)
where
(5.37)
90 Chapter .5. ILF for Adaptive Control
where
h (Z) _ W;1'
11 - ()
esG x
ie, [
0
0' ~
~
i=1
asg(Xn-l, 0' + VI) .
aXi XH 1
h (Z) _ _
12 - ()
1_1e s
[
0' ~
~ asgO(Xn-l, 0' + VI) .
aXi Xt+1
esG X 0 i=1
where bounded continuous function k(t) :::: k* > 0 with constant k*, W*T =
[wy, W;T], and by applying Lemma 2.6,
= [[S1' (,)]
/g x ,
[_1_11 [0
G(X) 0 es ~
~ as;;(Xn-1, Oe s + vd
ax. .
X t +1
'i=1 t
+vSg(Xn_l,oes+Vl)]do]T] (.5.41)
h (Z)=fg(,)
10 JO x
_1_1e,[0' ~asgO(Xn-I,O'+VI),.
+ esG ()
X
a
0
~
i=1 Xi
X,+I
-fg(,)
- JO X
_1_1
+ G ()
X 0
1 [O ~asgO(Xn-I,OeS+vI),.
es ~
i=1
a Xt+1
Xi
It is clear that that both S(Z) and hLO(Z) are smooth and well defined.
Now, we are ready to present the following corollary that provides a desired
feedback control for achieving the control objective.
Now, we are ready to present the following corollary that provides a desired
feedback control for achieving the control objective.
.. k(t)e s ..T ( - () ()
u nn = - a{x) - W S Z) - hLO Z sgn es , (5.45)
where smooth f1tnction hLO(Z) is defined in (5.44), then the system tracking
error converges to zero asymptotically.
Proof Taking the time derivative of v" given in (5.4) along (5,3) and by fol-
lowing exactly the derivations in the proof of Lemma 5.1, we have
(5.46)
92 Chapter 5. ILF for Adaptive Control
with hlO(Z) defined in (5.42). Substituting the controller given in (5.45), (5.46)
becomes
(5.49)
with a compact subset ne to be specified later. Because the function f(x)f g(x)+
h1 (Z) in (5.18) is smooth, we have the following approximation
f(x)
g(x) + h1 () .. T ( )
Z = W S Z + J-tl, 'r/Z E flz (5.50)
5.2 Matching SISO Nonlinear Systems 93
with III the function approximation error and W* the ideal weights defined as
with constant design parameter e > 0 and the initial weight IIW(O)II :s; Wm ,
then
(i) all of the signals in the closed-loop system are bounded and there exists a
computable time T ~ 0 such that for all t ~ T, the system state x remains
in
Proof It should be noticed that condition x E Loo plays an important role for
achieving the stable tracking control in Theorem .5.5. In the following, we first
establish the boundedness of the system states.
(i) Consider a Lyapunov function candidate V1 = e;/2. Substituting (5.51)
into (5.3), we have
Because g(x)k(t) ;:::: gals (as k(t) ;:::: lis obtained from (5.53)), the above in-
equality implies that
(5.58)
It follows from (2.23) that the bounded ness of e s implies the boundedness of the
system state x. Because IIW(O)II S W m , the adaptive law (5.52) with projection
algorithm guarantees IIW(t) II S W m , '<It;:::: O. We conclude that all of the signals
in the closed-loop system are bounded.
From (5.58), we have
Substituting controller (5.51) into the above equation and using the adaptive
law (5.52), we obtain
v = es [ -k(t)es-WTS(Z)+J.Ll] +'Y-1WTW
v~es[ -k(t)eS+J.Ll], Vt ~ T
(5.62)
Since Ve = Joe, a/g(xn-b a + vddn :::; e~/2go (followed by g(x) > go) and
IIW(t)1I :::; 2wm (as IIW(t)1I :::; W m , "It ~ 0), it is shown that
V(T) = Ve(T) + -y- 1 IlW(T)1!2 /2:::; e;(T)/2go + 2-y-lw~
It is clear that the last three terms of the above inequality are non-negative
constants and computable. Because the tracking error el = H(s)e s with H(s)
a proper stable transfer function, using Lemma 2.10 we conclude that (5.55)
holds with computable constants Cl, C2 ~ o. <>
Remark 5.1. We can see from Theorem 5.5 that the adaptive controller (5.51)
contains two parts for different purposes. The first part -k{t)e s can be viewed
as a bounding control term which is used to guarantee the system states expo-
nentially converge to a designed compact set lle for any initial states. It has
been shown that after a computable transient time T, function approximation
(5.50) is valid on Z E Hz, "It ~ T. The second part in controller (5.51) is an
adaptive term which is applied for learning the unknown nonlinearity of the
system to achieve adaptive tracking control.
Remark 5.2. Noting (5.58), we have es (t)2 :::; e~(O) + 3g 2 , "It ~ O. Equation
(2.23) implies that the tracking error
5.2 Matching SISO Nonlinear Systems 97
Therefore, the upper bound of tracking error depends on the initial state errors
and the design parameters ,\ and E. As time tends to infinity, the mean square
tracking error given in (5.55) satisfies
lim -
t---+(X)
lit
t 0
er(T)dT :::; E2CI lim -
t---+(X)
lit
t 0
J11(T)dT
which indicates the mean square error depending on the function approximation
error ILL and the design parameter E only.
Remark 5.3. Although decreasing E may reduce the tracking error signifi-
cantly, we do not suggest the use of a very small E in practical applications
because this may lead to a high-gain controller and requires powerful actua-
tors. The main purpose of setting the design parameter E is to adjust the size of
ng . From a practical perspective, the size of the compact set nz (depending on
ng and nd ) is very important in the construction of the function approximator.
For example, in the case of applying RBF neural networks, the selection of the
number and the centers of the basis functions is based on a region of interest
that corresponds to n z. In general, the larger the set nz is, the more NN nodes
are needed for achieving the same approximation accuracy.
Simulation Illustration
To show the effectiveness, the adaptive controller (5.51) and (5.52) are used to
control the pendulum plant shown in Figure 4.5 with the same parameters as
in Section 4.2.3. The plant can be expressed in the form of system (5.1) with
f(x) = 0.5sinxI(1 + 0.5cosxdx~ -lOsinxI(l + cos Xl)
0.2.5(2 + COSXI)2
1
g( x) - ---,------,- (5.65)
- 0.25(2 + cos XI)2
The initial states of the pendulum plant are chosen as [Xl (0), x2(0)f = [7f /6, OlT,
and the control objective is to make the output y track the reference Yd =
~ sin(t).
We first determine the bounds f(x) and go in Assumption 5.1. From (5.65)
we have If(x)1 < (0.75x~ + 10)/0.2.5 = 3x~ + 40 and Ig(x)1 ~ ~. Suppose
that only the conservative bounds f(x) = 3x~ + 40 and go = 4/9 are known
for the pendulum plant, and no other information about f (x) and g( x) is
available. In the following simulation, we choose ,\ = 0.5 and E = 2. A high-
order neural networks (HONNs) is used as the function approximator in the
adaptive controller (5 ..51), whose the basis functions described as
+e
e- Zj
-z (5.66)
J
JEIi
98 Chapter 5. rLF for Adaptive Control
where the input vector Z = [Xl, X2, e s , v, VI]T, {Jr, h,··· , It} is a collection of
l not-ordered subsets of {1, 2, ... ,5} and dj(i) are non-negative integers. In
this simulation test, a second order neural network is selected with the element
Si(Z) in (5.66) are chosen as S(Zl)' S(Z2)' S(Z3), S(Z4)' S(Z5), and the possible
combinations of them (i.e., S(Zi)S(Zj) with i, j = 1,2,3,4,5). The total number
of the NN nodes is 20. The parameters in the adaptive law (5.52) are taken as
Wm = 10.0, 'Y = 6.0 and the initial NN weight W(O) = 0.0.
o~r~
, 0.5
0.6
,
I
0.4
I
......... 0.4
0.2
-0.2
....
0.1
-0.4 .... .....
t •••
. f
-O.S
,
f··· , ......... I· .....
-0.8
0
;'
10 20 30 .. 50
sec
60 ~ ~
5 .
-5 ..
-10 ..
°0L---~'0--~L---~30--~"----~~--~~60 -15oL---~'0----'2O----30":-----':"-----:50L----!60
The tracking result in Figure 5.1(a) indicates that the system output Y
does not follow the reference Yd effectively during the initial 10 seconds. This
is because less information for the plant is available and the initial NN weights
are set to zero. After the NN on-line learning phase during the first 30 seconds,
5.2 Matching S1SO Nonlinear Systems 99
we can see that an almost perfect tracking is achieved and the output tracking
error shown in Figure 5.1(b) converges to a small region. The histories of the
NN weight norm IIWII and the control effort are shown in Figures 5.1(c) and
5.1(d), respectively. These simulation results confirm the effectiveness of the
proposed adaptive control scheme.
* 1 g(x)f{x) _
umax = -=--() [-k{t)e s - () - g{x)hmax{Z)] (5.67)
9 x g x
(5.68)
where Oz is given as (5.49) and J-tl is the function approximation error. With
the function approximation (5.68), we present the adaptive controller as follows
1 'T 'T
(5.69)
u = g(x) [-k(t)e s - W S(V Z)]
where neural network WT SeVT Z) is to approximate W*T S(V*T Z). The adap-
ti ve laws are taken as
(5.70)
(5.71)
with constant design parameter c: > 0, then for bounded initial conditions
x(O), W(O) and V(O), all signals in the closed-loop system are bounded.
100 Chapter 5. ILF for Adaptive Control
Proof The proof includes two parts. We first suppose that there exists a
compact set Oz such that Z E Oz, Vt > 0, and (5.68) holds. Then we show that
this compact set Oz do exist for bounded initial conditions.
Consider the Lyapunov function candidate
V = Ve+WT')',·;/W+tr{VT')';1V}
- f(x) - T -1 ~ - T -1 ~
= g(x)es( g(x) + hmax(Z) + u} + W ')'w W + tr{V ')'V V} (5.74)
Since
2
11V*IIFIIZWTS'IiFlesl ::; ~ IIZWTS'II} + ~IIV*II} (5.79)
(5.82)
with constant
(5.84)
Therefore,
(5.85)
where constant /\1 = min{go/c, lTw//\max(')'.;;;-1), lTv r\max (')';1)}. Solving in-
equality (5.84), we obtain
where positive constant Co = V{O). It follows from (5.21), (5.73) and (5.86)
that
(5.87)
This confirms that for bounded initial conditions, all signals e s , ltV and V of
the closed-loop system are bounded and there do exist a compact set Oz such
that Z E Oz for all time. <)
u = -ke-s
- W T S ( Z ) - hlO(Z)
A
(5.88)
a(x)
where ltV is the estimate of W* and design parameter k > O.
102 Chapter 5. ILF for Adaptive Control
Theorem 5.7. For system (5.1) satisfying Ass'umption 5.1, if weighting func-
t'ion a(x) renders the Ve in (5.4) satisfying Conditions (1) and (2) in Definition
5.1, then controller (5.88) with adaptive law (5.92) guarantees the boundedness
of all the signals in the closed-loop system and the globally asymptotic tracking,
i.e., limt-><Xl yet) = Yd(t).
Proof For the Ve given in (5.4), its derivative is (5.46). Substituting (5.88) into
(5.46), we obtain
(5.89)
The system stability is not clear at this stage because the last term in (5.89) is
indefinite and contains unknown W. To remove such an uncertainty, parameter
adaptive tuning is introduced for W. To construct the adaptive law, augment
Ve as follows
(5.90)
(5.91)
W = rS(Z)a(x)e s (5.92)
which leads to
V= -ke; ~ O. (5.93)
Since function g.;l(x) E c1, (5.4) shows that Ve is a C 1 function of x and Xd.
This guarantees that Ve(O) E L<Xl for any bounded initial values x(O) and Xd(O).
Integrating (5.93), we have fooo ke;(r)dr ::; V(O) < 00 and 0 ::; 'let) ::; V(O),
n
This implies that es E L2 L<Xl and Wet) is bounded. Consequently, u and es
n
are also bounded. Since es E L2 Loo and es E L oo , we conclude limt-><Xl e s = 0
by Barbalat's Lemma (Lemma 2.2). It follows from (2.23) that ::/; E Loo and
the tracking error converges to zero asymptotically. <>
As shown earlier, a key step in the design procedure is the choice of WF a( x)
and WCLF Ve. It should be pointed out that for a given system, different WF
can be found to construct different WCLF. Therefore, the resulting controller is
not unique and the control performance also varies with the choice of WCLFs,
This brings the designer some degrees of freedom in controller design. In the
following, we show that for controller (5.88) with a suitably chosen WF a(x),
transient performance of the closed-loop system can be guaranteed.
5.2 Matching SISO Nonlinear Systems 103
Theorem 5.S. !48}For the closed-loop adaptive system (5.1), (5.88) and (5.92),
if WF a(x) is chosen such that g;;l(x) ::; Co with Co a positive constant, then
(i) the L2 transient bound of the filtered error
with computable constants k o , /\0 > 0 which depend on the design para-
meter /\.
Proof (i) If a(x) is chosen such that 0 < g;;l(x) ::; Co, then
(5.96)
Therefore
we have
Remark 5.4. The Loo bound of the error vector ( in Theorem 5.8 is obtained
for high-order systems (n ~ 2). For a first-order system, to get an explicit
bound of the tracking error, an additional condition g.;;l(x) ~ Cl is needed for
the choice of a(x). In this case (e s = el for n = 1)
(.5.100)
Noticing Ve(t) ~ V(t) ~ V(O), we have er(t) ~ 2V(0)/Cl' From Ve(O) <
coei(0)/2, the Leo tracking bound for the first-order system can be found
(5.101)
Remark 5.5. Theorem 5.8 shows that different choices of WF a(x) may pro-
duce different control performance. As g.;;I(X) = [W*TSg(x) + go(x)]a(x) with
known functions Sg(x) and go(x), it is not difficult to design a WF a(x) to make
o < Cl ~ g.;;I(X) ~ co. For example, if g-l(x) = exp( -x;,)(WI + W2:l;;') with
constant parameters WI, W2 > 0, then one may take a(x) = exp(x;')/(l + x;')
which leads to
Remark 5.6. From a practical point of view, Assumption 5.1 holds on the
whole space might be a strong restriction for many physical plants. If Assump-
tion 5.1 holds only on a compact subset n c R n , the proposed approach is
still applicable if the controller parameters are designed appropriately. The
reason is that by suitably choosing the design parameters, upper bounds of
the states (derived from (5.95) and (5.101» are adjustable by the designer,
and subsequently can be guaranteed within the given compact set n in which
Assumption 5.1 is satisfied for all time.
5.3 Backstepping Adaptive NN Design 105
Simulation Studies
To show the controller design procedure and validate the effectiveness of the
developed scheme, we consider a second-order system
Xl = X2
. xi +u
x2 = --;----;~::-::-::--_::_:_:=---''''" (5.102)
exp( -xi)(Wl + W2X§)
with unknown parameters WI, W 2 > O. The objective is to control the output
11 = Xl to follow the reference 1Id(t) = sin(0.5t). Plant (5.102) can be expressed
in the form of system (5.32) with !g(x) = xi, and g-l(x) = exp( -xi)(Wl +
Wzx§). Comparing with (5.33)-(5.34), we have !g(x) = xi, W = [WI Wz]T
and Sg(x) = [exp(-xr) exp(-xi)x~V. In view of Remark 5.5, we choose the
WF a(x) = exp(xi)/(l + x~). It follows from (5.41), that
= ~(
esO! x
) [tan-Ixz-tan-lvl es-tan-lxz+tan-lvl]T
with VI = Yd - '\(XI -1Id), V = -Yd + ,\(xz - Yd), and h(Z) = xi. Then, (5.88)
and (5.92) suggest the following controller
In the simulation, the true values of the system parameters are [WI WzV =
[2.0 0.5]T and the initial condition is [Xl(O) xZ(O)]T = [0.5 o.of. The para-
meters of the adaptive controller are ,\ = 1.0, k = 1.0, II = 10.0, ,2
= 25.0
and [WI (0) W2(0)]T = [0.0 O.O]T. The simulation result given in Figure 5.2(a)
indicates that the output tracking error converges to zero asymptotically. The
boundedness of the estimated parameters and control input is shown in Figures
5.2(b) and 5.2(c), respectively.
10 15 00 25 30 35 40
Assumption 5.3. There exist constants giO > 0 and known smooth functions
iJi(Xi) such that 9;0 S; !gi(Xi)! S; lJi(Xi), '<Ix ERn.
We further assume that the signs of gi(Xi) are known. In the following,
we only consider the case of giO S; gi(Xi) S; gi(Xi), '<Ix ERn. Assumption 5.3
is a mild condition because gi(Xi) being away from zeros are controllability
conditions of system (5.103), which is made in most of control schemes [104,
166]. J:<or a given practical system, the upper bounds of gi (Xi) are not difficult to
determine by choosing functions !}i(Xi) large enough. It should be emphasized
that the low bounds giO are only required for analytical purposes, their true
values are not necessarily known.
The control objective is to design an adaptive controller for strict-feedback
system (5.103) such that the output y follows a desired trajectory Yd, while all
signals in the closed-loop systems are bounded. Define
-
Xd(i+l) = [
Yd,.Yd, ... , Yd' , ~ = 1,"', n
(i)]T'
Assumption 5.4. The desired trajectory vectors Xd'i with i = 2, ... , n+ 1 are
continuous and available, and Xdi E Odi C Ri with Odi known compact sets.
108 Chapter 5. ILF for Adaptive Control
(5.104)
(5.105)
Applying Lemma 2.6, the function Vzl (Zl' Yd) given in (5.105) can be re-written
as
Noting that 1 ::; glc!(Ozl + Yd) ::; 9l(OZl + Yd)/glO (Assumption 5.3), we have
(5.106)
Therefore, Vzl is a positive definite and radially unbounded function with re-
spect to Zl. In the following, Lyapunov function candidate Vzl shall be used
for constructing a desired feedback control (DFC).
Lemma 5.2. For the first-order system (5.104), if the DFC is chosen as
(5.107)
(5.108)
with Zl = [Xl, Yd, YdjT E R3 and k(t) > 0, then the system tracking error
converges to zero asymptotically.
Proof Taking Vz1 given in (5.105) as the Lyapunov function candidate, its time
5.3 Backstepping Adaptive NN Design 109
-1
= Zlglrnax (
Xl Zl +
).
1 ()" a
o
Z1 -1 (
a ()"
glrnax
Yd
+ Yd ) Yd. d.()"
-1
= Zlglrnax (
Xl Zl + Yd
)..
1 0
Z1
()"
aglrnax
-1 (
()" + Yd ) d
a(}" ()"
= Zlg1';wx(Xl)[gl(X1)U1 + h(X1) - Yd]
+Yd [(}"gl,,~ax (()" + Yd)I:1-1Z~1~ax(()" + Yd)d(}"]
= Zl [91 (X1)U1 + gl,,~ax(x1)h (Xl) - Yd t1 gl~ax(()" + Yd)d(}"]
Zl Jo
= Zl [91 (XdUl + gl,,~ax(xl)fdxd- Yd 11 gl~ax(OZl + Yd)dO] (05.109)
Substituting U1 = ui into (05.109), we obtain
. 2
Vz1 = -k(t)Zl ~ 0
(05.110)
where 21 = [Z[ l]T and o, z1 is a compact subset to be specified later.
Assumption 5.5. For NN approximation (05.110), there exist ideal constant
weights W{ and vtsuch that 111'11 ~ iiI with constant Ii! > 0 for all Zl E o, z1 '
The following adaptive controller can be taken for the first-order system
(5.104)
(5.111)
(5.113)
(5.114)
with rlU = rru > 0, r v = r:, > 0, and a lU, a v > 0, then for bounded 'initial
conditions Xl(O), WI (0) and V1 (0), all signals in the closed-loop system are
bounded, and the vector Z 1 remains in
Vs1 (.5.117)
By noting (3.58) in Lemma 3.9, (5.112) and the properties that
-r
2WI WI 2: IIWIII 2 -lIwtll 2
A
-
-T 2 2
2tr{VI VI} 2: 11V11I F - IlVt II F
A
-
5.3 Backstepping Adaptive NN Design 111
(5.119)
(5.120)
with constant
i sI
T ::; - 910 Vzl _ O"w IIW1 11 2 _ O"v II VI II} + Cl
Cl 2 2
Therefore
(5.122)
where constant
where positive constant Co = Vsl (0). It follows from (5.106), (5.116) and (5.123)
that
Z2( t)
2 -< Vzl (t) -< V.sl (t) -< c0 e-
_1_ A1t +~
C
/\1 '
Vt >_ 0 (5.124)
This confirms that for bounded initial conditions, all signals Zl, Wl and (T1 of
the closed-loop system are bounded and there do exist a compact set Ozl such
that Zl E Ozl for all time. <>
Remark 5.7. Most of the available NN-based controllers in the literature are
based on feeback linearization techniques, whose structures are usually in the
form 1./,1 = [-il(xd + V1l!§1(Xl), where VI is the new control variable. To
avoid singularity problem when §l (Xl) -- 0, several modified adaptive meth-
ods were provided [99,171,179, 186J. However, either discontinuous controllers
or adaptive laws has to be used, which become the main obstacle for using
backstepping method to extend these schemes to strict-feedback nonlinear sys-
tems. The proposed controller (5.111) and updating laws (5.113)-(5.114) are
smooth, hence, the use of backstepping design is possible for extending the
approach to high-order systems.
It follows from Lemma 3.9 that the function estimation error can be expressed
as
(5.126)
5.3 Backstepping Adaptive NN Design 113
(5.127)
(5.128)
Step 2
(5.129)
Again, by viewing X3 as a virtual control, we may design a control input (}:2 for
(5.129). Define a new variable Z3 = X3 - (}:2, we have
(5.130)
114 Chapter 5. ILF for Adaptive Control
where
(5.133)
we obtain
. 2 -
Vs2 = -kl(t)zl -1/Jlzl + Z191(xd z2
+ Z2 [92(X2)(Z3 + 0!2) + h 2(Z2)] (5.134)
where
5.3 Backstepping Adaptive NN Design 115
If Ci2 is chosen as
where
A similar procedure is employed recursively for each step. Consider the equa-
tion of system (5.103) for i = k given by
(5.135)
Note that
116 Chapter 5. ILF for Adaptive Control
where
(5.136)
with Wj and 1)j,p to be designed as smooth functions of Wj, -0;, Xd(j+l) and
Xj,j = 1,2"" ,k -1 later. Following the similar procedure of (5.132)-(5.134)
in Step 2, we obtain
k-l
where
By letting
(.5.138)
where
(5.139)
with constant Ck > 0, and noting (5.126) and the recursive design procedure,
we obtain
k
Vsk = - L [kj(t)z] + '!f;jZj] + zk?h(Xk)Zk+l
j=l
Step n
(5.140)
where
(5.142)
where
(5.143)
have been used, which is the key point of the proposed design method. Ac-
cording to Assumption 5.3, we know that
Vzi = 11
z; Og~ax(Xi-1,OZi + Qi-ddO ~ z; 11 1OdO = (5.145)
z?
Vzi = :::; -~
giO
11
0
OfJi(Xi-1,OZi + Qi-l)dO (5.146)
Next, we shall provide the adaptive laws such that the stability and control
performance of the closed-loop system can be guaranteed.
Theorem 5.10. Consider the closed-loop system consisting of strict-feedback
system (5.103) satisfying Assumption 5.3, controller (5. 142} and the NN weight
updating laws
(5.147)
(5.148)
where r wj = r~j > 0, r vj = r0 > 0, and Uwj,Uvj > 0, j = 1,2"" ,n, are
constant design parameters. Then for bounded initial conditions,
(i) all signals in the closed-loop system are bounded, and the vectors Zj re-
main in the compact sets
(5.149)
lim -
t->oo
lit
t 0
2c'
z;(r)dr::; __1_ ~:::>i
1 + gjO i=1
n
(5.150)
2
L +~L
n n
z;(t) :::; 2Vs(0)e-Ast Ci, "It ~ 0 (5.151)
i=1 / s i=1
r
Proof (i) For simplicity, let w = wj,r rv
= r vj , (jw = (jwj and (jv = (jvj for
j = 1,2, ... ,n. Consider the Lyapunov function candidate
Vs = Vsn +"21 ~ [ - T -1 -
~ Wj rw Wj + tr{Vj rv Vj}
- T -1 - ]
(5.152)
J=l
By taking its time derivative along (5.144) and noting (5.126), it follows that
-WTr~lWj - tr{Vlr;lVj}]
Substituting (5.147) and (5.148) into the above equation and using the fact
that
we obtain
Noting (5.139), (5.143), (5.127) and following the same procedure in the proof
of Theorem 5.9 (from (5.117)-(5.121)), we further obtain
(5.153)
where
(5.1.56)
where constant
1
L 2L
n n
\I~(t) 2: Vsn = V zj 2: zJ(t) (5.159)
j=l j=1
Inequalities (5.156), (5.158) and (5.159) confirm that all the signals x, Zj, Wj
and 'Cj in the closed-loop system are bounded.
;s
Let Co = 2Vs(0) + L~l Ci, we conclude that there do exist compact sets
fl zj defined in (5.149) such that the vectors Zj E nzj for all time.
(ii) It follows from tiJ(Xj) 2: 9jo (Assumption 5.3) that
j = 1,2"" , n (.5.160)
which implies that inequality (S.lS0) holds. The error bound (S.lS1) can be
derived from (S.1.56) and (S.159) directly. <)
Remark 5.8. Since the function approximation property (.5.11O) of neural net-
works is only guaranteed within a compact set, the stability result proposed
is semi-global, i.e., for any compact set there exists a controller with sufficient
large number of NN nodes such that all the signals of the closed-loop system are
bounded if the initial states are in this compact set. It should be pointed out
that the sizes of the compact sets S1 zj are not available because they depend
on the unknown parameters Wj* and "1*.
(5.161)
where
(.5.162)
(5.163)
(ao = Yd, Wo = Yd and Zo = 0 are used for notational convenience) and the
NN weight updating laws are chosen as {5.147} and {5.148}, then, for bounded
initial conditions,
(i) the boundedness of the closed-loop system is guaranteed, and there exists
122 Chapter 5. ILF for Adaptive Control
lim - (
1 t ( n Z2) dT::; 2:
2: -2 n
Ci (5.165)
t->oo t io j=1 €j i=1
Proof (i) Noticing that controller (5.161)-(5.163) is the same as the controller
given in section 5.3.2 by taking Q:i(Xi) = iJi(xd = 1 (except for (5.145) and its
corresponding parts), therefore all derivations (5.152)-(5.158) in Theorem 5.10
are valid here, and we may continue the proof as follows. From (5.153), by
taking Q:i(Xi) = 9i(Xi) = 1 we obtain
(5.166)
s t,Cj}
According to an extension of Lyapunov theorem [137]' we conclude that all
the signals of the system are uniformly ultimately bounded and there exists a
constant Co such that I:~1 z;(t)/€i ::; Co. Therefore, the compact sets Ozj
defined in (5.164) do exist such that the vectors Zj E zj , 'tit 2 O. n
(ii) It follows from (5.166) that
.S _ 2: -2+ L
n 2 n
v.<- Z '
CJ (5.167)
j=1 €j j=1
Remark 5.9. Comparing the results of Theorems 5.10 and 5.11, we can see
that the upper bounds of Zj cannot be explicitly captured in Theorem 5.11 due
to the lack of the knowledge for gi(Xi). Nevertheless, the semi-global stability
of the closed-loop system is still guaranteed. It is shown from (5.165) that the
mean square tracking performance can also be achieved and the tracking error
is adjustable by tuning the design parameters EO j, O"w and O"u·
I;l :
with the output y = Xl and the initial condition [Xl (0), X2(0)jT = [0, ojT.
Clearly, Plants I;1-I;3 are in strict-feedback form and satisfy Assumption 5.3. In
the following simulation, we suppose that no information for the upper bounds
of gi(Xi) are available, and adaptive NN controller (5.161)-(5.163) are used to
achieve the control objectives. As plants I;1-I;3 are second order systems, the
following controller is chosen according to Theorem 5.11
(5.168)
124 Chapter .5. ILF for Adaptive Control
and NN weights WI, W2 and VI, V2 are updated by (5.147) and (5.148), respec-
tively.
05
-0.5
-I
(a) Output Y("-") follows Yd("- -") (b) Control input u(t)
4.5
4 ..
0.'
as
0.'
2.5
0.'
~~-+'0--~'0~~'~0-~4Cl-~'0~~'~O-~'0~~ \ W 20 • 4Cl 50 00 ro 00
(c) IIW111 ("-") and IIW2 11 ("- -") (d) I!VIIIF ("-") and 1!V211F ("- -")
............
-0.5 .•... . .. !.
I
-1 ·1 .. ·,· ....
- • .5 : • '
o w ro ~ ~ ro 00 ro 00 ffi ro ~ g ro 00 ro 00
(a) Output y("-") follows Yd("- -") (b) Control input u(t)
0.7
0.6
0.5
0.4 ...
0.'
~~~,0--~ro~~.~0--~~--~oo--~00L-~ro~00 ~ ffi 00 ~ ~ ~ 00 ro 00
(c) II WI II ("-") and IIW211 ("- -") (d) 111\IIF ("-") and IIV2 11F ("- -")
Figure 5.4: Simulation results for Plant E2 with Yd = 0.5[sin(t) + sin(0.5t)J
Figure 5.3 shows the simulation results for plant EI (which is known as
Van der Pol system [177]) with the reference signal Yd = sin(t). From Figure
5.3(a), we can see that a large tracking error exists during the first 20 seconds.
126 Chapter 5. ILF for Adaptive Control
This is due to the lack of the knowledge for the plant nonlinearities and the
initial weights are chosen randomly. Through the NN learning phase, better
tracking performance is obtained after 30 seconds. Hence, the proposed adap-
tive controller possesses the abilities of learning and controlling the nonlinear
system. The responses of the control input u and NN weight estimates II Wj II
and II~IIF' j = 1,2 are shown in Figures 5.3(b)-5.3(d), respectively.
'.5 2 .
0.5
-2 .
-4
-0.5
-, -6
:------:::'o-----:.20,-------,"':':---7:""-----=50':-----!OO·
-1.50 -80L---:::------=20,---ao:!::----="",---::.:-50---:!OO·
(a) Output y("-") follows Yd("- -") (b) Control input u(t)
~~--7:'0----::20':----:::"'-~""~-~50-----:.00·
reference Yd = sin(t) + O.5sin(3t). Figure 5.5 shows the simulation results and
confirms the effectiveness of the developed method.
Although the dynamics, operation regions and the reference signals of three
plants El - E3 are quite different, the simulation results indicate that the
proposed controller has strong adaptability and achieves better control perfor-
mance for all these nonlinear systems. This verifies the discussion in Remark
5.10 that the controller design is almost independent of plant nonlinearities.
functions are chosen to develop the stable neural controllers and adaptive laws.
As mentioned earlier, the couplings that exist among the inputs and outputs
of MIMO nonlinear systems are the main difficulty in multi variable adaptive
controL For the systems without any input coupling, decentralized adaptive
controllers were presented for a class of nonlinear systems with state inter-
connections [81,168]. Global stability and asymptotic tracking were obtained
in [168,170], provided that the system interconnections are bounded by a pth-
order polynomial in states. Recently, this decentralized adaptive control idea
was further extended to large-scale nonlinear systems transformable to a de-
centralized strict-feedback form in [81] under the similar interconnection con-
ditions.
In this section, we shall restrict our attention to a class of MIMO nonlinear
systems having a triangular structure in control inputs [51]. Using this trian-
gular property, integral Lyapunov functions are used to construct a Lyapunov-
based control structure, which does not try to cancel the "decoupling matrL'C"
to approximately linearize the system. The control singularity problem dis-
cussed above is eliminated without using projection algorithms or restricting
the initial values of parameter estimates. The system stability is guaranteed
without any restriction on system interconnections. Explicit transient bounds
of output tracking errors are given analytically and several possible schemes
are provided for improving control performance of the adaptive system.
Xl,il = Xl,i , +1
Xl,PI = h(X) + 9l,1(xd u l
X2,i2 = X2,i2+I
X2,P2 = h(X) + 92,I(X)UI + 92,2(XI,X2)U2, 1 ~ i j ~ pj-1
(5.169)
1 ~j ~m
Xj,ij = Xj,ij+1
Xj,Pj = fj(X) + 9j,1 (X)UI + 9j,2(X)U2 + ... + 9j,j(X1, X2, ... , Xj)Uj
Yj = Xj,1
where X = [xi, xI, ... , x~]T with Xj = [Xj,l, Xj,2,'" ,Xj,pj]T E RPj, Uj E R
and Yj E R are the state variables, the system inputs and outputs, respectively;
fj(-) and 9j,i(-) are unknown smooth functions; j, ij, Pj and m are positive
integers. It should be noticed that MIMO nonlinear system (5.169) is fully
interconnected because the interconnection terms fjO contain all the system
states X, which is different from the hierarchical or triangular systems defined
in [177]. The control objective is to design a stable adaptive controller such that
5.4 NN Control for MIMO Nonlinear Systems 129
the system outputs Yj follow the desired trajectories Ydj. For ease of analysis,
let Xj = [xI, xL··· ,xJJT E Rnj with nj = I:i=l Pi·
Assumption 5.6. Functions 9j,j(X j ) are bounded away from zero with known
signs, and there exist known smooth functions gj(Xj ) such that 19j,j(Xj )1 ~
f}j(X j ), \lXj E Rnj.
Remark 5.11. Assumption 5.6 implies that smooth functions 9j,j(Xj ) are
strictly either positive or negative definite. From now onwards, we shall as-
sume that 0 < 90j ~ 9j,j(Xj ) ~ f}j(Xj ) where the unknown constants 90j are
only used for analytical purpose, their true values are not necessarily known for
controller design. System (5.169) possesses a special triangular form in control
inputs, which covers a large class of plants including the decentralized systems
studied in [168,170].
Define Xdj, ej and filtered tracking errors esj as
where Aj = [A/
p'-l
,(Pj - 1) \3
p'-2
l ' •• 1(pj - T
1)Aj] .
Remark 5.12. It has been shown in Section 2.5.1 that definition (5.170) has
the following properties:
(ii) if the magnitudes of esj are bounded by constants C j > 0, then ej(t) E
Dcj for all ej(O) E Dcj with
(5.171)
(iii) the tracking errors ej,l = Hj(s)esj with Hj(s) proper stable transfer
functions.
Assumption 5.7. The desired trajectory vectors Xdj = [x~,y~j)]T are con-
tinuous and available, and Xdj E Ddj C RPj+l with Ddj being compact subsets.
130 Chapter 5. ILF for Adaptive Control
(5.172)
(5.173)
(5.175)
and therefore
(5.176)
Thus, V zj are positive definite with respect to esj. Remark 5.12 shows that the
boundedness of system states and convergence of tracking errors are completely
determined by those of filtered tracking errors esj. In the following, we shall
study the system stability and performance through investigating the properties
of esj' It will be shown later that integral type functions Vzj are applied for
constructing a Lyapunov-based control structure which does not contain the
inverses of the unknown functions 9j,j (Xj), and the subsequent adaptive design
avoids the use of estimates fJ;}(X j ).
5.4 NN Control for MIMO Nonlinear Systems 131
VZj = eSjgj';'ax(Xj)eSj
+ l
o
es;
0"
[8gj';'ax(.Xj,0"+Vj) '-
8
Xj
Xj +
8gj';'axCXj,0"+Vj).]
8
Vj
Vj do- (5.177)
Vj [eSjg;;'~ax(Xj) - JI~ ]
=
o gjn~axCXj, 0" + vj)dO"
with
II j-l{8gj;'aXCXj,oeSj+Vj)[
+ Jo 0 ~ 8Xi,Pi Ji(X) + t;
i-I
gi,k(X)Uk+gi,i(Xi)Ui
J} dO
and z·,1 -- tXT , //.J, //'.J, U 1, U 2, '" , U·'J-l ]T E R n ",+j+l .
Substituting (5.174) into (5.177) and noting (5.178) and (5.179), equalities
(5.177) can be expressed as
(5.180)
132 Chapter 5. ILF for Adaptive Control
where
j-1
hj(Zj) = 9jn;ax(Xj ) [fj(X) + L9j,i(Xj )Ui]
11
i=l
In the ideal case, i.e., all nonlinearities of system (5.169) are known exactly, a
possible controller takes the form
with k j > O.
Substituting Uj = uj into (5.180) and noting (5.182), we have
Therefore, Vzj are Lyapunov functions and esj -+ 0 as t -+ 00. In the case that
no exact knowledge for the system nonlinearities is available, such a controller
cannot be implemented due to the unknown functions hj(Zj). Nevertheless,
hj (Zj) are continuous functions, and therefore may be approximated by the
neural networks provided in Chapter 3.
n
where Zj = [ZJ, IV are the input vectors; zj are compact sets to be specified
later; Wj* and Vj* are ideal weights; and lJ.tj(Zj) I :S Pj with constants Pj > O.
Assumption 5.8. For continuous functions hj(Zj), there exist constants Wj*
and Vj* such that function approximation (5.183) hold and lJ.tj(Zj)1 :::; fl.j with
constants fl.j > Ofor all Z E n z.
Le t W' j -- [" ' ] T E RIoJ an d V'j -- ["Vj,l, Vj,2, ... ,Vj,lj
Wj,l, Wj,2, ... , Wj,lj '] E
R(nm.+j+2)xlj denote the estimates of Wj* and Vj*, respectively.
5.4 NN Control for MIMO Nonlinear Systems 133
It follows from Lemma 3.9 that the NN estimation errors may be written
as
~,
Sj,k = S'(AT z) S(Za)
Vj,k j = -a:;-
I _"r z.' k = 1,2"" ,ij
a. Za.- 1J j,k'
j = 1,2"" ,m (5.186)
(5.188)
(5.189)
where r wj = r;:j > 0, r Vj = r~ > 0, and O'wj, O'vj > 0 are constant design
parameters. In the above adaptive algorithms, O'-modification [74] terms are
introduced to improve the robustness of the controller in the presence of the
NN approximation errors.
The following theorem summarises the stability and control performance of
the closed-loop adaptive systems.
134 Chapter 5. ILF for Adaptive Control
where Q z1 = {(X, VI, VI) IX E Qx, Xd2 E f!d2} for notational conve-
nience, and
(ii) mean square tracking performance
~
t Jo
t e]' 1(r)dr<- 2cjajcj
2
1 + gOj
+~(2cjajVsj(0) +bj ), W>O (5.191)
t 1 + gOj
and Loo tracking error bounds
(5.192)
~sj
T 1 [Wj
= V zj + 2 - T r -1 Wj
-
+ tr{Vj rvj Vj} ]
- T -1 - (.5.193)
tuj
Noting (5.184) and (5.186), their time derivatives along (5.180) are
.
Vsj = [
-esj kj (t)eSj + Wj
-T'
(Sj + Sj Vj Zj) + Wj Sj Vj Zj + dUj
','T- 'T',-T-
- /1j (Zj)
]
2 +
-esj[<pj (Zj) - <Pj(Zj)] + Wj rwj Wj + tr { Vj- T rvj-1 Vj
- T -1 ~ ~ }
Substituting (5.188) and (5.189) into the above equations and using the facts
'T"-T- _ -T- ' T ' , + .
that Wj Sj Vj Zj - tr{Vj ZjWj Sj} and <Pj (Zj) - <Pj(Zj) 20, we obtam
5.4 NN Control for MIMO Nonlinear Systems 135
Vsj S - e;J
E]
(1 + ior (%(Xj
1
,8es j + vj)d8 + IIZjWfSjll~ + IISjVIZjI\2)
+ (IIVj*IIFIIZj wf SjIIF + IIWi 1\ 1\ Sj "CjTZj 1\ + IWj*\I + IJLj(Zj)l) le.jl
_ a;j (II WjIl2 -IIWj*\l2) _ a;l (\lij\l~ -IIVj*II~) (5.194)
Since
2
-, T'I
\l Vj*11 F\lZjWj esj - 'Tj Sjl
SjllFlesjl ::; 'E.'""ZjW 'I 12F + 4:"Vj
Ej *IF12 (5.195)
J
2
(IW;h + IJLj(Zj)\)lesj! ::; ;;j. + Ej [IWj*lr + JL7(Zj)]
J
'I'T-
\l W11111 S
2
e sj 'I'T- 2
jVj ZJlllesjl ::; -IISjVj Ej
Zjll + -IIW)'II 2
(5.196)
Ej 4
and IJLj(Zj)1 S fij, inequality (5.194) can be re-written as
J
sj
v.. 5J. <- -eE-
2 1
2+
. [-
11
0
-
(jg-J'(XJ, (je SJ. + v.)de]
J
aWj
- -IIW'II~
2 J
- ')
avj - 2
-TI\YjIlF +Cj (5.197)
with constants
+
a2
;J I\Vj*II} (5.198)
where constants
(5.200)
n .=
tv]
{(w ],
V)
]
IIIW'11 2<
] -
2[Vsj(0)
\.
/\11'z,'z,n
+ Cj//\d
(r- 1)
wj
,
We conclude that for bounded initial conditions, there exist compact sets n zj
defined in (5.190) such that the vectors Zj E n zj for all time, which means
that all the signals X, Wj, ~ and Uj are bounded.
(ii) As J01 OgAXj , Oesj + I/j )dO ~ gOj /2, inequality (5.197) can be further
written as
. 2
v;,j:S -e Sj (1+goj)/2cj+cj
Integrating them over [0, t]leads to
i o
t e2sj ( T ) dT ~ -2c
-j - [ Vsj(O )
1 + gOj
+ tCj ] , j = 1,2,···,711 (5.203)
Since tracking errors ej,l = Hj(s)esj with stable transfer functions Hj(s) (see
Remark 5.12), by applying Lemma 2.10 we obtain
(5.204)
5.4 NN Control for MIMO Nonlinear Systems 137
·
11m
t-+oo
I1t
-
t 0
2 ()d l' < ----"--"-"'-
e . 1 l'
3,
2€jajcj
- 1 + gOj
which indicates that the mean square tracking error performance converges to
the £j -neighborhoods of the origin whose sizes are adjustable by the design
parameters £j, O'wj and O'vj.
Remark 5.14. Noticing (5.191) and (5.192), we can see that the transient
responses of tracking errors are affected by the bounds of VSj(O) significantly.
It is shown from (5.200) that larger adaptation gains r wj and r vj may result in
smaller \'sj(O), hence fast adaptations are helpful for improving the transient
performance. However, in practice, we do not suggest the use of high adaptation
gains because such a choice may result in a variation of high-gain control [147],
and therefore increases the bandwidth of the adaptive system. Any small noise
in the measurements might be amplified and cause large oscillations in the
control outputs.
Remark 5.15. It should be mentioned that the integrals in control gain (5.187)
might not be solved analytically for some functions Yj(X j ) and may make the
controller implementation difficult. This problem can be dealt with by suit-
ably choosing the design functions Yj(Xj). Since the choices of Yj(Xj ) are
only required to be larger than Igj,j(Xj)l, the designer has the freedom to find
suitable Yj(Xj ) such that the integrals are analytically solvable. As an alter-
native scheme, one can also use on-line numerical approximation to calculate
the integral, which however requires more computational power in practical
applications.
Remark 5.16. It is worth noting that Theorem 5.12 only shows boundedness
of the system states and the responses of output tracking errors; convergence
of the NN weight estimates is not guaranteed due to the lack of persistent
excitation (PE) condition. In adaptive control systems, PE condition is impor-
tant for parameter convergence and system robustness, however it is usually
a strong requirement in practical applications [74]. The works [40] [38] have
studied PE conditions of linearly parametrized networks and provided several
practically applicable conditions for system identification and adaptive NN con-
trol. For multilayer NNs, no theoretical result is available in the literature to
138 Chapter 5. ILF for Adaptive Control
Remark 5.17. Position definite functions gj;'ax(Xj ) = ?iJ (Xj)/gj,j (Xj) have
been used in the construction of Lyapunov function candidates Vzj (esj, X j , Vj)
as shown in (5.175). In fact, gja1(Xj) = a(Xj)/gj,j(Xj ) with any weighting
function a(Xj ) can also be used without any difficulty. For example, when
a(x) = 1, we have gjl(Xj ) = l/gj,j(Xj ).
5.5 Conclusion
The main feature of this chapter is the construction of integral Lyapunov func-
tion for different control problems. Firstly, the integral Lyapunov function is
introduced and three desired feedback controllers (DFC) are introduced for a
class of SISO nonlinear systems based on different choice of weighting functions
a(x). Then, by using NN or LIP parameterization, three kinds of stable adap-
tive control schemes were developed for different DFCs. The proposed design
completely solved the possible singularity problem in feedback linearization de-
sign. Secondly, using backstepping technique and MNNs, the proposed method
was extended to the adaptive control of a class of strict-feedback nonlinear
systems with completely unknown system functions. The developed control
scheme guarantees the uniform ultimate boundedness of the closed-loop adap-
tive systems. Finally, a stable adaptive control scheme for multivariable nonlin-
ear systems with a triangular control structure was presented. It is shown that
(i) by utilising the system triangular property, integral type Lyapunov functions
are introduced for deriving the control structure and adaptive laws without the
need of estimating the "decoupling matrix" of multivariable nonlinear system,
and (ii) without imposing any constraints on the system interconnections, the
developed controller guarantees the stability of the closed-loop system and
the mean square tracking errors converging to small residual sets which are
adjustable by tuning the design parameters. Transient performance of the re-
sulting closed-loop systems is analytically quantified by the mean square and
Loo tracking error bound criteria.
Chapter 6
Non-affine Nonlinear
Systems
6.1 Introduction
In the last decade, there has been a large amount of activity in adaptive control
of nonlinear systems using feedback linearization techniques ( [104,128,165] and
the references therein). The common assumptions in the literature are that the
plant under study is affine, i.e., model is linear in the input variables and the
nonlinearities are linearly parameterized by unknown parameters. However,
many practical systems, e.g., chemical reactions and PH neutralization, are
inherently nonlinear, whose input variables may not be expressed in an affine
form. The purpose of this chapter is to study adaptive NN control problem for
a general SISO nonlinear system
X = f(x,u)
{ (6.1)
y = h(x)
form as follows
x= f(x, x a )
Xa = V
y = h(x)
and many control methods for affine nonlinear system can then be used. How-
ever, due to the introduction of an additional integrator, the dimension of the
system is increased and control efforts are not direct and immediate either,
subsequently may not control the system effectively. In this chapter, we are
interested in control design for system (6.1) directly, which is not only acad-
emically challenging but also of practical interest.
For controller design of general nonlinear systems, several researchers have
suggested to use neural networks as emulators of inverse systems [85,109, 155J.
The main idea is that for a system with finite relative degree, the mapping
between a system input and the system output is one-to-one, thus allowing
the construction of a "left-inverse" of the nonlinear system using NN. Hence,
if the controller is an "inverse operator" of the nonlinear system, the reference
input to the controller will produce a control input to the plant, which will
in turn produce an output identical to the reference input. Based on implicit
function theory, the NN control methods proposed in [59,60,85,86] have been
used to emulate the "inverse controller" to achieve tracking control objectives.
Nevertheless, no rigorous stability proofs of the closed-loop systems were given
for on-line adaptive NN control due to the difficulties in analysis of discrete-time
nonlinear adaptive systems.
In this chapter, adaptive control techniques are investigated for non-affine
nonlinear systems expressed in continuous-time framework. Both state feed-
back and output feedback control are studied by combining NN parametriza-
tion and implicit function theory. The stability of the closed-loop system is
proved using Lyapunov stability theory. The output tracking errors converge
to adjustable neighborhoods of the origin.
Higher-order Lie derivatives are defined recursively as L7h = Lf(L~-l h), k > l.
6.2 System Description and Properties 141
a[L}h] a[L/h]
a;;-=O, i=0,1, ... ,p-1, -au
-#0 (6.2)
{ ~n~i == ~i+I'
b(~, u)
i=1, ... ,n-1
(6.3)
y = 6
with b(~, u) = Ljh(x) and x = <I>-I(~). Define the operation domain of normal
system (6.3) as
bu = a[b(~,u)] (6.4)
au
According to Assumption 6.1, it can be shown that
which implies that the smooth function bu is strictly either positive or negative
for all (~, u) E [T. From now onward, without losing generality, we shall assume
bu ~ bo > O.
Assumption 6.2. There exists a smooth function bl(x) such that 0 < bo ::;
bu ::; bl (x) for all (x, u) E U.
Remark 6.1. From (6.4), we know that bu can be viewed as the control gain of
the normal system (6.3). Assumption 6.2 means that the plant gain is bounded
by a positive function of x, which does not pose a strong restriction upon the
142 Chapter 6. Non-affine Nonlinear Systems
(6.5)
(6.6)
where A = [>'1, '\2, ... ,,\n_l]T is an appropriately chosen coefficient vector such
that sn-l+'\n_lSn-2+ . . +'\1 is Hurwitz. Let the vector ( = [~l' ~2' ... ,~n-ljT,
then a state representation for the mapping (6.6) is
(6.7)
(6.8)
U
*( Z ) = ex C(C)
1", V, V
b1(x)
=- - es - Yd(n) + [0 (6.9)
E
Z = [~T, vJT E Oz C R n +1
such that for all (~(O)) E cl>o, the error equation (6.8) can be expressed in the
form
(6.10)
6.2 System Description and Properties 143
(6.11)
Using Implicit Function Theorem (Lemma 2.8), we know that there exists a
continuous trajectory dependent function QC(~, v), which guarantees
Under the action of u*(z), (6.11) and (6.13) imply that (6.10) holds. As
bl~X) > 0, (6.10) is asymptotically stable, i.e., limHoo lesl = O. Because
sn-1 + /\n_1 Sn - 2 + ... + /\1 is Hurwitz, we have limHoo Iy(t) - Yd(t)1 = 0
asymptotically.
It should be noticed that the above result is obtained under the condition
e
of E <I>(nx ), Vt ~ O. We will specify the initial states such that this condition
rt~d
is satisfied. It follows from (6.10) that es(t) = es(O)e- Jo • 7'. Because
b1ix) > 0, we have the bound les(t)1 :::; les(O)I. Define the following compact
set
Then, for all ~(O) E <1>0, we have' E <I>(n x ), Vt ~ o. This completes the proof.
<>
Different choices for bI (x) will result in different desired error dynamics. By
choosing bI(x) = 1, we have the following lemma.
144 Chapter 6. Non-affine Nonlinear Systems
Lemma 6.2. Consider system (6.3) sat·isfying Ass·umptions 6.1 - 6.3. For a
given small constant c > 0, there eX'ist a compact subset <Po C <p(Ox) and an
unique ideal implicit feedback linearization control (IFLC) inpttt
such that for all ~(O) E <Po, the error equation (6.8) can be written in a linear
form
(6.17)
- [.T
Z - x ,es,VI ]T (6.18)
Remark 6.2. In Lemma 6.1, we suppose that Ou is large enough such that
u" E OU' If the region of stu is not large enough, some restrictions should be
addressed upon the desired trajectory (d and design parameter c.
Remark 6.3. It is shown from Lemma 6.2 that the magnitude c of the desired
signal (d affects the size of the allowed initial state region <I>o. This is reasonable
because if the reference ~d is very large and out of the region <I>(Ox), then <I>o
will become an empty set and the tracking problem cannot be solved.
6.2 System Description and Properties 145
Lemmas 6.1 and 6.2 only suggest the existence ofIDFC u", it does not pro-
vide the method of constructing it. Later, neural networks shall be introduced
to construct u" for achieving tracking control.
Note that the choice for v and subsequently the error dynamics are not
necessarily in the above form only. Other choices are also possible. Define
with 'Y being any small positive constant. As'Y - 0, sat(es ) approaches a step
transition from -1 at e s = 0- to 1 at es = 0+ continuously.
Lemma 6.3. Consider system (6.3) satisfying Assumptions 6.1 - 6.3, ~(o) E
O~ C R n and ~d E Od C R n , where O~ and Od are two compact sets. There
exists an ideal control input
(6.25)
where kv and kvs are positive constants. Subsequently, (6.25) leads to limt--+oo
Iy(t) - Yd{t)1 = o.
Proof Plus and minus kves + kvs sat(e s ) to the right-hand side of the error
equation (6.8), we obtain
(6.26)
From Assumption 6.1, we know that ob(~, u)/ou ~ bo > 0 for all (~, u) E Rn+l.
Considering the fact that ov / ou = 0, we obtain
o[b(~, u) + v] b 0
au ~ 0 >
Using Implicit Function Theorem (Lemma 2.8), for every ~(O) E O~, ~d E Od
and every value of ~, there exists a continuous ideal control input u"(z) with
z = [er, vjT E R n +1 such that
b(~, u") +v = 0 (6.27)
Under the action of u", (6.26) and (6.27) imply that (6.25) holds. Define a
2
Lyapunov function candidate V = ~. Differentiating V along system {6.25}
146 Chapter 6. Non-affine Nonlinear Systems
(6.28)
Lemma 6.4. Suppose the system output y(t) and its first n derivatives are
bounded, so that Iy(k) I < Yk with posit'ive constants Yk. Consider the following
linear system
(6.29)
where E is any small positive constant and the parameters :\1 to :\n-1 are chosen
such that the polynomial sn + :\lSn-1 + .. jn-1s + 1 is Hurwitz. Then,
(i)
(6.30)
(ii) there exist positive constants t* and hk only depending on Yk, E and ,\;, i =
1,2, ... ,n - 1 s1tch that for all t > t* we have 11jJ(k) I ::; hk, k = 2,3, ... , n.
6.3 Controller Design Based on LPNN 147
Proof The proof can be found in [10]. For completeness, it is given here. From
the last equation in (6.29), we have
(6.31)
(6.32)
in (6.32) is bounded by o}j for each j. Further, since ly(j)1 < Yj, there exist
constants D j , which is independent of E, such that, for each j, the second term
in (6.32)
nonlinear systems based on the u*(z) in (6.24). Since u*(z) in (6.24) is a con-
tinuous function, on a compact set fl. C R n + 1 , there exist ideal weights W*
so that the ideal input u*(z) can be approximated by an ideal RBF neural
network u~n(z) = W*TS(z). Thus, we have
u*(z) = u~n(z) + C:u(z) (6.33)
where c: u (z) is called the NN approximation error.
Lemma 6.S. Suppose system (6.3) satisfies Assumptions 6.1 - 6.3 and As-
sumption 3.1 with (E, u) in a compact set f!. Then the following inequality
holds
(6.34)
where C 1 and C 2 are positive constants and c: = SUP.Efl. {1C:u(z)I}.
Proof Since b(E, u) is a smooth function with respect to u (Assumption 6.2),
let u = u*, according to (6.33), we have u - u~n = C:u(z) and the Taylor series
expansion of b(E, u) at u~n = W*T S(z) as
k = 1,2, ...
L
00
O u'" =:::
Ab[il
U u"
(U -
*
U nn
)i-1
nn nn
i=2
As ~btkJ are continuous functions, from Lemma 2.1, we know that ~bt1J and
OU;',n a;'; Lipschitz on the compact set f!. Since II W* II is bounded and ~~n (z)
is a function of z, i.e., the function of E, v and Ed (noting z = [E T , v]T from
(6.24», by letting c: = sUPzEfl. {1C:u(z)I}, there exist three constants a1, a2 and
a3 such that
Ib(E,u*) - b(E,u~n)1 = I~brlnn C:u(z) + Ou' C:u(z)l::; a1c:IIEIl + a2c:llvll + a3C:
nn
From equations (6.5), (6.6) and Assumption 6.3, we can derive IIEII ::; d 1 +d2 iesl
and Ilvll ::; d3 + d4 1e s l with positive constants d1 to d4 over the compact set f!.
Hence there exist positive constants C 1 and C2 such that
(6.35)
with C 1 = a1(d 2 + d4 ) and C 2 = a1d1 + a2d3 + a3. From (6.27) and (6.35), we
have
6.3 Controller Design Based on LPNN 149
with W being the estimates of the NN weight W* and S(z) being the known
basis function vector. In order to establish the error system, taking the Taylor
series expansion of b(~, u~n) at it.nn(z) = WT S(z), we have
(6.37)
where
O·U nn = ~
L-, .1.b[!)
U nn
[WT S(z)]i-l
i=2
From (6.26) and (6.37), we obtain the error system for u = unn
es = b(x, Unn ) + v - kves - kvs sat(e s )
= b(~, u~n) + v + .1.b~ln WTS(z) - OUnn WTS(z)
-kves - kvs sat(es ) (6.38)
LollWl1 + Lllesl + L2
IOunn! S (6.39)
l.1.b~lJ S L311WII + L41esl + L5 (6.40)
To guarantee the closed-loop stability and the tracking error y(t) - Yd(t) to be
suitably small, consider the weight update law
where lio, lil, li2 and /j are positive constants. The first term on the right-
hand side of (6.41) is a modified backpropagation algorithm and the second
term corresponds to the e-modification [134] usually used in robust adaptive
control, which is applied for improving the robustness of the controller in the
presence of the NN approximation error.
Theorem 6.1. For system (6.3), if the controller is g'iven by (6.36) and the
ne1tral network weights are updated by (6.41), with
(1) Assumptions 6.1 - 6.3 and Assumption 3.1 being satisfied, and
(2) the existence of two compact sets Dw and Des such that W(O) E Dw and
es(O) E Des'
then, for a s·uitably chosen design parameter kv, the filtered tracking error e s ,
neural network weight Wand all system states are SGUUB. In addition, the
tracking error can be made arbitrarily small by increasing the controller gains
and neural network node number.
Proof Consider the Lyapunov function candidate as
(6.42)
Using IIW*I/ ::; Wm and IIWII ::; IIWI/ + W m , (6.43) can be written as
as = sup {IIS(z)ll}
zEn.
(6.48)
152 Chapter 6. Non-affine Nonlinear Systems
(c:C2)2
f32le s l + kvsesIes + -v:;;;-
2
V < -(31 e s -
where
Define
where c:o and (30 are positive constants. Since 01 to 05, 0, as, and C1 are
positive constants, we know that ko is a positive constant. If choosing kv > ko,
we can guarantee that (31 > (30 and (32 > (30.
Since, for e s E R, esexp(-e s /,/,) has a maximum value of ,/,/eo at es = ,,/,
with eo being the natural exponential (eo = 2.7183). From (6.22), we have
kvsesIes ::; kvs'/' / eo· Define set
Since C 2 , "/ and eo are positive constants, kv > ko, (31 > (30 and f'h > (30, we
conclude that 8 es is a compact set. V is negative as long as es(t) is outside
the compact set 8 es ' According to a standard Lyapunov theorem [135], we
conclude that the filtered error es(t) is bounded and will converg~ to 8 es '
Since e s is bounded, from Lemma 3.4, (6.41) guarantee that W is bounded
in a compact set, i.e., WE 8 w in (3.25). Now define
If we initialize W(O) inside Dtv and es(O) inside De., there exists a constant
T such that all trajectories will converge to 8 and remain in 8 for all t > T.
This implies that the closed-loop system is SGUUB. The filtered tracking error
will converge to the small compact set 8 es which is a (c:,,)-neighborhood of
the origin. Since sn-1 + '\n_1Sn-2 + ... + /\1 is Hurwitz, y(t) - Yd(t) -1 8 es
as es(t) -1 8 es ' Because "/ can be chosen as any small positive constant, and
c: can be as small as desired by increasing the number of neural nodes l, we
conclude that arbitrarily small tracking error can be achieved. <)
6.3 Controller Design Based on LPNN 153
(6.52)
(6.53)
where E, "'0, "'1, "'2 and 6 are positive constants. The closed-loop error equation
(6.26) becomes
(6.56)
154 Chapter 6. Non-affine Nonlinear Systems
To make the proof of the main theorem easier to follow, the following lemma
is firstly provided.
Lemma 6.6. For the ideal neural network control u~n = W· T S(z), the non-
linear function b(c;, u~n) can be expressed as
[.] , T
where OU~n = I:~2 Ab~~n [u~n - W S(2)](i-l). In addition, if z is in a com-
pact set fl. C Rn+l, there exist positive constants Lo to L5 that
(6.58)
(6.59)
Let 11. = 11.~n = W*T S(z), we know that equation (6.57) holds. Since Abh~ ,
(k = 2,3, ... ,), are continuous functions, we know from Lemma 2.1 that ab~f~)
and OU:: n are Lipschitz on the compact set Dz . Since u~n is a function of W,
c;and ~d, there exist positive constants a4 to ag such that
Proof The proof is similar to that of [10]. We first assume that the system
signals remain in a compact set, then, we show that, by properly choosing
controller parameters, the system trajectories in fact remain in the compact
set. Choose the Lyapunov function candidate as
(6.60)
Define
where 0!1 to 0!2 are positive constants, 0!3 to 0!7 are positive definite and
bounded functions, and 0!4a, 0!4b, 0!4c, 0!5a, 0!5b, 0!5c, 0!6a, and 0!7a are properly
6.3 Controller Design Based on LPNN 157
[( ~
8
-2IesIl18(z)1I IIWII- O!~
48 ) 2 - 1682
O!t]
~
-E [(IIWlllesl- 0!7a 2- ( 2
2) 0!7 a ) 2] (6.70)
where
(6.71)
(6.72)
Define
where eo and {30 are positive constants. Since £lI to £l3, £l5e and £l6a are
bounded, £ls, {j and C I are positive constants, we know that ko is a positive
constant. Therefore, if choosing kv > ko, {6.71} and (6.72) show that {31 > {30
and {32 > 130. Define
Since £l8, £lg, "I, f and eo are positive constants, kv > ko, {31 > 130 and f32 > 130,
the set 8 e is a compact set. V is negative as long as es(t} is outside the compact
set 8 e . Hence, es and es are bounded.
The boundedness of W can be derived by following the same procedure as
in the proof of Lemma 3.4 by replacing e s with es . In fact, it can be shown
that W is bounded in
8 W.-
·- { W(t) : IIWII S sup
A A
eE9.
[1(2 + lesl- T"'0)
1
If we initialize W(O) inside D w , e(O) inside De, ~(O) inside Dt;, and choose a
large enough kv guaranteeing (31 > (30 and (32 > (30, then there exists a constant
T such that all trajectories will converge to 6 and remain in 6 for all t > T.
This implies that the closed-loop system is SGUUB. The filtered tracking error
will converge to the small compact set 6 e which is a (c, E, I)-neighborhood of
the origin.
Since sn-1 + /\n_1Sn-2+, ... ,+/\1 is Hurwitz, y(t) - Yd(t) ~ 6 e as es(t) ~
6 e . In addition, because c and I can be made arbitrarily small by increasing
the number of neural nodes l and the state observer gain 1/ E can be designed
arbitrarily large, we conclude that arbitrarily small tracking error is achievable.
<>
Remark 6.6. The high-gain observer (6.29) may exhibit a peaking phenom-
enon in the transient behaviour. The input saturation method introduced
in [37,82] may be used to overcome such a problem. Thus during the short
transient period when the state estimates exhibit peaking, the controller satu-
rates to prevent peaking from being transmitted to the plant.
Remark 6.7. The adaptive output feedback NN controller proposed here is
easy to implement because it is simply a state feedback design with a linear
high-gain observer without a priori knowledge of the nonlinear systems. Un-
like exact linearization approach [77]- [169], it is not necessary to search for a
nonlinear transformation and an explicit control function.
6=6
~2 = Er + 0.15u 3 + 0.1(1 + E~)u + sin(O.lu) (6.74)
Y= 6
Since the nonlinearities in the plant is an implicit function with respect to u,
it is impossible to get an explicit controller from system feedback linearization.
In this example, we suppose that there is no a priori knowledge of the system
nonlinearities except for some of its properties. As 8b(~, u)/8u. = 0.45u 2 +
0.1 +0.1~~ + 0.1 cos(O.lu) > 0 for all (~, u) E R n +l, Assumption 6.2 is satisfied.
The tracking objective is to make the output y(t) follow a desired reference
Yd(t) = sin(t) + cos(0.5t). The initial conditions are E(O) = [0.6,0.5jT.
The neural network controller u(t) = WTS(z) has been chosen with l = 8,
J1i = 0.0 and al = 0.1 for i = 1,2"" ,l. Other controller parameters are
chosen as A = 10.0, kv = 2.0, I = 0.03. The initial conditions for neural
networks are W(O) = 0.0.
160 Chapter 6. Non-affine Nonlinear Systems
E6 = 6
E6 = 6
E~3 = -b 1 6 - b2 6 - 6 + yet)
with the parameters E = 0.01, bl = 1.0, b2 = 3.0 and the initial condition
';(0) = [O.O,O.O,O.O]T. The estimate of vector z is z = [xl,6!E,iY. We use
the proposed output feedback adaptive NN controller to control the system.
In order to avoid the peaking phenomenon, the saturation of the control input
u(t) is ±4.0.
Figures 6.2 (a)-(d) illustrate the simulation results of the adaptive output
feedback controller. It can be seen that, after a short period of peaking shown
in Figure 6.2 (b), the tracking error and the state estimate error becomes small
and the saturation mechanism in Figure 6.2 (c) becomes idle. The plots indicate
the satisfactory tracking performance with bounded closed-loop system signals.
Assumption 6.4. There exist positive constants bo, bl and b2 such that bo ::;
bu :::; bI and Ibul :::; b2 for all (.;, u.) E 0.
6.4 Controller Design Based on MNN 161
10
-2
-I
-1,5
-2 ~L_--~--~~----,~~---toI~
o 10 15 20
Tlrnesec TwtlflSEIC
HI
,,,,,,,,,,
15 20 25 30
'.5r--~-~-~----~-----,
ro
15
10
O~'--------------------------~
10 15
T!mesoc
20 25 -50~---:----;:10---'C::-5-----:'O:-----::25~-----:~
Tmesec
'O,----~-~--,-----~-~--_____,
I.
16
14
12
-I
-,
~~-~-~'O--~15-~ro~-~25~~~
Time sec
Remark 6.S. From the definition of bu., we know that bu. can be viewed as the
control gain of normal system (6.3). The requirement for bu. ~ bo means that
the control gain of the system is larger than a positive constant. Many feedback
linearization methods for affine nonlinear systems need such an assumption
!10, 77, 175]. We also require the absolute values of bu. and bu. being bounded
by positive constants b1 and b2 , respectively. In general, this does not pose a
strong restriction upon the class of systems. The reason is that if the controller
is continuous, the situation in which a finite input causes an infinitely large
effect upon the system rarely happen in physical systems due to the smoothness
of bu..
with Vi = !Vil, Vi2,' .. ,Vi2n+2]T, i = 1,2, ... ,l. Because the ideal IFLC input
u*(z) defined in (6.20) is a continuous function on the compact set Oz, for an
arbitrary constant eN > 0, there exists an integer l (the number of hidden
neurons) and ideal constant weight matrices W* E R! x 1 and V* E R(2n+2) x I,
such that
(6.76)
where eu.(z) is called the NN approximation error satisfying leu. (z)1 :s; eN, 'r/z E
Oz, and the term Uk(Z) in (6.76) is a prior continuous controller (possibly PI,
PID or some other type of controller), which is perhaps specified via heuristics
or past experience with the application of conventional direct control. The
ideal constant weights W* and V* are defined as
SUbstituting (6.76) and (6.78) into (6.81) and noting (6.77), we obtain
u>. s = -~b-le
b-1e c: u>. s + WT(S - S'VTz) + WTS'VTz
-kslesle s - C:u(z) + du (6.82)
(i) all initial states ~(O) E <1>0, (W(O), V(O)) E no, and
(6.85)
.
VI = es [1 I -T~
- -gb;;;J. es + W (8 -
~/~T
8 V z) + W~T~/-T
8 V z + du - kslesles - cu(z) ]
1 d(b- ) I
+ _-..-!:L
2 dt e 2 s
+ wTr-Iw
-
w
:.
+ t r {V- T r -v 1V}
:.
b2/2b6 and b;;'; 2: b11 (Assumption 6.4), inequality (6.86) can be further writ-
ten as
131 = tl + a1 (6.89)
(6.91)
(6.92)
Since D, t and ks are positive constants, we know that ee, e w and e are
compact sets. Thus, VI < 0 as long as VI is outside the compact set e.
According to a standard Lyapunov theorem [135], we conclude that e s , Wand
V are bounded. It follows from (6.91) and (6.92) that VI is strictly negative as
long as e s is outside the compact set ee. Therefore, there exists a constant To
such that for t > To, e s converges to ee.
From the state representation (6.7) of es , we can conclude that there exist
constants ko > 0 and '\0 > 0 such that lIeAtll ::; koe- Aot . The solution for (can
6.4 Controller Design Based on MNN 167
nD l : = {(e s , Mr, V) I V1 :S D 1 }
E: = inf{D1 I nDl J e}
Clearly, all trajectories starting in nE will remain in nE for all time. We can
also see that E is a function of c, 10, ks and 101, and
(i) can be made arbitrarily small by increasing the number of the MNN
101
nodes l, and
Remark 6.10. It can be seen from (6.92) that the larger the number of NN
nodes is, the smaller the tracking error will be. Therefore, if a high tracking
accuracy is required, small €, large ks and l should be chosen. However, the
multilayer neural networks with a larger hidden node number l usually make the
controller more complex and increase the computation burden for the control
system. Hence, there is a design trade-off between MNN complexity and system
tracking error accuracy.
Remark 6.11. Theorem 6.3 shows that the boundedness of the system states
and the convergence of tracking error are guaranteed without the requirement
of PE condition, which is very important for practical applications because PE
condition is very difficult to check/guarantee in adaptive systems. The results
indicate that although convergence of the NN weight estimates is not achieved
in the absence of PE condition, the tracking error is guaranteed to converge to
an €-neighborhood of zero.
Assumption 6.5. There exist a constant bo > 0 and a smooth function bl~X) >
o such that bu 2: bo and Ibu/bul :::;
bl~X), Vex, u) E Ox x R with constant € > O.
When only the output y = 6 is measurable and the rest of the output
derivatives are not available, the high-gain observer (6.29) is used to estimate
6, 6, ... , ~n to implement the output feedback control.
Similarly, define the following variables
where
Lemma 6.4 shows that \1fJ(k) \ are bounded by the constants hk' hence \}I, eso
and Zo are all bounded.
We shall use the MNN WTS(VTi) to approximate unknown u*(z) or the
difference between u*(z) and any a prior controluk(z). Because the ideal IFLC
input u*(z) defined in (6.20) is a continuous function on the compact set Oz, for
an arbitrary constant £ON > 0, there exists an integer l (the number of hidden
neurons) and ideal constant weight matrices W" E Rl x 1 and V" E R(2n+2) x I ,
such that
(6.97)
where eu(z) is called the NN approximation error satisfying \eu(z)\ ::; ION, "i/z E
Oz, and the term Uk(Z) in (6.97) is a prior continuous controller (possibly PI,
PID or some other type of controller), which is perhaps specified via heuristics
or past experience with the application of conventional direct control. The
ideal constant weights W" and V* are defined as
where Ow = {(W, V) II\WI\ ::; W rn , I\VI\F ::; vm} with constants W m, Vrn >0
specified by the designer. From Lemma 3.6, we have
U = U nn + Uk + Ub (6.101)
with
U nn = WTS(VTz) (6.102)
where 10, ks > 0 are design parameters. The above controller contains three
parts for different purposes. The first part U nn is introduced to approximate
the ideal NN control W . .TS(V",T z) in (6.97) for achieving adaptive tracking
control. The second part Uk is a priori control term based on a nominal model
or past control experience to improve the control performance. If no knowledge
170 Chapter 6. Non-affine Nonlinear Systems
for the plants is available, Uk can be simply set to zero. The third part Ub is a
bounding control term, which is applied for limiting the upper bounds of the
system states such that the NN approximation (6.97) holds on the compact set
fl •.
The MNN weight updating laws are taken as
V= -rv[iWTS~es+c5v(1+lesl}V] (6.105)
where r w = r~ > 0, r v = r; > O,<5w > 0 and <5v > 0 are constant design
parameters. From Lemma 3.10, we know that the above learning algorithms
guarantee that Wet), Vet) E Leo for bounded initial weights W(O) and V(O)
and bounded s . e
Using the Mean Value Theorem (Theorem 2.5), there exists a ,\ (0 < ,\ < 1)
such that
where bux = 8(b(~, u)1/8u!u=ux with u).. = ,\u + (1 - >.}u*. Considering (6.13)
and (6.106), the error system (6.11) can be written as
(6.107)
Theorem 6.4. For system (6.1) with Assumptions 6.1, 6.4 and 6.5 being sat-
isfied, high-gain observer (6.29), controller (6.101) and adaptive laws (6.104)
and (6.105), there exist a compact set <Po C <Pe, and positive constants c*, k;,
and 1;* such that for any bottnded W(O) and V(O), ·if
(i) the initial state ~(O) E <Po,
(ii) the observer (6.29) is turned on time t'" in advance, and
then all the signals in the closed-loop system are bounded, the system state
~ E <Pe, "It 2: 0, and the tracking error converges to an (E, I;)-neighborhood of
the origin.
6.4 Controller Design Based on MNN 171
Proof The proof contains two steps. We first assume that ~ E <I>~ holds for all
time, which ensures that neural network approximation (6.97), Assumptions
6.1 and 6.5 are valid. In this case, we prove the tracking error converging to an
(c, E)-neighborhood of the origin. Later, for a proper choice of the reference
signal Yd(t) and controller parameters, we show that ~ do remain in the compact
set <I>~ for all time.
Step 1 Consider the Lyapunov function candidate
(6.109)
1 d(b;;;) - T - T
W r w W + tr{V rv V}
2 -1':' -·1':'
+Ubes + [d u -lLiles + 2~es +
V.1 = - [b-c-bu>.
1 (x) -1 bu >.] 2 [
+ 2b 2 es + Ubes + du -lLdes
u>.
-(1 + le s l)(8w WTW + 8vtr{VTV})
+Eeso[WTS~VTz+ WT(So - S~VTi)] (6.110)
Since bu > 0 and Ibulbul ::; b1lx) (Assumption 6.5), we know thatlbu>.1/2b~>. ::;
b1ix)b;;>.I. Using 2WTW ~ IIWl1 2 -IIW"11 2 and 2tr{VTV} ~ llii"ll} -IIV"II},
we have
Since
2 2
VmllzWTS~IIFlesl ~ ;;IIzWTS~II} + c~m
2 2
wmiiSolllesl ~ ~~ II Soli 2 + c~m
2 2
WmllS~VTzillesl ~ ;~IIS~VTzI12 + c~m
2
IlLtilesI ~ :: + CIL{
Ie' sI("2Wm
Ow 2 + 2Vm
Ov 2) ~ 4c e~ Ow 2 Ou 2 2 I (OW 2 Ov 2)
+ c( "2Wm + 2Vm) + Eesol "2Wm + 2Vm
VI ~ -~(1 +Eesoes)(IIzWTS~II}+IISoI12+IIS~VTzll2+1)
-kslesl(lesl - Eleso!) (Iunnl + 1) + ~(Ow + 2c + Eowlesoi)w;,
1
+2"(Ou +c+EOvIesoDvm 2 c
+clLl2 + 4 (Owwm2 +OvVm2 )2
+E2e;oC~)zWTS~II} + LI Soll2 LIIS~VTzI12)
+ (6.112)
6.4 Controller Design Based on MNN 173
2
VI :::; - :; (11~WT8~1I} + 118011 2 + 118~VT~112 + 1)
-kslesl(lesl- Elesol) (Iunnl + 1)
1 2 1 ) 2
+2'(Ow + 210 + Eowlesol)wm + 2'(Ov + 10 + Eovlesol Vm
+EILI2 + 4
10 (0 2
wWm + 0vVm2)2
2 2 [1 1 " T' 2 1 1 '2 "T' 2 1]
+E eso (£' + 20)lliW S~IIF + (£' + 0)(118011 + IIS~V ill) + £'
2 2 2
= -(:; - (30) - (:; - (31)llzWT8~11} - (:; - (32)(118011 2 + 118~VT~112)
-kslesl(lesl- Elesol) (Iunnl + 1) (6.113)
with
Now let (3s(E, E) = max {2~, 2v'Ef31, 2v'c:f32, Elesol} and define
(6.116)
Since ow, ov, Vm , Wm,E and 10 are positive constants and eso and ILl are bounded,
it follows that f3s(£, E) is also bounded, which means that 8 e is a compact set.
From (6.113)-(6.116), it is shown that VI is strictly negative as long as e s is
outside the set 8 e . Therefore the filtered error es is bounded, and there exists
a constant To 2=: t* such that for t 2=: To, the filtered tracking error es converges
to 8 e which is an (£, E)-neighborhood of the origin.
From the state representation (6.7) of es , it is easy to find two constants
ko > 0 and AO > 0 such that IleAstl1 :::; koe->'ot [74]. The solution for ( can be
written as ((t) = eAst((O) + J~ eAs(t-T)bsesdT. It follows that
(6.117)
174 Chapter 6. Non-affine Nonlinear Systems
. bl (x) 2 ..
Vb = - - - e s + bu>.(u - U )es (6.119)
c
Using (6.97), (6.101)-(6.103), we have
Since every element of S(V .. T z) is not larger than one, we know that
(6.121)
Therefore,
Vi < - bl(x)
b - c
2 _
es
b k (I
u" s U nn es 10
[9
1+1)1 I les l _ Iunnl + wmv'l + l~tlJ
ks(lunnl + 1)
(6.122)
6.4 Controller Design Based on MNN 175
eIlo := {~(O) I {~ 11I~(t)1I < kall((O)II} C eIle , les(O)1 < Ro(E, ks)} (6.125)
k; := inf {ks
k.ER+
I{~ III~II < kall((O)1I + kbRo(O, k s) + c,
E*:= SUP{E
fER+
I {~III~II ~kall((O)II+kbRo(E,ks)+c,
~(O) E eIlo, c ~ c*, ks 2: k; } C eIle} (6.127)
In summary, for all initial states ~(O) E eIlo, the desired signal IItdll ~ c ~ co.,
if control parameters ks and E are chosen such that ks 2: k; and E ~ E*,
e
then the system states will stay in eIle for all time. The boundedness of e
€
guarantees that the observer states are bounded (see Lemma 6.4). Since
the NN weights have been proven bounded for any bounded W(O) and YeO)
(see Lemma 3.10), we conclude that all signals of the closed-loop system are
bounded. This completes the proof. <>
Remark 6.12. It should be mentioned that for the MNNs with a chosen NN
node number l, constraint sets (W, V) E Ow and z E Oz, if the estimation of the
upper bound for ILL can be obtained a prior (it is always possible by choosing
a large value, although this choice might be conservative), the initial condition
eIlo and the constants c*, k; and E* can be explicitly calculated using (6.125)-
(6.127). As will be shown in the simulation example, it is possible to obtain
the conditions for achieving a stable closed-loop system when implementing the
proposed controller.
176 Chapter 6. Non-affine Nonlinear Systems
where variables C a and Ta are the concentration and temperature of the tank,
respectively; the coolant flow rate qe is the control input; and the parameters
of the system are given in Table 6.1.
6.4 Controller Design Based on MNN 177
Cae Tr
Ca Ta
v
(6.130)
Obviously, the model is not in affine form, because the control input, U, appears
nonlinearly in (6.130).
From the parameters given in Table 6.1 and the irreversible exothermic
property of the chemical process, we obtain the operating region of the states
and control input as follows
(6.131)
where constant hI is the highest temperature of the reactor and constant h2 is
the maximum value of the coolant flow rate. Since
Y= Lfh = 1- Xl
10 4
oy = 0
- aOxle-"X2,
ou '
and
178 Chapter 6. Non-affine Nonlinear Systems
a')=~
~ Peepe
= 6.987 X 102
we have
1 - e-~ - a2 e-~ = 1 - (1 + a2)e-(1+~)e > 0
U U
by noticing the fact that (1 + w)e-(1+ w ) < e- l for all w > O. Consider the
parameters shown in Table 6.1 and the operating region (6.131), we know
that there indeed exist positive constants bo and bl such that bo :S bu :S b1 .
Therefore the plant is of relative degree 2 and Assumptions 6.1, 6.4 and 6.5
are satisfied under operating condition (6.131). The above analysis is just to
check the validity of the assumptions made. Since the exact model of CSTR is
difficult to be known in practice, in the following simulation study, we design
an adaptive NN controller without the prior model of the CSTR. The control
objective is to make the concentration, y, track the set-point step change signal
r(t).
To show the effectiveness of the proposed method, two controllers are stud-
ied for comparison. A fixed-gain proportional plus integral (PI) control law is
first used as follows
Upi = ke(r - y) k
+~
Ti
lot (r - y)dT
0
(6.132)
with kc = 440L 2 /mol min and Ti = 0.25 min [116J. The parameters of the
PI controller are selected to give an adequate response for step changes r(t) in
6.4 Controller Design Based on MNN 179
the set-points of ±0.02 mol/l about the nominal product concentration of 0.1
mol/I.
In order to obtain a smooth reference signal, we use a linear reference model
to shape the discontinuous reference signal r(t) to get the desired signals Yd,
iJd and Yd. The reference model is taken as
where Wn = 5.0 rad/min and (n = 1.0 with the initial conditions Yd(O)
0.1, Yd(O) = 0.0, Yd(O) = 0.0.
(6.133)
The input vector of MNNs is chosen as z = [y, Y, Yd, Yd, v, IV. The gain
ks = 100.0 and the NN node number l = 10 are used in the simulation.
The parameters in the weight updating laws (6.83) and (6.84) are chosen as
r w = diag{2.0}, rv = diag{100.0}, Ow = 0.2 and Ov = 4.0 X 10- 4 . The initial
conditions are x(O) = [0.1, 440.0V, and W(O) = 0.0. The initial weights \/(0)
are taken randomly in the interval [-1, 1J.
The dashed lines in Figures 6.4 (a) and 6.4 (b) indicate the output and input
of the system using PI controller (6.132), respectively. The PI controller cannot
provide a good tracking response due to the effects of the nonlinearities. The
solid lines in Figures 6.4 (a) -6.4 (b) show the output and input of the system
using adaptive MNN controller (6.133). Comparing the set-point responses in
Figures 6.4 (a), the adaptive NN controller presents better control performance
than that of the PI controller. It should be noticed that though we choose the
initial weights W(O) = 0.0 and \/(0) randomly, the set-point tracking of the NN
controller is still guaranteed during the initial period. This is due to the choice
of the PI control, V'pi, as the prior control term. Figure 6.4 (b) indicates that a
stronger control action is needed when using NN controller (6.133), which is not
surprising from a practical point of view. The boundedness of the NN weight
estimates are indicated in Figure 6.4 (c). To show the effectiveness of neural
180 Chapter 6. Non-affine Nonlinear Systems
networks further, the responses of the bounding controller and neural network
outputs are shown in Figure 6.4 (d). It is found that the magnitude of Ub is
much smaller than that of neural networks. Therefore, the neural networks
plays a major role in improving the control performance.
112 vlI1n
110 .,
:1/ I\_~I n
111\ !
11 1,,1,' .
108 ~~+~ i1 I\-:-ll~
III \ IIII
106 ,'". I.
I
III>
'I
,", \'\
"• I ,
98 "
0.08
96
O,070!---:5:---~'0'-----':'5---:"'~--=25':--~30,-!.T~me~m~il35 9<$0 5 10 15 20 25 30 35
(a) Output tracking (Yd" ", (b) Control input (PI control" "
PI control "- -" and NN control "-") and NN control "-")
,mil
2
o17'1~
", I,
- 2 "1/11
II
II
"" '
..
-4
,,,,,,
I
/ ' 'I
1,1
,I'
, 1,1
-a
,,,,,,
'I'
3 ' ...... ~' "I
-' I
-12 "i\
i
-14 i
-16
Tinemifl Time min
10 15 ~ ~ ~ 35 -180 5 10 15 20 2S 3() 35
The adaptive controller (6.102)-(6.103) is taken with Uk = Upi and the para-
meters Al = 3.0 and e = 0.1. Since the time derivative of the concentration Y
is not available, it is estimated by a second-order high-gain observer
with ).1 = 1.0 and the initial condition 1l'(0) = [0.1, O.ojT.
The MNNs used in the controller contains 10 nodes and the parameters
in the weight updating laws (6.104) and (6.105) are r w = diag{10.0}, rv =
diag{100.0}, 6w = 2.0 X 10- 3 and 6v = 2.0 X 10- 4 . The initial weights W(O) =
0.0 and the elements of V(O) are taken randomly in the interval [-0.5,0.5].
Next we utilize Theorem 6.4 to specify the design parameters t and ks such
that the system output IXll < 0.8. Because the plant is a second order system
and A = Al = 3.0, we know that ( = Xl - Yd and ko = 1 and Ao = 3.0. Then,
from (6.117) and (6.123), the upper bound of the state Xl is found
(6.134)
Since Xl (0) = Yd(O) = 0.1 and the set-point is changed from 0.08mol/l to
0.12mol/l, it is shown that ((0) = 0 and IYd(t)1 ~ 0.12. In order to obtain
the bound of IXl(t)l, we make two suppositions: (i) for the MNNs with NN
node number l = 10 and the constraint condition IIW"II ~ Wm = 3.0, the NN
approximation error 1J.L11 ~ 2.0, and (ii) a very conservative bound lesol ~ 20.0.
Then, it follows from (6.122) and (6.124) that for all ks > k; = 6.8011 and
t < f* = 0.0102, the state IXll < 0.8 holds. In the simulation, we take ks = 7.0
and f = 0.01.
The comparison studies for PI controller (6.132) and the adaptive output
feedback controller (6.102) are given in Figure 6.5. Comparing the tracking
results of two controllers in Figures 6.5(a) and 6.5(b), the adaptive NN con-
troller presents better tracking performance than that of the PI controller. It
should be noticed that though the initial weights W(O) = 0.0 and V(O) are
chosen randomly, the set-point tracking of the NN controller is still guaranteed
during the initial period. This is due to the choice of the PI control, Upi, as
the prior control term in the adaptive controller. Figure 6.5(b) indicates that
the control efforts of the two controllers and the boundedness of the NN weight
estimates is indicated in Figures 6.5(c). In Figure 6.5(d), a small estimation
error of output derivative iJ - 1l'2/f is shown owing to the high-gain observer.
182 Chapter 6. Non-affine Nonlinear Systems
...
0.12
\~~" ~ 110
~ ~~ ~
5
0, 2~'
'; I~~
;~I ,Bi;
'I" ,~ ,~r,
~,
.' 'H'"
~ ',II, ~
, ) ~ I~ :,,'\ ;~h~ ~ ~ ~ ~"
./,
,I
~ ~
1"1,','1
II(~' ~
.,
11 I I"~ ~
1 '1 11
I ,
'11'
I~
I~Y '"
'"~ ",111 "
0.11 5 ;)' ,~ '/ ,:~ ," " /',' 1,1 III'IV
',: I~ \ I ~ :i:~
"',
II',
~ ~:
" t
, I~ \'
:,~
:',: I
I "
" I"
:~
"r ,,,
01 I " " '06 'I
I • ':
0.105 '04
O. , 102
0.095
,OIl
0.09
,
\
"I
\ ,
I
,
I
,
\
I 98 \
0.085 I
',\.
I,
,\ ,\, I'
moo
008
o 10 W ~ ~ ~ ro
'\'
ro ~
96
0 10 ro ~ ~ ~ ro ro 00
0.02 mol/11m,"
0.01
011. ~ ~ l ~ l ~l ~
iii , , ' , I r I
~.Ol
-0.02
-<1.03
-0.05
-<1.06
-<1.7
~~~,,~-7.20~-7.30-~~-~50--'~O-~ro~~~ Time min
zo
(C) NN weights IIWII ("-") -R080 to
7r2/ E
and IIVIIF ("- -")
6.5 Conclusion
In this chapter, adaptive NN control for non affine nonlinear system has been
discussed. Three different implicit desired feedback control laws were first pre-
sented, and then adaptive state feedback control and output feedback control
using LPNN and MNN were illustated respectively. Simulation results have
shown the effectiveness of the proposed methods.
Chapter 7
Triangular Nonlinear
Systems
7 .1 Introduction
In the past decade, adaptive backstepping [92J has become one of the most
popular design methods for adaptive nonlinear control because it can guar-
antee global stabilities, tracking, and transient performance for the broad
class of strict-feedback systems with unknown parameters [104J. For the case
when both unknown nonlinear functions and/or parametric uncertainty are
present in the systems, adaptive neural control schemes were recently pre-
sented [147,151, 192J. In Chapter 5, by using backstepping design and intro-
ducing a novel Lyapunov function, we presented an adaptive neural control
scheme for nonlinear strict-feedback systems in the following form
In Section 7.5, direct adaptive neural control methods are proposed for
some classes of uncertain multi-input/multi-output (MIMO) nonlinear systems
in block-triangular forms, with uncertain nonlinear functions and/or unknown
constant parameters. For MIMO nonlinear systems, the control problem is very
complicated due to the couplings among various inputs and outputs. Moreover,
when there exist uncertain parameters and/or unknown nonlinear functions
in the input coupling matrix, it is in general very difficult to deal with the
problem of controlling MIMO nonlinear systems. With the employment of
NNs to approximate all the uncertain nonlinear functions in the controllers
design, and by utilizing the systems' block-triangular property, the developed
schemes achieve semi-global uniform ultimate boundedness of all the signals
in the closed-loop of MIMO nonlinear systems. The proposed direct adaptive
NN control schemes can be applied to the classes of MIMO nonlinear systems
without repeating the complex controller design procedure for different system
nonlinearities.
are. introduced and play an important role in overcoming the singularity prob-
lem. However, due to the integral operation, this approach is complicated and
is regarded as tedious to use in practice.
It should be noted that in system (7.3), the affine term gn(Xn-l) is assumed
to be independent of state X n . By exploiting this property, the developed neural
control scheme avoids the controller singularity problem completely, and the
stability of the resulting adaptive system is guaranteed without the require-
ment for integral Lyapunov functions. On the other hand, since differentiation
of the virtual controls has to be performed in backstepping design, the differen-
tiation requirement becomes a technical problem when designing controllers for
the uncertain strict-feedback system (7.1). The derivatives of virtual controls
will involve the uncertainties IiO and gin, and thus are unknown nonlinear
functions which are unavailable for implementation. Though these derivatives
are smooth functions of the system states and neural weight estimates, and can
also be approximated by NNs, the approximation might become computation-
ally unacceptable when the large number of neural weight estimates are taken
as inputs to the NNs. By introducing some intermediate variables as inputs
to NNs, where the intermediate variables are available through the computa-
tion of system states and neural weight estimates, the NN approximation can
be implemented by using a minimal number of NN inputs. In summary, the
differentiation requirement can be satisfied by using NNs to approximate all
the uncertainties, including the derivatives of virtual controls, while interme-
diate variables are introduced to make the NN approximation computationally
feasible.
With the proposed neural control scheme, semi-global uniform ultimate
boundedness of all the signals in the closed-loop are guaranteed, and the output
of the system is proven to converge to a small neighborhood of the desired
trajectory. The control performance of the closed-loop system is guaranteed
by suitably choosing the design parameters. The proposed direct adaptive NN
controller can be applied to a large class of nonlinear strict-feedback systems in
the form (7.3) without repeating the complex controller design procedure for
different system nonlinearities.
The control objective is to design a direct adaptive NN controller for system
(7.3) such that (i) all the signals in the closed-loop remain semi-globally uni-
formly ultimately bounded, and (ii) the output y follows a desired trajectory
Yd in (7.2).
It should be noted that in the following derivation of the direct adaptive
NN controller, NN approximation is only guaranteed within a compact set.
Accordingly, the stability results obtained are semi-global in the sense that,
as long as the states of the system (7.3) remain within the compact set, i.e.,
xn Ene Rn, 'r:/t ~ 0, where n can be made as large as desired, there exists a
controller with sufficiently large number of NN nodes such that all the signals
7.2 Special Systems in Strict-Feedback Form 187
n-l ( ) n-l ( )
. (-
gn Xn-l
) '"' agn Xn-l.
= L.. ax
'"' agn Xn-l
Xk = L.. a
[ (-)
gk Xk Xk+l + Ik (-Xk )1
k=l k k=l Xk
Clearly, they only depend on states xn . Because Ii (.) and gi (-) are assumed
to be smooth functions, they are therefore bounded within the compact set O.
Accordingly, we have the following assumption.
Assumption 7.2. There exist constants gid > 0 such that IthOI ::; gid, 'l:fxn E
Oc Rn.
It should be noted that the bounds gil,giO and gid in Assumptions 7.1-7.2
are only used for analytical purpose. They are not used for controller design.
For illustration purpose, the following RBF NN is used to approximate the
continuous function h(Z) : Rq -- R,
where the input vector Z E Oz C Rq, weight vector W = [WI, W2, ... ,W1V E
Rl, the NN node number l > 1; and 8(Z) = [8l(Z),'" ,81(Z)V with 8i(Z)
given by (3.4).
It has been proven that neural network (7.4) can approximate any con-
tinuous function over a compact set Oz C Rq to arbitrary any accuracy as
Step 1
Define Zl = Xl - Xdl. Its derivative is
(7.6)
91 gl 91
(7.7)
withCI 2:: :JM- > 0 being a design constant, such that lIZl = -C1 Z r - 82 Zf S
~o
'1
~
-CIZf + fM-Zf S O. Therefore V is a Lyapunov function, and Zl = 0 is
910
Z1
asymptotically stable.
7.2 Special Systems in Strict-Feedback Form 189
Since h(xI) and gl(XI) are unknown smooth functions, the desired feedback
control at cannot be implemented in practice. From (7.7), it can be seen that
the unknown part gl(~d (h(xl) - Xdl) is a smooth function of Xl and Xdl.
Denote
(7.8)
where Wi denotes the ideal constant weights, and IEll :::; E1 is the approximation
error with constant Ei > O. Since Wi is unknown, let WI be the estimate of
Wi-
Remark 7.i. The principle for designing the neural network WT8l(Zl} is to
use as few neurons as possible to approximate the unknown functions hI (Zt}.
Since Xdl = fdl (Xd) is available through the computation of Xd, we use Xdl E R
instead of Xd E R m as an input to NN wi
8 1 (Zl), i.e., we define Zl = [Xl, Xdl]T,
rather than Zl = [Xl, Xd]T to approximate hI (Zl)' Thus, minimal inputs to
the neural network wi 81 (Zl) are employed to implement the approximation.
Since X2 is only taken as a virtual control, not as the real control input for
the zl-subsystem, by introducing the error variable Z2 = X2 - al and choosing
the virtual control
(7.9)
(7.10)
Adding and subtracting gl(xl}ai in the right hand of (7.1O), and noting (7.7)
and (7.8), we have
(7.11)
1 2 1 - T -1 -
VI = 2g l (Xl)Zl + 2Wl r l WI (7.12)
(7.14)
(7.15)
(7.16)
(7.17)
where the coupling term ZIZ2 will be canceled in the next step.
7.2 Special Systems in Strict-Feedback Form 191
Step 2
The derivative of Z2 = X2 - al is Z2 = h(X2) + g2(X2)X3 - 0: 1. By viewing
X3 as a virtual control input to stabilize the (ZI' z2)-subsystem, there exists a
desired feedback control
(7.19)
:X
where ¢1 = ~Xd+ I [ri (81 (Zl)Zl -0'1 W1 )] is introduced as an intermediate
variable which is computable. Since h(Xl) and gl(Xl) are unknown, Ql is in
fact a scalar unknown nonlinear function. Let
(7.20)
(7.21)
Then, we have
(7.23)
(7.26)
where a2 > 0 is a small constant. Let C2 = C20 + C21, where C20 and C2I > O.
By using (7.17), (7.23) and (7.26), and with some completion of squares and
straightforward derivation similar to those employed in Step 1, the derivative
of V2 becomes
• t::,. 92d
where C20 1S chosen such that c;o = C20 - -2- > O.
2920
Step i (3 ~ i ~ n - 1)
The derivative of Zi = Xi - ai-I is Zi = !i(Xi) + 9i(Xi)Xi+1 - ai-I. By viewing
Xi+1 as a virtual control input to stabilize the (ZI,'" ,zi)-subsystem, there
exists a desired feedback control a; = Xi+1
(7.28)
where
(7.29)
7.2 Special Systems in Strict-Feedback Form 193
, . h 'l'i-l
WIt A.
= " i - l 8a.' 1 .
L.,.,k=l 8~~ Xd + "L.,.,k=l
i-l 8a.'
8~~
1 [r k {8k (Zk ) Zk - (J'k Wk
A
)1 computable.
Let
(7.30)
Note that the number of inputs to the NN is kept minimal by the introduction of
intermediate variables ~,
OXl
... , ~a.;_l
X,,_l
,<Pi-I, as discussed in Remark 7.2. By
employing an RBF NN W? 8i (Zi) to approximate hi(Zi), at can be expressed
as
(7.32)
Define the error variable Zi+I = Xi+1 - ai and choose the virtual control
(7.33)
Then, we have
(7.36)
where (J'i > 0 is a small constant. Let Ci = Cia + Ci1, where Cia and Cil > O.
By using (7.27), (7.34) and (7.36), and with some completion of squares and
straightforward derivation similar to those employed in the former steps, the
derivative of Vi becomes
h
were Cio '1
IS C losen such1 *!;;
t lat cio := Cia - 9id2 > O.
-2
9iO
194 Chapter 7. Triangular Nonlinear Systems
Step n
This is the final step. The derivative of Zn = Xn - an-I is in = fn{:xn) +
gn(Xn-I)U - an-I. To stabilize the whole system (ZI,'" , zn), there exists a
desired feedback control
(7.38)
where
8;;-1
n-1
a n-1 = L (gk(Xk)Xk+1 + fk(Xk» + <Pn-1 (7.39)
k=1 k
WI
A. ",n-l 8a n - l . ",n-l 8an - l [r k (S k (Z kZk-UkVYk
'th 'f'n-I=L.Jk=l---ax;/Xd+L.Jk=l8W;;" ) tIT )] t ble.
compua
Let
(7.40)
[ -T 8an-l aan-l ]T 2n
Zn = Xn , -8--'"'' -a--,¢n-l
Xl Xn-l
c R (7.41)
(7.42)
we have
(7.44)
where Un > 0 is a small constant. Let Cn = Cno + Cn 1, where CnO and Cn1 > O.
By using (7.37), (7.44) and (7.46), and with some completion of squares and
straightforward derivation similar to those employed in the former steps, the
derivative of Vi becomes
(7.47)
h
were . c1losen suc11 t 1lat
Cno IS "L:;.
cnO : = cnO - 2'"2
9nd > 0 .
9n o
1IW;"1I 2
Let 8 ~ Ln
Uk 2k +L
n
4€k
,,2
. If we choose ckO such that CkO ~ 27 ,
k=l k=l Ckl 9kO
. c1loose CkO SUCl
l.e., 1 t hat CkO 7 + -2-'
> -2- 9kd k = h7'IS a POSI't'Ive
1, " ' , n, were
gkO 29 ko
constant, and choose Uk and rk such that Uk ~ 7Amax{r;1}, k = 1"" ,n,
then from (7.47) we have the following inequality
[L..t
~ 1 2 ~ w[r;l Wk] ;:
::; -7 2 Zk + L..t 2 +u
k=l gk k=l
::; -7Vn + 8 (7.48)
Theorem 7.1. [57} Consider the closed-loop system consisting of the plant
(7.3), the reference model {7.2}, the controller {7.43} and the NN weight up-
dating laws {7.14}, {7.26}, {7.36} and {7.46}. Assume there exists sufficiently
large compact sets Oi E R 2i , i = 1, ... ,n such that Zi E Oi for all t ~ O. Then,
for bounded initial conditions,
(i) all signals in the closed-loop system remain bounded, and the states xn
and the neural weight estimates W[, .. " W! eventually converge to the
compact set
(7.49)
(ii) the output tracking error y(t) - Yd(t) converges to a small neighborhood
around zero by appropriately choosing design parameters.
196 Chapter 7. Triangular Nonlinear Systems
Proof (i) From (7.48), using the Boundedness Theorem (e.g., Theorem 2.14
in [156]), we have that all Zi and Wi (i = 1,··· , n) are uniformly ultimately
bounded. Since ZI = Xl -Xdl and Xdl are bounded, we have that Xl is bounded.
From Zi = Xi - ai-I, i = 2, ... , n, and the definitions of virtual controls ai
(7.9)(7.22)(7.33), we have that Xi, i = 2,··· ,n remain bounded. Using (7.43),
we conclude that control u. is also bounded. Thus, all the signals in the closed-
loop system remain bounded.
To provide some estimates of the regions of attraction of equation (7.48),
we consider the following two conditions.
1) if
(7.50)
where Osl is given in (7.49), then according to Theorem 2.14 in [156], all
the states :En and the neural weights WI, ... , Wn will remain in Osl, i.e.,
2) if
where 0;1 denotes the complimentary set of Osl, then lin remains nega-
tive definite until the stateS:En and the neural weights WI,'" , Wn enter
and stay inside Osl, i.e.,
Thus, for bounded initial conditions, all signals in the closed-loop system
remain bounded, and the states :En and the neural weights WI," . , Wn will
eventually converge to the compact set 0sl'
(ii) Let p ~ oh > 0, then (7.48) satisfies
0::; Vn(t) < P + (Vn(O) - p)exp( -it) (7.51)
From (7.51), we have
1
L -z~
n
< p + (Vn(O) - p)exp( -it)
k=1 2gk
(7.52)
that is,
n
I>~ < 2g* p + 2g*Vn(O)exp( -"It) (7.54)
k=1
which implies that given fL > ~, there exists T such that for all t 2: T, the
tracking error satisfies
(7.55)
where fL is the size of a small residual set which depends on the NN approxi-
mation error Ei and controller parameters Ci, ai and rio It is easily seen that
the increase in the control gain Ci, adaptive gain r i and NN node number lj
will result in a better tracking performance. <>
Remark 7.3. In the above analysis, it is clear that the uniform ultimate
bounded ness. of all the signals are guaranteed by choosing Ci = CiQ + Cil large
enough, such that c:o ~ CiQ - gi~ > O. Moreover, it can be seen that i) increas-
2giO
ing CiQ might lead to larger "I, and increasing Cil will reduce 0, thus, increasing
Ci will lead to smaller ns1 ; ii) decreasing ai will help to reduce 0, and increas-
ing the NN node number lj will help to reduce <, both of which will help to
reduce the size of n s1 . However, increasing Ci will lead to a high gain control
scheme. On the other hand, though ai is required to be chosen as a small pos-
itive constant when applying a-modification [74], a very small ai may not be
enough to prevent the NN weight estimates from drifting to very large values in
the presence of the NN approximation errors, where the large Wi might result
in a variation of a high-gain control. Therefore, in practical applications, the
design parameters should be adjusted carefully for achieving suitable transient
performance and control action.
Remark 7.4. Note that in backstepping design, differentiation of the virtual
controls has to be performed. For the control of uncertain strict-feedback sys-
tem (7.1), the differentiation requirement becomes a technical problem due to
the presence of uncertainties. The derivative of virtual control O:i-l (7.29)
involves the unknown nonlinear functions h,'" ,fi-l0 and gl,'" ,gi-l(-)
(i = 2"" ,n), and thus is a unknown nonlinear function which is unavailable
for implementation. Because derivative ai-l appears in the desired virtual
control o:t (7.28), an RBF NN can be employed to approximate all the un-
known part of o:t (7.28), including the unknown ai-I. However, since ai-l is a
smooth function of system states Xi and neural weight estimates WI, . .. ,Wi-I,
as can be seen from (7.29), approximating the unknown nonlinear function ai-l
might become computationally unacceptable when the large number of neural
weight estimates WI,'" , Wi - 1 are taken as inputs to the RBF NN, which will
198 Chapter 7. Triangular Nonlinear Systems
Remark 7.5. The adaptive NN controller (7.43) with adaptation laws (7.14),
(7.26), (7.36) and (7.46) are highly structural, and independent of the complex-
ities of the system nonlinearities. Thus, it can be applied to other similar plants
without repeating the complex controller design procedure for different system
nonlinearities. In addition, such a structural property is particularly suitable
for parallel processing and hardware implementation in practical applications.
Xl = O.5Xl + (1 + o.lxDx2
X2 = XlX2 + [2 + COS(Xl)JU (7.56)
Y= Xl
and
Xl = O.lx~ + [2 + sin{xl)]x2
X2 = O.2x~ + O.3(e X1 + e-X1)u (7.57)
Y = Xl
7.2 Special Systems in Strict-Feedback Form 199
The control objective is to guarantee (i) all the signals in the closed-loop system
remain bounded, and (ii) the outputs Y of the two systems follow a desired
trajectory Yd generated from the following van der Pol oscillator system
Xdl = Xd2
As shown in [177], the phase-plane trajectories of the van der Pol oscillator,
starting from an initial state other than {O,O}, approach a limit cycle when
,6 > 0.
~~~~~~~-~10~,,~7.14--~~'~'~20' ~L-~~--~~'--~10~1~'~~~~1~'~20'
Tllne(Seeonds) TlIlIttsecards)
"
-, At os
,
:
I
,1
t l ~
\I ..
" '
-60 ~---;----;--~---!o"---;--~---!o--~1B-.-J" "0 10 12 16 18. 20
Tunt/Seconds) '" TlmelSecon:jsl '"
(c) Control input u(t} (d) IIWIiI ("-") and IIW2 11 ("- -")
Figure 7.1: Responses of adaptive NN controller for plant {7 ..57}
Clearly, systems {7.57} and (7.58) are in strict-feedback form (7.3) and
200 Chapter 7. Triangular Nonlinear Systems
satisfies Assumptions 7.1 and 7.2. As systems (7.57) and (7.58) are of second
order, the adaptive NN controller is chosen according to (7.43) as follows
(7.59)
(7.60)
II ,'1 ,\
I'
" ""
IS
I'
I'
II
I' ," ,
I:
~~~-7-~~'--'~fD:--'':-'~''~~'':--'':-'~'' '!-',-"-!,--,-:-,-'~~'-~":--'':-2~,,~~,,:--,':-.~2D'
Tlrne(Seoondsj ,.. Tme(Secordsl,..
(c) Control input u(t) (d) IIW11I ("-") and IIW2 11 ("- -")
where Xi = [Xl,'" ,Xi]T E Ri, i = 1,'" ,n are the state vector, y E R,u E
R are system output and input, respectively; h(Xi) and Fi(Xi) are known
smooth nonlinear functions, c,oi(xd are unknown smooth nonlinear functions;
and 0i are unknown constant parameters, the affine terms gi of. 0, i = 1, ... , n
202 Chapter 7. Triangular Nonlinear Systems
are unknown constant parameters whose signs are known. Without losing
generality, it is assumed that 9i > 0, i = 1, ... ,n.
The partially known strict-feedback system (7.61) is motivated by previ-
ous work on the control of uncertain strict-feedback systems [147,150,151].
The structural properties of system (7.61) include the unknown constant affine
terms 9i and the a priori information represented by fieri) and Fi(xd. In
practice, there are many systems falling into this category, such as induc-
tion motors [113], Rossler system [158], Chua's circuits [27], etc. Compared
with the systems considered in Chapter 5 and Section 7.2, the expression of
!;(Xi) +(JT Fi(Xi) +'Pi(Xi) in system (7.61) is more general in the sense that both
known and unknown information are included, and both parametric and non-
parametric uncertainties are contained. The known information represented by
!;(Xi) and Fi(Xi) might be available through the laws of physics, properties of
materials, or any identification methods. The unknown constants gi, (Ji, and
the unknown nonlinearities 'Pi(Xi) might come from the inaccurately modeling
or modeling .reduction.
It is noticed that though the neural control approaches in Chapter 5 and
Section 7.2 are also applicable to system (7.61) when setting the affine terms
9iC) in (7.3) as unknown constants, the a priori information of the system
dynamics was not incorporated into controller design due to the fact that less
properties of the systems were employed. In the presented scheme, by exploit-
ing the system structural properties of constant affine terms 9i and known
nonlinear functions fiC) and Fi (·), neural networks are only used to approxi-
mate the unknown nonlinear functions which are parts of the overall nonlinear
functions in the virtual controls and the practical control. The benefits of
using the structural properties include: (i) the number of neurons employed
for approximation is much less than those used in Chapter 5 and Section 7.2,
since for the neural network inputs, Zi E R 2i - 1 as to be shown in this sec-
tion, while Zi E R2i in Section 7.2 and Zi E R2i+l in Chapter 5; (ii) better
control performance might be achieved, because neural networks are used only
to approximate the unknown functions which might have lower nonlinearity,
less complexity and smaller magnitudes than the overall nonlinear functions,
and thus might produce smaller approximation errors; and (iii) with the help
of NN approximation, the overparametrization problem caused by parametric
uncertainties in adaptive backstepping design is avoided without using tuning
functions [103]. The presented scheme is much simplified since it requires nei-
ther integral operation as in Chapter 5, nor tuning functions to deal with the
partial derivative terms as in [103]. Semi-global uniform ultimate boundedness
of all the signals in the closed-loop is guaranteed and the output of the sys-
tem is proven to converge to a small neighborhood of the desired trajectory.
The control performance of the closed-loop system is guaranteed by suitably
choosing the design parameters.
7.3 Partially Known Nonlinear Systems 203
Step 1
Define Zl = Xl - Xdl. Its derivative is given by
(7.62)
When viewing X2 as a virtual control input, there exists a desired feedback
control
(7.65)
204 Chapter 7. Triangular Nonlinear Systems
with TJ1 being the estimates of 1]1 and W1 being the estimates of Wi, we have
i1 = gl(Z2 + a1) + of F 1 (X1) + c,ol(X1) + h(X1) - !d1(Xd)
= gl[Z2 - C1Z1 - (ih -1]l)T Fl'/l - (WI - wtf Sl(Zd + E1] (7.66)
Consider the Lyapunov function candidate
1 1 -T
VI
2
= 2g 1Zl + 2171 r 1'/1-1 TIl_ + 21 W1
- T -1 -
r W W1 1 (7.67)
where ih = r,1 - TIl, WI = W1 - Wi, r 1'/1 = r~l > 0, rW1 = r[v1 > 0. (Through
out this section, we shall define ih = TJi -1]i and Wi = Wi - Wt, i = 1, ... , n.)
The derivative of VI is
V·1 -- Zl Z1 + 1]1
-Tr- 1 ~ + W- Tr- 1 W~
1'/1 1]1 1 W, 1
gl
= Zl Z2 - + Zl E1 - TIl-TD
C1 Z 1
2
rl'/l Zl + TIl 1'/1 TIl
-Tr-1~
. < ~ ~
V1 "'1"'2 -
C z2
10 1 -
0'1'/1/l i
2
h11 2 O'W , //W1 // 2
2
+0'1'/l//Tll112 + O'wl llwtll 2 + Ei
2
(7.72)
2 2 4Cll
7.3 Partially Known Nonlinear Systems 205
Step 2
The derivative of Z2 = X2 - al is
(7.73)
(7.74)
where C2is a positive constant to be specified later, and til is a function of Xl,
Xd, "1 and WI. Thus, al can be expressed as
(7.75)
where
¢1 =
(7.76)
(7.77)
OT
where 1]2 = [g~ , i2 jT is an unknown constant vector, F'Y}2 = [F!, (12 - ¢l}]T
is a known function vector. Note that since ¢1 is computable through the
computation of X2, Xd, "1
and Wl, as given in equation (7.76), it is included in
the known function vector F'Y}2.
For the desired feedback control a2' let
(7.78)
206 Chapter 7. Triangular Nonlinear Systems
where Wi is the vector of ideal constant weights, and IE21 ::; E:; is the approxi-
mation error with constant E:; > O.
Remark 7.6. Note that in this step, Eq. (7.77) can be written as
a; = -Zl - C2Z2 - ~
"If F'12 - 92 [<P2(X2) - aaaxll <PI (xd] (7.80)
(JT (JT a a T
where "12 = [~, :2'~' ~Vand Fm = [Fi, (f2-¢1), -~Fl(XI)T, -~X2] ,
and the unknown nonlinear function in (7.80)
h;(Z2) = ~
92
[<P2(X2) - aaaXll <P1(Xd] (7.81)
Remark 7.7. Though the unknown function h 2{Z2) is a function of X2, Xd, r,l
and WI, neural network weights WI are not recommended to be taken as inputs
to the NN because of the curve of dimensionality of RBF NN for the possibility
of large size of WI. By defining intermediate variable ~,
aXl
which is computable
since 0:1 is available, the NN approximation W! 8 2 (Z2) of the unknown function
h 2{Z2) can be computed by using the minimal number of NN inputs Z2 =
[X2' ¥x7jT . The same idea of choosing the inputs of NN is also used in the
following design steps.
Since 1)2 andW; are unknown, 0:2 cannot be realized in practice. Define
Z3 = X3 - 0:2 and let
(7.82)
with r,2 being the estimates of 1)2 and W2 being the estimates of W;, we have
(7.83)
(7.84)
(7.85)
(7.87)
208 Chapter 7. Triangular Nonlinear Systems
-T'
-aW2 W 2 W 2 = -aW2 W 2 (W2
-T -
+ W 2* ) ::; -aw211 W
- 2
211 + aW211 W
-
21111 W 2* I
< _ aw2 liW2 11 2 + aw 11W2'11 2 2
- 2 2
-a'72flf il2 = -a'12flfUh + 1];) ::; -a'721Ifl21i 2+ a'72 IIfl211 111]211
< _ a'72llfl2112 + a'721i1]211 2
2 2
(7.88)
(7.89)
(7.90)
When viewing Xi+l as a virtual control input to stabilize the (ZI,'" , zd-
subsystem, there exists a desired feedback control
Q*
't = -z· 1 - c·z· - ~[cp.(x.)
f,- 1. t + f)T R(x·)
t t " + f·(x)
" t - a· 1]
,t 1. 1- (7.91)
9i
(7.92)
7.3 Partially Known Nonlinear Systems 209
where
~ 8;i-l
-~g. [IPi(Xi) - k=l Xk
(gkXk+l -/- Or Fk(xk) -/- IPk(Xk))] (7.94)
where'TJi = [~, l]T is an unknown constant vector, Fry, = [Fl, (Ii - ¢i-1)]T
g. fl·
is a known functIOn vector.
For the desired feedback control aT, let
(7.96)
where Wt denotes the vector of ideal constant weights, and IEil < < is the
approximation error with constant Et > o.
Define zH 1 = XH 1 - ai and let
(7.97)
with TJi being the estimates of'TJi and Wi being the estimates of Wt, we have
- T - T ~
-Wi Si(Zi)Zi + Wi fw,1 Wi (7.100)
(7.102)
(7.104)
Step n
This is the final step. The derivative of Zn = Xn - Qn-l is
(7.105)
7.3 Partially Known Nonlinear Systems 211
To stabilize the whole system (ZI,'" , zn), there exists a desired feedback con-
trol
where
<Pn-I
(7.108)
u* = Tp,'In
-Zn-I - CnZn - TJn
-~ ['Pn(Xn) -
9n
I: a~n-l
k=1 Xk
(9kXk+l -\- 'Pk(Xk) + Or FdXk))] (7.109)
(F
where TJn = [-ZL,....L]T
gn (In
is an unknown constant vector, F'In = [FJ, Un - <Pn-l)]T
is a known functlOn vector.
For the desired feedback control u*, let
(7.111)
212 Chapter 7. Triangular Nonlinear Systems
where W~ denotes the vector of ideal constant weights, and I€nl ::; €~ is the
approximation error with constant €~ > O. Let
(7.112)
where r,n being the estimates of TIn and Wn being the estimates of W~. Then,
we have
The derivative of Vn is
. . znzn -T -1 ~ - T -1 ~
Vn = Vn- 1 + -gn- +Tlnr.,·on TIn + Wn rWn Wn
= v..n-1 - Zn-1Zn - CnZn + Zn(n - TIn .L"l1n Zn
2 -T z;>
+ TIn
-Tr- 1 ~
I1n TIn
- T - T -1 ~
-Wn Sn(Zn)Zn + Wn r Wn W n (7.115)
(7.117)
(7.118)
7.3 Partially Known Nonlinear Systems 213
(7.119)
Theorem 7.2. Consider the partially known nonlinear system (7.61). Assume
there exists sufficiently large compact sets Oi E R 2i - l , i = 1,' .. ,n such that
Zi E Oi for all t :::: O. Consider the adaptation laws for r,i and Wi, i = 1, ... ,n
as in (7.69), (7.86), (7.101) and (7.116). Then the control law (7.112) guar-
antees (i) ultimate uniform boundedness of all the signals in the closed-loop
system, including the states of the slave system x = [Xl, ... ,Xn]T, the control
u and parameter estimates r,i and NN weights Wi, and (ii) the output tracking
error yet) -Y,d(t) converges to a small neighborhood around zero by appropriately
choosing design parameters [178].
n II 112 n IIW*112 n *2
Proof Let b := ""'
L
a fJk
2
Tfk + ""'
L
aWk
2
k + ""' ~.
L4ckl
If we choose CkO
k=l k=l k=l
such that CkO > -2' ,k = 1, ... ,n, where, is a positive constant, and choose
gkO
a fJk' aWk and r fJk' rWk such that
(7.121)
control u is also bounded. Thus, all the signals in the closed-loop system remain
bounded.
From (7.121), we have
n 1
~ _z2 < p + (Vn(O) - p)exp( -/,t)
~ 2g k k
(7.122)
(7.123)
i.e.,
n
LZ~ < 2g*p+2g*Vn(O)exp(-/'t) (7.124)
k=1
which implies that given J.L > y'2g*p, there exists T > 0 such that for all t ~ T,
the tracking error satisfies
(7.12.5)
where J.L is the size of a small residual set which depends on the NN approxi-
mation error Ei and controller parameters Ci, a"li' anl, and r "Ii' rwi . It can be
seen that i) increasing CiO might lead to larger /', and increasing Cil will reduce
b, thus, increasing Ci will lead to smaller p; ii) decreasing ai will help to reduce
b, and increasing the NN node number li will help to reduce <,
both of which
will help to reduce p. Thus, the increase in the control gain Ci and NN node
number li, and the decrease in ai will result in a better tracking performance.
o
Remark 7.8. In the literature of adaptive NN control, since the estimates gi
for unknown constants gi usually appear in the denominators of the controllers,
parameter projection algorithms are usually used to avoid the possible singu-
larity problem. In this section, by using NNs to approximate the unknown
functions as a whole in the desired virtual control aT and desired feedback
control u*, the ith-order subsystem is stabilized with respect to Lyapunov
functions Vi, Therefore, controller singularity is avoided since there is no need
to explicitly estimate the affine terms gi in the controller.
Remark 7.9. Note that in the design steps, neural networks W? Si(Zi) (i =
I, ... , n) are only used to approximate the unknown nonlinear functions hI (ZI) =
7.3 Partially Known Nonlinear Systems 215
il<f'l(Xl) and hi(Zi) (i = 2"" ,n), which are parts of the overall nonlin-
il
ear functions [<f'l(Xl) + OfFl(XI) + !1(Xl) - fdl(Xd)] in oi, and t[<f'i(Xi) +
Or Fi(Xi) + h(Xi) - Qi-l] in or
(i = 2, ... ,n, o~ = u*). In other words, neural
networks are only used to approximate unknown functions hi(Zi) (i = 1"" ,n)
which might have lower nonlinearity, less complexity and smaller magnitude.
We claim that better tracking performance might be achieved, because approx-
imating nonlinear functions hi ( Zi) rather than the overall nonlinear functions
might produce smaller NN approximation errors fi, which will lead to smaller iJ
and p. This can be seen as one of benefits of exploiting the structural properties
and the known system information, and it will be investigated by simulation
studies. Another benefit is that the neurons employed for approximation is
much less than those used in Chapter 5 and Section 7.2. From the point of
view of implementation, this makes the scheme computationally feasible.
The reference model is chosen as the famous van der Pol oscillator (see, e.g,
[177]) given by
Xdl = Xd2
Xd2 = -Xdl + .8(1 - X~1)Xd2
Yd = Xdl (7.127)
As shown in [177], the phase-plane trajectories of the van der Pol oscillator,
starting from an initial state other than (0,0), approach a limit cycle when
.8 > O.
The objective is twofold. The first is to guarantee that all the signals in
the closed-loop system remain bounded, and the output Y follows a desired
trajectory Yd generated from van der Pol oscillator (7.127). The second is to
investigate whether the adaptive neural controller (7.112) will achieve better
tracking performance in case a priori information is available.
Clearly, plant (7.126) is in strict-feedback form (7.61). As system (7.126) is
of second order, the adaptive NN controller is chosen according to (7.112) as
(7.128)
216 Chapter 7. Triangular Nonlinear Systems
(7.129)
information, while <PI (Xl) = sin Xl! <P2(X2) = XIX2 are unknown nonlinear
functions, and (h = 0.5, ()2 = [-0.3,O.2jT, gl = 1.2, g2 = 1.7 are unknown
constants. All the design parameters, including the centers and widths of neural
networks, are kept the same for both cases.
It can be seen from Figures 7.3 and 7.4 that in both cases, the neural
controller (7.128) achieves boundedness of all the signals in the closed-loop.
Moreover, the output y of system (7.126) converges to a small neighborhood
of the desired trajectory Yd. It is interesting to notice that the tracking perfor-
mance of case (ii) is much better than that of case (i), as shown in Figure 7.3
(a)-(b), respectively. Since all the parameters are kept unchanged, the improve-
ment of the tracking performance is not resulted from the increase of the NN
node numbers li (i = 1,2), but possibly from the reduction of NN approxima-
tion errors Ei (i = 1,2), which is achieved by exploiting the system structural
properties and the known system information, as mentioned in Remark 7.9.
~1 : (7.131)
218 Chapter 7. Triangular Nonlinear Systems
-,
-2
, ,
,, ,,,
I:
..
-2 ,
./
)"
~,~~~~--~--~.--~,~,--~~~~~~,,--~--~.
TIII'ie(Seccrdt)
-., • HI
TmetSeclrldsJ
12 10 11 20
(a) Xl of case (i) ("- -"), Xl of (b) State X2: case (i) ("- -")
case (ii) ("_") and Xdl (" ... ") and case (ii) ("_")
-,
-100~~:-~--~----:-'---:,':-,---7l2o----::"~-::"---,,!-;.---:!..
l .... (SeciordI1
" "
" "
""
,11
,"'"
"- ---", ,,- ...... ,
" 15 III
...... - ... - ~ ' . . . ,h/\/ '\ .... , .... r '\ ... \
"
',~~,:-~.--~.----:-.--_,,!_;,---:,~,---7":----::II~-::"--~"· °0 2 4 II 8 10 12 14 111 1. 20
lime(SeoordII "T,,"e(Seoon;ls)
(a) Case (i): !!W!! ("-"}j 111111 ("- _") (b) Case (ii): IIWII ("-")j 111111 ("- _")
Figure 7.4: The boundedness of parameters
7.4 Pure-feedback Nonlinear Systems 219
and
Xi = /i(Xi, Xi+!), 1 :S i :S n - 2
{ ~n-l = f~-l(Xn-l~ + gn-l(Xn-l)Xn (7.132)
Xn = fn(xn) + gn(xn)u, n ;::: 3
y= Xl
(7.133)
Clearly, they only depend on states x n . Because IiO, i = 1"" ,n -1, gn-IO
and gnO are assumed to be smooth functions, 9i(-), i = 1, ... ,n are therefore
bounded within a compact set. Accordingly, the following assumption is made
for 9i(-), i = 1, ... , n.
Assumption 7.5. There exist constants 9id > 0 such that Igi(-)I ~ gid, VXn E
Slxn C Rn where Slxn is a compact region, i = 1, . " ,n.
In this section, direct adaptive NN design is presented by using RBF NN.
At each recursive step i, an intermediate desired feedback control a:
is first
shown to exist which possesses some desired stabilizing properties, and then
7.4 Pure-feedback Nonlinear Systems 221
Step 1
Define Zl = Xl - Xdl. Its derivative is
(7.134)
From Assumption 7.4, we know that a h (Xl, X2) / aX2 > glO > 0 for all (Xl, X2) E
R2. Define VI as
(7.135)
Using Mean Value Theorem (Theorem 2.5), there exists ).1 (O < '\1 < 1) such
that
(7.137)
where glAl := gl (Xl, X2Al) with X2Al = '\lX2+(1-'\1)ni. Note that Assumption
7.4 on gl (Xl, X2) is still valid for glAl' Since glAl is a function of Xl, X2 and ni,
and from (7.135) and (7.136), it can be seen that n! is a function of Xl and Xd,
the derivation of ilIAl is
+L
2 n n
L
Tn
uglAl' UglAl .
= --Xk --Xdk
aXk aXdk
k=l k=1
2 n m n
= ~ a;;-
~ uglAl _
!k(Xk+1
)
+ ~ a;;- fdk (Xd)
~ uglAl
k=l k k=1 dk
VXn E OXn and Xd E Od, where gld is the same as in Assumption 7.5 for 1.(71 I·
In other words, Assumption 7.5 on 91 (Xl, X2) is valid for 91>."
Combining (7.134)-(7.137) yields
il = -klZl + g1)." (X2 - an (7.138)
(7.139)
where Wi denotes the ideal constant weights, and Itll ~ lOt is the approximation
error with constant ti > o.
Let WI be the estimate of Wt. Define Z2 = X2 - al and let
, T
al = -ClZl + WI Sl(Zl) (7.140)
(7.141 )
where WI = WI - Wr
Through out this section, define () = () - (-)*.
Consider the Lyapunov function candidate
(7.142)
The derivative of VI is
. . 2
V·
1=
ZIZI _ g1)." Zl
22
+ W-1Tf-lw~
1
1
g1>" g1>"
kl 2
---Zl + ZlZ2 - 2
ClZl -
91>.1
2
-2-Zl - ZItI
glA , 2g lA1
-T
+Wl SI(Zl)Zl + WI
-T
f1 IWI
'
(7.143)
Consider the following adaptation law
(7.144)
where 0"1 > 0 is a small constant. Let C1 = ClO + Cll, with ClO and Cll > O.
Then, equation (7.143) becomes
T'r
VI = kl 2
---ZI + Z1 z2 - (91).1 ) 2
ClO + - 2 - ZI -
2
CllZl
glA, 2g1>.1
- T '
-ZItI - 0"1 WI WI (7.145)
7.4 Pure-feedback Nonlinear Systems 223
(7.146)
Step 2
The derivative of Z2 = X2 - al is
(7.148)
From Assumption 7.4, we know that 8h(x2, X3)/8X3 > 920 > 0 for all
X3 E R3. Define
(7.149)
(7.150)
Using Mean Value Theorem (Theorem 2.5), there exists ,\2(0 < A2 < 1)
such that
(7.151)
where 92>'2 := 92(X2, X3>'2) with X3>'2 = '\2X3 + (1- A2)a;. Note that Assump-
tions 7.4 on 92(X2, X3) is still valid for 92>'2' Since 92>'2 is a function of Xl, X2, X3
224 Chapter 7. Triangular Nonlinear Systems
(7.152)
(7.153)
where
is computable.
By employing an RBF NN wi S2(Z2) to approximate a 2(x2, V2), where
Z2 = [X2, ~, <PI, Z2jT E n 2 C R 5, a2 can be expressed as
where Wi denotes the ideal constant weights, and 1(021 ::; (2 is the approximation
error with constant (2 > O.
Remark 7.10. From the definitions of V2 in (7.149), and al in (7.153), it can
be seen that a2(x2, V2) in (7.151) is a function of X2, Xd, Z2 and WI. However,
7.4 Pure-feedback Nonlinear Systems 225
(7.156)
(7.157)
(7.158)
(7.159)
(7.160)
where a2 > 0 is a small constant. Let C2 = C20 + C21, where C20 and C21 > o.
Then, equation (7.159) becomes
(7.161)
226 Chapter 7. Triangular Nonlinear Systems
(7.162)
* := C20
C20 - 2"2
g2d
>0
g20
(7.163)
Step i (3 ~ i ~ n - 2)
The derivative of Zi = X·i - ai-l is
(7.164)
From Assumption 7.4, we know that 8ji(Xi,XHl)/8xHl > giO > 0 for all
XH 1 E RH 1. Define
(7.165)
According to Lemma 2.8, by viewing XHI as a virtual control input, for every
value of Xi and Vi, there exists a smooth ideal control input XHI = a;(Xi' Vi)
such that
(7.166)
7.4 Pure-feedback Nonlinear Systems 227
Using Mean Value Theorem (Theorem 2.5), there exists /\ (0 < .Ai < 1) such
that
(7.167)
where gi)'i := gi(Xi, X(i+l),x,) with X(i+l),xi = /\Xi+l + (1 - /\)a;' Note that
Assumptions 7.4-7.5 on gi(Xi,Xi+l) is still valid for gi,xi. Combining (7.164)-
(7.167) yields
(7.168)
(7.169)
where
(7.170)
is computable.
By employing an RBF NN wl Si(Zi) to approximate a;(Xi' Vi)' where
Z i-Xi,
-[- ax;:-' ... , 8Xi-1 A-. n C R 2i + l ,a * can b e expresse d as
]T E Hi
80<'-1 80<i-1
,'!'i-l,Zi i
(7.171)
where Wt denotes the ideal constant weights, and lEi I ::; e:; is the approximation
error with constant e:; > o.
Remark 7.11. Following the same idea in Remark 7.10, the number of inputs
to the NN is kept minimal by the introduction of the intermediate variable
--ax;:-'
80<i-1
... , ~
80<'_1
an d '!'i-l
A-.
as 1·t·
s mpu t s.
Since wt is unknown, a; cannot be realized in practice. Let Wi be the
estimate of W/. Define Zi+l = Xi+l - ai and let
(7.172)
-k·z·
1. 1. + g.,
'tl\i [Z·+l
1. - z·1.- 1 - c·z·
1. t + W·
-T
S·(Z·)
1. t t - E·]t (7.173)
228 Chapter 7. Triangular Nonlinear Systems
= \i:.
~-
1 - ~z2 -
. t
z·2- l Z'~ + Z'Z'+1
2 'l
- c·z 2-
2 'l 2g'i>"i
2 z22 - Z,·E.~
g.>"i g';>'i
- T
+Wi Si(Zi)Zi + Wi
- T -1"
r i Wi (7.175)
Wi =
= rd-Si(Zi)Zi - ai Wi]
Wi (7.176)
where ai > 0 is a small constant. Let Ci = CiQ + Cil, where CiQ and Cil > O.
Then, equation (7.175) becomes
Vi. = Vi-I
. ki 2
- -.-Zi - Zi-l Zi + ZiZi+l -
(9i>"i ) 2
+ 2T
CiO Zi
~~ ~~
2 - T "
-CilZi - ZiEi - aiWi Wi (7.177)
By completion of squares, the following inequalities hold
-T" -T - - 2 -
-aiWi Wi = + Wt) ~ -ailiWill
-aiWi (Wi +ailiWillllwtll
aillWdl2 ai II wt 112
< - 2 + 2
(7.178)
i
~aj IIW*112
j i
~ Ej
*2
+ L.,.. + L.,.. -4C '1 (7.179)
j=1
2 j=1 J
7.4 Pure-feedback Nonlinear Systems 229
Step n-1
The derivative of Zn-l = Xn-l - O!n-2 is Zn-l = !n-l(xn-d + gn-l(xn-dxn -
Ct n -2· By viewing Xn as a virtual control to stabilize the (Zl,'" ,Zn-l)-
subsystem, there exists a desired feedback control
where Cn-l is a positive constant to be specified later, gn-l(Xn-l) and !n-l (Xn-l)
are unknown smooth functions of xn-l, and Ct n -2 is a function of X n -2, Xd and
WI,'" ,Wn - 2 . Therefore, Ct n -2 can be expressed as
(7.181)
where
(7.182)
is computable.
By employing an RBF NN W n - 1 T Sn-l (Zn-l) to approximate the following
unknown part of O!~-l in (7.180)
(7.183)
n -2,/..]T E n e R 2 (n-l)
w1lere Z n-l -- [-Xn-l, --ax;--'
8Ct n -2
... , 8Ct
8X n -2 ' 'f'n-2 Hn-l ,
..
O!n-l can
be expressed as
(7.184)
where W~_l denotes the ideal constant weights, and lEn-II ::; E~_l is the ap-
proximation error with constant E~_l > O.
Since W~_l is unknown, 0!~_1 cannot be realized in practice. Let Wn - 1 be
the estimate of W~_l' Define Zn = Xn - O!n-l and let
(7.185)
1 2 1 - T -1-
Vn - 1 = Vn - 2 + 2gn-l (-Xn-l )Zn-1 + -2Wn-1rn-1Wn-1 (7.187)
The derivative of Vn - 1 is
where an-I> 0 is a small constant. Let Cn-l = C(n-l)O + C(n-l)l, where C(n-l)O
and C(n-l)l > O. Then, equation (7.188) becomes
. . (9n-l ) 2
1 = Vn - 2 - Zn-2 Zn-l
Vn - + Zn-lZn - C(n-l)O +~ Zn-l
gn-l
(7.190)
(7.191)
Because
-
(
C(n-l)O +~
gn-1 ) 2
Zn-l S -
[
C(n-l)O -
g(n-1)d] 2
2 2 Zn-l
gn-l g(n-l)O
".- g(n-l)d 0
c(n-l)O . - C(n-1)O - 2 2 > (7.192)
g(n-l)O
7.4 Pure-feedback Nonlinear Systems 231
(7.193)
Step n
This is the final step. The derivative of Zn = Xn - an-l is
(7.194)
To stabilize the whole system (Zl' ... ,zn), there exists a desired feedback con-
trol
(7.19.5)
where
(7.197)
is computable.
By employing an RBF NN Wn T Sn(Zn) to approximate the following un-
known part ofu* in (7.19.5)
(7.198)
h
were Zn = [X- n , --,,-,'
OCXn-l
.. OCXn-l A.
, -,,--, n e R2n ,u *
'Vn-l JT EHn can be expresse d
UXI UXn-l
(7.199)
where W~ denotes the ideal constant weights, and IEnl :S E~ is the approxima-
tion error with constant E~ > O.
232 Chapter 7. Triangular Nonlinear Systems
(7.200)
(7.201)
The derivative of Vn is
. . Znzn iJnz~ - T -1 ~
Vn = Vn- 1 +- --
gn
-2-
2gn
+ Wn rn Wn
. 2 % 2
= Vn- 1 - Zn-1 Zn - CnZn - - 2 Zn - Zncn
2g n
-T
+Wn Sn(Zn)Zn + Wn
-T
r;;: lW' n (7.203)
(7.204)
where (Tn > 0 is a small constant. Let Cn = Cno + Cn1, where CnO and Cn1 > o.
Then, equation (7.203) becomes
'r _ .
~n - Vn- 1 - Zn-1 Zn - (CnO + -gn2 )Zn2 - 2
Cn1Zn - Zncn - (TnWn Wn
- T '
(7.205)
2gn
By completion of squares, the following inequalities hold
(7.206)
Because -(cnO + in? )z~ ::; -{CnO - 9~d )z~, by choosing CnO such that
gn 2gno
n n II - 112 n .IIW*112 n *2
< _ '"" C* _2
~Jo"'J
_ '"" (Yj
~
Wj
2
+ '""
~
(YJ
2
j + '""
~4c·
l (7.207)
j=l j=l j=l j=l J1
Theorem 7.3. Consider the closed-loop system consisting of the plant (7.131),
the reference model (7.2), the controller (7.200) and the NN we·ight updating
laws (7.144), (7.160), (7.176) and (7.204). Assume there exists sufficiently
large compact sets Oi E R 2 H1,i = 1,'" ,n - 2 and 0i E R 2i ,i = n -l,n
such that Zi E Oi for all t ;::: O. Then, all signals in the closed-loop system
remain bottnded, and the output tracking error y(t) - Yd(t) converges to a small
neighborhood: around zero by appropriately choosing design parameters (56).
[*2
Proof Let ~ := Ln
k=l
(Y
k
IIW*1I2
2
k +L
n
_k_.
k=l 4Ck1
If ckO is chosen such that
'Y
ckO ;::: - - ,
29kO
.
l.e., CkO > 'Y
-2- + 9kd
-2-'
k = 1,'" ,n, I
Wlere . a pOSItive
'Y IS . . cons t an t ,and (Yk
9kO 29 ko
and rk are chosen such that (Yk ;::: 'Y>'max{r;l}, k = 1" .. ,n, then from (7.207)
the following inequality holds
(7.208)
(7.209)
(7.210)
(7.211)
that is,
n
L z~ < 2g* p + 2g*Vn(0)exp( --yt) (7.212)
k=l
which implies that given J1, > ";2g* p, there exists T such that for all t 2: T, the
tracking error satisfies
(7.213)
where J1, is the size of a small residual set which depends on the NN approxi-
mation error €i and controller parameters Ci, O'i and rio It is easily seen that
the increase in the control gain Ci, adaptive gain r i and NN node number l will
result in a better tracking performance. <>
Remark 7.12. Note that in the proposed adaptive NN control scheme, im-
plicit function theorem plays an important role to assert the existence of
the continuous desired virtual controls ai and desired practical control u".
Note also that in general, implicit functions of the desired virtual controls are
very difficult to be solved, even in the case when the system nonlinearities
Ii (Xi, Xi+ d and In (Xi, u) are known functions. With the help of NN, there
is no need to solve the implicit functions for the explicit virtual controls and
the practical controller to cancel the unknown functions in each backstepping
design procedure.
Remark 7.13. In the above analysis, it is clear that the larger the Cil is,
the smaller the 8 is, which implies better tracking performance. However,
increasing Ci will lead to a high gain control scheme. On the other hand,
when giO and gid are known, the design constants Ci,O'i and fi can be chosen
appropriately such that (i) the uniform ultimate boundedness of all the signals
is guaranteed and (ii) the tracking error yet) - Yd(t) to a small residual set can
be achieved.
7.4 Pure-feedback Nonlinear Systems 235
Remark 7.14. The adaptive NN controller (7.200) with adaptation laws (7.144),
(7.160), (7.176) and (7.204) are highly structural, and independent of the com-
plexities of the system nonlinearities. Thus, it can be applied to other similar
plants without repeating the complex controller design procedure for different
system nonlinearities. In addition, such a structural property is particularly
suitable for parallel processing and hardware implementation in practical ap-
plications.
(7.214)
where
_ [- 8an -2 Oan-2 IT
Zn-l - Xn-l! --!;I-'''' '-8--' <Pn-2 E f!n-l C R
2(n-l)
uXI Xn-2
Zn -- [-Xn , an-I, oan-l --!:I-''''
Oan-l
'-!;I--' <Pn-l
IT E f!n C R
2n+1
UXI UXn-l
i-I !;I i-I
'"' Uai-l (_) oai-l. '"' oai-l ; .
<Pi-l = ~ -!;I--/k Xk + -!;I--Xd + ~ -,-Wk, 2::; z::; n (7.215)
k=l UXk UXd k=l OWk
where Ci > 0 are design constants, RBF NNs Wi T 8 i (Zi) are used to approxi-
mate the unknown functions in the controller design, with Wi being the esti-
236 Chapter 7. Triangular Nonlinear Systems
(7.216)
where r i = rr > 0, and (Ji > 0, i = 1,··· , n are positive constant design
parameters.
Theorem 7.4. Consider the closed-loop system consisting of the plant (7.132),
the reference model (7.2), the controller (7.214) and the NN weight ttpdating
laws (7.216). Assume there exists sufficiently large compact sets ni s'uch that
n
Zi E i for all t 2: O. Then, all signals in the closed-loop system remain
bounded, and the outp'ut tracking error y(t) - Yd(t) converges to a small neigh-
borhood around zero by appropriately choosing design parameters [56/.
Proof From the first to the (n - 1)th step, the proof is carried out along
the same lines as that of Theorem 7.3, thus it is omitted here. In the last
step, integra~ Lyapunov function is employed in controller design to avoid the
possible singularity problem caused by gnUi n ) in case the quadratic Lyapunov
function candidate is chosen.
Step n
The derivative of Zn = Xn - a n -1 is
(7.217)
Define Y;;';,ax (xn) = ?in (xn) I Yn (Xn), and a smooth scalar function
(7.219)
(7.220)
7.4 Pure-feedback Nonlinear Systems 237
(7.221)
where
(7.223)
is computable.
Using the fact that
we obtain
(7.224)
where
(7.225)
with
238 Chapter 7. Triangular Nonlinear Systems
To stabilize the whole system (Zl,'" ,Zn), there exists a desired feedback
control
(7.226)
(7.227)
where W: denotes the ideal constant weights, and Ifni ~ f~ is the approxima-
tion error with constant f~ > O.
Since W: is unknown, u· cannot be realized in practice. Let Wn be the
estimate of W:, and the controller be chosen as
where an > 0 is a small constant. Let Cn = CnO + Cnl> where Cnl > 0 and
(7.231)
(7.233)
7.4 Pure-feedback Nonlinear Systems 239
+" n
f:r
a
J
IIW*1I2
2
J
n
f:r
+"_J_
4Cj1
E*2
< _"
n-1
C~Oz2
n
_ C Oz2 _ ,,<Tj
IIWj 112
LJJ nnL 2
j=1 j=1
+"n
f:r
<T.IIW*II2
J
2
J
n E~2
f:r
+"_J_
4Cj1
(7.234)
Let J :="L
n 'a \\W*1I2
k
2
k +"
L
n
_k_.
4Ck1
E*2 {
If ckO is chosen such that c kO 2: --, i.e.,
2gkO
k=1 k=l
CkO > ~+
2gko
9\d , k = 1, . "
2g ko
,n - 1, where { is a positive constant, and ak and
fk are chosen such that ak 2: {'\max{f;;-1}, k = 1" .. ,n, then from (7.234)
the following inequality holds
(7.236)
Therefore, following the same procedure in the proof of Theorem 7.3, it can
be concluded that all the signals in the closed-loop system, including Xi, Wi
(i = 1"" ,n) and u remain bounded. Moreover, the output tracking error
y(t) - Yd(t) converges to a small neighborhood around zero by appropriately
choosing design parameters. <>
240 Chapter 7. Triangular Nonlinear Systems
(7.237)
The reference model is taken as the famous van der Pol oscillator (see, e.g.,
[177])
Xdl = Xd2
Xd2 = -Xdl + ,6(1 - X~1)Xd2
Yd = Xdl (7.238)
As shown in [177], the phase-plane trajectories of the van der Pol oscillator,
starting from an initial state other than (O,O), approach a limit cycle when
,6 > O.
The control objective is to design controller for system ~Sl such that (i)
all the signals in the closed-loop system remain bounded, and (ii) the output
of system ~Sl follows the desired trajectory Yd generated from the van der Pol
oscillator.
Clearly, system ~Sl is in the pure-feedback form (7.131). As system ~Sl
is of third order, the adaptive NN controller is chosen according to (7.200) as
follows
(7.239)
and NN weights WI, W 2 and W3 are updated by (7.144), (7.189) and (7.204)
correspondingly.
7.4 Pure-feedback Nonlinear Systems 241
In practice, the selection of the centers and widths of RBF has a great
influence on the performance of the designed controller. According to [161]'
Gaussian RBF NNs arranged on a regular lattice on R n can uniformly approx-
imate sufficiently smooth functions on closed, bounded subsets. Accordingly, in
the following simulation studies, the centers and widths are chosen on a regular
lattice in the respective compact sets. Specifically, neural networks WTSl(ZI)
contains 64 nodes (i.e., II = 64), with centers j-tl (I = 1, ... ,h) evenly spaced
in [-15, 1.5J x [-10,lOJ x [-15,15]' and widths TJI = 10 (I = 1"" ,h). Neural
networks wi S2(Z2) contains 256 nodes (i.e., h = 256), with centers P.l (I =
1, ... ,1 2 ) evenly spaced in [-15, 15J x [-15,15] x [-15,15] x [-90,90]' and widths
TJI = 20 (I = 1"" ,12)' Neural networks wi S3(Z3) contains 1024 nodes
(i.e., 13 = 1024), with centers j-tl (l = 1,··· , h) evenly spaced in [-15,15]
x [-15, 15J x [-15, 15J x [-30, 30J x [-15, 15J x [-50, 150J, and widths TJI = 30
(I = 1,· .. , h). The design parameters of the above controller are Cl = 5, C2 =
2, C3 = 2, r 1 = r 2 = r3 = diag{2.0}, 0'1 = 0'2 = 0'3 = 0.2. The initial condi-
tions [Xl(0),;J:2(0),X3(0)]T = [1.4,0.3,0.lJ T and [Xdl(0),Xd2(0)V = [1.5,0.2JT.
The initial weights W1 (0) = 0.0, W2(0) = 0.0, W3(0) = 0.0.
Figures 7.5(a)-7.7(a) and Figure 7.8 show the simulation results of applying
controller (7.239) to system ESl (7.237) for tracking desired signal Yd with f3 =
0.2. From Figure l(a), it can be seen that fairly good tracking performance is
obtained. The boundedness of system states X2 and X3, NN weights WI, W2, W3
and control signal u are shown in Figures 7.6 (a), 7.7 (a) and 7.8 respectively.
In comparison, Figures 7.5 (b)-7.7 (b) show the results of applying con-
troller (7.239) to the same system ESl with the adaptation of neural networks
being turned off, i.e., Wi(t) = 0, i = 1,2,3, \it 2: O. From Figure 7.5 (b), it can
be seen that the output tracking performance is much worse compared with
Figure 7.5 (a), though the boundedness of all the signals is achieved, as shown
in Figures 7.5 (b )-7.7 (b). As a matter of fact, it can be proven that when
the neural adaptation is turned off, the proposed controller can achieve local
stability by choosing the control gains Ci, i = 1"" , n large enough. Since
high gain controller is known to be undesirable for its expensive implemen-
tation and excitation of unmodeled dynamics, the comparison of Figures 7 ..5
(a)- 7.7 (a) with Figures 7.5 (b)-7.7 (b) shows that the adaptive NN controller
(7.200) can achieve good control performance for uncertain pure-feedback sys-
tems (7.131). On the other hand, when the neural adaptation is completely
turned off, the controller (7.200) becomes relatively simple. Therefore, there
is a design trade-off between the controller complexity and the improvement
of control performance. Simulation studies for system E2 can be conducted
similarly, and it is omitted here for conciseness.
242 Chapter 7. Triangular Nonlinear Systems
,~
2 "
1
..
Time (Secon::lll) Time (Se.:ondll
.• , ,
,~
" .;\ ~ :: ~ 'I~~~~:~
1\
\,
J,/I
/\t*·i.
2
It • ,,\, - - ~':r:{\!:, ~, ~
I
~J:Y:'r:~::~~'~
..;r......
';E:j'
."
-.,
~'~-2~~'--~'--~'--~1O~~12~~1~'~1~'--'~'~"·
Time (SecCllldil
.~ o 1 " II e 10
TIIIIII!Sc0ftd5I
12 !4 HI l' 20
(7.241)
.. =f",(X1('
:i;3,1.; 1,1] ) ... 'm,
I 1.j -ejl'
X (.'Ii -ejm )
+ gj,i; (X1,(i;-e;d"" ,Xm ,(i;-e;m)Xj,i;+1
1 ::; i j :s Pj - 1
Xj,p; = fj,p; (X, Ul, ... ,Uj-l) + gj,p; (Xl,Pl-l. ... 'Xm,Pm -1)Uj
Yj = Xj,l, 1::; j ::; TIl
where Xj,i;, i j = 1, ... ,Pj are the states of the jth subsystem, X = [XT,Pl' ... ,
x~,p".lT with Xj,ij = [Xj,l,'" ,Xj,ij]T E Ri; (i j = 1,," ,Pj and j = 1"" ,TIl)
representing the state variables of the whole system, Uj E Rand Yj E Rare
the system inputs and outputs, respectively; hi; (-) and gj,i; (-) (i j = 1, ... ,Pj
and j = 1"" ,TIl) are unknown nonlinear smooth functions, j, ij, Pj and TIl
are positive integers, and {!jl is defined as
{!jl ~ Pj - PI (7.242)
Remark 7.15. Note that {ljl is the order difference between the jth and lth
subsystems. The introduction of the notation {ljl (j, l = 1"" ,TIl) is very im-
portant in analyzing the MIMO system (7.241). To help readers to understand
the notation better, let us consider the following two cases:
(i) When j = l, we have (Jjj = 0. Accordingly, Xj,(ij -e;j) = Xj,i;, which is
exactly the state variables of the jth subsystem.
(ii) When j of. l, we have two situations to consider. If i j - (!jl 0, then the :s
corresponding variable vector Xl,(ij-ejz) does not exist, and does not ap-
pear in the functions in (7.241). If i j - (ljl > 0, then XI,(i,-ej,) represents
the maximum state variables of the lth subsystem which are embedded
in the jth subsystem.
!
Remark 7.16. As examples, consider the following two block-triangular MIMO
systems
XI,1 = h,1 (X1,1, X2,1) + gl,l (XI,l. X2,I)Xl,2
:i;1,2 = h,2(X) + gl,2(Xl,l, X2,I)U1
Es1 : X2,1 = h,1(XI,I,X2,1) + g2,1(Xl,I,X2,I)X2,2 (7.243)
X2,2 = h,2(X,Ul) + g2,2(XI,I,X2,I)U2
Yj = Xj,l, j = 1,2
7.,5 MIMO Nonlinear Systems 24,5
Assumption 7.7. The signs of gj,i j (-) are known, and there exist constants
gj,ij2: gJ'iJ > 0, i j = 1, ... ,Pj, j = 1, ... ,Tn such that gj,i j 2: Igj,ij () I 2: gJ'i
- I - J
.
I
The above assumption implies that smooth functions gj,i j () are strictly
either positive or negative. Without losing generality, we shall assume gj,ij >
9j,ij 0 2: flj,i j '
The derivatives of gj,iJ) are given by
a .. (.)
..
gJ"j
. (-x · -
1,('j-{}jI), " ·,m,('j-{}jm)
X· ) L L
=
m ij-{}jl
ax
gJ,>J X. l,k
1=1 k=l l,k
m ij-{}jl a .. (.)
=L L ~;'] [gl,k()Xl,k+1 + Ak(')J,
1=1 k=l l,k
i j = 1, ... , Pj - 1, j = 1, .. , , Tn (7.246)
m Pl-1 a ()
. (-
gj,Pj X1,Pl-1,'"
-)
,Xj,Pm-1 = "L" ' "L" ' a Xl,k
gj,Pj' .
1=1 k=l Xl,k
a. (.)
LL
m Pl-1
= ~:j [gl,k(-)XI,k+1 + fl,k OJ, j = 1, ... , Tn (7.247)
1=1 k=l I,k
Clearly, they only depend on states X. Because fj,ij (.) and gj,i j (.) are assumed
to be smooth functions, they are therefore bounded within the compact set n.
Thus, we have the following assumption.
and prove the stability of the closed-loop MIMO system in a nested iterative
manner.
For the controller design of the jth subsystem of (7.241), an intermediate
desired feedback control a J*, i.J is first shown to exist which possesses some de-
sired stabilizing properties at the recursive ijth step, and then the ijth-order
sub-subsystem of the jth subsystem is stabilized with respect to a Lyapunov
function Vj,ij by the design of a stabilizing function aj,ij' where an RBF NN is
employed to approximate the unknown part in intermediate desired feedback
control aj,ij. The control law Uj for the jth subsystem is designed in the pjth
step.
Step 1
where (!jl = Pj - PI, l = 1,··· ,Tn. If 1 - (!jl :s; 0, then the corresponding
variable vector Xj,(l-I!jl) does not exist.
By viewing Xj,2 as a virtual control input, apparently there exists a desired
feedback control
(7.249)
(7.250)
(7.251)
where Wj~l denotes the ideal constant weights, and !c:j,ll :s; fi,l is the approxi-
mation error with constant fi,l > o.
Since W1,J is unknown, aj,l cannot be realized in practice. Define Zj,2 =
Xj,2 - aj,l, and let
(7.252)
248 Chapter 7. Triangular Nonlinear Systems
Then, we have
(7.253)
(7.254)
2 9j,1 2
= Zj,l Z j,2 - Cj,l Z j,l - 2g2 Zj,l + Zj,lEj,l
),1
(7.255)
where O"j,l > 0 and rWj,l = r~'j'l > 0 are design constants, Wj,l = Wj,l - Wj~l'
Let Cj,l = Cj,lO + Cj,l1, with Cj,lO and Cj,ll > O. Then, equation (7.255)
becomes
(7.257)
. d
Because - (CJ' 10 + -gj,l
2 - ) ZJ'2 1 ::; - ( CJ' 10 - gj,l ) ZJ'2 1 , bY Cl
-n- '
loosmg CJ' 10 suc11 t 1lat
, 2g.' , 2g~ , ,
J,l -j,l
7.5 MIMO Nonlinear Systems 249
d
cj,lO ~ Cj,lO- gj~1 > 0, we have the following inequality
2g'
-), l
O'j,II1 Wj,111 2
2
+ ),1 IIW*j,1 112 + ~
0" 10*2
(7.259)
2 4Cj,1l
(7.261 )
(7.262)
where
(7.263)
is computable.
For the desired feedback control aj,ij' let hj,ij (Zj,i j ) = 9/ij (fj,ij - aj,ij -1)
( 0-
oaj i_l)T
'J , <Pj,ij-l
]T E rlj,ij (7.264)
X m ,(ij-l-l!jm)
250 Chapter 7. Triangular Nonlinear Systems
(7.265)
where WJ*, •..3 denotes the ideal constant weights, and IEJ' ' i J I :::; EJ* i'J is the approx-
I
(7.266)
Then, we have
tion of known information, the NN approximation Wj~j Sj,ij (Zj,ij) of the un-
known function hj,i j (Zj,ij) can be computed by using the minimal number of
NN inputs Zj,ij as given in (7.264). The same idea of choosing the inputs of
NN is also used in the following design steps.
Consider the Lyapunov function candidate
(7.268)
(7.269)
7.5 MIMO Nonlinear Systems 251
(7.270)
where aj,ij > 0 and rWj.ij = r~'j'ij > 0 are design constants, Wj,ij = Wj,ij -
WJ\, .
J
Let Cj,ij = Cj,ijO + Cj,ij1, where Cj,ijO and Cj,ij1 > O. Then, equation (7.269)
becomes
(7.271)
. d
Because - (cJ' i J 0 + 2g. gj,ij ) 2
- 2 - ZJ' " .
< - (CJ' i'O -
gj,ij ) 2
- 2- ZJ' " . ,
b y cl '
loosmg 1
CJ' i'O SUC 1
, . 'J - , J 2g ,J ' J
J,~j -j,ij
d
that cJ\.o ~ cJ' iJO- gj'1,i j > 0, we have the following inequality
'J ' 2rr:.
-J,~j
(7.272)
Step Pj
To stabilize the jth subsystem (Zj,l,' .. ,Zj,Pj)' there exists a desired feedback
control
(7.273)
(7.274)
where
is computable.
For the desired feedback control uj, let hj,pj(Zj,pJ = g/pj (hpJ - O:j,p;-l)
denote the unknown part of uj, where
Zj,p; = [X
T
, Ul," . , Uj-l,
( 8Ci.j,P;-l )T
8
Xl,(Pl-l)
,"',
(8Ci.j,Pj_l)T
8
Xm,(p",-l)
' <Pj,pj-l
1 T
E OJ,Pj (7.275)
(7.276)
where Wj~p; denotes the ideal constant weights, and I€j,pj I ::; €j,Pi is the ap-
proximation error with constant €j,Pi > O.
Since Wj~Pi is unknown, uj cannot be realized in practice. Consider
(7.277)
Then, we have
7.5 MIMO Nonlinear Systems 253
1 2 1- T -1-
VJ·, pj =VJ·, p j -
l+--Z,
2g.. J,Pj 2 J,Pj f.J,Pj W J·"p .
+-W. (7.279)
J,P,
.
Vj,Pj -V.
- J,pj-1
+ z·J,Pj i·J,Pj
gj,Pj
(7.280)
(7.281)
where aj,pj > 0 and fWj,pj = f~j'Pj > 0 are design constants, Wj,Pj = Wj,pj -
W;'Pj'
Let Cj,Pj = Cj,Pjo+Cj,pj1, where Cj,PjO and Cj,pjl > O. Then, equation (7.280)
becomes
(7.282)
(7.283)
Because - ( +
Cj pO
gj,Pj ) 2
- 2 - Zj p' _
< - (Cj,PjO -
gl,Pj ) 2
- b l' 1
22 Zj,Pj' Yc lOosmg Cj,PjO suc 1
, , 2g. ' , g.
l,pj -l,Pj
254 Chapter 7. Triangular Nonlinear Systems
d
that cj,pjo ~ Cj,PjO- g~Pj > 0, we have the following inequality
2-J,Pj
y,
(7.284)
Let
6. ~""
J
Pj
L--
k=l
aj,k
IIW" 112
2
j,k +"" Pj
~
L-- 4c
k=l
.. 2
J,kl
(7.285)
If we choose ..
cj,kO such th a t ..
cJ,kO > -""I-J
, ·l.e., Cj,kO > _""J.
21
gd k
+_J,_
2'" k =
- 2g
-J, k 9-], k g'"'.k
-J,
1, ... , Pj, where 1'j is a positive constant, and choose aj,k and rj,k such that
aj,k 2: 1'F\max{ri.n, k = 1,··· , n, then from (7.284) we have the following
inequality
Pi Pi
" " aj,k
IIW-j,k II" +
Vj,Pj <- " ".. _2
L-- Cj,kO"'j,k - L-- 2
U
UJ
Ji: .
k=l k=l
Pj Pj WT r-1w
< _ "" ~z2 _"" 1'j j,k j,k j,k +6
L--2g. ],k L-- 2 J
k=l -J,k k=l
< _ .. ["" __ z2
- 1'J
Pj
L--
1
2g. k J,k
+"Pj WT r- l-V- k]
L--
j,k
2
1
j,k J, +6.
J
k=l J, k=l
where l' = minbl, ... , 1'm} and 6 = 2::7=1 6j are positive constants.
Therefore, according to Theorem 2.14 in [156], all Zj,ij and Wj,i j (j =
1, ... , m, i j = 1, ... , pj) are uniformly ultimately bounded. Note that because
7.5 MIMO Nonlinear Systems 255
(i) all signals in the closed-loop system remain bounded, and the states X and
the neural weights Wj = [WD,'" ,WDjIT (j = 1"" ,m) eventually
converge to the compact set
(7.288)
(ii) the output tracking errors Yj(t) -Ydj(t) (j = 1, ... ,m) converge to a small
neighborhood around zero by appropriately choosing design parameters.
Proof (i) According to Theorem 2.14 in [156], we know from (7.286), that all
Zj,i jand Wj,i; (j = 1"" ,m, i j = 1"" ,Pj) are uniformly ultimately bounded.
Since Zj,l = Xj,l - Xdj and Xd are bounded, we have that the first state Xj,l of
all the subsystems (j = 1, ... ,m) remain bounded. To prove Xj,2 = Zj,2 + <Xj,l,
j = 1, ... ,m remain bounded, we need to show that <Xj,l is bounded. Note that
because <Xj,l in (7.252) is a function of Zj,l, Zj,l and Wj,l, i.e., a function of
Xl,(l-e;d' .. " Xm,(l-e;",)' Xd and Wj,l, we may not conclude the boundedness
of <Xj,l immediately in case some of the states Xl,(l-e;d' l = 1", . ,m, l =t j
have not been proven bounded yet. As stated in Remark 7.15, (!jl is the order
difference between the jth and lth subsystems, and XI,(ire;l) represents the
state variables of the lth subsystem which are embedded in the jth subsystem.
For clarity of the presentation, we consider the following two cases:
1) All the subsystems are of the same order, i.e., Pl = P2 = ... = Pm·
Accordingly, for the order difference between the jth and lth subsystems,
(!jl = Pj - PI = 0, \lj, l = 1"" ,m, and XI,(i;-e;l) = Xl,i;, l = 1"" ,m,
which means that <Xj,l in (7.252) is a function of Xl,l, " ' , Xm,l, Xd and
Wj,l' Since the first state Xj,l of all the subsystems (j = 1,,,, ,m)
have been proven bounded, we conclude that <Xj,l! j = 1"" ,m remain
bounded, which in turn leads to the boundedness of Xj,2, j = 1"" ,m.
256 Chapter 7. Triangular Nonlinear Systems
Following the same way, we can prove one by one that all O'.j,ij-1 and
j = 1, ... ,Tn, i j = 3, ... ,Pj remain bounded. Therefore, the states
Xj,ij'
X of the interconnected MIMO system (7.241) remain bounded.
2) Not all the subsystems are of the same order, i.e., there exist at least two
subsystems (e.g., the jth and the lth subsystems), such that Pj =f. PI, and
{!jl = Pj - PI =f. O. Note that in this case, we may have i j - {!jl ::; O. As
shown in Remark 7.105, if i j - (!jl ::; 0, then the corresponding variable
vector Xl,(ij-{!jl) does not exist, and does not appear in the functions in
(7.241). If i j -(!jl > 0, then state variables Xl,(ij-Iljl) of the lth subsystem
will be embedded in the jth subsystem.
In this case, we cannot prove the boundedness of O'.j,l, j = 1, ... ,Tn in
one single step as in case 1). We can only prove the boundedness of some
O'.q,l in those qth subsystem(s) first, where q is (or q's are) determined by
(i q = i j = 2,j = 1"" ,Tn in this step)
(7.289)
with
By choosing the qth subsystem(s) in this way, we can pick only those
O'.q,l which are functions of bounded variables Xd, Wq ,l, and X q,l' Ac-
cordingly, we conclude that O'.q,l are bounded. From the bounded ness of
Zq,2, and O'.q,l, we obtain the boundedness of X q ,2. Next, we will proceed
to prove the boundedness of O'.q,i q -1, and subsequently Xq,i q by defini-
tion, where (q, iq) are determined again by (7.289). In fact, according
to the block-triangular structure of system (7.241), checking condition
(7.289) indicates which virtual control O'.q,i q -1 has the minimum variable
embedding from other subsystems and thus is a function of state vari-
ables which have all been proven bounded already. Accordingly, O'.q,i q -1
can be proven bounded next. In this way, we can prove one by one that
all O'.j,i j -1 in (7.266) and subsequently, XJ,i j , j = 1, ... , Tn, i j = 2, ... , Pj
remain bounded. Thus, the states X of the interconnected MIMO system
(7.241) remain bounded.
1) if
(7.291)
(7.292)
2) if
(7.293)
where 0~1 denotes the complimentary set of Osl, then V remains negative
definite until the states X and the neural weights Wj = [Wl,l' ... , WDj]T
(j = 1,···· ,rn) enter and then stay inside 0 81 , i.e.,
(7.294)
Thus, for bounded initial conditions, all signals in the closed-loop system re-
main bounded, and the states X and the neural weights Wj = [WD, ... , WDj]T
(j = 1, ... ,rn) will eventually converge to the compact set Osl.
(ii) Let Wj ~ ojhj > 0, then (7.279) satisfies
(7.295)
(7.296)
(7.297)
that is,
Pj
L Z;'k < 2gjwj + 2gj \Ij,Pj (O)exp( -"Ijt) (7.298)
k=l
258 Chapter 7. Triangular Nonlinear Systems
which implies that given Vj > J2g{cuj, there exists T such that for all t ~ T,
the tracking error satisfies
(7.299)
where Vj is the size of a small residual set which depends on the NN approx-
imation error €j,ij and controller parameters Cj,ij' (Fj,ij and rj,ij. It is easily
seen that the increase in the control gain Cj,ij' adaptation gain r j,ij and NN
node number l will result in a better tracking performance. <>
Remark 7.18. For MIMO system (7.241) with many subsystems, it might
become very involved when trying to find out which state to be proven bounded
next. By introducing the checking condition (7.289), the determination of q
and i q , and subsequently, aq,iq-I and Xq,i q can be easily done. To help readers
understand the above stability analysis for the whole closed-loop system better,
let us consider systems ~Sl (7.243) and ~S2 (7.244) as examples for cases 1)
and 2), respectively.
still need to prove the boundedness of the state variable X2,3 first. By
using (7.289) again, we have (q,i q ) = (1,2) and (q,i q ) = (2,4). At this
time, we can prove the bounded ness of al,l (7.252) and a2,3 (7.266), and
subsequently, XI,2 and X2,4 simultaneously. Therefore, the states X of
the overall system (7.244) are proven bounded. The sequence of the state
variables being proven bounded is shown as follows
Remark 7.19. In the above analysis, it can be seen from (7.286)-(7.288) that
the size of [lsI depends on Wj,b Ej,b and all design parameters. Since there
is no analytical result in the NN literature to quantify the relationship of the
NN node numbers lj, the ideal neural network weights WJ*, i.' J
and the bounding
approximation error EJ": , i"J an explicit expression of the stability condition is not
available at present. However, it is clear that i) increasing Cj,ijO might lead
to larger ,j, and subsequently larger " and increasing Cj,ijl will reduce OJ,ij'
and subsequently 0, thus, increasing Cj,ij will lead to smaller [lSI; ii) decreasing
aj,ij will help to reduce OJ,ij' and increasing the NN node number lj will help
to reduce EJ":, i J., both of which will help to reduce the size of [lSI.
Remark 7.20. The adaptive neural controller (7.277) with adaptation laws
(7.256) is highly structural, and independent of the complexities of the system
nonlinearities. Thus, it can be applied to other similar plants without repeating
the complex controller design procedure for different system nonlinearities. In
addition, such a structural property is particularly suitable for parallel process-
ing and hardware implementation in practical applications.
Remark 7.21. Using the methodologies in this section and Section 7.3, adap-
tive neural controller can be designed without any difficulty for an even larger
class of nonlinear systems in the following form
XI,il = h,il (XI,(il-{lll)" .. , Xm,(i! -(lim»)
+ eLl FI,il (XI,(il -(lll)' ... , Xm,(il-i!lm»)
+ C1,il(XI,(il-{lll)"" ,Xm,(il-{llm») + gl,ilXI,il+1
1 :s: i l :s: PI - 1
XI,Pl = h,Pl (X) + eLl FI,Pl (X) + CI,Pl (X) + gl,Pl UI
where Xj,ij' i j = 1, ... , Pj are the states of the jth subsystem, X = [Xf,Pl' ... ,
x~"p",jT with Xj,i j = [Xj,l, ... , Xj,ij jT E Rij (i j = 1,· .. , Pj and j = 1, ... ,m)
representing the state variables of the whole system, Uj E Rand Yj E R
are the system inputs and outputs, respectively; hi j (-) are unknown smooth
functions, Fj,ij () and Gj,ij () are known smooth nonlinear functions, gj,ij and
()j,ij are unknown constant parameters, j, ij, Pj and m are positive integers,
and (!jl ~ Pj - PI is the order difference between the jth and lth subsystems,
as discussed in Remark 7.15.
System E2 (7.300) has unknown nonlinear functions hij (.), parametric un-
certainties gj,ij' ()j,ij) and known nonlinear functions Fj,ij (.) and Gj,ij O. This
problem formulation is very general in the sense that any a priori information
represented by Fj,ij (-) and Gj,i j 0, which may be available through the laws of
physics, properties of materials, or any identification methods, can be employed
and incorporated into the controller design.
7.6 Conclusion
In this chapter we have shown how to design adaptive neural controllers for
several classes of nonlinear systems in triangular forms. Three cases of SISO
nonlinear systems were studied: (i) a special nonlinear strict-feedback system
where the affine term gn(xn-d is independent of state Xn; (ii) partially known
nonlinear strict-feedback systems, and (iii) nonlinear pure-feedback system with
unknown nonlinear functions.
For special nonlinear strict-feedback system, an adaptive neural control
scheme was proposed by using its nice property of gn(Xn-l) being independent
of state Xn in Section 7.2, and the controller singularity problem is overcome
completely without using integral Lyapunov function as done in Chapter 5. In
Section 7.3, by exploiting the a priori information of the system under study,
a direct adaptive NN controller was presented for a class of partially known
nonlinear system in strict-feedback form, which can simplifY the design meth-
ods of Section 7.2 and Chapter 5. Using implicit function theorem, adaptive
neural control problem was first solved for pure-feedback nonlinear systems in
Section 7.4. The extension from SISO nonlinear systems in triangular forms
to MIMO nonlinear systems in block-triangular forms have also been consid-
ered in Section 7.5. All the developed adaptive neural control schemes in this
chapter achieve semi-global uniform ultimate boundedness of all the signals in
the closed loop. Moreover, the output of the system was proven to converge
to a small neighborhood of the desired trajectory. The control performance
of the closed-loop system can be guaranteed by suitably choosing the design
parameters.
Chapter 8
Conclusion
8.1 Conclusion
Stability and transient performance are two important aspects in adaptive
neural network control systems. In this book, constructive and systematic
methods have been developed for stable adaptive NN control of nonlinear
dynamic systems, while transient performances of these control schemes are
rigorously analyzed. The problems mentioned in Chapter 1 associated with
parameter drift, controllability problem, transient behaviour, choice of design
parameters, and effects of initial conditions and reference signals, have been
fully addressed in the controller design chapters after the relevant mathematics
presented in Chapter 2, and the comprehensive treatment of neural network
tunning algorithms in Chapter 3.
In Chapter 4, for a class of nonlinear systems in a Brunovsky controller
form, a regionally stable adaptive controller was firstly proposed. The ex-
plicit conditions on the initial states, design parameters and allowed reference
signals have been investigated for guaranteeing stability of the neural control
systems. The system transient performance was analytically quantified by the
mean square and Loo error criteria. Then, a semi-global adaptive controller
has been developed by combining VSC technique and multilayer NNs. When
the bounds of the system nonlinearities are available, a globally stable neural
control approach is readily available.
The novel integral Lyapunov function in Chapter 5 plays an important role
in the adaptive control design of the book. Such an integral Lyapunov function
was used to construct new adaptive NN control structures and adaptive laws
which completely solve the control singularity problem existing in feedback lin-
earization adaptive control approaches. This Lyapunov function has also been
successfully applied for constructing stable adaptive controller for a class of
and integral, maybe blend together for the best result. For example, at one
step, one type of Lyapunov functions can be used, at the next step, another
type can be considered by exploiting the properties of a physical system. It is
believed that pursuing different ways to utilize the structure properties of dif-
ferent systems would be beneficial in deriving stable control schemes for larger
class of multivariable nonlinear systems. Though only neural networks are used
as function approximators in the book, other function approximators can also
be used without any difficulty, they include polynomials, splines, fuzzy systems
and wavelet networks as explained earlier in the book. Interested readers may
directly extend the results in their studies by using other approximators for
different performance.
[l1J L. Blum, F. Cucker, M. Shub, and S. Smale. Complexity and Real Com-
putation. Springer-Verlag, New York, 1998.
[20J T. Y. Chai and T. Zhang. A new model reference robust adaptive con-
troller in the presence of unmodeled dynamics and bounded disturbances.
Automatica, 30:865-869, 1994.
[33] X. Dong, G. Chen, and L. Chen. Adaptive control of the uncertain duffing
oscillator. Int. J. Bifurcation and Chaos, 7:1651-1658,1997.
[40] J. Farrell and T. Baker. On the effects of the training sample density in
passive learning control. In Proc. American Control Conf., pages 872-
877, Seattle, WA, June 1995.
[49] S. S. Ge, C. C. Hang, and T. Zhang. A direct method for robust adaptive
nonlinear control with guaranteed transient performance. Systems and
Control Letters, 37:275-284, 1999.
[51] S. S. Ge, C. C. Hang, and T. Zhang. Stable adaptive control for nonlinear
multivariable systems with a triangular control structure. IEEE Trans.
on Automatic Control, 45:1221-1225, June, 2000.
[58] F. Girosi and T. Poggio. Networks and the best approximation property.
Artificial Intelligence Lab. Memo, Oct., 1989.
[146] T. Poggio and F. Girosi. Networks for approximation and learning. Proc.
of IEEE, 78:1481-1497, 1990.
[156J Z. Qu. Robust control of nonlinear uncertain systems. John Wiley and
Sons, New York, 1998.
[192] T. Zhang, S. S. Ge, and C. C. Hang. Adaptive neural network control for
strict-feedback nonlinear systems using backstepping design. Alttomatica,
36:1835-1846,2000.