(The Springer International Series on Asian Studies in Computer and Information Science 13) Shuzhi S. Ge, Chang C. Hang, Tong H. Lee, Tao Zhang (auth.) - Stable Adaptive Neural Network Control-Springe.pdf

STABLE ADAPTIVE NEURAL
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:
FULLY TUNED RADIAL BASIS FUNCTION NEURAL NETWORKS FOR FLIGHT

CONTROL by N. Sundararajan, P. Saratchandran and Yan Li; ISBN: 0-7923-7518-1
NONLINEAR CONTROL SYSTEMS AND POWER SYSTEM DYNAMICS by

Qiang Lu, Yuanzhang Sun, Shengwei Mei; ISBN: 0-7923-7312-X
DATA MANAGEMENT AND INTERNET COMPUTING FOR IMAGE/PATTERN

ANALYSIS David Zhang, Xiobo Li and Zhiyong Liu; ISBN: 0-7923-7456-8
COMMON WAVEFORM ANALYSIS: A New and Practical Generalization of Fourier

Analysis, by Yuchuan Wei and Qishan Zhang; ISBN: 0-7923-7905-5
DOMAIN MODELING-BASED SOFTWARE ENGINEERING: A Formal Approach, by

Ruqian Lu and Zhi Jin; ISBN: 0-7923-7889-X
AUTOMATED BIOMETRICS: Technologies and Systems, by David D. Zhang;

ISBN: 0-7923-7856-3
FUZZY LOGIC AND SOFT COMPUTING, by Guoqing Chen, Mingsheng Ying

Kai-Yuan Cai; ISBN: 0-7923-8650-7
INTELLIGENT BUILDING SYSTEMS, by Albert Ting-pat So, Wai Lok Chan;

ISBN: 0-7923-8491-1
PERFORMANCE EV ALUA TION, PREDICTION AND VISUALIZATION OF

PARALLEL SYSTEMS by Xingfu Wu; ISBN: 0-7923-8462-8
ROBUST MODEL-BASED FAULT DIAGNOSIS FOR DYNAMIC SYSTEMS by Jie

Chen and Ron J. Patton; ISBN: 0-7923-8411-3
STABLE ADAPTIVE NEURAL
NETWORK CONTROL
by
S. S. Ge
C. C. Hang
T. H. Lee
T. Zhang
Department 0/Electrical Engineering

National University a/Singapore
SPRINGER SCIENCE+BUSINESS MEDIA, LLC

Library of Congress Cataloging-in-Publication Data
Stable adaptive neural network control/by S. S. Ge ... ret al.J.

p. cm.-- (The Kluwer international series on Asian studies in computer and
information science; 13)
Includes bibliographical references and index.
ISBN 978-1-4419-4932-5 ISBN 978-1-4757-6577-9 (eBook)
DOI 10.1007/978-1-4757-6577-9
1. Adaptive control systems. 2. Neural networks (Computer science) 1. Ge, S.S.
(Shuzhi S.) II. Series.
TJ217.S7362001
629.8'3--dc21 2001050337
Copyright © 2002 by Springer Science+Business· Media New York

Originally published by Kluwer Academic Publishers in 2002
Softcover reprint of the hardcover 1st edition 2002
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system
or transmitted in any form or by any means, mechanical, photo-copying, recording, or otherwise,
without the prior written permission of the publisher, Springer Science+Business Media, LLC.
Printed on acidjree paper.

To
Jinlan, Sao Chin, Ruth, Yuan
and
our loved ones

SERIES EDITOR'S
ACKNOWLEDGMENTS
I am pleased to acknowledge the assistance to the editorial work by

Beijing University of Aeronautics and Astronautics and the National
Natural Science Foundation of China
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
3 Neural Networks and Function Approximation 27

3.1 Introduction. 27
3.2 Function Approximation . 27
3.3 Linearly Parametrized Neural Networks 29
3.4 Non-linearly Parametrized Networks 35
3.5 Neural Networks for Control Applications 44
3.6 Conclusion 46
x Contents
4 SISO Nonlinear Systems 47

4.1 Introduction.............. 47
4.2 NN Control with Regional Stability. 49
4.2.1 Desired Feedback Control . . 49
4.2.2 HONN Controller Design Based on (4.7) . 51
4.2.3 MNN Control Based on (4.10) . . . . . . 59
4.3 VSC - Semi-Global Stability . . . . . . . . . . 70
4.3.1 VSC-based Adaptive NN Control Design. 73
4.3.2 Elimination for Controller Chattering 77
4.3.3 Simulation Study. 79
4.4 Conclusion . . . . . . . 79
5 ILF for Adaptive Control 81

5.1 Introduction............. 81
5.2 Matching SISO Nonlinear Systems 82
5.2.1 Integral Lyapunov Function 83
5.2.2 Choice of Weighting Function Q(x) 83
5.2.3 Adaptive NN Control Based on DFCs 92
.5.3 Backstepping Adaptive NN Design . . . . . . 10.5
.5.3.1 Adaptive Design for a First-order System 108
.5.3.2 Design for nth-order Systems . . . . . . . 112
.5.3.3 Controller Design with Reduced Knowledge 121
5.3.4 Simulation Studies . . . . . . . . . 123
.5.4 NN Control for MIMO Nonlinear Systems . . . . . 127
5.4.1 System Description. . . . . . . . . . . . . . 128
5.4.2 Lyapunov Function Design and Control Structure 130
.5.4.3 Adaptive MIMO Control Using MNNs 132
.5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6 Non-affine Nonlinear Systems 139

6.1 Introduction................. 139
6.2 System Description and Properties . . . . 140
6.2.1 Implicit Desired Feedback Control 142
6.2.2 High-gain Observer. . . . . 146
6.3 Controller Design Based on LPNN 147
6.3.1 State Feedback Control . 149
6.3.2 Output Feedback Control . 153
6.3.3 Simulation Study. . . . . . 159
6.4 Controller Design Based on MNN . 160
6.4.1 State Feedback Control . 163
6.4.2 Output Feedback Control 168
6.4.3 Application to CSTR .. 176
Contents xi
6.5 Conclusion ......... . 182
7 Triangular Nonlinear Systems 183

7.1 Introduction................. 183
7.2 Special Systems in Strict-Feedback Form. 185
7.2.1 Direct Adaptive NN Control 188
7.2.2 Simulation studies . . . . . . . . 198
7.3 Partially Known Nonlinear Systems .. 201
7.3.1 Adaptive Neural Control Design 203
7.3.2 Numerical Simulation . . . . . . 215
7.4 Pure-feedback Nonlinear Systems . . . . 217
7.4.1 Direct Adaptive NN Control for El . 220
7.4.2 Direct Adaptive NN Control for E2 . 235
7.4.3 Simulation studies 240
7.5 MIMO Nonlinear Systems 242
7.6 Conclusion . . . . . . . . 260
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
proven through Lyapunov stability analysis. Transient performance of adaptive

neural control systems is also essential for the control applications. It has been
shown that poor initial conditions may result in unacceptably poor transient
behaviour in adaptive systems. It is highly desirable for a control engineer
to have an estimate of the transient performance before a neural network con-
troller is put into practice. In this book, for different neural control designs, the
effects of controller parameters, initial conditions and reference signals on sys-
tem stability and control performance are investigated for providing valuable
insights into performance improvement and design trade-offs.
The main special features of this book are as follows: (i) singularity-free
controllers are presented for a class of nonlinear SISO systems by exploiting
its property of 8b(x}j8xn = 0, (ii) through the introduction of integral Lya-
punov function (ILF) candidates, novel design methodologies are introduced
to solve the control problems for a wide class of nonlinear adaptive control
problems without encountering the controller singularity problem, (iii) besides
affine nonlinear systems, controller design for nonaffine nonlinear systems are
also be treated using implicit function theorem, neural network approxima-
tion and adaptive control techniques, and (iv) most of the results presented
are analytical with repeatable design algorithms because closed-loop stability
is proven mathematically, and detailed performance analysis of the proposed
adaptive neural controllers is performed rigorously.
The book starts with a brief introduction of adaptive control, neural network
control, and the possible instability mechanisms in adaptive neural control
systems in Chapter 1.
For completeness, Chapter 2 gives a brief summary of the basic mathe-
matical tools of norms, stability theorems, implicit function and mean value
theorems and properties of integrations, which are used for controller design,
stability and performance analysis in the subsequent chapters of the book.
Chapter 3 presents two classes of function approximators, namely, linearly
parameterized neural networks (LPNN) and non-linearly parameterized (multi-
layer) neural networks (MNN) for function approximation. Main properties of
such two kinds of networks are discussed for control applications. In addition,
their advantages and shortcomings are studied when they are used in system
identification and adaptive control design.
In Chapter 4, a regionally stable NN design is firstly proposed for nonlin-
ear systems in a Brunovsky form. The control performance of the systems is
analytically quantified by the mean square criterion and Loo criterion. Then, a
semi-global NN controller is provided by using variable structure control tech-
nique. Furthermore, the transient behaviour of the adaptive neural system has
been investigated, and several methods are provided for improving the system
response.
In Chapter 5, by introducing an integral Lyapunov function, adaptive NN
Preface xv
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.
Shuzhi S. Ge, Chang C. Hang, Tong H. Lee and Tao Zhang

Department of Electrical fj Computer Engineering
The National University of Singapore
Chapter 1
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-
S. S. Ge et al., Stable Adaptive Neural Network Control

© Springer Science+Business Media New York 2002
2 Chapter 1. Introduction
work control only. For other linear-in-the-parameters function approximators

such as polynomials [151]' splines [17], fuzzy systems [179J and wavelet net-
works [190]' the controllers presented in this book using linear-in-the-parameter
NN can be replaced by these function approximators without any difficulty, thus
ommitted.
1.2 Adaptive Control

Research in adaptive control has a long history of intense activities that involves
rigorous problem formulation, stability proof, robustness design, performance
analysis and applications. The advances in stability theory and the progress
of control theory in the 1960s improved the understanding of adaptive control
and contributed to a strong interest in this field. By the early 1980's, several
adaptive approaches have been proven to provide stable operation and asymp-
totic tracking [35,36,62,108,132, 137J. The adaptive control problem since then,
was rigorously formulated and several leading researchers have laid the theo-
retical foundations for many basic adaptive schemes [5,61,63,73,74,137,164].
In the mid 1980s, research of adaptive control mainly focused on robustness
problem in the presence of unmodeled dynamics and/or bounded disturbances.
A number of redesigns and modifications were proposed and analyzed to im-
prove the robustness of the adaptive controllers, e.g., applying normalization
techniques in controller design [75,101,154] and the modification of adaptation
laws using projection method [61,164]' dead zone modifications [35,101,145]'
(-modification [134]' and a-modification [20,73].
In spite of the development of stability and robustness in adaptive control,
transient performance of the closed-loop systems is not clear, which is actually
more important in practical applications. It has been shown that poor ini-
tial parameter estimates may result in unacceptably poor transient behaviour.
Performance improvement has been considered in a number of ways, including
the use of dominant rich reference input [75], multiple models [136, 139], fast
adaptation [185], and the use of compensators [31, 172J. Research reveals that
a trade-off between performance and robustness has to be made for adaptive
control systems. Recently, motivated from nonlinear system theory [91]' non-
linear design method has been applied to linear adaptive control [105,193,194]
and a new design approach has been developed to achieve better transient per-
formance than the traditional schemes. The design of adaptive controller with
improved transient performance is a current research topic.
Interest in adaptive control of nonlinear systems was stimulated by the
development of the differential-geometric theory of nonlinear feedback con-
trol [771. Motivated by the early success of adaptive control of linear systems,
the extension to nonlinear systems has been investigated from the end of 1980s
1.3 Neural Network Control 3
to early 1990s. Several important results were presented based on feedback

linearization techniques [18,90,153,165,174,175]. The robustness problem of
adaptive nonlinear control has also been considered for nonlinear systems with
unmodeled dynamics and/or external disturbances [2,87,90,174]. However, in
order to obtain global stability, some restrictions on the plants had to be made
such as matching condition, extended matching condition, or growth condi-
tions on system nonlinearities [18,90,133,174]. The extended matching barrier
was finally broken in [91,103] by introducing a novel recursive design proce-
dure called adaptive backstepping, and globally stable and asymptotic track-
ing adaptive controllers were developed for parametric strict-feedback systems.
One of the distinguished features of adaptive backstepping design is that for
a class of strict-feedback systems, not only global stability can be achieved,
but also the transient performance is guaranteed and explicitly analyzed [104].
Recently, robust adaptive controllers have been studied for a class of strict-
feedback systems by combining robust backstepping design with robust control
strategy [72,84,114,129,143,147,182,184]' which guaranteed global uniform
ultimate boundedness in the presence of parametric uncertainties or unknown
nonlinear functions. Adaptive output feedback control for nonlinear systems
has also been studied in [2,82,83,93,95,126,127]. The current research of
adaptive nonlinear control focuses on broadening the class of the nonlinear
systems being handled, studying the convergence of the estimated parame-
ters [39,102,115,119,120], and improving performance and robustness of the
adaptive systems [181,195].
In all of the above-mentioned adaptive control techniques, a key assump-
tion is linear parameterization, i.e, nonlinearities of the studied plants are in
the forms of linear in the parameters. For many practical systems, nonlin-
ear parametrization is common, e.g., fermentation processes [14], bio-reactor
processes [13,15] and friction dynamics. Several results were presented for dif-
ferent kinds of nonlinearly parametrized systems [3,12-16,43,124]. The works
in [14,16] provide an interesting design approach by appropriately parame-
trizing the nonlinearly parametrized plant and choosing a suitable Lyapunov
function. Recently, a new adaptive control scheme was investigated for nonlin-
ear systems with convex/concave parametrization through the introduction of
a tuning function and an adaptive law based on a min-max strategy [3,124].
1.3 Neural Network Control

Since McCulloch and Pitts [130] introduced the idea of studying the computa-
tional abilities of networks composed of simple models of neurons in the 1940s,
neural network techniques have undergone great developments and have been
successfully applied in many fields such as learning, pattern recognition, signal
processing, modelling and system control. The approximation abilities of neural

networks have been proven in many research works [7-9,25,26,30,44,58,69,76,
88,97,131,141,144,146,160]. Their major advantages of highly parallel struc-
ture, learning ability, nonlinear function approximation, fault tolerance, and
efficient analog VLSI implementation for real-time applications, greatly moti-
vate the usage of neural networks in nonlinear system identification and control.
The early works of neural network applications for control design were reported
in [1,183]. The popularization of the backpropagation algorithm [159] in the late
1980s greatly boosted the development of neural control, and many neural con-
trol approaches have been developed [17,21,65,70,85,86,109,117,138-140,155].
Most early works on neural control describe innovative ideas and demonstrate
neural controllers through simulation or by particular experimental examples,
but fall short of analytical analysis on stability, robustness and convergence of
the closed-loop neural control systems. The theoretical difficulty raised mainly
from the nonlinearly parametrized networks used in the approximation. The
analytical results obtained in [22-24, 121J showed that using multilayer neural
networks as function approximators, stability and convergence of the resulting
control systems can be guaranteed when the initial network weights are chosen
sufficiently close to the ideal weights. This implies that for achieving a stable
neural control system using the gradient learning algorithms such as back-
propagation (BP), sufficient off-line training must be performed before neural
network controllers are put into the systems.
In order to avoid the above difficulties in constructing stable neural sys-
tem, Lyapunov stability theory has been applied in developing control struc-
ture and deriving network weight updating laws. Several research groups have
involved in the developments of stable adaptive neural network control tech-
niques. Specifically, Polycarpou and Ioannou [147,148,152] provided an unified
framework for identification and control of nonlinear dynamic systems, in which
the parametric method of both adaptive nonlinear control and adaptive linear
control theory can be applied to perform the stability analysis. Sanner and Slo-
tine [161-163J have done in-depth treatment in the approximation of Gaussian
radial basis function (RBF) networks and the stability theory to adaptive con-
trol using sliding mode control design. Multilayer NNs based control methods
have been successfully applied to robotic control for achieving stable adaptive
NN systems [29,42,80,98,106,110-112, 186J. Their design approach has also
been extended to discrete-time nonlinear systems [79]. Through introducing
the Ge-Lee operator for ease of stability analysis and presentation, a system-
atic and coherent treatments of the common problems in robot control is given
in [52]. Other stable approximation-based schemes include Lyapunov-based
training design [188,189]' and fuzzy control techniques [34,171,179,180].
Due to the lack of suitable analytical tool and control methodology, the
problem of adaptive control for non-affine nonlinear systems has not been fully
1.4 Instability Mechanisms in Adaptive Neural Control Systems 5
investigated in the literature. In an effort to solve the problem, several re-

searchers have suggested the use of neural networks to approximate the inverse
of the systems [109,116, 155J. The main idea is that for a system with finite
relative degree, the mapping between the system input and the system output
is one-to-one, thus allowing the construction of an "inverse operator" of the
nonlinear system using NNs. Using the implicit function theory, the NN con-
trol methods [59,60) have been developed to emulate the implicit function of
the unknown dynamical relationship between the input and output of the sys-
tem. The results show that for achieving adaptive tracking control, uniformly
persistent excitation (PE) condition is required. At the present stage, for 110n-
affine unknown nonlinear systems, adaptive control approach with guaranteed
stability is not available in the literature.
1.4 Instability Mechanisms in Adaptive Neural

Control Systems
In spite of the development of neural network control techniques and their
successful applications, there still remain several fundamental problems about
stability, robustness, and performance issues of neural network adaptive sys-
tems yet to be further investigated. To begin with, the instability mechanisms
in neural control design are firstly presented for motivating the research of this
work, then the objectives and contributions of the book are provided.
The emergence of Lyapunov-based neural control design makes it possible to
use the available adaptive linear/nonlinear control theories, and therefore share
some similar features in parameter estimation, robustness, stability and perfor-
mance. However, in comparison with the available parametric adaptive control
approaches, there are two major differences between model-based control and
neural network control, (i) the approximation property of NNs is guaranteed
only over a compact, and (ii) in general, function approximation errors always
exist. These special features result in that neural network control has its own
additional stability and robustness problems that need to be addressed. Some
of the instability mechanisms of adaptive neural control have been studied by
Polycarpou and Ioannou [148J. We will discuss this problem in the following
five aspects:
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
designs by introducing projection algorithm [148,171,179] and modified

adaptive laws [29,42,78-80,98,106,110-112,147,186].
2. Controllability Problem
Up to now, most available neural network control is based on feedback
linearization techniques. For a nonlinear system in the form of
i; = f(x) + g(x)u (1.1)
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.
3. Transient Behaviour Problem

A common feature of adaptive control systems is that a bad transient
performance may happen in the early stages of adaptation when large
initial parameter estimation errors exist. Such a bad transient behaviour
may cause additional problems for neural network systems. In most of
NN-based methods, controllers and adaptive algorithms are derived un-
der the assumption that states x(t) remain in some compact set n for
all time, where n represents the region that the function approximations
of f(x) and g(x) hold. If x(t) leaves the region n due to the bad tran-
sient behaviour, then the modelling error may be large, which could force
x(t) even further out from n, eventually causing instability of the neural
adaptive systems.
In order to solve the above transient behaviour problem, two methods
have been applied in the literature. The first is to the use of an off-
line identification procedure so that during the on-line control phase, the
estimates j(x, Of) and flex, Og) represent good approximations of f(x) and
g(x), respectively. Another commonly used method is the introduction of
1.4 Instability Mechanisms in Adaptive Neural Control Systems 7
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].
4. Initial conditions and reference signals

As discussed above, bad estimates of initial weights may cause large track-
ing error and lead to an unstable closed-loop system. For on-line adaptive
NN control, applicable conditions on the allowed initial estimation errors
are not available in the literature. The results in [22-24,121] show that
the initial values of the NN weights are required sufficiently close to the
ideal values [89].
The compact set on which neural networks work, guarantees the function
approximation accuracy. This property imposes a limit on the initial
condition of system states during the NN controller development. For
a designed NN controller, the allowable maximum operation range for
initial system states is determined. If the initial states are out some
given set, the NN approximation is not valid, and the system might be
unstable. The investigation on such a condition has not been attentively
performed in the current NN research. Similarly, suitable conditions for
desired reference signals should be derived for guaranteeing stable neural
control systems.
5. Choice of design parameters

For most of NN control schemes, design parameters not only affect the
transient performance but also determine the system stability. From an
implementation point of view, a good control design should provide both
the range of controller parameters and its relationship with robustness
and performance. Without a suitable guidance for the choice of design
parameters, a control scheme may result in either bad transient perfor-
mance or an unstable closed-loop system in practical applications.
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.
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.
The third objective of this book is to solve the above-mentioned control

problems existing in adaptive feedback linearization strategy. By introducing
a novel kind of integral Lyapunov functions, a new NN control structure is
developed, in which the cancellation of the nonlinearity g(x} is not needed,
therefore, completely avoid the control singularity problem.
The fourth objective is to enlarge the class of nonlinear systems handled by

Lyapunov-based adaptive control techniques. The newly proposed Lyapunov
function is used to solve an adaptive control problem for a class of nonlinearly
parametrized plants. By combining neural networks, integral Lyapunov func-
tion, and adaptive backstepping design, stable neural control is achieved for
strict-feedback nonlinear systems. The proposed techniques are also shown to
be efficient in adaptive NN control for the multivariable nonlinear systems in
triangular form.
The final objective is to develop neural adaptive approaches for non-affine

nonlinear systems. The motivation of this design lies in the facts that many
practical systems such as chemical reactions, pH neutralization and distillation
columns are inherently nonlinear with input variables appearing nonlinearly.
The available physical laws are not sufficient in deriving a relative represen-
tative model for system control. Even if there are some rough models of the
systems, the model parameters may not be precisely known in most applica-
tions. From a practical point of view, adaptive control of non-affine nonlinear
systems is an important and challenging problem.
1.5 Outline of the Book 9
1.5 Outline of the Book

The book starts with a brief introduction of adaptive control, neural network
control, and the possible instability mechanisms in adaptive neural control
systems in Chapter 1. For completeness, Chapter 2 gives a brief summary of
the basic mathematical tools of norms, stability theorems, implicit function
and mean value theorems and properties of integrations, which are used for
controller design, stability and performance analysis in the subsequent chapters
of the book.
Chapter 3 presents two classes of function approximators, namely, linearly
parametrized neural networks (LPNN) and non-linearly parametrized (multi-
layer) neural networks (MNN) for function approximation. Main properties
and common used tunning methods of such two kinds of networks are dis-
cussed for control application. In addition, their advantages and shortcomings
are studied when they are used in system identification and adaptive control
design.
In Chapter 4, a regionally stable NN design is firstly proposed for nonlin-
ear systems in a Brunovsky controller form. The control performance of the
systems is analytically quantified by the mean square criterion and Leo crite-
rion. Then, a semi-global NN controller is provided by using variable structure
control technique. The transient behaviour of the adaptive neural system are
also investigated, and several methods are provided for improving the system
response.
In Chapter 5, by introducing an integral Lyapunov function, adaptive NN
controller is firstly developed for a class of SISO nonlinear systems, which
completely avoid the possible controller singularity problem. Then, adaptive
control design is developed for strict-feedback nonlinear systems through com-
bining multilayer NNs with backstepping technique. It is proven that under
certain conditions, the semi-globally uniformly ultimate boundedness is achiev-
able for the closed-loop adaptive systems. Finally, adaptive NN Control is pre-
sented for a class of MIMO nonlinear systems having triangular structure in
control inputs using multi-layer neural networks.
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. Closed-loop stability is established by Lyapunov's
stability theory, and their effectiveness are verified by numerical simulation.
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 independent
of X n . This nice properties can be exploited for better controller design. Sec-
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 Mathematical Preliminaries

Stability analysis is all about the magnitudes of all the closed-loop signals of
a system, and some metric or norms are needed to measure them. Controller
design is carried out using those norms that are easy to manipulate mathe-
matically, rather than they truly represent real signals and their properties.
Despite this reservation, all norms adopted do have some physical significance.

12 Chapter 2. Mathematical Preliminaries
2.2.1 Norms for Vectors and Signals

Norms of Vectors
The class of Lp-norms is defined by
for 1 :S p < 00 (2.1)
(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)
Induced Norm of Matrices

Definition 2.1. For an m x n matrix A, the induced p-norm of A is defined
by [94]
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)
For p = 1,2,00, we have the corresponding induced norms as follows:

m
IIAIII = max L laijl (column sum)

J i=1
IIAII2 = m~x JAi(AT A)

m
IIAIIe<> = m~x L laijl (row sum)

j=l
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]
II A IIJ,.= L a~ = tr(A T A) (2,6)

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
N(X,E) = {z E R n Illz -xii < E} (2.7)
such that N(x, E) C S.

A set is closed if and only if its complement in R n is open; bottnded if there
is r > 0 such that Ilxll < r for all XES; and compact if it is closed and
bounded; convex if, for every x, yES, and every real number e, 0 < e < 1,
the point ex + (1 - O)y E S [94J.
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
IIx - yll < 8 ===> IIf(x) - f(y)11 < E (2.8)
A function f is continuous on a set of S if it is continuous at every point

of S, and it is uniformly continuous on S if given E > 0, there is beE) > 0
(dependent only on E), such that the inequality holds for all x, yES.
Definition 2.3. Let U be an open subset of Rn+1. A mapping f(x) : U ---4 R
is said to be Lipschitz on U if there exists a positive constant L 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/,
Differentiable Function
A function 1: R --4 R is said to be differentiable at a point x if the limit
j(x) = lim l(x + t5x) - l(x) (2.9)

ox-->a t5x
exists. A function 1: R n --4 R m is continuously differentiable at point x (on
a
a set S) if the partial derivatives Iii OXj exist and continuous at x (at every
point of S) for 1 ~ i ~ m, 1 ~ j ~ n, and the Jacobian matrix is defined as
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
II flip = (10 00 If(t)IP dt) lip, for p E [1, (0) (2.11)
1111100 = sup If(t)l, for p = 00 (2.12)

tE[a,oo)
By letting p = 1,2,00, the corresponding normed spaces are called L 1 , L 2 ,

L oo , respectively. More precisely, let f(t) be a function on [0,(0) of the signal
spaces, they are defined as
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)
Loo ~ {I: R+ --4 R 11111100 = sup

tE[a,oo)
Ifl < 00, bounded Signal} (2.15)
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.
Lemma 2.2 (Barbalat's Lemma). Let f(t) be a differentiable function, if

limt->oo f(t) = k < 00 and j(t) is uniformly continuous, then
lim j(t)
t->oo
=0 (2.16)
Corollary 2.1. If f(t) is uniformly continuous, such that
lim
t->oo
it
0
f(7) d7 (2.17)
ex'ists and is finite, then f(t) -t 0 as t - t 00.
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.
2.2.2 Properties of Matrix

Recall that a scalar function a(t) is said to be
• positive if a(t) ;::: 0 for all t; and
• strictly positive if a(t) > 0 or for some E> 0, a(t) 2: E for all t.
Definition 2.5. A square matrix: A E R nxn is
• positive semi-definite (denoted by A;::: 0) if xTAx;::: 0, \:Ix ERn;
• positive definite (denoted by A > 0) if xT Ax > 0, \:I x E Rn, x of- 0, or if

for some f3 > 0, xT Ax;::: f3x T X = f3llxl1 2 for all x (equivalently, x T Ax;::: f3
for all x such that Ilxll = 1);
• negative semi-definite if - A is positive semi-definite;
• negative definite if -A is positive definite;
• indefinite if A is positive for some x E R n and negative for other x ERn;
• symmetric if AT = A;
• skew-symmetric if AT = -A;
• symmetric positive definite (semi-definite) if A > 0(2: 0) and A = AT;

and
• A time-varying matrix A(t) is uniformly positive definite if there exists

O! > 0 such that A(t) 2: cd.
Some of the most useful facts on real matrices are summarised below [52].
1. A = AT E R nxn is positive (semi-)definite if /\i(A) > 0 (2: 0).

2. A = AT E R nxn is negative (semi-)definite if /\(A) < 0 (::::; 0).
3. The eigenvalues of a symmetric matrix are all real.
4. A necessary condition for a square matrLx A to be positive definite is

that its diagonal elements be strictly positive. Sylvester's Theorem states
that a necessary and sufficient condition for a symmetric matrLx A to be
positive definite is that all its principal minors (i.e. all, all a12 - a12a21,
... , det A) be strictly positive.
5. If A = AT > 0 (2: 0) and B = BT > 0 (2: 0) then A + B > 0 (2: 0)

and has all eigenvalues real positive but it is not true in general that
AB > 0 (2: 0). If and only if A and B are commutative, i.e. AB = BA,
then AB > 0 (2: 0).
6. A symmetric positive semi-definite matrix A can be decomposed as A =

U T AU where U is a unitary matri-x and satisfies UTU = I, and A is a
diagonal matrix containing the eigenvalues of the matri-x A. Therefore,
we have
2.3 Concepts of Stability

Consider the autonomous non-linear dynamic system described by
x= f(x), x, fERn (2.18)
Definition 2.6. A state x* is an equilibrium state (or equilibrium point) of

the system, if once x(t) equal to x*, it will remain equal to x* for ever. Math-
ematically, that means the vector x* satisfies
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
IIx(t)11 < E, 1ft ~ 0

Otherwise, the equilibrium point is unstable.
Definition 2.8. The solution of (2.18) is semi-globally uniformly ultimately
bounded (SGUUB), if for any S1, a compact subset of R n and all x(to) = Xo EO,
there exist a J-I- > 0 and a number T(J-I-, xo) such that IIx(t)11 < fJ for all t > to+T.
2.4 Lyapunov Stability Theorem

Definition 2.9. Continuous function a(r): R -t R belongs to class K if
• a(O) = 0;
• a(r) - t 00 as r - t 00;
• a(r) > 0 Ifr > 0; and

• a(r) is nondecreasing, i.e., a(rd 2: a(r2)' Ifrl > r2·
Definition 2.10. A continuous function V(x, t): R n x R+ -t R is
• locally positive definite if there exists a class K function a(.) such that
'V(x, t) ~ a(llxll)
for all t 2: 0 and in the neighbourhood N of the origin Rn;
• positive definite if N = Rn;
• (locally) negative definite if -Vis (locally) positive definite; and
• (locally) decrescent if there exists a class K function 13(-) such that
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
v= dV(x,t) = O~T(X,t) + [oV(X,t)]T f(t,x)

dt at ax
Theorem 2.4 (Lyapunov Theorem). Given the non-linear dynamic system
x= j(x, t), x(O) = Xo
with an eq'uilibriwn point at the origin, and let N be a neighbourhood of the

origin, i. e. N = {x: IIxll ::; E, with E > O}, then, the orig'in 0 is
• stable in the sense of Lyap'unov if for x E N, there exists a scalar function

V(x, t) SItch that V(x, t) > 0 and V(x, t) ::; 0;
• 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;
• asymptotically stable if for x E N, there exists a scalar f'ltnction 1/(x, t)

such that V(x, t) > 0 and V(x, t) < 0;
• globally asymptotically stable if for x E Rn (i. e. N = R n ), there exists a

scalar funct'ion V(x, t) s'uch that V(x, t) > 0 and V(x, t) < 0;
• uniformly asymptotically stable if for x E R n (i. e. N = Rn), there exists a

scalar function '\/(x, t) such that V(x, t) > 0 and decrescent and V(x, t) <
O·,
• globally, uniformly, asymptot'ically stable if for N = R n , there exists

a scalar function V(:1;, t) s'uch that V(x, t) > 0 and decrescent and is
radially ItnboItnded (i.e., V(x, t) -; 00 uniformly in time as Ilxll -; 00)
and i-'(x, t) < 0;
• exponentially stable if there exist positive constants a, f3, , such that,

\fx EN, allxli 2 ::; V(x, t) ::; f3lixll 2 and V(x, t) ::; -,lixIl 2 ; and
• globally exponentially stable there exist positive constants a, f3, , such

that, \fx ERn, allxl1 2 ::; V(x, t) ::; f3llxl1 2 and V(x, t) ::; -,llxI1 2 .
The function V(x, t) in Theorem 2.4 is called a Lyapunov function. The

theorem provides sufficient conditions for the origin to be stable. No conclusion
on the stability and instability can be drawn if a particular choice of Lyapunov
candidate does not meet the conditions on ~T.
A Lyapunov function is not unique, i.e. there exist many Lyapunov func-
tions for the same system. However, for a given system, specific choices of
Lyapunov functions may yield more precise results than others. For controller
design, different choices of Lyapunov functions may result in different forms of
controller (and different performance albeit stable).
2 ..5 Useful Theorems and Formula 19
2.5 Useful Theorems and Formula

2.5.1 Sliding Surface
Define a filtered tracking error es as
es = [AT 1Je (2.20)

where e = [el,e2,'" ,enJ T ERn, A = [/\1,/\2,'" ,An_IJ T is chosen such that
polynomial sn-l + /\n_ISn-2 + ... + Al is Hurwitz. With these definition, the
tracking error may be expressed as el = H(s)e s with H(s) being a proper
stable transfer function, which implies that el (t) - t 0 as e s - t O.
Definition (2.20) has the following properties:
°
(i) es = defines a time-varying hyperplane in R n on which the tracking
error el asymptotically converges to zero,
(ii) a state representation of (2.20) can be expressed as
,=
.
As' + bse s, ,= T
[el e2 ... en-I] ,n:::: 2
where
;1 E R(n-l)x(n-l),bs = [~1
-A n - l 1
with As being a stable matrix. In addition, two constants ko > 0 and

/\0 > 0 can be found such that II eAst I ::; koe->'ot [74J. The solution for'
IS
(2.22)
If A is chosen as A = [/\1, /\2, ... "\n-l]T = [/\n-l, (n - 1)/\n-2, ... , (n - l)A]T

with A > 0, then the filtered tracking error es becomes
(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)
~ 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}/>..
2.5.2 Mean Value Theorem

Theorem 2.5 (Mean Value Theorem). Assume that f(x) has a derivative
(finite or infinite) at each point of an open interval (a, b), and assume also that
it is continuous at both endpoints a and b. Then there is a point ~ E (a, b) such
that f4J
f(b) - f(a) = !,(~)(b - a)
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(y) - f(x) = 8~~) Ix=z (y - x)
2.5.3 Integral Formula

Theorem 2.1 (Change of Variable in a Riemann Integral). Assume that
function g(x) has a continuous derivative g'(x) on an interval [a, bJ. Assume
that function f(*) is continuous on [g(a),g(b)J and define F(x) as
F(x) = 19 (X)
g(a)
f(t)dt, Vx E [a,b]
Then, for each x in [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
Theorem 2.8 (Change of Variable in a Riemann-Stieltjes Integral).

Let f(x) be integrable with respect to a{x) on [a, bJ (i.e., J:
f{x)da{x) exists),
and g(x) be a strictly monotonic continuous function defined on an interval S
having endpoints c and d. Assume that a = g{c) and b = g{d), and let h{x)
and fi(x) be the composite functions defined as
h{x) = f(g(x», fi(x) = a(g(x», Tlx E S
Then, h(x) is integrable with respect to fi{x) on Sand
1 9 (d)
f{t)da{t) =
lb f{g{x»da{g(x»
g(c) a
Theorem 2.9 (Integration by Parts). If f{x) is integrable on x E [g(a), g{b)],

then g(x) is integrable on [J{a), f{b)] and
1 b
f(x)dg(x) = f(x)g(x)
Ib
a -1 b
g{x)df(x)
which is known as the integration by parts.
Lemma 2.3 (Riemann Integral Inequality). Assume f(x) and g{x) are
integrable on [a, b], then, we have the following inequalities:
1. If f{x):s g{x) on [a,b], then
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
lb If{x)ldx '5: ({b - a) lb f 2{X)dX) !

4. Let k1 > 1, k2 > 1 and k1 + ~ = 1, then we have Holder's inequality

(74! 1 2
lib f(x)g(X)dXI ~ (i b lJ(x W1 dX) rk (i b

If(xW2dX) ~
Equality holds only when the sign of f(x)g(x) is fixed and If(xWl
clg(XW2 (c is a constant) holds.
Lemma 2.4 (Riemann-Stieltjes Integral Inequalities). Assume thatfunc-
tion a(x) is increasing on la, b], f(x) and g(x) are integrable with respect to
a( x) on [a, b], the following ineq1lalities are in order:
1. Comparison Theorem if f(x) ~ g(x) for all x in [a, b], then we have
ib f(x)da(x) ~ lb g(x)da(x)
2. Cauchy-Schwartz inequality i4!
Lemma 2.5 (Differentiation of integrals). The commonly 'used differenti-

ation formttlas of integrals are listed below:
1. In general, if 1jJ(y) and ¢>(y) are differentiable, and for b ~ y :S B, a :S

1jJ(y) :S A, a :S ¢>(y) :S A, then
d l¢(Y) l¢(Y)
-d f(x, y)dx = f(¢>(y), y)¢>'(y) - f(7jJ(y), y}1//(y) + j~(x, y)dx
Y '1(y) '1(y)
2. For the special cases, we have the following equations
: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(Zl, () = F(Z2, () + (Zl - Z2) 11 (aF~::, (») d,\ (2.25)
where z>. = ,\Zl + (1 - ,\)Z2.

which proves (2.25). <>
2.5.4 Implicit Function Theorem

Theorem 2.10 (Implicit Function Theorem). Assume that f(x, y) : R n x
R m ---t R m is continuously differentiable at each point (x, y) of an open set
S C Rn x R'nl. Let (xo, Yo) be a point in S for which f(xo, Yo) = 0 and for
which Jacobian matrix [8f(x, y)/8y](xo, Yo) is nonsingular. Then there exist
neighborhoods U C Rn of Xo and V C Rrn of Yo such that, for each x E U, the
eqttation f(x, y) = 0 has a unique solution y E V. Moreover, this soltttion can
be given as y = g(x) where g(x) is continuously differentiable at x = Xo /96j.
Lemma 2.7. Assume that f(x, y) : R n x Rrn ---t R m is continuottsly differen-

tiable at each point (x, y) of an open set S C R n x Rrn. Let f(x, y) = 0 and
Jacobian matri~; [8f(x, y)/8y](x, y) i- 0 for (x, y) E S. Then there exists a
ttnique continuous function y = g(x) such that f(x, y) = O.
Proof Using the Implicit Function Theorem 2.10, for every point Xi E S, there
exists a neighborhood Ui of Xi and a unique local function y = gi(X) E C l such
that f(x, y) = 0, \:Ix E Ui holds.
In the following, we shall show that we can patch up a unique continuous
function over S, i.e., there exists a unique continuous function y = g(x) such
that f(x,g(x) = 0, \:Ix E S. Let us prove the continuity and uniqueness
separately by contradiction. First, let us assume that there is no continuous
function such that f(X,gi(X» = 0, \:Ix E S. That is to say that for all the
solutions that satisfy f(x, gi(X») = 0, \:Ix E S, there is at least one point, say
Xl E S, such that anyone of gi(Xl) is discontinuous, i.e., the following equalities
hold
f(X,gi(Xl)) =0
lim gi(X) = gi(X;-)
x-).x~
Because [8f(x, y)j8y](x, y) =J 0 holds for all xES, according to Implicit

Function Theorem 2.10, there exists a neighborhood Ul of Xl and a unique
continuous function gl (x) such that
which implies
lim gi(X) = lim gl(X) = gl(Xl)

1
X---?X l
X--i'X
i.e., gi(x 1) = gi(Xi) = gl(Xl). This is in contradiction with the proposition

that none of functions gi(X) is continuous. Therefore, a continuous function
g(x) does exist, which guarantees
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
f{x, g2(X)) = 0, '<:Ix E U2 C S
which implies
This is in contradiction with the proposition that the continuous function is

not unique. In summary, we know that we can patch up a function over the
compact set which is unique and continuous. Therefore, we claim that there
exists a unique continuous function g(x), which guarantees
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.
Lemma 2.8. Ass'ume that f(x, y) : R n x R ....... R is continuously differentiable

\f(x, y) E Rn x R, and there exists a positive constant d such that
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) + 8 f~: y) Iy=y! (I~I - 0) > f(x, 0) + Icl 2: 0
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
<)
2.5.5 Input-Output Stability

Lemma 2.9. [104/ Let v and p be real-valued functions defined on R+, and
let band c be positive constants. If they satisfy the differential inequality
iJ(t) :::; -cv(t) + bp(t)2, v(O) 2: 0 (2.26)

then the following integral ineq1tality holds:
(2.27)
Proof Upon multiplication of (2.26) by eel, it becomes
! (v(t)e ct ) :::; bp(t)2e ct (2.28)
Integrating (2.28) over [0, t], we arrive at (2.27). <>

Lemma 2.10. Let H(s) be a proper stable transfer function and y = H(s)u.
If U E L(XJ, then
(2.29)
for some constants k 1 , k2 2 0 and any t 2 0, T > O. Furthermore, if H(s) is

strictly proper, then kl :::; aIlH(s)ll~ for some constant a > O.
Proof See the proof of Lemma 4.8.2 in [74]. <>
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
Neural Networks and

Function Approximation
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.
3.2 Function Approximation

The development of mathematical analysis during the past two hundred years
has lead to the discovery and study of important classes of approximating
functions. Under certain conditions, it has been proven that several approxima-
tion methods such as polynomials, trigonometric series, orthogonal functions,
splines, neural networks, and fuzzy systems, etc., have function approximation
abilities. Function approximation problem is in general defined as follows.
Definition 3.1 (Function Approximation). If f(x) : Rn ~ R is a contin-

uous function defined on a compact set n, and f(W,x) : Rl X R n ~ R is
an approximating function that depends continuously on Wand x, then, the

28 Chapter 3. Neural Networks and Function Approximation
approximation problem is to determine the optimal parameters W*, for some

metric (or distance function) d, such that
d(f(W*,x),f(x)) $ E (3.1)
for an acceptable small E [157J.

All the developed approximation approaches have their own characteristic
theory and history. Conditions under which the method is preferable are known
for a given approximation problem, but no one method is absolutely the best.
Actually, in most practical problems, physical considerations determine the
choice of the approximation functions.
Artificial neural networks (ANNs) are inspired by biological neural net-
works, which usually consist of a number of simple processing elements, call
neurons, that are interconnected to each other. In most cases, one or more
layers of neurons are considered that connected in a feedback or recurrent way.
The following features of ANNs make them particularly attractive and promis-
ing for applications to modelling and control of nonlinear systems [70, 173].
(i) Universal approximation abilities

A multilayer NNs with one or more hidden layers can approximate any
continuous nonlinear function arbitrarily well over a compact set, pro-
vided sufficient hidden neurons are available.
(ii) Parallel distributed processing abilities
The network has a highly parallel structure and consists of many simple
elements, which is attractive from the viewpoint of implementation.
(iii) Learning and adaptation

The intelligence of ANNs comes from their generalization abilities with
respect to fresh, unknown data. On-line learning and adaptation of ANNs
are possible.
(iv) Natural fault tolerance and feasibility for hardware implementation

The connected architecture of numerous NN nodes makes that NN has
the potential for fault tolerance. Moreover, it is easy for VLSI realization
which will benefit the real-time implementation.
An extensive literature exists related to the approximation capabilities of dif-

ferent types of neural networks and function approximators [7,8,25,26,30,34,
44,58,69,76,88,88,89,97,118,131,144,146,160,173, 1791. This chapter dis-
cusses two basic types of artificial neural networks, (i) linearly parametrized
neural networks (LPNNs) in which the adjustable parameters appear linearly,
and (ii) multilayer neural networks (MNNs) in which the adjustable parameters
appear nonlinearly. For other linear-in-the-parameters function approximators

such as polynomials [151]' splines [171, fuzzy systems [1791 and wavelet net-
works [190], among others, the controllers presented in this book using LPNN
can be changed to these function approximators without any difficulty, thus
is ommitted for clarity. Throughout this book, we use the following notation
for expressing the neural network weights: W expresses the general weights in
the introduction of neural network structures; W" denotes the optimal neural
network weights; W is the estimate of W*; and tV = W - W*.
3.3 Linearly Parametrized Neural Networks

The Radial Basis Function (RBF) networks can be considered as a two-layer
network in which the hidden layer performs a fixed nonlinear transformation
with no adjustable parameters, i.e., the input space is mapped into a new
space. The output layer then combines the outputs in the latter space linearly.
Therefore, they belong to a class of linearly parameterized neural networks
(LPNN), and can be simply described as
(3.2)
with the input vector z E Oz C R n , weight vector W E RI, weight number l,

and basis function vector
(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
i = l,2, ... ,l{3.4)
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)
or Inverse Hardy'S multiquadric form
(3.6)
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)
where St is a bOltnded fltnction vector.

z
Proof Substituting = Z - E1/; into (3.4), we have
i = 1,2, ... , l (3.8)
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
where Z E Oz C Rn+l, the integer l is the number of NN nodes, {h,I3,"',Iz}

is a collection of l not-ordered subsets of {1,2, ... ,n+ I} and dj(k) are non-
negative integers, W is the adjustable synaptic weight vector. The activation
function s(Zj) is a monotone increasing and differentiable sigmoidal function.
It can be chosen as logistic, hyperbolic tangent or other functions. It is shown
in the literature [100, 144] that HONNs satisfy the conditions of the Stone-
Weierstrass Theorem and can therefore approximate any continuous function
over a compact set.
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)
Universal approximation results in [141,161,179] indicate that, for any con-

tinuous function g(z) : R n - t R if I is sufficiently large, then there exists an
ideal constant weight vector W* such that
max Ig(z) - gnn(W*,

zEn.
z)1 < j.L, z E Oz
with an arbitrary constant p > 0 and compact fl •. Formally, W* can be defined

as
W* := arg min {sup /g(z) - WTS(Z)/}

WERl zEn.
Assumption 3.1. On a compact set Oz C Rn+l, the ideal neural network

weights W* satisfies
/IW*/1 ~wm (3.13)
where Wm is a positive constant.
Common Tuning Algorithms

Next, let us present several commonly used tuning algorithms.
Lemma 3.3. If e s is bo·unded, then the learning algorithm
(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)
Differentiating (3.1.5) along (3.14), and noting W= W- W*, we obtain
(3.16)
Since IIW"'II :s; W m, and the fact that
2WTW = IIWII 2+ IIWII 2-IIW"'1I2 ~ IIWI1 2-IIW"'1I2 (3.17)

we have
.
VW = -esW -T
S - (1 + leslm)8wW -T"
W
:s; leslllWIlIiSII- 1 + ~eslm 8w(IIW1I 2 - w~)
:s; - 1 + ~eslm [8wIl WII2] + lesillWilliSIl + ao(es)
< _ 1 + lesl m[(JJ:IIWII-

- 2 w
leslllSIl
~(l+leslm)
)2 _a(e)]
s
(3.18)
where
(3.19)
(3.20)
which are all well defined at e s = O. For bounded e s E 6 e• C R, let a =

le.IIISIl } ,an d d efi ne
sUPe.ER { a (e s )} an d a- e• = sUPe.ER { .5 w {l+le.I"')
(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
The time derivative of Vw along (3.14) is

·"T
VW = -W Se s- 8w(1 + leslm)W W
"T"
(3.22)
Opening the bracket, we have
Since IWTSI ::; ~IIWII2 + IIr~2, we have
V.U ::; -81O IlW1I 2 - lesl ln C~w IIWII 2_ lesl(l-In) ":~2)

Because every element of S is not larger than one, we know that IISI1 2 ::; l with
l the NN hidden-layer node number. Therefore, Vw < 0 once
411 I(l-In)
IIWI1 2> e;82 (3.23)
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
W= - (KoIIWIl + Kllesl + K2 )S(z)es- 8(IIWII + lesl + 1) IIS(z)lIlesIW (3.24)
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
V = _WT(KoIIWII + KIle s l + K2)S(z)e s - 8(IIWII + lesl + 1)IIS(z)lIl esIIl WI1 2

~ (KoIIWII + Kllesl + K2) IIS(z)IIIIWIIJesl
-8(IIWII + 1)IIS(z)lllesIIIWI12
r
lesl +
::; -IIS(z)IIIIWlllesl{ 6[IIWII + ~ (1 + lesl- ~O)

-~(1 + lesl- ~of -Kllesl-K2}
Define
810 : = {W(t) IIWII::; essE~es [~( 1 + lesl- T)

+ ~(1+le81-K;)2+~(Kllesl+K2)]} (3.25)
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 . <>
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)
Its time derivative is
(3.28)
The boundedness of W can be easily seen for different cases as below.

(1) If IIWII < W m, the boundedness of W is obvious.
(2) If IIWII = Wm and WTS(Z)es :S 0, we have
V = WTS(Z)e s :S 0 (3.29)
V is non-increasing, and accordingly, we have IIWII :S W m .

(3) If IIWII = Wm and WT S(Z)e s > 0, we have
(3.30)
which means that V is kept constant.
In summary, II WII ::; W m , 'r/t 2: 0 is guaranteed. 0

It should be noted that, from a mathematical perspective, neural networks
represent just one class of function approximation. Other function approxima-
tors such as B-splines [141], wavelet-based approximation schemes and fuzzy
systems [179] are also popularly applied function approximators in practical
applications. An overview and comparison study of different function approx-
imation approaches can be found in [26,88].
In some of RBF networks such as Gaussian RBF networks discussed above,
if the parameters J.Li and ai of basis functions are also adjustable, then the
network is referred as "Variable Neural Networks" [89,122,123]. In this case,
the variable RBF networks are in fact nonlinearly parametrized approxima-

tors, which are very similar to multilayer neural networks as will be detailed
in the next section. To avoid any confusion, we denote the term "RBF neural
networks" as linearly parametrized networks with determined radial basis func-
tions.
3.4 Non-linearly Parametrized Networks

Multilayer neural networks, which are also called multilayer perceptions in the
literature, are one of the most widely used neural networks in system modelling
and control. It is a static feedforward network that consists of a number of lay-
ers, and each layer consists of a number of McCulloch-Pitts neurons [130j. Once
these have been selected, only the adjustable weights have to be determined to
specify the networks completely. Since each node of any layer is connected to
all the nodes of the following layer, it follows that a change in a single parame-
ter at anyone layer will generally affect all the outputs in the following layers.
The structure of MNNs with two hidden layers is shown in Figure 3.1, which
can be expressed in the following form
gnn (z) ~ t, [Wi' (t, v;'z, +9.;) 1+9. (3.31)
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:
Hidden-layer
Zl
Z2 • (
g(z)
Figure 3.1: Three-layer neural networks
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/
nlax Ig(z) - gnn(Z)1

zEn.
< J.L (3.35)
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)
with Vi = [Vil,Vi2,'" ,vi(n+1)lT, i = 1,2,,,, ,i. The term zn+! = 1 in input

vector Z allows one to include the threshold vector [Ovl,Ov2,'" ,OvltV as the
last column of V T , so that V contains both the weights and thresholds of
the first-to-second layer connections, Then the three-layer NNs (3.31) can be
expressed as
gnn(W, V,z) = WTS(VTZ) (3.38)

S(VTZ) = [S(V[Z), s(vfz), ... , s(vTz), l]T
W = [WI, W2,'" , WI+If E R I+ I
where the last element in S(VT z) incorporates the threshold Ow as Wl+l of
weight W. Any tuning of Wand V then includes tuning of the thresholds as
well [112]. Then in three-layer NNs (3.38), the total number of the hidden-layer
neurons is l + 1 and the number of input-layer neurons is n + 1. According to
Theorem 3.1, there exist ideal constants W* and V* such that
max Ig(z) - gnn(W*, V*,

zEn.
z)1 < /1- 5. fl
with constant fl > 0 for all Z E ~z. That is,
(3.39)
with £(z) satisfying maxzEn. 1£(z)1 < fL, Vz E ~z.

The ideal weights satisfying the above approximation are not unique. To
avoid confusion, the ideal weights W* and V* are defined as
(W*, V*) : = arg min {sup IWTS(VTz) - g(z)l} (3.40)

(w,v) zEn.
In general, W* and V* are unknown and need to be estimated in function

approximation. Let Wand V be the estimates of W* and V*, respectively,
and the weight estimation errors be TV = W - W* and V = V - V*.
Assumption 3.2. On the compact set n., the ideal neural network weights
W*, V* and the NN approximation error are bounded by
(3.41)
with W m , Vm and £1 being positive constants.

As multilayer NNs are nonlinearly parametrized function approximators,
i.e., the hidden layer weight V* appears in a nonlinear fashion, it is very hard
to derive the tunning algorithms in comparison with the linearly paramtrized
functions. When applying function approximators WT S(VT z) for solving the
estimation problem, it is highly desirable to have a linearly parametrized form
in terms of TV and V.
Lemma 3.6. The linearly parameterized approximation error of NN approxi-
mator (3.38) is given by
WTS(VTz) - W*TS(V*T z) = WT(S - S'VTz) + WTS'VTz + du (3.42)
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
Proof The Taylor series expansion of S(V"T z) about yT Z can be written as

(3.44)
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)
where the residual term is given by

du = WT S'V*T z - W .. T o ((iT z)2 (3.46)
Noting that W= W- W* and V = Y- V .. , (3.42) implies that

du = WTS - W .. TS(V*T z ) - (W - W*)T(S - S'yT z ) - WTS'(V - V*)T z
= WTS'V*T z - W*TS'yT z + W .. T[S - S(V .. Tz)]
Since every element of the vector S - S(V*T z) is bounded by 1, we have
I
W .. T[S - S(V*T z )] ~ L Iw;1 = IW*h
i=l
Considering WTS'V .. Tz = tr{V*TzWTS'} ~ IIV*IIpllzWTS'IIp, we conclude

that (3.43) holds. <>
Lemma 3.7. If sigmoid activation function (3.34) is used, then there exist
positive constants C1 to C2 such that the residual term d.u is bounded by /50/
(3.47)
Proof The derivative of the sigmoid activation function (3.34) with respect to
Zais
(3.48)
It is easy to check that 0 :S s'(Za) :S 0.2.5 and IZas'(za)1 :S 0.2239, 't/za E R.

Using the properties of the Frobenius norm in (2.6), we have
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
Ilo(VT Z)211:S IIS'VTzll + IIS(V*T z) - SWTz)11

:sIIS'VTzll + IIS'V*Tzll + IIS(V*Tz) - SWTz)11
:S IIL9'VTzii + IIS'IIFIIV*IIFllzll + IIS(V*T z) - S(VTz)11 (3 ..50)
Considering (3.49), IIV*IIF :S Vm (Assumption 3.2) and IIS(V*Tz)-SWTz)11 :S
l, we have
(3 ..51)
By (3.46), it can be seen that
Idul :S IWTS'V*T zl + IW*T O(VTz)21

:S vmllWllllS'llFllzll +wm IlO(VT z)211 (3.52)
Using (3.49)-(3.51), inequality (3 ..52) can be further written as
By choosing Co = 1.2239wm l, C1 = 0.25wm vm and C2 = 0.2.5v m l, it follows that

(3.47) holds. <>
For uniformity, let Yd E en denote the desired trajectory, and define the
following signals
where A = ['\1, '\2, ... , '\n-1jT is an appropriately chosen coefficient vector such
that sn-1 + '\n_1Sn-2 + ... + '\1 is Hurwitz.
Assumption 3.3. The desired trajectory vector ~d is continuous and avail-

able, and II~dll :S c with c being a known bound.
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)
Proof According to the definitions of z and z, we have
(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)
Common Tuning Algorithms

Next, we shall present several common tuning algorithms for the weights of
multilayer neural networks, in which the first terms of the right-hand sides of
the algorithms are the modified backpropagation algorithms and the last terms
of them correspond to a combination of a-modification [74] and e1-modification
[134], which are introduced to improve the robustness in the presence of the
NN approximation error.
Lemma 3.10. If e s is bounded, then the updated learning algorithms
w= -rw [(S - S'VTz)e s + 6w (1 + leslm)W] (3.62)
V= -ru [zWTS'e s + 6v (1 + leslmW], (3.63)
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)
Differentiating (3,64) along (3,62)-(3,63), we obtain
,:rw = _WT(S - S'VTz)e s - tr{VTzWTS'}e s

- (1 + lesl m) (bwWTW + bvtr{VTV} )
Vw = e s [ - WTS + (WT - W*T)S'VTz - WTS'VTz]
-(1 + lesl m) (bwWTW + bvtr{VTV} )

= es [ - WT S - W*T S'V T z] - (1 + lesl m) (6 wWTW + 6vtr{V T V})
::; lesl [IIWIlIISIlF + I W*IIIIS'VTzlI] - (1 + lesl m)(bwWTW + bvtr{VTV})
Using the properties of MNNs (see (3.49) in the proof of Lemma 3,7 and As-
sumption 3.2)
and the fact that
2WTW = IIWII 2 + IIWII 2 -IIW"1I 2 ~ IIWII 2 -IIW"1I 2 (3.65)

2tr{VTV} = IIVII} + IIVII} -IIV"II} ~ IIVII} -IIV"II} (3.66)
we obtain
Vw ::; lesl(lllWIl + 0.2239lwm)

_ 1 + ~eslm [bw(IIWII 2 _ w~) + bv(IIVII} - v~)]
::; _ 1 + ~eslm [bwll WlI 2 + bvIIVII }] + lleslllWIl + ao(e s )
::; -
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}
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
Vtt' = -WT(8 - 8'V T z)e s - W T 8'VT ze s - (1 + le s lm)(6wIIW/l2 + 6vl!v/I~)

:::; -6w1lW1I 2 - 6vl!Vll} - lesl m(6w11W112 + 8vl!VlI} _lesl(l-m)lwT 81)
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.
Lemma 3.11. For the weights adaptation algorithms
TV = -I'w [(8 - 8'V T Z)es + 8wieslW] (3.72)
V= -fv[ZWT8Ies+6vlesl",T] (3.73)
where I'w > 0, f v = f; > 0, if W(O) E 8 w with
(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
Its time derivative along (3.72) is

It can be checked that the elements of S'VT Z are bounded by
with Zai = 'YvTZ.

Thus, the elements Is( vT Z) - s~vr ZI of vector S - S'VTZ are not larger
than 1.224. It follows that
I
IWT(S - S'VTZ)I :::; L 1.2241wil :::; 1.224v'lllwll
i=l
Therefore
Vw :s -leslllWIl (bwllWIl - 1.224v'i)

It follows from (3.74) that V.U :s 0 once IIWII ~ 1.224Vl/bw. Hence, for the
initial condition W(O) E 8 w , the estimated weight W(t) E 8 w , Vt ~ O.
Consider the positive function
Vv = ~tr{VTr;lV}
and using the property that WTS'VTZ = tr{VTZWTS'}, its time derivative
along (3.73) may be written as
Vv :s -Iesl [bvliVll} -IIWIIIIS'VTZII]

From IIS'VT ZII :::; 0.224v'l and IIWII :::; 1.2l~VI, we further have
Vv :::; -l5vlesl [IIVII~ _ 0~~~2l]

which proves that Vv :::; 0 once IIVIIF ~ °tt21. Therefore, V(t) E Loa if V(O)
is bounded. (> W v
3.5 Neural Networks for Control Applications

As shown in the previous two sections, both MNNs and RBF NNs have the
capability of approximating any continuous function on a compact set. An
immediate question arisen concerns the choice of neural networks to be used for
a given practical control problem. In this section, main properties of such two
kinds of networks shall be discussed, and their advantages and shortcomings
for system identification and adaptive control design will also be addressed
through comparison study.
3.5 Neural Networks for Control Applications 45
(i) Approximation properties

The number of basis functions for RBF networks needed to approximate
a given function is a critical factor in solving identification and control
problem. Because such a number tends to increase exponentially with
the dimension of the input space, the approximation approach becomes
practically infeasible when the dimensionality of the input space is high
which often is referred to as "the curse of dimensionality' [142,161].
It has been shown in [7] that for linear approximators, the interacted
square approximation error cannot be made (uniformly) smaller than
order (1/1)2/n. This result gives a hard lower bound to the approximation
error for RBF networks. MNNs are global approximators due to the
nature of the global active neuron functions within the networks. The
research work by Barron [7J shows that multilayer NNs reduces the effect
of the curse of dimensionality problem under certain conditions.
(ii) Learning speed
A disadvantage of multilayer NNs lies that the NN weights appear in a
nonlinear fashion, which makes analysis of networks more complex. Fur-
ther, the adjustment of a single weight of the networks affects the output
globally. Hence, in general, all the weights have to be adjusted simul-
taneously for each training data set. Available simulation study showed
that slow convergence rate observed in the phase of MNN learning. In
contrast to the above, the activation functions of RBF networks are lo-
calised, thus these networks store information locally in a transparent
fashion. The adaptation in one part of the input space does not affect
knowledge stored in a different area, i.e., they have spatially localised
learning capability. Therefore, if the basis functions are correctly chosen,
the learning speed of RBF NNs is in general better than that of MNNs.
(iii) Pr·iori knowledge required
The RBF neural networks are local approximators where each basis func-
tion can only locally affect the network output, therefore, the construc-
tion of the RBF NNs and the resulting approximation heavily depend
on the compact set they work on. For example, the network structure,
the number of basis functions, their location and shape, must be chosen
a priori by considering the working space which is in general unknown
in the case of neural control systems. This can often lead to an over or
under determined networks.
In contrast, MNNs are global approximators due to the nature of the
global active neuron functions within the networks, and a priori design
for basis functions like in RBF neural networks is not needed. The tuning
of the weights in hidden layers plays similar roles as adjusting the location
and shape of basis functions of RBF NNs. Such an advantage of MNNs is
especially important for on-line adaptive NN control, in which the input

signals are generated from closed-loop systems and are generally unknown
a priori.
(iv) Stable controller design of neuml adaptive system
The RBF neural networks have some useful properties which render them
suitable to be applied to on-line non-linear adaptive modelling and con-
trol. This is because they belong to a class of linearly parametrized
networks where the network output is related to the adjustable weights
in a linear manner. Thus, on-line learning rules can be used to update the
weights and the convergence results can be derived using the available lin-
ear adaptive techniques. The MNNs are often referred to as non-linearly
paramterized networks, which means that the network output is related
to the adjustable weights in a non-linear fashion. This property often
makes analysis of systems containing multilayer networks difficult and
the results obtained more conservative.
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
SISO Nonlinear Systems
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
u = [-a(x) + vl/b(x) (4.2)
where v is a new control variable. In an effort to solve the problem of unknown

nonlinearly parameterized a(x) and b(x), adaptive control schemes based on
function approximation techniques have been studied [24,45,52,99,161,171,
179,186]. In these approaches, the nonlinearity b(x) is usually approximated
by neural networks or fuzzy systems b(x, ltV) (where ltV denotes the estimated
weights). Therefore, additional precautions should be made to avoid possible
singularities of the controllers (i.e., b(x, ltV) =I 0). To cope with such a diffi-
culty, the work [24] suggests that the initial values of the NN weights are chosen
sufficiently close to the ideal values. Hence, off-line training phases are needed

48 Chapter 4. SIS a Nonlinear Systems
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
b(x) = d[b(x)J = 8b(x) x = ~ 8b(x) Xi+l (4.3)

dt 8x ~ 8x·
i==l •
which only depends on states x. Such a property is utilised to design novel

adaptive controllers while avoiding the singular problem discussed earlier. In
fact, many practical systems, such as pendulum plants [6,19], single link robots
with flexible joints [128], magnetic levitation systems [125], can be transformed
to system (4.1) and possess such a property.
By utilizing such a nice property of the studied systems, adaptive NN con-
trollers are presented in this Chapter which avoid the possible control singu-
larity problem in adaptive nonlinear control completely. In the first part, re-
gionally stable NN controllers are presented based on Lyapunov analysis with
guaranteed transient performance. The second part develops a semi-globally
stable controller by using MNNs and variable structure control (VSC) tech-
niques. The controller singularity problem is avoided and the stability of the
closed-loop system is guaranteed without the requirement for off-line training.
The control objective is to design a controller such that the output y follows
a desired continuous trajectory Yd E en. For ease of analysis, let Xd=[Yd, Yd,
. .. y(n-l)J T and Xd=[X T y(n)J T
'd' d' d .
Assumption 4.1. The desired trajectories IIxdll < c with known constant
c> O.
4.2 NN Control with Regional Stability 49
Define the tracking error and the filtered error as
e = x - Xd = [e1,e2,'" ,en]T (4.4)

es = [AT l]e (4.5)
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
fo S = a(x) + b(x)u + v + d(t) (4.6)
4.2 NN Control with Regional Stability

In the section, we shall investigate the control problem based on Lyapunov
analysis under mild assumptions in a compact set.
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.
Assumption 4.3. ab(x)jax n = 0, \:Ix E fl.

Note that the compact set fl denotes an operation region of interest in which
the system works on. This implies that the obtained result is regionally stable
in the sense that all the states must be guaranteed within the compact set fl.
4.2.1 Desired Feedback Control

To begin with, we first consider the case that the nonlinear functions a(x) and
b(x) are known exactly without any disturbance, i.e., d(t) = 0, and present
two desired feedback control (DFC) laws, u*, which make the system output y
follow the desired trajectory Yd asymptotically.
Lemma 4.1. Consider system {4.1} satisfying Assumptions 4.1-4.3, d(t) = o.

If the DFC input is chosen as /49/
* 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)
Choosing a Lyapunov function candidate Vs = 2:tx) and differentiating it along

2
(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)
Since bo < b(x) :S b1 , according to Lyapunov theorem, equation (4.9) implies

that limt-Hxl les(t)1 = O. Subsequently, we have limt->oo lIe(t)1I = 0 . <>
From (4.9), it is shown that the smaller the parameter c:, the more negative
the Vs. Hence, the convergence rate of the tracking error can be adjusted by
tuning the design parameter t:.
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.
Lemma 4.2. Consider system (4.1) satisfying Assumptions 4.1-4.3, d(t) = O.

If a desired feedback control (DFC) is chosen as f191}
u
*
= -
1
b(x) [a(x) + v] -
[1 + 1
-;;
b(x) ]
t:b 2 (x) - 2b2(x) es (4.10)
with constant c: > 0, then limt-+oo lIell = O.

Proof Substituting the DFC u = u* (4.10) into (4.6), under the assumption of
d(t) = 0, we have the closed-loop system
.=
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)
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)
which is unknown and can be approximated by a neural network as will be

detailed later.
Remark 4.3. Note that for the stability of the system without any distur-
bance, there is no need to have gb~(x) in the presented DFCs (4.7) and (4.10)
as can be seen from (4.9) and (4.12). However, it plays an important role in
handling the cross-terms of d(t) in the proof of closed-loop stability later as in
(4.26).
In general, both LPNN and MNN can be used for function approximation.
To illustrate the main ideas, they are both used for controller design.
4.2.2 HONN Controller Design Based on (4.7)

When nonlinear functions a(x) and b(x) are unknown, DFC u* (4.7) is not
realisable, and we only know that it is a function of x, e s and 1/
U = u*(z), eS]T EHzC

z= [x T ,e s ,-,1/ .. Rn+3 (4.15)
c
where compact set !.lz is defined as
!.lz = {(x,e s ,e;,I/)! xE!.l;xdE!.ld;e s = [AT 1]e;

1/ = _y~n) + [0 AT]e} (4.16)
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
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*
u*(z) = W*T S(z) + Ji-l, Vz E Oz (4.17)
where J.L1 is called the NN approximation error satisfying IJ.L1 I S Ji-o·

In general, the ideal NN weight W* is unknown and needs to be estimated
in controller design. We propose the direct adaptive controller
(4.18)
and the robust updating algorithm for NN weights W as

(4.19)
where constant (I > 0 and r = rT > O.

In adaptive law (4.19), (I-modification is used to improve the robustness of
the adaptive controller in the presence of the NN approximation error and the
external disturbance. Substituting controller (4.18) into (4.6), we obtain the
closed-loop error equation
£Os = a(x) + l/ + b(x)WTS(z) + d(t) (4.20)
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)
Substituting (4.7) into (4.21) leads to
(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
Theorem 4.1 (Stability). For the closed-loop adaptive system consisting of

plant (4.1), controller (4.18) and NN weight adaptive law (4.19), there exist a
compact set 0 0 C 0, and positive constants l"', c'" and /\ '" such that if
(i) the initial condition x(O) E 0 0 , and
(ii) the design parameters are chosen such that 1 ~ l*, c s:; c*, and /\max < /\ '"
with /\max the largest eigenvalue of r- 1 ,
then, W(t) E Loo and x(t) EO, 'tit ~ 0 /49}.
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)
Differentiating (4.23) along (4.19) and (4.22) yields
(4.24)
Using the facts that
2WTW = IIwl12 + IIWI1 2 -IIW"'11 2~ IIwll 2 -IIW*112 (4.25)

d(t) e; c 2
(4.26)
b(x) e s s:; cb2(x) + 4d(t)
2
lJ.tlesl s:; 2c~(x) + ;J.trb(x) (4.27)
and noting that Ip,d s:; J.to, Id(t)1 s:; do and 0 < b(x) < b1 , the following
inequality holds
(4.28)
Considering (4.23) and wTr-1w s:; AmaxllWl12, we obtain

,11
Vs s:; - - Vs + -ao
ao 2
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
((t) = eAst((O) + fat eAs(t-r)bses(T)dT
Accordingly, it follows that
11((t)11 :S koll((O)IIe- Aot + ko fat e-Ao(t-r)les(T)ldT (4.31)
Considering (4.30) and e-~ J2Vs(0)bl :S J2Vs(0)bl, we have
11((t)II :S koll((O)IIe- Aot + k o( J2Vs(0)bl + J aoaob l ) fat e-Ao(t-r)dT

k
:S koll((O)11 + _,0 (J2\1~(0)bl + JaoaOb1), Vt 2': 0 (4.32)
/0
Noting e s = [AT 1Je and e = [(T enV, we obtain
IIell :S 11(11 + lenl :S (1 + IIAI!)II(II + lesl

Substituting (4.30) and (4.32) into the above inequality leads to
1
Ilell :S kAII((O)II + (1 + ~~) [b~!~~~) + /\maxIIW(0)112] 2.jb;

+ J aoaob 1 (1 + kA) /\0
(4.33)
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)
Step 2 In the following, we shall find the conditions such that x E n, Vt 2: O.

Firstly, define a compact set
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)
c. = e~% {cl {xlllx - c,

xdll $ R( Amax, x(O), W(O») , Amax < A" } CO,
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
no = {X(O) I{x Ilix - xdll $ R( c, '-\max, x(O), W(O») } CO, Xd E Od}
Then, we say that no

is the largest region of initial states for achieving a stable
closed-loop system of the given controller.
56 Chapter 4. S1S0 Nonlinear Systems
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
le1(t)1 :S ko [11(O)lle-,xot + te-,xstJ2\1,;(0)b 1 + 7] (4.38)
(ii) the Loo tracking error bO'und is
(4.39)
where /\8 = minpo, 2!o}' ao = Vf:J-L5 + ~d5 + a1lW*1I 2 , ao = max{f:,

~ }, ko and AO are positive constants defined in Theorem 4.1, and
positive constant If.s(0) depends on x(O) and W(O),
Proof If x(O) E no, I 2': 1*, /\max < >'*, and f: :S f:*, Theorem 4.1 en-
sures that x E n, Vt 2': 0, subsequently NN approximation (4.7), Assump-
tions 4.2 and 4.3 are valid. Therefore, inequalities (4.24)-(4.31) hold. Since
( = [e1' e2,'" ,en _1]T, it follows from (4.30) and (4.31) that
h(t)1 :S 11(t)1I :S koll(O)lle-,xot
+koJ2V8(0)b 1 t e-,xO(t-T)e-2~odT+ ko~
Jo /0
(4.40)
For the integral in the above inequality, we have
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
It e-,xo(t-r)e -*0 dT :S te-,xst

Substituting the above inequality into (4.40), we arrive at (4.38). Because

lel(t)1 ~ 11((t)ll, the Leo tracking error bound (4.39) can be obtained from
(4.32) directly. <>
Remark 4.6. It is shown from (4.38) and (4.39) that large initial errors e(O)
and W(O) may lead to a large tracking error during the initial period of adap-
tation. In view of (4.38) and (4.39), output tracking error may be reduced by
choosing small design parameters £ and /\max (i.e., by increasing the controller
gain and adaptation gain r).
Remark 4.7. Taking the limits of both sides of (4.38) and noting the facts
that limt--.oo e- Aot = 0 and limt--.= te- Ast = 0, we have
(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.
Simulation Study: Pendulum of Variable Length

To verify the effectiveness of the proposed approach, the developed adaptive
controller is applied to a pendulum plant with a variable length l (e) as shown
in Figure 4.1 [187]. The plant dynamics can be expressed in the form of system
(4.1) with
(4.43)
1
b(x) = ml6~(x1) (4.44)
where 9 is the gravity acceleration, u = T, d(t) = cos(3t) cos Xl, ~(xd =

(COSXl +lo/ll)2lUl6, and x = [Xl,X2V = [e,e]T. The initial states [Xl(O),
i
X2(0)]T = [O,O]T, and the reference signal Yd = sin(t). The operation range
of the system is chosen as
58 Chapter 4. S1SO Nonlinear Systems
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
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 ~ ~ $ ~ ~ ~
(a) Tracking error Y - Yd (b) Control input u(t)
10
:IS 45 50
(c) HONN weight IIWII (d) States Xl ("--") and X2 ("- -")
Figure 4.2: Responses of the robust adaptive NN controller ( c 0.25 and

r = diag{2.0})
4.2.3 MNN Control Based on (4.10)

In this section, controller design based on MNN parameterization is presented
based on the DFC input u* in (4.10). In view of the property that b(x) depends
on X only, the DFC input u* (4.10) may be expressed as a function of x, e s and
Figure 4.3: Responses of the robust adaptive NN controller( E. 0.25 and

r = diag{6.0})
(a) Tracking error Y - Yd (b) Control input u( t)
Figure 4.4: Responses of the robust adaptive NN controller( E. = 0.05 and

r = diag{2.0})
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.,
* = -l'w[(S - S'VTz)e s + 6wleslW} (4.50)
~ = -rv [iWTS'e s + 6vlesiV} (4.51)
where I'w > 0, rv rr

= > O.
Substituting (4.48) into (4.6) leads to
es = a(x) + v + b(x) [WTS(VTz) - ~ + ur ] + d(t) (4.52)

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
+b(x)[wTs(~rTz) +ur -u~(z)] +d(t) (4.53)
Substituting (4.47) into the above equation yields
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,
its time derivative along (4.54) is

d(t) ]
Vs = - C
2
es _ _ e_
2
s _
cb2(x)
+ [wTS(frT '-",) + u r _ W*TS(V*T z ) - II.
r-l
+ --
b(x)
e
S
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
00 -_ - - + y r111W*11 + J.Lo + -do

1.2241 (4.56)
Ow bo
Since e; = 2b(x)Vs 2:: 2boVs, inequality (4.55) shows that
Multiplying by e~t, it can be re-written as

d ~ EO 2 ~t
dt (Vse t) ::::
E 2"0 oe •
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)
It follows from bo :::: b(x) :::: bl that
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)
whose bound can be found as
11((t)1i :::: ko li((O)lie->-'ot + ko lot e->-'o(t-r)les(T)ldT
:::: koli((O)11 +~: ~~ [e~(O) +10206], 'lit 2:: 0 (4.61)

From (4.5) and e = [(? enjT, it is shown that en = es - AT (,. Therefore,
IIxll S (1 + IIAII)II('II + lesl + Ilxdll S R(c, c:, x(O)) , (4.62)
where
- - b1
+ c:2aol + c
2
R(c, c:, x(O)) = koll(,(O)1I + kl" bo [e~(O)
with ko = ko(l + IIAII) and kl = 1 + ko//\o.

Step 2 To complete the proof, we show that under the following conditions,
states x indeed remain in 0 for all time. Define
00 = {x(O) I {x IlIxll < R(O, 0, x(O))} co} (4.63)
For x(O) E 0 0 , define a positive constant
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
c:* = sup {c:1 {x

cER+
I II xII S R(c, c:, x(O)) } co}
In summary, for all initial conditions x(O) E 0 0 and W(O) E 8 w , and the
desired signal IIXdll :s
c·, if the design parameter c: c:*, then the system :s
states x stay in 0 for all time. This completes the proof. <>
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
multilayer NNs, no theoretical result is available in the literature to guarantee

such a condition. The control objective here is to make the system output track
the given reference under a bounded control action, the convergence of the NN
weight estimates is not necessary.
For a given neural controller in the form of (4.48), if the reference signal
Xd and parameter e have been determined and condition (ii) in Theorem 4.3 is
satisfied, a compact set may be defined
no := {x(O) I {x I IIxll :S R( c, e, x(O)) } c ~ , 'dt :::: 0 }

Then, we say that no
is the maximum region of x(O) for this controller, i.e.,
under the given control objective, all the system starting from x(O) E can no
be stabilized in 0 by the proposed controller. On the other hand, if the initial
condition x(O) E 00 and the design parameter e have been determined and
conditions (i) and (ii) in Theorem 4.3 are satisfied, we may define
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
(i) the mean sq-uare tracking error bound is
'ilt >0 (4.64)
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
(ii) the Loo tracking error bOltnd is
(4.65)
Proof Theorem 4.3 has shown that x E 0, Vt 2: O. Therefore, NN approxima-

tion (4.47) is valid. Consider the Lyapunov function candidate
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
2WTW = IIWII 2 + IIWll 2-IIW*1I2 2: IIWII 2-IIW*1I2 (4.68)

2tr{VTV} = IJVII~ + IJVII~ -IIV*II~ 2: IJVII~ -IIV*II~ (4.69)
and noting (3.43) and (4.49), we obtain
t\ :s - e;c (IIZWT8'11 2 + 118'VTZII 2+ 1) _-.iL + d(t) es

F cb (x) b(x)
2
_0; 1es1(IIW1I2 -IIW*1I2) _ 0; lesl(IJVII~ -IIV*II~)

+Iesl (IIV*IIFIIZWT 8'IIF + IIW*1I1I8'VTZIl + IIW*1I1 + IJ-Ld) (4.70)
2
11V*IIFIIZWT8'IIFlesl :s e; IIZWT8'1I~ + ~IIV*II~
(0; IIW*1I2 + 0; 11V*1I~)lesl :s ~~ + cO;uIlW*1I4 +co;IIV*II}
3
(IIW*lIl+IJ-LII)lesl:S 8c e;+c(IW*li+ 2J-Lr)
2
IIW*IIIIS'VTZiliesl :S e; IIS'VTZII 2+ ~IIW*1I2
d(t) e; c 2
b(x) es :S cb 2 (x) + 4d (t)
and IJ-Ld :S J-Lo, Id(t)1 :S do, inequality (4.70) can be further written as
(4.71)
with
f30 = ~IIW*112 + ~IW*II} + Ilw*lli + o;uII W *114 + o~IW*II}

d2
+2/16 + ~
4
(4.72)
Now, integrating (4.71) over [0, t]leads to
it e;(T)dT :s -it 2sVl dT + 2s2f3ot:s 2sVl (0) -+ 2s2f3ot (4.73)
with
(4.74)
Since the tracking error el = H(s)e s , we have
it ei(T)dT :s Cl it e;(T)dT -+ C2 :s 2SCl Vl(O) + 2S2clf30t + C2 (4.75)
with computable constants Cl, C2 > 0 [74]. Dividing (4.75) by t, we arrive at

(4.64). Because lel(t)1 :s
11((t)lI, the Loo tracking error bound (4.65) can be
obtained from (4.61) directly. <>
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
which confirms that the tracking error converges to an s-neighborhood of origin

from the mean square point of view.
Remark 4.12. Theorem 4.4 also provides some possible methods for improv-
ing the control performance. Both performance criteria (4.64) and (4.65) can
be improved if the initial values ((0) and es(O) are set to zeros by appropriately
initializing the reference trajectory :/;d(O) using the method developed in [104].
The NN weight errors W(O) and V (0) are, in general, not possible to be ze-
ros because the ideal weights W* and V* are not available. Nevertheless, it
follows from (4.74) that the constant V1 (0) may be reduced by choosing large
adaptation gains r tv and r v'
Remark 4.13. In view of (4.72), a smaller f30 can be obtained by choosing

smaller btu and bv , and may result in a smaller tracking error. It is worth notic-
ing from Lemma 3.11 that too small btu and bv may not be enough to prevent
the NN weight estimates from drifting to very large values in the presence of the
NN approximation error or external disturbances. In this case, IV and V might
increase to very large values, which result in a variation of a high-gain control.
Therefore, in practical applications parameters btu and bv should be adjusted
carefully for achieving suitable transient performance and control action.
Simulation Study: Inverted Pendulum
e
m
u
M
Figure 4.5: Inverted pendulum plant
To verify the effectiveness of the proposed approach, a simulation study

for an inverted pendulum control is performed. The model of the inverted
pendulum [19] shown in Figure 4 ..5 can be written in the form of system (4.1)
with
. mLx~ sin(2xt} COS Xl
gsmxl - 2(M+m)
a(x) = b(x) - M+m
L(1 _ mcos 2 Xl) - L(1 _ mcos 2 Xl)
3 M+m 3 M+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;
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
The parameters of adaptive laws (4.50) and (4.51) are chosen as
f w = diag{1O.0},fv = diag{20.0}, 8w = 0.5, and 8v = 0.25
such that (Xl, X2) E S1, 'Vt ~ O.

Because the plant is a second order system, we have ( = Xl - Yd. If choose
A = 3.0, then ko and /\0 in (4.61) are ko = 1.0 and /\0 = 3.0. From (4.61), the
bound of Xl is calculated as
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
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.
4.3 VSC - Semi-Global Stability

In the previous section, regional stability results are obtained because Assump-
tions 4.2-4.3 are only valid over a compact set Sl. In the following, we consider
the adaptive control problem for system (4.1) which satisfies the following as-
sumptions on the whole space.
Assumption 4.5. The sign of b(x) is known and there exist two constants bo
and b I such that bI ~ Ib(x)1 ~ bo > 0 for all X E R n
Assumption 4.6. ab(x)/ax n = 0, \:Ix ERn.
The adaptive control approach developed in the previous section applies a

robust control term U r in (4.48), which is very similar to a high gain controller
4.3 VSC - Semi-Global Stability 71
0.1,
0.05
-0.05
-0.1
-0.15
" "
0 10 20 40 50 70 80 90 100
(a) y("-") and Yd ( H_

-" ) (b) Xl("-") and xz("- -")
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
(c) Tracking error Y - Yd

80 90 .00 0 10 20
'" 40 50
(d) Control input u(t)

60 70 80 90 .00
1.2
0.
0.6
0.4
02
~ " 40 50 W 70 80 90 _
(e) MNN weight IIWII (f) MNN weight I!VIIF

Figure 4.6: Control performance when c = 0.1 and l =5
72 Chapter 4. S1S0 Nonlinear Systems
(a) Tracking error Y - Yd (b) Control input u( t)
Figure 4.7: Control performance when £ = 0.1 and l = 10.
(b) Control input u(t)
Figure 4.8: Control performance when E = 0.05 and l = 5.

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 ~
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.
4.3.1 VSC-based Adaptive NN Control Design

In this section, a variable structure control scheme is used for the adaptive
NN controller design. The main objective of applying VSC is to eliminate the
effects of the residual term du of MNN approximation (3.42) such that a semi-
globally stability can be achieved. Our design is based on the DFC input ui(z)
given in (4.46).
According to the approximation property of neural networks in Chapter 2,
the smooth function ui(z) may be approximated by multilayer neural networks
(3.31) as below
ui(z) = W*TS(V*T z) + J.Ll , (4.77)
where z = [zT 1jT.
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)
where
Ie = [kr, k~, kilT = [IIV*IIF, IIW*II, IIW*lh + Ito + ~:r

1f; = [llzWTS'IIF, lis - S'VTzll, It
We present the NN adaptive controller as
U = -kse s + W T 8(VT-z) + U vsc
A A
(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
Slx= {X(t) IXnl :$ koilAIl (11((0)11 +
1I((t)11 :$ ko (11((0)11 + (4.85)

where ( = [Xl, X2," . ,Xn_l]T and ko, '\0 and VI (0) are positive constants.
Proof (i) Consider the Lyapunov function candidate
VI = Hb~;) + wTr~lw + tr{VTr;;-lV} + k T r k k] 1 (4.86)
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)
and 0 :S VI (t) :S VI (0), vt 2:: O. We know that es E L2 Leo, the estimates n

W, (I, k and states x are all bounded. This implies that the control input u
in (4.79) is also bounded, which leads to es E Leo. Since e s E L2 Leo and n
es E Leo, by Barbalat's lemma we conclude es - t 0 as t - t 00, which means
that the tracking error converges to zero asymptotically.
(ii) Considering (4.59)-(4.60) and applying the following Schwartz inequality
[74J
(4.89)
we have
1I((t)11 :S koll((O)lle- Aot + ko (it e-2AO(t-T)dT) t (it e;(T)dT) t
:S koll((O)lle- Aot + v~:o (it e;(T)dT) t

Integrating (4.87) over [0, t], we obtain
(4.90)
with
Therefore,
11((t)II ~ koll((O}11 + ko (4.91)
Considering VI(t) ~ t-TI(O) and (4.86), we have
(4.92)
From (4 ..5) and x = [e xn]T, it is shown that

(4.93)
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 .
4.3.2 Elimination for Controller Chattering

It should be pointed out that controller (4.79) is discontinuous due to the use
of sign function sgn(e s }. The switching term U vsc (4.80) introduces a high-
frequency signal to the plant which may lead to control chattering in practice
and excite possible unmodeled dynamics. To avoid such a problem, a smoothed
VSC control can be applied
(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)
K= rk(1jjes -akk) (4.99)
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.
Proof Differentiating (4.86) along (4.81) and (4.97)-{4.99), we obtain

. 2 e; d{t)
VI = -kses - cb 2 (x) + uses + [du + b{x) - p·LJe s
+l<T1,ble s l - awWTW - avtr{VTV} - akl<T k
It follows from (4.78) and (4.94) that
') e2
VI ::; -kse; - cb 2 (x) - J('T 1jJes
-
+ J(* T 1ple l + J(- T 1jJes
s
-
-T' -T' -T'

-a w W W - avtr{V V} - akJ( J(
*T -
::; -kse s + J( (1jJle s l -1jJe s )
2
-T' -T' -T'

-awW W - avtr{V V} - akJ( J( (4.100)
Using the property (4.96), we have
J(*T (1jJle s l- {;e s ) ~ O.2785Es(k~ + k; + k;)

Since
2WTW ~ IIWII 2- IIW*112
2tr{VT~"} 2: liVll} - IIV*II}
2KT k 2: IIKI12 -11J(*11 2
inequality (4.100) can be further written as
(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
D= {(e s, W, V) I kse; + a; IIWll + ~v liVll} + a; IIkl1

2 2 ::; f30}
According to an extension of Lyapunov theorem [137], we conclude that all the
system signals are uniformly ultimately bounded, and there exists a compact
s1 x such that x E s1 x , Vt 2: O. Since VI ::; 0 once e; 2: f3o/ks, the tracking error
will converge to a small neighbourhood of zero, whose size is adjustable by the
design parameters ks, cs, a w , au and ak. <>
Remark 4.16. In the above modification, the bounds of the system states
x are not obtained explicitly. In order to guarantee the NN approximation
hold for all time, the node number l of neural networks should be chosen large
enough such that Assumption 4.4 is satisfied on all possible operation space.
4.4 Conclusion 79
4.3.3 Simulation Study

The proposed VSC-based adaptive controller is applied to a pendulum plant
with a variable length l(O) [6] as shown in Figure 4.1. The plant dynamics is
expressed in the form of system (4.1) with a(x) and b(x) given as (4.43) and
(4.44).
The initial states of the system are [Xl (0), X2(0)jT = [0, ojT, and the control
objective is to make the output y track the reference Yd = ~ sin(t). Clearly,
Assumptions 4.5-4.6 are satisfied because b(x) only depends on Xl and away
from zero for all Xl E R.
MNNs used in the controller contain 10 hidden neurons with 'Y = 1.0 in
the activation functions. The parameters of controller (4.79) and adaptive
laws (4.97)-(4.99) are chosen as ks = 1.0, A = 1.0, r w = rv = diag{lO.O},
rk = diag{0.2}, CYw = 0.02, CYv = 0.002, CYk = 0.2 and es = 0.01. The initial
conditions are taken that k(O) = 0.0, and the elements of W(O) and V(O) are
chosen randomly in the interval [-0.5,0.5].
Figures 4.1O(a) and 4.1O(b) show that the transient response of the system
output is quite well and the tracking error Y - Yd converges to a small region
around zero after ten seconds. The control input u( t) given in Figure 4.10( c)
does not exhibit chattering due to the use of the smooth switch function tanh(·).
The boundedness of the NN weight estimates W, V and the sliding control gains
are shown in Figures 4.10(d)-4.10(f).
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.
06,---.--.---.--.---,----,
0.2
-0.2
.,"
...... 1. .. ~ I
-0.4
\ :I
\]
-O.60~--:---,':---.L----L---.L---'---"o
(a) y ("-") follows Yd ("- -") (b) Tracking error Y - Yd
(c) Control input u(t) (d) NN weight IIWII

'.5,---,----..----.--.---,-----,
2.5
1.5
O.SO':-------':---":-----'-15----'-ro----'-25---".1O
(e) NN weight IIVIIF

Figure 4.10: Responses of VSC-based adaptive NN controller
Chapter 5
ILF for Adaptive Control
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

82 Chapter .5. 1LF for Adaptive Control
strict-feedback nonlinear systems and MIMO nonlinear systems. It is shown

that asymptotic tracking is achieved with guaranteed control performance. As-
sumptions 4.3 and 4.6 in Chapter 4 are removed due to the use of the integral
Lyapunov function.
5.2 Matching SISO Nonlinear Systems

To illustrate the main idea, consider the following matching 8180 nonlinear
systems in Brunovsky form
Xi = Xi+l, i = 1,2, ... ,n - 1

Xn = f(x) + g(x)u (5.1)
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)
where constant /\ > 0 and A = [/\n-l (n - 1)/\n-2 ... (n - l)/\jT.

From (5.1) and (5.2), the time derivative of es can be written as
es = f(x) + g(x)u + II (5.3)
where II = _y~n) + [0 AT]e.

5.2.1 Integral Lyapunov Function

For convenience of analysis, let Xn-1 =
[Xl x2 ... Xn _1]T, 1.11 = y~n-1) -
[AT Ole. Then it can be seen from (5.2) that Xn = es +1.11. Let g;;l(x) = ;t:?,
where a(x): R n - t R+ is a smooth weighting function to be specified later.
For notation g;;l(x), g-l(X) indicates the i[xy, and the subscript (*)c. denotes
the multiplication function. For simplicity, denote
Definition 5.1 (Weighted Control Lyapunov Function). For a bounded

reference vector Xd, a scalar integral function [48]
\I~ = foes ag;-1(xn_1' a + v1)da (5.4)
is called a weighted control Lyapunov function (WCLF) for system (5.1), if

there exist a smooth function a(x) and a control input u such that Ve satisfies:
(1) Ve is positive definite in the filtered error e 81
(2) Ve is radially unbounded with respect to e s , i.e., Ve - t 00 as lesl - t 00,
and
(3) Ve < 0, \fes 1: 0.
In addition, a(x) is called a weighting function (WF).
For example, if g-l(x) = exp( -X n )(Ol + x;) with constant 0 1 > 0, then we
may choose a(x) = exp(x n ) which leads to
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 .
5.2.2 Choice of Weighting Function a(x)

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.
In fact, if we choose a(x) = 1'g(x), assuming that g(x) is known, then
g;-l(X) = 1', subsequently, the WCLF (5.4) becomes
Ve = l' i
o
e, l'
ada = -e s
2
2
(5.5)
84 Chapter 5. ILF for Adaptive Control
which is exactly the conventional quadratic function.

Let us first show that for system (5.1) satisfying Assumption 5.1, there
indeed exist a WF a(x) and a control input u such that Ve defined in (5.4) is
a WCLF. For convenience of analysis, let
then p(O) = O. Define
(5.7)
then, by the fundamental theorem of calculus, we have
lim h(Z) = 8p(e s )j8e s I (5.8)

es->O a(x) es=O
Therefore, h(Z) is smooth and well defined.

Now, we are ready to present the following lemma that provides a desired
feedback control for achieving the control objective.
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
u* = _ k(t)e s _ f(x) _ h(Z) (5,9)

a(x) g(x)
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)
Because 8g;;1(x n _1,0- + V1)/8v1 = 8g;;1(xn_1'0- + v1)/80- and v = -1i1' it

follows that
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-
~ -y [ug;'(xn-" u + y,{ - 1" g;'(xn-" u + y,)du1

= -ve s g;;1(x) +v l e
• g;;1(Xn_1' (1' + 111)00
Substituting the above equation into (5.10), we obtain
. [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
leo k*e~(r)dr ~ lIe(O) - Ve(oo) ~ Ve(O)

and hence €s E L2. Since x, Xd E Leo, it is clear from (5.2) that es E Leo.
Using the conditions of x, Xd, es E Leo, and the smoothness of system functions
l(x), g(x) and h(Z), it can be seen from (5.3), that es E Leo. According
to Barbalat's Lemma (Lemma 2.2), e s E L2 and es E L oo , we conclude that
e s --+ 0 as t --+ 00, subsequently, we have limt-+oo e1(t) = O. <)
The apparent and convenient choices for a(x) are 1 and y(x) for general
nonlinear system. For any other forms, it depends on the properties of the
physical system under study.
DFC1: Weighting Function a(x) =1

When we choose a(x) = 1, g;;l(x) = g-1(x) and equation (5.4) becomes the
following smooth integral scalar function
Ve = l e
• {1'g-1(X n _1, {1' + IIdoo (5.13)
By Mean Value Theorem (Theorem 2.5), lie can be re-written as
v:e = A8e~ , As E (0 ,1 ) (5.14)

g(X n-1, Ases + vJ)
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
+vg-1(xn_l,a+Vl)]da, PI(O) =0 (5.15)
hl(Z) = PI(e S ) , Z = [T
X ,e s , v, VI
]T (5.16)
es
(5.17)
which shows that hI (Z) is smooth and well defined.

Accordingly, we have the following corollary that provides a desired feedback
control for achieving the control objective.
Corollary 5.1. For system (5.1) with Assumptions 5.1-5.2 satisfied, if x E
L oo , and the desired contin·/tous Lyapuno-v-based controller is chosen as
* f(x)
U 1 = -k(t)e s - g(x) - h1(Z) (5.18)
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.
DFC2: Weighting Function a(x) = g(x)

Another convenient choice for a(x) is the upper bound of g(x), i.e., g(x). Let
g;;l(x) = g;;.~x(x) = ~ and a smooth scalar function
Ve = foes ag;;'~x(Xn-I' a + vI)da (5.19)
Applying Lemma 2.6, function Ve(es,xn-l,Vl) can be re-written as
r1[)Ve(W,Xn-l, vr} I es2 ior Ogmax(Xn-l,

1
Ve = e s io
-1 - )
OW dO = Oe s + v1)dO (.5.20
o w=(Je s 0
Noting that 1 :S g;;'~x(Xn-l' Oe s +V1) :S g(Xn-l, ()es + V1)/gO (Assumption 5.1),

we have
(5.21)
.5.2 Matching SISO Nonlinear Systems 87
Therefore, Ve is a positive definite and radially unbounded function with respect

to e s . The corresponding equations from (.5.6)-(.5.8) becomes
(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)
= g~x) 11 g;"~x(Xn-l' (Jes + vl)d(J (.5.29)
Thus, (5.23) can be further written as
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
+ g~x) 11 g;"~x(Xn-l' (Je s + vl)d(J (5.30)

88 Chapter .5. ILF for Adaptive Control
Similarly, we have the following corollary that provides a desired feedback con-
trol for achieving the control objective.
Corollary 5.2. For system (S.l)with Assumptions 5.1-5.2 satisfied, if x E

L(X)1 and if the DFC is chosen as
'" __ k(t)e s _ f(x) _ h (Z) (5.31 )

u max - g(x) g(x) max
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.
DFC3: Choice of a(x} for LIP parametrization

Most available adaptive controllers deal with control problem of systems with
nonlinearities being linear in the unknown parameters. In practice, however,
nonlinear parameterization is very common in many physical plants. Adaptive
control for nonlinearly parametrized systems is an interesting and challenging
problem in control community. The global output feedback control for systems
containing nonlinear parametrizations is investigated in [127], which is designed
using high-gain adaptation and applicable to set-point regulation problem. For
a class of first-order nonlinearly parametrized models similar to those arising
in fermentation processes, an interesting adaptive control scheme is provided
with three unknown parameters (two of them do not enter linearly) [14]. The
key points of this design method lie in the appropriate parametrization of the
plant and the suitable choice of a Lyapunov function with a cubic term for
developing the stable adaptive controller.
If g(x) > 0 is linear in the unknown constant parameters, a smooth function
a(x} can be found such that Ve satisfies Conditions (1) and (2) in Definition
5.1. For the sake of convenience of analysis, system (5.1) can also be written
as
Xi = Xi+l, i=1,2,··· ,n-1
:En = g(x) [fg(x) + u] = r:(x) [fg(x) + u] (5.32)
Y= Xl
where f 9 (x) = ful

g(x) .
One of the benefits of converting description (5.1) into (5.32) might be that,
for some systems, fg(x) and g-l(X) E C l can be expressed in the linear-in-the-
parameter (LIP) form for ease of controller design. For clarity, suppose that
we have the following descriptions
fg{x) = WYStg{x) + J6{x) (5.33)

g-l{X) = W;T Sg{x) + go (x) (5.34)
where, depending on different cases, we have
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)
From (5.6), we have
(5.36)
where
From (5.7), we have
(5.37)
where
h (Z) _ W;1'
11 - ()
esG x
ie, [
0
0' ~
~
i=1
asg(Xn-l, 0' + VI) .
aXi XH 1
+vS;; (Xn- l , 0' + VI)] dO' (.5.38)
h (Z) _ _
12 - ()
1_1e s
[
0' ~
~ asgO(Xn-l, 0' + VI) .
aXi Xt+1
esG X 0 i=1
+vS;;O(Xn-I,O' + VI)] dO' (.5.39)
From (.5.9), we have the corresponding controller
where bounded continuous function k(t) :::: k* > 0 with constant k*, W*T =
[wy, W;T], and by applying Lemma 2.6,
ST(Z) = [[Sf (x)],

9
[~()
esG X
t, [0' ~ asg(xn-;/, 0' + vd
Jo i=1 Xi
Xi+1
+vS;; (Xn-I, 0' + VI)] dO'] T]
= [[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
+vS;;O (Xn-I, 0' + VI)] dO'
-fg(,)
- JO X
_1_1
+ G ()
X 0
1 [O ~asgO(Xn-I,OeS+vI),.
es ~
i=1
a Xt+1
Xi
+vsgo (X n-l, Oe s + VI)] dO (5.42)

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
Corollary 5.3. For system (5.1) 'with Assumptions 5,1-5.2 satisfied, if x E

L oo , and the desired continuous Lyapunov-based controller is given by (5.40),
weighting function a{x) renders Ve in (5.4) satisfying Conditions (1) and (2) in
Definition 5.1, then the system tracking error converges to zero asymptotically.
If equations (5.33)-{5.34) represent linear function approximators, then the
residual function approximation errors !J(x) and 9o{x) are unknown, subse-
quently, hLO is not computable, and controller (5.40) is not feasible even if Wj
and W; are known. Noticing that !8{x) and 90{X) are small on XED, assume
that
1J8(x)1 < f8(x), 190(x)1 < go(x), IS~O(x)1 < S~O(x)

I
OSgO(x)
OXi I .. '90()._
< /::Yoi x, 2 - 1,2, ... ,n - 1 (5.43)
then we have hLO(Z) 2: hLO(Z) if hLO(Z) is chosen as
hLO(Z) 2: f8(x) + a(x) 11 [Be ~ S~o

s (X n -l, Be s + Vl)Xi+1
+vS&O (Xn-I, Be s + VI)] dB (5.44)
Now, we are ready to present the following corollary that provides a desired
Corollary 5.4. For system (5.1) with Assumptions 5.1-5.2 satisfied, if x E

L oo , weighting function a(x) renders Ve in (5.4) satisfying Conditions (1) and
(2) in Definition 5.1, and the desired continuous Lyapunov-based controller is
given by
.. 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)
with hlO(Z) defined in (5.42). Substituting the controller given in (5.45), (5.46)
becomes
Ve = -k(t)e~ + a(x)es[hlO(Z) - hlO(Z) sgn(e s )] :s -k*e~ :s 0 (5.47)

Integrating it, we have
and hence e s E L 2 . Since x, Xd E Loo, it is clear from (5.2) that e s E Loo.

Using the conditions of x, Xd, es E Loo, and the smoothness of system functions
f(x), g(x) and hlO(Z), it can be seen from (5.3), that e. s E Loo. According
to Barbalat's Lemma (Lemma 2.2), e s E L3 and e. s E Loo, we conclude that
e s -+ 0 as t -+ 00, subsequently, we have limt-+oo e1(t) = O. 0
5.2.3 Adaptive NN Control Based on DFCs

Apparently, different choices of a(x) lead to different DFCs. To illustrate the
main design approaches, LPNN, MNN and LIP parametrization techniques are
used to construct controllers for the three DFCs presented though they are not
exclusive.
LPNN Control Based on DFC (5.18)

In the case that the nonlinear functions f(x) and g(x) are unknown, the desired
controller u" defined in (5.18) is not available. A function approximator can
then be used to construct the desired control u" for realizing adaptive tracking
control. Any linearly parameterized approximator of the form
(5.48)
where the input vector Z E Rn+3, weight vector W E RI, weight number 1,
and basis function vector S(Z) = [S1 (Z), S2(Z), ... , Sl(Z)]T E RI, can be used
to approximate the unknown nonlinearity in the desired controller (5.18). As
discussed in Chapter 3, if 1is sufficiently large, then flnn(W, Z) can approximate
any continuous function to any desired accuracy over a compact set. We first
define a compact set
(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)
with III the function approximation error and W* the ideal weights defined as
W* := arg min {sup IWTS(Z) - f(x)g-l(x) - h 1

WEn", ZEn.
(z)l}
where Ow = {W IIIWIl $ w m } with a constant Wm > 0 specified by the
designer.
Because any smooth function is bounded over a compact set, it follows that
~f:? +h1 (Z) is bounded for all Z E Oz. Therefore, the approximation error III is
also bounded. The magnitude of III depends on the choices of the basis function
S(Z), the number l and the constraint set Ow. In general, the larger the weight
number l and the constraint set Ow are, the smaller the approximation error
will be.
With the function approximation (5.50), we present the adaptive controller
as follows
U = -k(t)e s - WT S(Z) (5.51)
,,/S(Z)es, if IIW" < Wm or
"W" = Wm and WTS(Z)e s <0
W= { - (5.52)
WWTS(Z) . A AT
,,/S(Z)e s - "/ , - -es , If!!WII
IIwII2 = Wm and W S(Z)e s >0
with the adaptive gain "/ > 0 (Lemma 3.5).
Theorem 5.5. For the closed-loop adaptive system consisting of plant (5.1),
controller (5.51) and adaptive law (5.52), if the gain in the controller is chosen
as
1
k(t) = - 1 + 2[f2(x)
1 + vZ] + [WTS(Z)
A ] Z
(5.53)
e go
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
(ii) the integral of the tracking error er is bounded by

lot er(T)dT $ c lot 1l~(T)dT +
2 Cl Cz (5.55)
with computable constants Cl, Cz ~ o.

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
es = f(x) + v + g(x)( -k(t)e s - WT S(Z)) (5.56)
The time derivative of V1 along (5.56) is
VI = esg(x) [ - k(t)e s - WTS(Z) + (f(x) + v)g-1(x)]

s -g(x)k(t) { e; -l~;~ [(If(x)1 + Ivl)g-1(x) + IWTS(Z)I]}
Noting (5.53), g(x) ;:::: go and the fact that (a + b + c)2 S 3(a 2 + b2 + c 2 ), we
have
Since V3sle s I S e; 12 + 3s 2 /2 and e; = 2V1 , it is shown that
Using the comparison principle in [107], we have
2V1 (t) - 3s 2 S [2\/1 (0) - 3s 2 ] e- f~ g(x(r»k(r)dT
Because g(x)k(t) ;:::: gals (as k(t) ;:::: lis obtained from (5.53)), the above in-
equality implies that
Therefore, the filtered error e s is bounded by
(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
les(t)1 S 2s, '<It;:::: T1 = max [0, 2€ In les(O)I]

go s
Equation (2.23) shows that the state errors
Hence, the system state x EO", Vt ~ T.

(ii) For system (5.1) with t ~ T, the function approximation (5.50) holds
for all Z E Oz due to x E 0" and :Ed E Od (Assumption 5.2). Taking a positive
definite function
-1
F = V.e + L2 WTW ' W = W - W·
where Fe is given by (5.13). Its time derivative is
In view of (5.50), we obtain
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
es [ - k(t)e s + J.Ll], if IIWII < Wm or

IIWII = Wm and
WTS(Z)e s ~ 0
WTWWT S(Z) •
es [ - k(t)e s + J.Ll ] - IIwll 2 es, if IIWII = Wm and
WTS(Z)e s > 0
When IIWII = Wm and WTS(Z)e s > 0, we know that WT\~~~S(Z)es ~ 0

(since 2WTW = IIWl1 2 + IIWI1 2 -IIW*112 ~ 0). Hence,
v~es[ -k(t)eS+J.Ll], Vt ~ T
By noting k(t) ~ 1/e: and J.Lles ~ e;/2e: + e:J.Lf /2, we have

2 2
V<
-
- es
e:
+ ,..."Ie s <
-
es + :'J1,21
_ 2e: 2 '
Vt >_ T (5.60)
Integrating (5.60) over [T, t]leads to
£ e;(T)dT :::; -it 2gVdT + g2 £JL~(T)dT

:::; 2gV(T) + g2£t JLr(T)dT (5.61)
Integrating (5.58) over [0, T], we obtain
(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~
Combining (5.61)-(5.62), and applying the inequality
e~(T) :::; e~(O)e-goT/e + 3g 2 (5.63)
which is obtained from (5.58), we have
Jrto e~(T)dT :::; g2 Jto JLl(T)dT + ~

go
[e~(O) + 3g 2] + 4q-lw~ + 3g 2T (5.64)
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
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
Si(Z) = II [s(Zj)]dj(i), i = 1,2,,,, ,l, with s(Zj) =

eZj
e
Z
J
_
+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 . J ..... 0.3
0.2
-0.2
....
0.1
-0.4 .... .....
t •••
. f
-O.S
,
f··· , ......... I· .....
-0.8
0
;'
10 20 30 .. 50
sec
60 ~ ~
(a) Output y("-") follows Yd("- -") (b) Tracking error Y - Yd
5 .
-5 ..
-10 ..
°0L---~'0--~L---~30--~"----~~--~~60 -15oL---~'0----'2O----30":-----':"-----:50L----!60
(c) Weight IIWII (d) Control input u(t)
Figure 5.1: Responses of adaptive NN control for the pendulum plant
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.
MNN Control Based on DFC (5.31)

For comparison, let us investigate controller design based on the desired con-
troller u* given by (5.31). Rewrite (5.31) as
* 1 g(x)f{x) _
umax = -=--() [-k{t)e s - () - g{x)hmax{Z)] (5.67)
9 x g x
Because the unknown function g(;t~Jx) + g(x)hmax(Z) in (5.67) is smooth, we

have the following approximation based on multilayer neural network
(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)
where IW = i£ > 0, IV = iI > 0, and O"w,O"v > O.

Theorem 5.6. For the closed-loop adaptive system consisting of plant (5.1),
controller (5.69) and adaptive laws (5. 70)-{5. 71), if the gain in the controller
is taken as
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.
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')';1 V}] (5.73)
where Ve is defined by (5.19). Its time derivative is
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)
In view of (5.69), we have
V = es[-k(t}e s - WTS(VTZ} + hmax(Z}1 + WT')'~1W + tr{VT')';1V}(5.75}

Using (5.68) and Lemma 3.9, we obtain
V = es[-k(t}e s - WT(S - S'VTZ} - WTS'VTZ - d,. + J.L1

+WT ')',~1W + tr{VT')';1V} (5.76)
Considering adaptive laws (5.70) and (5.71), and the fact that WTS'VTZ =
tr{VTZWTS'}, the above equation can be further written as
V = -k(t}e~ + (J.L - du}e s - O"wWTW - O"vtr{VTV} (5.77)
By noting (3.58), i.e., Idul ::; IIV*IIFIIZWTS'IIF + IIW*IIIIS'VTZIl + IW*b,
(5.72) and the properties that 2WTW ~ IIWll 2- IIW*1I2 and 2tr{VTV} ~
liVll} - IIV* II}, the following inequality follows
V ::; - :~ (1 + 11 Og(Xn-1' Oes + vddO + IlzwTs'lI} + IIS'V T z II2 )
+(lIV*IIFllzwTS'IiF + IIW*IIIIS'VTZIl + IW"11 + 1J.tl)lesl

- ')'; (11W11 2-IIW*1I2) - ~ (liVlI} -IIV*II}) (5.78)
Since
2
11V*IIFIIZWTS'IiFlesl ::; ~ IIZWTS'II} + ~IIV*II} (5.79)
(lW*h + IJ.LDlesl ::; 2;e + e(lW*12 + J.L2)

2
(5.80)
2
IIW*IIIIS'VTZIl ::; : IIS'VT ZII 2 + ~IIW*1I2 (5.81)
5.2 Matching S1S0 Nonlinear Systems 101
and !-L ~ !-L1, inequality (5.78) can be rewritten as
(5.82)
with constant
C1 = c(~IIW*112 + ~IIV*II~ + IW*li + !-Li) + 1'; IIW*1I2 + 1'; 11V*1I~(5.83)

From (5.21) and (5.82), we further have
(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
V(t) ~ V(O)e- A1t + C1 rt e-A1(t-r)dT ~ coe- A1t + ~1 , "it 20 (5.86)

Jo /\1
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. <)
Adaptive Control Based on DFC (5.40)

Now let us consider the case that fg(x) and g-1(X) can be expressed in the
linear-in-the-parameter (LIP) form as given by (5.33) and (5.34). If W* =
[WiT, W;TV were known, then the desired feedback control law (5.40) could
be directly utilized. In the case of unknown parameter W*, we employ its
certainty-equivalence controller as
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.
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)
with gain matrix r = rT > 0, The time derivative of V along (5.89) is
(5.91)
To eliminate W from V, the adaptive law can be chosen as
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.
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
(ii) for systems with n 2: 2, the Loo tracking error bottnd
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)
Integrating (5.93) over [0, tJ and applying (5.96), we obtain
iot ke;(r)dr ::; - io(t Vdr = V(O) - V(t)
: ; c; e;(O) + ~WT(0)r-1W(0), Vt 2: 0 (5.97)
from which L2 bound (5.94) can be concluded.

(ii) For the systems with order n 2: 2, Subsection 2.5.1 shows that ~ =
As( + bse s with stable matri.x As. It is not difficult to find two constants
ko, /\0 > 0 which depend on the design parameter A such that IleA.tll ::; koe-)"ot
[74J. The solution for ( can be written as
Therefore
Applying the following Schwartz inequality [74J

we have
Using (5.97), the inequality (5.95) follows. <>
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
S(Z)=_v_[t s 1 da t s (a+vI)Z da]T

esa(x) Jo 1 + (a + vl)2 Jo 1 + (a + Vl)2
= ~(
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
u = -k(l + x~)exp( -xDe s - WT S(Z) - xi
with adaptive laws

WI = IlV( tan- l Xz - tan- l VI)
Wz = IZV( es - tan- l Xz + tan- l VI)
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.
5.3 Backstepping Adaptive NN Design

In the early stage of the research, in order to obtain globally stable adaptive
systems, some restrictions on the plants had to be made such as matching
10 15 00 25 30 35 40
(a) Tracking error Y - Yd (b) W1 ( " - " ) and Wz ("- -")
(c) Control input u( t)
Figure 5.2: Responses of the adaptive system

condition, extended matching condition, or growth conditions on system non-

linearities [18,90, 133, 174J. In an attempt to overcome these restrictions,
a globally stable and asymptotic tracking adaptive controller was developed
for parametric strict-feedback systems using recursive adaptive backstepping
procedures [91], and the overparameterization problem was successfully elimi-
nated in [103J through the tuning function method. In this section, after the
introduction of the basic principles of backstepping design, adaptive backstep-
ping control approaches is presented for a class of strict-feedback nonlinear
systems based on on-line MNN function approximators parameterization and
the integral Lyapunov function [192J.
The control problem is considered for nonlinear systems transformable to
the following strict-feedback canonical form
~i ~ h(Xi.) + gi(Xi)Xi+1 1S;iS;n-1

{ Xn - fn(x) + gn(x)u (5.103)
Y = Xl
where Xi = [Xl,X2,'" ,XiJ T , X = [Xl,X2,'" ,xnJT ERn, U E R , Y E R

are the state variables, system input and output, respectively; hO and gi(')'
i = 1,2, ... , n are smooth functions and may not be linearly parameterized. For
this control problem, one of the main difficulties comes from unknown functions
giO. When gi(-) are known exactly, the control scheme provided in [147] can
be applied directly. Without such a knowledge, no effective method is available
in the literature.
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.
5.3.1 Adaptive Design for a First-order System

To illustrate the design method, consider a first-order system
(5.104)
where Ul is the control input.

Define Zl = Xl -Yd, g~ax(Xl) = 9l(Xl)/gl (Xl) and a smooth scalar function
(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)
where the smooth function is defined by
(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
derivative along (5.104) is
-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
Therefore, Vz1 is a Lyapunov function and the tracking error Zl -7 0 as t -7 00

asymptotically. <>
In the case of unknown nonlinearities h(Xl) and gl(Xl), MNNs can be
used to approximate the unknown function h 1 (ZI) in the desired controller ui
as follows
(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)
where neural network virl SI (lit 2 1 ) is introduced to approximate the nonlinear

function hI (Zl). We now specify the gain kl (t) and the adaptive algorithms for
adjusting the NN weights to guarantee the stability of the resulting systems.
Theorem 5.9. Consider the closed-loop system consisting of the first-order
plant (5.104) and controller (5.111), if the gain
k1(t) = :1 (1 +11 091 (OZl +Yd)dO +N1) (5.112)

with constant Cl > 0, Nl = IIZIW!Sill} + IlsivtZll12, and the NN weights

are updated by
(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
with positive constants Co, Cl and )1}.

Proof The proof includes two parts. We first suppose that there exists a
compact set Ozl such that ZI E Ozl, Vt 2: 0, and the function approximation
(5.110) with Assumption 5.4 hold. Later, we show that this compact set Ozl
do exist for bounded initial conditions.
V~1 = Vz1 + ~ [w!r;,/w1 + tr{vtr;IVd] (5.116)
where WI = WI - Wi, VI = VI - Vt and the positive function Vz1 is defined by

(5.105). Taking its time derivative along (5.109) and noting (5.111), we have
Vs1 = Zl [ - kl (t)ZI - W! 8 1(lit Zl) + hI (ZI)] + w!r~l W1 + tr{Vtr;1 11 d

Using (5.110) and Lemma 3.9, we obtain
VS1 = ZI [- kl(t)Zl - W[(Sl - SiVtZl) - w!siiiTzl - dUl + J.Ll]

-T l~ -T l~
+Wl r~ WI + tr{Vl r; Vd
Considering adaptive laws (5.113) and (5.114), and the fact that
AT;, - T - - T - AT;,
WI ~1 VI ZI = tr{Vl Zl WI ~I}
the above equation can be further written as
= -kl(t)Zl + (J.LI - du1 )ZI - alUWI WI - avtr{VI Vd

. 2 -T A -T A
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
-
the following inequality follows
VSI ::; - ;~ (1+ 11 Ogl(OZl +Yd)dO+ IIZ1W!S~II} + Ils~vtzlI12)

+(llVtIIFIIZIW!S~IIF + Ilwtlllls~vtZlll + Iwtl1 + 1J.l11) hi
-0'; (IIWl112 -llwtI12) - ~v (IIVIII} -IIVn~) (5.118)
(5.119)
(5.120)
with constant
CI = CI GII Wtli 2 + ~llVtli} + IWtli + iii) + 0"2~ I Wtl1 2 + ~; IlVtll}

From (5.106) and (5.121), we have
i sI
T ::; - 910 Vzl _ O"w IIW1 11 2 _ O"v II VI II} + Cl
Cl 2 2
Therefore
(5.122)
where constant
Solving inequality (5.122) using Lemma 2.9, we obtain

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.
5.3.2 Design for nth-order Systems

In this section, the design idea in Section 5.3.1 is extended to system (5.103)
of n-dimensions through the adaptive backstepping procedure [91,104J. Be-
fore proceeding, we present some notations used later. Define positive definite
functions g~ax(Xi) = !Ji(Xi)/gi(Xi), i = 2,··· , n, and let hi(Zi) be smooth
functions on compact sets Ozi for input vectors Zi. According to the NN ap-
proximation properties shown in Chapter 3, we have
VZi E OZ'i, i = 1,2, ... , n (5.125)

where Wt and V:* are ideal constant weights, and IJ-li I ~ Pi with constants
Pi > O. Let Wi and l% be defined as
It follows from Lemma 3.9 that the function estimation error can be expressed
as
(5.126)
where Si = SieVtZi); S: = diag{s~l;S~'),··· ,s~l} with S~j = s/(f/[jZi) =

d[s(za)l/dzalza==ilTzi,j = 1,2"" ,Ii: an~fli > l;'~nd the r~sidual te;ms dUi
'.3
are bounded by
(5.127)
The backstepping procedure contains n steps. At each step, an intermediate

control function (}:k shall be developed using an appropriate Lyapunov function
~k and the method similar to that of Section 5.3.1. The difference from Section
5.3.1 is that the weight tuning laws are not given explicitly at the steps (k <
n) until the final step (k = n). For simplicity, define M = II Zi wl S:II} +
IISrC'l Zi11 2 , i = 1,,,, ,n.
Step 1
Let us firstly consider the equation in (5.103) when i = 1, i.e.,
By viewing X2 as a virtual control, we choose a new error variable Z2 = X2 - (}:1
with (}:1 = U1 defined in (5.111), then
Taking V~l given in (5.105) as a Lyapunov function candidate, similar to that

of (5.109), its time derivative can be expressed as
l1z1 = Zl [91(xd(Z2 + (}:1) + h1(Zdl

Using (5.111) and (5.126), we have
(5.128)
Step 2
The equation in system (5.103) for i = 2 is given by
(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)
Choosing a positive definite function
Vs2 = VzI + 1 a92~ax

Z2
(Xl, a + O!l)da (5.131)
its time derivative becomes
Using (5.128), (5.130) and the facts that
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
If Ci2 is chosen as
where
with constant £'2 > 0, and noting (5.126), then we obtain

2
1182 = - L [kj(t)z; + 1f;jZj] + z21h(x2)z3

j=l
Step k (3:::; k :::; n - 1)
A similar procedure is employed recursively for each step. Consider the equa-
tion of system (5.103) for i = k given by
(5.135)
Similar to Step 2, by taking Xk+l as a virtual control, a control function Cik

can be found for (5.135). Let zk+l = Xk+l - Cik, we have
Choosing the following positive function
its time derivative becomes
Note that
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
Vsk = - L [kj(t)z; + 1/JjZj] + Zk-l!Jk-l(Xk-l)Zk

j=l
+Zk [flk(Xk)(Zk+l + O:k) + hk(Zk)] (5.137)
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
This is the final step. Consider Zn = Xn - O:n-l, the time derivative of Zn is

Taking the following Lyapunov function candidate
(5.140)
and following the same procedure as in step k, we obtain
where
Now we are ready to present the controller as
(5.142)
where
(5.143)
with constant cn > O.

Substituting controller (5.142) into (5.141) and using the approximation
(5.126), we finally have
n
Vsn = - L [kj(t)z; + 'ljJjZj] (5.144)
j=l
It is worth noticing that at each step the positive definite functions

have been used, which is the key point of the proposed design method. Ac-
cording to Assumption 5.3, we know that
1 ::; g~ax(Xi-1' U + Qi-1) ::; lJi(Xi-1, U + Qi-1)/ giO

and the following properties hold
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)
(ii) the following inequalities hold
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
with positive constants Ci, Vs(O) and '\s.

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
. ~[ 2 -T" ","T- "T",-T-

VS = -~ kj(t)Zj+Wj (Sj-Sj\lj Zj)Zj+WjSjVj ZjZj+(duj-J.1j)Zj
-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)
By considering (5.146), (5.152) and (5.153), we have

n
Vs :S ->'s Vs + 2: Cj (5.155)
j=l
where
It follows from (5.155) that
(5.1.56)
where constant
v. (0) ~ t, ([,(0) ug;;';.",Xj _, (0), u +aj _, (0))00
+~ [W[(O)r;;;l Wj(O) +tr{V[(O)r;l Vj (O)}] ) (5.157)
with g;~ax(Xj-1(0),a + Qj-1(0)) ::@: gln~ax(a + Yd(O)) for j = 1. Considering

(5.152), we know that
~ W 2 < 2Vs (t) ~ - ') 2Vs (t)

(5.158)
~ II JII - '\min(r;;;l)' ~
j=l
IIlftllF ~ />, m2n. (r-1)
v
It follows from (5.145) and (5.152) that
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
11 Ogj(Xj-1, OZj + Qj_1)dO 2: gjo/2
Then inequality (5.153) can be re-written as

Integrating the above inequality over [0, tj leads to
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*.
5.3.3 Controller Design with Reduced Knowledge

This section considers the neural control problem of system (5.103) when the
upper bounds 9i(Xi) in Assumption .5.3 are not available except that gi(Xi) ::::
giO > O.
Theorem 5.11. For strict-feedback system (5.103), if the controller is chosen

as
(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
a constant Co s'uch that the vectors Zj remain in the compact sets
where Co > 0 is defined in Theorem 5.10, and

(ii) the following ineq·uality holds
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)
Hence, 178 < 0 outside a compact set
D = {(ZI ... WI'"

, Zn, , Wn" t:'1 ... v.n ) I t, (:: + a~u !!Wj !!2 + ~u IIllill})
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
Integrating (5.167) over [0, tJ and dividing it by t, we arrive at (5.165). <>

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·
Remark 5.10. It is interesting to notice that adaptive NN controller (5.161)-

(5.163) with adaptive laws (.5.147)-(5.148) is highly structural, and independent
of the complexities of the system nonlinearities if the numbers of the NN nodes
are taken large enough. Such a structural property is particularly suitable for
parallel processing and hardware implementation in practical applications. In
addition, the controller, once implemented, can be applied to other similar
plants without repeating the complex controller design procedure for different
system nonlinearities. In this sense, it is a universal controller for a large class
of strict-feedback nonlinear plants satisfying the basic conditions (i.e., gi(Xi)
are bounded away from zeros and their signs are known).
5.3.4 Simulation Studies

To illustrate the design procedure and test the control performance of the
proposed method, we consider the following three second-order plants:
I;l :
{ Xl = Xle- O. 5x1 + (1 + Xi}X2

X2 = XIX~ + [3 + COS(XIX2)]U
Xl
{ X
= 0.5xr + In{lO + Xi)X2
2
= X1X2
l+X~+x~
+ (1 + e-X~-X~)u
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)
and NN weights WI, W2 and VI, V2 are updated by (5.147) and (5.148), respec-
tively.
05
-0.5
-I
-1.50'L--.L.'O---i20--30'---4Cl-'---L50--'OO~-7i-O- " , -10 10 20 30 40 so 60 70 80
(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 ("- -")
Figure 5.3: Simulation results for Plant El with Yd = sin(t)

Both neural networks W!Sj(V{Zj), j = 1,2 contain 10 hidden nodes (i.e.,

h = l2 = 10) and the coefficient in activation function s(z) = 1/(1 + e~''fZ)
is taken as 'Y = 3.0. The design parameters of the above controller are r vI =
r v2 = diag{1O.0}, r wI = r w2 = diag{1.0}, £1 = c2 = 1.0, (TwI = (Tw2 =
1 X 1O~2, (TvI = 1 X 1O~4, and (Tv2 = 1 X 1O~3. The initial weights WI (0) = 0.0,
W2 (0) = 0.0, and the elements of VI(O) and V2 (0) are taken randomly in
the interval [-0.5,0.5J. To show the learning ability of the proposed scheme,
without modifying any controller parameters, we apply controller (5.168) to
three plants EI-E3 whose dynamics and control objectives are quite different .
............
0.5 . .... ,':

••••
-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.
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·
(c) I! WI II ("-") and I!W2 11 ("- -")

Figure 5.5: Simulation results for Plant E3 with Yd = sin(t) + 0.5sin(3t)
Keeping all design parameters as before, we apply this NN controller to

plant E2 for tracking desired signal Yd = O.5[sin(t) + sin(O.5t)J. Again, the
initial response shown in Figure 5.4(a) is not satisfactory and a small tracking
error is achieved after several learning periods. Figure 5.4(b )-5.4( d) indicate
the bounded ness of control input and NN weight estimates. Finally, the same
adaptive controller is applied to Plant E3 for forcing the output Y track the
5.4 NN Control for MIMO Nonlinear Systems 127
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.
5.4 NN Control for MIMO Nonlinear Systems

Recent years have seen many significant developments in adaptive control for
single-input/single-output (SISO) nonlinear systems. For multi-input/multi-
output (MIMO) nonlinear systems, due to the couplings among various inputs
and outputs, the control problem is more complex and few results are avail-
able in the literature. One of the major difficulties comes from the uncertainty
in the input coupling matrix. Based on feedback linearization techniques, a
variety of adaptive controllers has been proposed for linearizable systems. In
these methods, for removing the couplings of system inputs, an estimate of
"decoupling matrix" is usually needed and required to be invertible during the
parameter adaptation. Therefore, additional efforts have to be made to avoid
the possible singularity problem when calculating the inverse of the estimated
"decoupling matrix". For example, projection algorithm is applied in [165J to
keep the estimated parameters inside a feasible set in which the singularity
problem does not happen. Although such a projection is a standard technique
in linear adaptive control, it usually requires a priori knowledge for the feasible
parameter set and no systematic procedure is available for constructing such a
set for a general plant [74]. In fact, even for SISO adaptive nonlinear control,
solving the control singularity problem is far from trivial (see [186] and the
references therein). For MIMO nonlinear systems, there is no effective scheme
available at the present stage. As an alternative approach, neural control de-
signs proposed in [23, 121J suggested that the initial values of neural network
(NN) weights are chosen sufficiently close to the ideal values such that the
inverse of the estimated "decoupling matrL'{" always exists, and therefore an
off-line training phase for NNs is necessary before the controller is put into the
closed-loop system. In [133], an interesting adaptive controller was developed
for multi-input parametric pure-feedback nonlinear systems. Global stability
and convergence of the tracking errors were achieved for all initial parameter
estimates lying in an open neighborhood of the true parameter space. For a
special class of MIMO nonlinear plants, rigid robot systems, adaptive neural
network controllers have been proposed using several nice properties of the
input decoupling matrL'{ of robotic systems [52,112]. Energy-type Lyapunov
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.
5.4.1 System Description

Consider MIMO nonlinear systems in a triangular form [.51]
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
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
with Aj >0 (5.170)
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:
(i) esj = 0 define time-varying hyperplans in RPj on which the tracking

errors ej,l converge to zeros asymptotically,
(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)
\lij = 1,2"" ,Pj, j = 1,2"" ,m, and
(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.
5.4.2 Lyapunov Function Design and Control Structure

Define positive definite functions g;:';'ax(Xj ) = gj(Xj)/gj,j(Xj ) and
(5.172)
(5.173)
with .\.\ Xl and nj-l = 'L{:;

Pi for j > 1. From system (5.169) and
definitions (5.170) and (5.173), the time derivatives of esj can be written as
j-l
esj = fj(X) + L9j,i(X)Ui + 9j,j(Xj )Uj - Vj (5.174)

i=l
'Lf:l (.) = 0 for j

. 1
where = l.
It is shown from (5.170) and (5.172) that Xj,Pi = esj + Vj. In the following,
we denote gj";'ax(Xj ) = gj1'l~axCXj, esj + Vj) for notational convenience. Define
smooth scalar functions
(5.175)
The second equality is obtained by changing the variable a = Besj. According

to Assumption 5.6 and Remark 5.11, we have
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 ).
Taking V zj (5.175) as the Lyapunov function candidates, their time deriv-

atives are
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)
The following two equalities hold
Vj [eSjg;;'~ax(Xj) - JI~ ]
=
o gjn~axCXj, 0" + vj)dO"
= vjesj [g;;'~ax(Xj) -11 gj';'axCXj , Oesj + vj)dO] (.5.178)
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)
where
j-1
hj(Zj) = 9jn;ax(Xj ) [fj(X) + L9j,i(Xj )Ui]
11
i=l
-Vj 9j:'ax(Xj , fJesj + Vj)dfJ + eSj¢j(Zj) (5.181)
¢+(z .. ) = {¢j(Zj), ¢j(Zj) ~ 0 (5.182)

J J 0, otherwise
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.
5.4.3 Adaptive MIMO Control Using MNNs

In the following, MNNs are used for function approximation, and the unknown
continuous functions hj(Zj) can be approximated as
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.
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'
and lj > 1; and the residual terms dUj are bounded by
ldujl :5liV/IIFIIZj WJSjIlF + II WIll II Sj vlZj II +IWIll (5.185)
We now present the adaptive NN controller
j = 1,2"" ,m (5.186)
where Zj = [X T ,vj,zij ,ul,U2,'" ,uj-l,ljT with Zl = [XT,Vl,zil, IjT for j =

1 and
kj(t) = ~ (1+ l\Jg;CXj,fJesjVj)dO+Ni) (5.187)
with Ni = IIZjWJSjll} + IISj"CjTZj Il2 and design parameters Cj > O. From

(5.186), it is clear that kj(t) can be viewed as the controller gains. Since
J; (Jg;('Xj, OeSj + vj)dO ;:: gOj/2 > 0 (using conditions gj(Xj) ;:: gOj in Remark
5.11), the integrals in kj(t) are always positive definite. It is shown later that
these integral terms are essential for capturing the explicit upper bounds of the
system signals. The following adaptive laws are used to update the NN weights
(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.
Theorem 5.12. Given nonlinear system (5.169) satisfying Ass'umptions 5.6-

5.8, controller (5.186) and weight updating laws (5.188) and (5.189), then for
bounded initial conditions,
(i) all signals in the closed-loop system are bounded, and there exist compact
sets Qx and Qwj such that the vectors Zj remain in
Qzj = {(X'Vj'Vj'U1'U2'''' ,Uj-1)lx E Qx,
(WI, Vd E Qwl> (W2' V2) E Qw2,'" ,(Wj - 1, 'Ci-d E Qw(j-1),
Xd(j+1) E Qd(j+1) } (5.190)
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)
with positive constants aj, bj , Cj, Vsj (0) and /\sj.

Proof (i) The proof includes two parts. We first suppose that there exist
compact sets H zj such that Zj E f!zj, Vt 2 0, and the NN approximation
(5.184) hold. Then, we prove that these compact sets f!zj do exist for bounded
initial conditions. Consider the augmented Lyapunov function candidates
~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
Noting (5.185), (5.187) and the properties that
2WTWj 2 IIWj l 2-IIWj*112, 2tr{il~} 2 II lI; II} -IIVj*II}

the following inequalities hold
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
Cj = CjGIIWj*1I 2 + ~IIV/II} + IWlii + fin + a;j II Wj* II 2
+
a2
;J I\Vj*II} (5.198)
Considering (5.176), (5.193) and (5.197), we further have
v"j(t) ::; VSj(O)e->'Sjt + +-Cj, Vt 2 0 (5.199)

/\5j
where constants
(5.200)
Since IIWj l12 ~ V~j(t)//\min(r~;) and IWill 2 ~ l!;,j(t)r\min(r;;}), the NN

weights (Wj , l7j) belong to the following compact sets
n .=
tv]
{(w ],
V)
]
IIIW'11 2<
] -
2[Vsj(0)
\.
/\11'z,'z,n
+ Cj//\d
(r- 1)
wj
,
Ilfjll~ ~ 2 [V8j (0) + ~!)..8j]} (5.201)

,\nin(rvj )
It is shown from (5.175), (5.176) and (5.193) that Vsj(t) ~ Vzj ~ e;j(t)/2.
According to Remark 5.12, the system states remain in the compact set
nx = {X Ilej,ij(t)1 :S 2ij-1 /\~j-Pj (2v;,j(0)e->,sjt + ~:~) 1/2,
i j = 1,2, ... , Pj, j = 1,2, ... , Tn, Xd(j+1) E nd(j+1)} (5.202)
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)
with computable constants aj, bj > O. Dividing (5.204) by t, we arrive at

(5.191). Considering (5.199) and lesj(t)1 :::; J2"\'sj(t), and using (5.171) for
i j = 1, inequality (5.192) follows. 0
Remark 5.13. It is shown from (5.191) that large initial errors esj(O), Wj(O)
and "Cj(0) may lead to large mean square tracking errors during the initial
period of adaptation. As time goes on, we have
·
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
guarantee/check such a condition. One of the advantages of the proposed con-

troller lies that the control objective is achieved without the requirement for
PE condition, which is very attractive for control applications.
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)
where x E Rn is the state vector, u E R is the input and y E R is the

output. The mapping fe, .) : R n +1 -4 R n is an unknown smooth vector field
and hO : R n -4 R is an unknown smooth function. The control objective is to
design a controller such that the system output y follows the desired trajectory
Yd. The main difficulty of this control problem is that the system input u does
not appear linearly, which makes the direct feedback linearization difficult.
Though it is true that by introducing an additional state, Xa = U = J~ v,
where v is the new control input, system (6.1) can be transformed into affine

140 Chapter 6. Non-affine Nonlinear Systems
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.
6.2 System Description and Properties

Let Lfh denote the Lie derivate of the function h(x) with respect to the vector
field f(x, u)
Lfh = a[h(x)] f(x, u)

ax
Higher-order Lie derivatives are defined recursively as L7h = Lf(L~-l h), k > l.
6.2 System Description and Properties 141
Let nx c Rn and nu c R be two compact sets such that x E nx and

u E nu. System (6.1) is said to have a strong relative degree p in U = nx x nu
if there exists a positive integer 1 ::; p < 00 such that
a[L}h] a[L/h]
a;;-=O, i=0,1, ... ,p-1, -au
-#0 (6.2)
for all (x, v.) E U [176].
Assumption 6.1. System (6.1) possesses a strong relative degree p = n,

V(x,u) E U.
Under Assumption 6.1, system (6.1) is feedback linearizable and the map-
ping (x) = [¢1(X),¢2(X), ... ,¢n(X)]T with ¢j(x) = LtIh,j = 1,2, ... ,n has
a Jacobian matrix which is nonsingular for all x E nx [77,128]. Therefore,
(x) is a diffeomorphism on n x . By setting ~ = (x), system (6.1) can be
transformed into a normal form
{ ~n~i == ~i+I'
b(~, u)
i=1, ... ,n-1
(6.3)
y = 6
with b(~, u) = Ljh(x) and x = -I(~). Define the operation domain of normal
system (6.3) as
Define the smooth function
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
class of systems. In addition, in the following design procedure we only need

the existence of Assumption 6.2, and b1 (x) is not required to be known a priori.
Define vector ~d, ~d and ~ as
(6.5)
and a filtered tracking error as
(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)
with A. and b as defined in (2.21).

From (6.3)-(6.6), the time derivative of the filtered tracking error can be
written as
(6.8)
Assumption 6.3. The desired trajectory vector ~d is continuous and avail-

able, and II~dll ::; c with c being a known bound.
6.2.1 Implicit Desired Feedback Control

We have the following lemma to establish the existence of an implicit desired
feedback control (IDFC), u*, which can bring the output of the system to the
desired trajectory Yd(t).
Lemma 6.1. Consider system (6.3) satisfying Assumptions 6.1 to 6.3. For a
given small constant e > 0, there exist a compact s'ubset cl>o C cl>(Ox) and a
continuo1£s input
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)
Subsequently, (6.10) leads to limt-+oo Iy(t) - Yd(t) I = 0 asymptotically.

Proof Adding and subtracting bl~X) e s to the right-hand side of (6.8), we obtain
(6.11)
Since 8vj8u = 0 and 8[b(~, u)l/8u ~ bo > 0, V(~, u) E [r, we have
8[b(~,u) + v] > b > 0 V(C u) E [r (6.12)

8u - 0 , <",
Using Implicit Function Theorem (Lemma 2.8), we know that there exists a
continuous trajectory dependent function QC(~, v), which guarantees
bee, QC(e, v» + /.I = 0, V(~, v) E G (6.13)

G = { (~, v) leE (nx ), lIe~lI:::; c}
If the IDFC input is chosen as u*(z) = QC(e,v),z = [e,vjT E nz c Rn +1 ,
then
b(~, u*(z)) + v = 0, V(e, v) E G (6.14)
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 (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
<1>0 = {e(o) I {e IleB(t)1 < les(O)1 , lI~dll :::; c} c (nx) ,

~(O) E (n x ) } (6.15)
Then, for all ~(O) E <1>0, we have' E (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.
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
u*(z) = aC(~, v), v = es _ y~n) + [0 AT1€ (6.16)

c
z = [e,vf E Oz C R n +1
such that for all ~(O) E <Po, the error equation (6.8) can be written in a linear
form
(6.17)
which leads to limt->oo Iy(t) - Yd(t)1 = O.

Proof The proof can be done similarly in Lemma 6.1 by just setting bl (x) = 1.
I)
Since ~ = (x), the continuous function aC(~, v) is also a function of x and
v. Re-write v = b'ix ) es + VI with VI = _y~n) + [0 AT]{ Taking the input of
the IDFC as
- [.T
Z - x ,es,VI ]T (6.18)
then the IDFC can be expressed as
u*(z) = a C (';, v), z E Oz C R n +2 (6.19)

Oz = {(x, e s , VI) I x E Ox, lI(dll:::; c}
Note that the IDFC can also be written as
u*(z) = a(';, v), (6.20)
where the compact set Oz is defined as
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 o. This is reasonable
because if the reference ~d is very large and out of the region (Ox), then o
will become an empty set and the tracking problem cannot be solved.
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
Ie = {exp( -esh), es > 0 (6.22)

• - exp(esh)' e s ::; 0
1 - exp{ -esh), es > 0
sat(es ) = sgn(es ) - Ie. = { -1 + exp (/) <0
e s 'Y , es _
(6.23)
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
u"(z) = a(~,v), v=kves+kvssat(es)-y~n)+[O AT]€ (6.24)

z = [e,v]T E Oz C R n +1
such that
(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}
yields V = -kve; - kvses sat(e s ). Considering e s sat(e s ) 2: 0, we have V ::;

-kve; ::; O. Since V ~ 0 and V ::; 0, this shows the closed-loop stability in the
sense of Lyapunov, thus, es is bounded. From (6.25), fo s is also bounded. For
~(O) E O~ and ~d E Od, where O~ and Od are two compact sets, therefore
(6.28)
Considering (6.23) and (6.25), we obtain that V is bounded. Using Barbalat's

Lemma [169J in connection with (6.28), we conclude that V goes to zero with
t - t 00 and hence limt-->oo e s = O. This implies that limt-->oo Iy(t) - Yd(t)1 = o.
o
6.2.2 High-gain Observer
When only the system output is measurable, we need an observer to estimate
the unavailable system states. In the following lemma, a high-gain observer
used in [10] shall be presented for estimating the output derivatives of system
(6.3).
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)
7rk+1 (k) - E'f/

./,(k+l) , k -- 1 , ... ,n - 1
7<-y
where
(6.30)
and 7/1(k) denote the kth derivative of 7/1,
(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.
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)
From (6.29) and the above equation yields

1r2. ..
- - y= -E7/J
E
By differentiating it and utilizing (6.29), item (i) follows.

The derivatives of the vector 1r = [1rl 1r2 ... 1rnV may be computed as
follows:
(6.32)
where A is the matrix corresponding to the homogeneous part of (6.29) and

independent of E, and b = [00 ... IV. Since ( belongs to the compact set {
and u is bounded, there exist constants Yj > 0 such that Iy(j) I :S Yj. Then,
for any 0 > 0, we may find a constant t* > 0 such that, for all t > t*, the first
term
E~ Aj exp ( ~t) [1r(0) + A-1by(0) + ... + Ej - 1 A -jby(j-l)(O)]
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)
~exp (~t) fat exp (~T) by(j)(T)dT < DjYj

Now fix an arbitrarily small 0*. Then for t > t*, we have 17/J(j) I ::; h j where
hj = B(Dj + o*)Yj with B the norm of the vector [1.\1 '" .\n-ll. As D j , 0*,
Yj, and B are independent of E, the proof is completed. 0
In Sections 6.3 and 6.4, we shall show the adaptive NN controller design
procedures based on the ideal control (6.24) and (6.20), respectively.
6.3 Controller Design Based on LPNN

For simplicity, in this section we will consider the use of the linearly parameter-
ized Gaussian RBF networks in the design of adaptive controller for non affine
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
b(E, u*) = b(E, u~n) + ~bt11nn C:u(z) + G u'nn C:u(z)

where
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
6.3.1 State Feedback Control

In this section, under the condition that the full state ~ of system (6.3) is
available for feedback, we proceed to design an adaptive controller using RBF
neural networks. Let the NN controller take the form
, ' T
U = unn(z) = W S(z) (6.36)
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
.1.b[.k) = 8(k)b(~, u) I k = 1,2,3· ..

Unn k!8u(k) U=U nn '
00
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)
Further, if 0 C Rn+l is a compact set, considering Assumption 6.1 and Lemma

2.1, we know that b(~, u) and .1.blUknn
) (k = 1,2, ... ,n) are Lipschitz for (~, u) E O.
Similar to the proof of Lemma 6.5, we can derive II~II :::; d1 + d2 1e s I with d1 and
d2 being positive constants over the compact set 0, and there exist positive
constants La to L5 such that
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
w = -("'oIIWII + fi:l!esl + "'2)S(z)es

-t5(IIWII + lesl + 1) IIS(z)lllesIW (6.41)
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)
Differentiating (6.42) along (6.38) and (6.41), we have
v = _WT (liollWII + lillesl + li2) S(z)e s - /j(IIWil + lesl + 1) IIS(z)II le s illWII 2

+es [ - kves - kvs sat(e s) + b(~, u~n) + 1/ + ~b~!n WT S(z)
- T
-DUnn W S(z) ]
Using (6.34), (6.39) and (6.40), we obtain the following inequality
v ::; - (KoIIWil + Kllesl + K2) WT S(z)e s - /j (IIWII + lesl + 1) IIS(z)llleslllWI1 2

-kv e; - kvses sat(e s ) + (€Cllesl + €C 2 ) lesl
+~b~ln WTS(z)es+ (L311WII + L41esl + L5) IW*TS(z)esl
+ (LoIIWII + Lllesl + L2) IIWlllesIIIS(z)11 (6.43)
From Assumption 6.2 and (6.40), we have

Using IIW*I/ ::; Wm and IIWII ::; IIWI/ + W m , (6.43) can be written as
v ::; [111:0 - L311/WI/ + 111:1 - L411esl + 111:2 - L51] IWTS(z)esl

-J(IIWII + lesl + 1) IIS(z)l/lesIIIWI1 2
-kv e; - kvs lesl + kves1e + (sC1les I + sC2) les I
s
+ (L3I1WII + L41esl + L5 )wmIIS(z)lllesl

+ (LoIIWII + Lllesl + L2) (IIWII + Wm) lesIIIS(z)1I (6.44)
Define the following positive constants
al = 111:0 - L31 + Lo (6.45)

a2 = 111:1 - L41 + L1 (6.46)
a3 = 111:2 - L51 + L2 + (Lo + L3)wm (6.47)
a4 = (L2 + L5)Wm
a5 = (Ll + L4)Wm
as = sup {IIS(z)ll}
zEn.
Equation (6.44) can be further written as
v ::; - (kv - sC1- a5aS) e; - (kvs - sC2- a4aS) les I + kvsesles

-IIS(z)1/ [Je;I/WI1 2+ Jle sll/WI/2 + JleslllWl/3]
+I/S(z)11 [alleslllWl12 + a2 e;I/WII + a3leslllWI/]
(6.48)
Thus, (6.48) can be expressed as
(c:C2)2
f32le s l + kvsesIes + -v:;;;-
2
V < -(31 e s -
where
(31 = k;s - c:C1 - (05 + ~!) as (6.49)
(32 = kv - (0 4 + 3~13 + ~J)os (6.50)
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. <)
Remark 6.4. If a high tracking accuracy is required, a large number of NN

nodes should be chosen such that c is small enough to achieve the desired
tracking performance. The parameters "'0,"'1,"'2 and 6 in adaptive law (6.41)
can also be designed to make a1,a2 and a3 in (6.45) - (6.47) small. Equation
(6.51) shows that, the larger /31 and /32 are, the smaller the tracking error will
be. Therefore the control performance are adjustable through the choices of
kv, "1, "'0, "':1, "'2 and 8.
Remark 6.5. The weight update law (6.41) is derived from the Lyapunov
method and the e-modification [134] term is introduced to achieve the robust-
ness in the presence of the NN approximation error. There is no requirement
for persistent excitation condition for tracking convergence. In addition, the
NN controller needs not to be trained off-line.
6.3.2 Output Feedback Control

When only the output y = 6 is measurable and the rest of the output deriva-
tives are not available, we use the high-gain observer (6.29) to estimate ~2' 6,
... , ~n to implement the output feedback control.
For ease of analysis in using observer (6.29), define the following variables
(6.52)
(6.53)
The NN controller based on observer (6.29) is given by
u = u~n(z) = WTS(z) (6.54)
The weight update law is chosen as
w= - ("'oIlWIl + "'llesl+ "'2 )S(2)e s
- [8 (IIWII + lesl + 1) IIS(2)11 + tlesl] leglW (6.55)
where E, "'0, "'1, "'2 and 6 are positive constants. The closed-loop error equation
(6.26) becomes
(6.56)
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
b(c;,u~n) = b(c;,u~n) + (Ab~L +OU~n) [W*TS(z) - WT S(2)] (6.57)
[.] , 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)
Proof The Taylor series expansion of b(c;, u) at u~n = WT S(2) is
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
IAb~lnn I :S a411 WII + a511c;II + a6

IOu~J :S a711WII + asllc;lI + ag
From (6.5) and (6.6), we can derive II~" :S d1 + d2 lesl with d 1 and d2 being
positive constants over the compact set D z , which means that positive constants
Lo to L5 exist such that (6.58) and (6.59) hold. <>
Theorem 6.2. Consider the closed-loop system consisting of system (6.3), ob-
server (6.29), controller (6.54) and adaptive law (6. 55}. Under the conditions
that
(1) Ass1t1nptions 6.1 - 6.3 and Assumption 3.1 hold,
(2) there exist three compact sets D w , Des and D~ such that W(O) E D w ,
es(O) E Des and c;(O) E D~.
for a sttitably chosen design parameter kv, the closed-loop system is SGUUB.
The tracking error can be made arbitrar-ily small by increasing the approxima-
tion accuracy of the neural networks and the high-gain llf. of the state ob-
server f47J.
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)
Differentiating (6.60) along (6.55) and (6.56) yields
v = _WT (lI:ollWII + II:lle l + s 11:2 )S(z)e s

-6(IIWII + lesl + 1)IIS(z)lIlesIIIWII2
-tit?s 12 II wl1 2 + es [ - kve - kvs sat(e) + f(x, u~n) + v]
For £ = Z - Eij;, from Lemma 3.1, we know that
S(£) = S(z) + ESt (6.61)
with St a bounded function vector.

From (6.52) and (6.57), the following equation follows
iT = -(KoIIWIl +lI:llesl +1I:2)WT S(z)e s

-o(IIWIl + lesl + 1)IIS(£)lIles IIlWII 2
_Ele s 12 11W11 2 - kv (e s - EATW) 2 -(e EATW) sat (e EATW)
kvs s - s -
+ [f(x, u.~n) + v] e - (~b~L + OU~n) [W*T S(z) - WT8(Z)] e
By using (6.23), we obtain
where eo is the natural exponential (i.e., eo = 2.7183). Considering (3.13),

v :-:; (IKO - L3111wII + IKl - L411esl + IK2 - L51) IW T S(2)e s l
-o(lIwll + lesl + 1)IIS(2)lIlesIIIWII2 + (cCllesl +cC2)lesl
_E!e s I2I1WII 2 - kv (e s - EATiJI f -kvs (Iesl -IEATiJll-i/eO)

+ [ (Lo + L3) IIwll + (Ll + L4) les! + L2 + L5] wm!IS(z)lIlesl
+ (LollwlI + L11esl + L2) IIWIIIIS(2)lIlesl (6.62)
From (6.52) and (6.61),
IIS(z)1I :-:; IIS(2)1I + Ell St II

legl = les - EATiJll :-:; lesl + IEATiJll
Define
0!1 = IKo - L31 + Lo (6.63)

0!2 = IK1 - L41 + L1 (6.64)
0!3 = IK2-L51+L2+(Lo+L3)Wm+2LlIEATiJll (6.65)
0!4 = kv(EA T iJI)2 + [(L1 + L4)IEA T iJll + L2 + L5]wmIlS(z)IIIEATiJll
+c(C1IEA TiJll + C 2) + kvsi/eo
= CO!4a + EO!4b + kvsi/eo (6.66)
0!5 = kvEATiJI+ [2(L1+L4)kATiJll+L2+L5]wmIlS(z)1I
+c(ClIEATiJll + C 2 )
= C0!5a + EO!5b + 0!5c (6.67)
0!6 = (L1 + L4)IIS(z)lIwm + cCl = CCI + 0!6a (6.68)
0!7 = (Lo + L3)wm1lSt1iE = EO!7a (6.69)
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
chosen bounded functions. Thus, (6.62) can be rewritten as
v ::; -(kv - 0!6)e; - (kvs -0!5)le s l + 0!4 + 0!711Wllles l- E(IIWlllesl)2

-118(£)11 [8e;II WII 2 + 81e s lllWII 2+ 8leslllwlI3]
+118(£)11 [0!11eslllwlI2 + 0!2 e;II WII + 0!3leslllwlI]

8 ~ 2
::; -(kv - 0!6)e; - 0!5)lesl + 0!4 - 2IesIl18(£)IIIIWIl
(kvs -
~ [~0!1 2 0!1 2]
-8Ie s IlIWII1I8(z)1I (IIWII - 28) - (28)
-8e 2118(£)11 [(IIWII- 0!2)2 _
s 28 28
t2)2]
8
-2IesI1l8(£)11 [~0!3 2 0!3 2]
IIWlllesl ) 2+0!711 W~ llle s l
(IIWII-T) -(T) -E (~
k
(kvs - 0!5c -
2 2
::; - ( ; - 0!6 - ~~ 118(£)II)e; - ~~ 118(£)11) lesl + 0!4
_k; [(I e I _ m5a ~
s W5b ) 2 _ (m5a ~ W5b ) 2]
[( ~
8
-2IesIl18(z)1I IIWII- O!~
48 ) 2 - 1682
O!t]
~
-E [(IIWlllesl- 0!7a 2- ( 2
2) 0!7 a ) 2] (6.70)
Consider (6.65) to (6.69), equation (6.70) becomes
v ::; - [k; - eel - 0!6a - ~; 118(£)11] e~
- [kvs - 0!5c - (3~18 + ~~) 11 8(£)11] lesl

(m5a + W5b)2 O!~a
+ 2kv + e0!4a + W4b + kvs"l/eo + 1: 4
::; -i3l e; - i32le s l+ mg + Ws + kvs"l/eo
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
e, ,~ {e.(t) , le.1 ,; max [ ~ (:~ +ead "'8) + ,AT""

1 (kvs"l
{32 --;;- + €Qg + fa8 ) + f AT]
111 } (6.73)
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
Since es is bounded, the set 8 w is also a compact set. Define

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.
6.3.3 Simulation Study

To illustrate the effectiveness of the proposed adaptive controller for unknown
non-affine nonlinear systems, consider a nonlinear plant
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.
State feedback results
When 6 and 6 are measurable, we choose the adaptive NN controller u(t) =

WT S( z) with the input vector z = [Xl, X2, v]T. The parameters in the weight
update law (6.41) are chosen as KO = "'1 = "'2 = 10.0 and is = 5.0. Figure 6.1 (a)
shows that the output y tracks the reference Yd effectively, and Figure 6.1 (b)
indicates the history of the control input u. The norm of the weight estimates
is also given in Figure 6.1 (c) to illustrate the boundedness of the NN weight
estimates. The results of the simulation show good transient performance and
the tracking error is small with all the signals in the closed-loop system being
bounded.
Output feedback results
When 6 is not measurable, a high-gain observer is designed as follows
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.
6.4 Controller Design Based on MNN

In this section, MNN shall be used to construct the adaptive controller for
system (6.3) based on the IDFC (6.20). Both adaptive state feedback con-
troller and adaptive output feedback controller are presented, and simulation
results on the control of CSTR system are used to show the effectiveness of the
proposed approaches.
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
(a) Tracking performance (b) Control input
HI
,,,,,,,,,,
15 20 25 30
(c) Norm of estimated weights IIWI!
Figure 6.1: Responses of the state feedback controller

'.5r--~-~-~----~-----,
ro
15
10
O~'--------------------------~
10 15
T!mesoc
20 25 -50~---:----;:10---'C::-5-----:'O:-----::25~-----:~
Tmesec
(a) Tracking performance (b) The estimate error X2 - X2
'O,----~-~--,-----~-~--_____,
I.
16
14
12
-I
-,
~~-~-~'O--~15-~ro~-~25~~~
Time sec
(c) Control input (d) Norm of estimated weights IIWII
Figure 6.2: Responses of output feedback controller

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..
6.4.1 State Feedback Control

We shall use the MNN WTS(VTz) to approximate unknown u*(z). Define
Z = [Zl' Z2,'" ,Z2n+2f = !zT, l]T E R 2n +2 (6.75)

W = [Wl,W2,'" ,wdT E R 1x1
V = [VI, V2, ... ,vlf E R(2n+2)x!
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
(W*, V*) : = arg min {sup iWTS(VTz) +Uk(Z) - u*(z)i}

(W,V) zEn.
From Lemma 3.6, we have
The coresponding MNN controller takes the form
U = WTS(VTz) + Uk(Z) + Ub (6.78)

Ub = -kslesles (6.79)
where ks > 0 is a design parameter. The first term WT S(VT z) in controller

(6.78) is used to approximate the ideal NN control in (6.76). The second term,
Uk (z ), is a prior control term based on prior model or past experience to improve
the initial control performance. If such a priori knowledge is not available, Uk
can be simply set to zero. The third term, Ub, given in (6.79), is called the
bounding control term, which is applied for guaranteeing the boundedness of
the system states. Using Mean Value Theorem (Theorem 2.5), there exists
/\ E (0,1) such that
b(~, u) = b(~, u*) + bu>. (u - u*) (6.80)
where bu>. = 8[b(cu)l/8ulu=u>. with u\ = /\u+ (1- /\)u*. Considering (6.14)

and (6.80), the error system (6.11) (where we let b1(x) = 1) can be rewritten
as
Since b1 ~ bu >. ~ bo > 0 (Assumptioin 6.4), we have
bu-1'e s = - -£1 bu- >.1 e s + u - 11,

* (6.81)
>.
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)
The MNN weight updating algorithms are chosen as follows
w= -r w [(S - S'VTz)e s + ow(1 + e;)W] (6.83)
V = -rv [zWTS'e s + OvV] (6.84)
where rw = r; > 0, rv = r; > 0, Ow > 0 and Ov > 0 are constant design

parameters. The first terms on the right-hand sides of (6.83) and (6.84) are the
modified back propagation algorithms and the last terms of them correspond
to the a-modification [74] terms in adaptive control, which are helpful to
guarantee bounded parameter estimates in the presence of NN approximation
error.
Theorem 6.3. For the closed-loop system (6.3), (6.78), (6.83) and (6.84),
there exist a compact subset flo C fl x , and positive constants c* and k; such
that if
(i) all initial states ~(O) E <1>0, (W(O), V(O)) E no, and
(ii) c ~ c'" and ks ~ k:,

then the tracking error converges to an c-neighborhood of the origin and all
the states and control input of the system remain in the compact set (SO/ o.
Proof The proof includes two parts. We first assume that the system tra-
jectories remain in the compact set 0, in which the NN approximation in
Assumption 3.2 is valid. With this assumption, we prove the tracking error
converging to an c-neighborhood of the origin. Later, we show that for a suit-
able reference signal Yd(t) and he proper choice of controller parameters, the
trajectories do remain in the compact set 0 for all time.
Part 1 Choosing the Lyapunov function candidate
(6.85)
and differentiating (6.85) along (6.82)-(6.84), we have
.
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}
:.
= -ks Ies Ies -

1 -1 2 buJ. 2
2 -gbuJ. es - 2b 2 es + [du - cu (z )] es
uJ.
- 8w (1 + e;yW·TW - 8v tr{VT V}
By completing the squares, it is shown that 2WT W ~ IIWII 2- IIW*1I2 and

2tr{VT V} ~ IIVII} -IIV"II}· Noticing (3.41) and (3.56), we obtain
Vi ~ -k s le s l3 - ~b~:e; - ;:{J. e; - 8; (e; + 1)(IIWI1 2-IIW"1I2)

-8; (IIVII} -/IV"'II}) + lesl (Ql + Q2le + Q311WII + Q411WII/esl + e
s/ 1)
:5 -k /es ~b;;;;e~ _ 8; e~IIWII2 _ 8; IIWII ~ I/VII}

s /3 - 2 -
+(~~J.I + Q2 + 8; IIW"'112)e; + (el + Ql)les l

UJ.
+Q311Wlllesl + Q4I1Wlle; + 8; IIW"'1I2 + 8; IIV"II} (6.86)
b2/2b6 and b;;'; 2: b11 (Assumption 6.4), inequality (6.86) can be further writ-
ten as
where 130 to f32 are positive constants defined by

b2 tJ w * 2 a~ (6.88)
130 = 2b6 + a2 + 211 W II + tJw
131 = tl + a1 (6.89)
Ih = tJw IIW*1I2 + tJv

2 2
IIV*II~ + tJa~ (6.90)
w
Nt " I-'Oe 2<ksI13+~11

2R 1 e s I<R
ks e s an d(~+13)1
s )2
o Icmg s _ 2' e s ks _ I-'Oe s2+(2,60+,61k
4k~,6o' we
have
Thus, inequality (6.87) can be written as
(6.91)
(6.92)
8w • ~ {(W, V) IIIWIl '" J~~, I!VIIF '" J~~ } (6.93)
e: = {(e s, W, V) I i lesl3+ t!l e; + tJ2w II Wll 2+ tJ; liVll~ ::; D} (6.94)
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
be expressed as ((t) = ((O)e At + J~ eA(t-7')be sdT. Noting lesl :S Em for t > T 1 ,

we conclude that, 'tit > To
where Em = min { ?!2D/ks , VEDb 1 } and koc'\otll((O)1I decays exponen-

tially. This confirms that the tracking error Y - Yd = t1
will converge to
an E-neighborhood of the origin.
Part 2 To complete the proof, we need to prove that for a suitable tracking
signal Yd(t) and the appropriate design parameters, the trajectories do remain
in the compact set U. From (6.6) and = t [e tnJT,
it is shown that = tn
e s - AT(. Considering (6.92) and (6.95), there exists positive constants ko and
k1 such that
(6.96)
From the analysis of Part 1, we know that i'i :S 0 as long as (e s , Mr, V) is

outside the compact set e. Let
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
(ii) inequality (6.96) implies that II~II :S koe-'\otll((O)1I + k1Em + II~dll,

Hence, there do exist a compact set no, and positive constants c* and k; such
that for all ~(O) E <1>0, ('W(O) , V(O)) E no, c:S c* and ks 2: k;, we have
(e s , lV, V) E OE => (~, u) E [r
This completes the proof. <>

Remark 6.9. The bounding control term, Ub, in controller (6.78) is used to
limit the upper bounds of the system states, therefore guarantees the validities
of the strong relative degree and the NN approximation in Assumptions 6.1
and 6.4, respectively. The developed adaptive method also accommodates the
existence of prior control term, Uk, such that control engineers can plug in the
proposed NN technique to work in parallel to improve the control performance.
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.
6.4.2 Output Feedback Control

In this subsection, we consider the adaptive output feedback control for non-
linear system (6.3) using MNN.
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
eso = [/\1, /\2,'" , An-I, 1j1}i

Zo = [O,1}iT,e s o,AT1}i,Of
1}i = [0, -;J;, -1//3), ... , _1jJ(n)J T
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
(W", V*) : = aTg min {sup IWTS(VT z) + Uk - u*(z)l} (6.98)

(W,V)En w zEn.
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
W T S(VT2) - W*TS(V*T z) = WT (80 - 8~VTi) + WT8~VTi + du (6.99)

Id.,.1 ::; vml\zWT 8~I\F + w m(1\80 1\ + 1\8~VTzl\) (6.100)
Consider the output feedback controller as follows
U = U nn + Uk + Ub (6.101)
with
U nn = WTS(VTz) (6.102)
Ub = - [~(l\zWT 8~1I~ + 1\8 1\2 + 118~VTz1l2 + 1)

0
+ks (\U nn \ + 1) ]es (6.103)
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
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
w= -rw[(So - S~VTi}es +<5w(1 + lesl)W] (6.104)
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
b(~, u) = b(~, u*) + bu), (u - u*) (6.106)
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)
Applying (6.97), (6.99) and (6.101), we obtain
-1 . b1 (x) _ - T' ' I 'T "- ' T ' I - T "-

bu x es = ---bu
E x
es+ub+W (So-SoV z)+W SoV z
+du - /-tt, Vz ESlz (6.108)
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
(iii) c ::; c*, ks 2: k; and I; ::; 1;*
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.
Proof The proof contains two steps. We first assume that ~ E ~ 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 ~ for all time.
Step 1 Consider the Lyapunov function candidate
(6.109)
Differentiating (6.109) along (6.108), we have
V1 = - b1 c(x) b-u~1 e2s + WT(S _ S'VTi)e s + WTS'VTie s

0 0 0
1 d(b;;;) - T - T
W r w W + tr{V rv V}
2 -1':' -·1':'
+Ubes + [d u -lLiles + 2~es +
Noting WTS~VTz = tr{VTzWTS~}, es = es + Eeso, (6.104) and (6.105), we

obtain
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
VI ::; Ubes - 8; (1 + le s l)(IIWl1 2 - IIW"112) - ~ (1 + lesl)(IIVII} -IIV"II})
+lesl(ldul + IILL!) + EleBol [IIVIIFllzW TS~IIF + IIWII(IISoll + IIS~VTzll)]

Because
- II I" T " 8v - 2 E2 2 " T ' , 2

EIeso III V FI zW SollF ::; 211V1IF + 28vesollzW SollF
ElesolllWlllISol1 ::; 8: IIWI1 2 + ;2

w
e;ollSoll2
ElesoiIlWIIIIS~VTzll ::; 8: IIWI1 2 + ;: e;oIIS~VTzI12

and considering IIW*II ~ Wm , IIV*IIF ~ V m , (6.100) and (6.103), we obtain
v, S; -e. (e, + €e,o) [ ~ (II.WTS;II} + IISoll' + II S; VT'II' + 1)

+k. (IUnol + 1)1+ (1+ li.))( "; w;' + "; v;')
+Iesl (Vm ll z-wT S~IIF + wmllSol1 + wmllS~ VTz-11 + IILL I)
2 2 (1 ,,-' T 2 1, 2 1 T"- 2)
+E eso 2o)zW SollF + Ow IISoil + Ow IISoV zll (6.111)
'I 'I'
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
inequality (6.111) can be written as
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)
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
{30 = ~ [(ow + 210 + Eowlesol)w~ + (ov + c: + EOvlesol)v~]

10 E2 e 2
+C:ILl + 4(OWW~, + ovv~)2 + ~ (6.114)
221 1 221 1
{31 = E eso (-
10
+ -2;
Uv
),(32 = E eso (- + r-)
c: w U
(6.115)
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)
Because les(t)1 :S f3s(c, t), "it ~ To, we have
1I((t)11 :S ko [11((0)11 + (TO eAOTles(r)ldr] e-Aot + ko (t e-AoCt-T)les(r)ldr

Jo J~
- k
:S ko[II((O)II+ ~:]e-Aot+ ",:f3s(c,t), "it~To (6.118)
where es = sUP09:STo les(t)l· Since es(t) is bounded, we know that ko [11((0)11 +

es / /\0 le- Aot decays exponentially. Inequality (6.118) implies that the tracking
error 6 = Y - Yd will converge to an (c, t)-neighborhood of the origin.
Step 2 To complete the proof, we need to show that for a proper choice of
the tracking signal Yd(t) and control parameters, the trajectory'; do remain in
the compact set <p(. Considering a positive function Vb = e;/2 and controller
(6.101), the time derivative of Vb along (6.107) is
. bl (x) 2 ..
Vb = - - - e s + bu>.(u - U )es (6.119)
c
Using (6.97), (6.101)-(6.103), we have
bl(x) 2 + bu" [ - ks ( Iunnl + l)(e + teso ) + U - W .. T S (V .. T_)

. = --c-es
Vb s nn Z
-~l] e s - b;>. es(e s + teso) (lIzW T 8~1I} + 1180 11 2 + 118~VTz1l2 + 1)(6.120)
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)
- b;>. lesl(lesl- tlesoD (1IzWT 8~11} + 1180 11 2 + 118~VTz112 + 1)

-bu>.ks Iesl( 10
lesl
- tlesol)(lunnl + 1)
I+w", v1+lell <

Since IU nnk.CJunnl+l) - -l
k.
1+ III. 1+1) 'u>.
(w myt,...1 b > 0 and !!.!..W.
g '
> 0 it is shown
that Vb :S 0 once lesl ~ Ro(t, ks) with
(6.122)
We know that if les(O)1 ~ Ro(E, k s ), then
les(t)1 ~ Ro(E, ks), '<:It 2: 0 (6.123)
It follows from (6.117) and (6.123) that /If(t)II ~ ka/l((O)/i +kbRo(E,ks),'<:It 2: 0

with ka = ko(l + IIAIi) and kb = ka/ >'0 + 1. Hence,
1I~(t)1I ~ IIt(t)II + lI~d(t)1I ~ kall((O)1I + kbRo(E, ks) + c, \;It ~ 0 (6.124)
We now provide the conditions which guarantees ~ E eIle, '<:It ~ O. Define the
compact set
eIlo := {~(O) I {~ 11I~(t)1I < kall((O)II} C eIle , les(O)1 < Ro(E, ks)} (6.125)
the positive constants
co. := c~~ {c I {~ III~II < kall((O)11 + c, ~(O) E eIlo} C eIle}
k; := inf {ks
k.ER+
I{~ III~II < kall((O)1I + kbRo(O, k s) + c,
~(O) E eIlo, c ~ c" } C eIle} (6.126)
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.
Remark 6.13. In the adaptive NN controller (6.101), two additional control

terms, the bounding control term Ub and the prior control term Uk, are provided.
The first one can be viewed as a supervisory control, which is introduced for
limiting the upper bounds of the system variables such that € E 4>~ holds.
The second one provides a chance that control engineers can use conventional
techniques to design an initial controller and then add the adaptive neural
networks to work in parallel to achieve high tracking accuracy. From (6.98) it
can be seen that the closer Uk and u* is, the smaller the ideal weight W* and V*
will be. Considering (6.114)-{6.116), one can see that smaller W* and V* will
lead to smaller output tracking error. Therefore if Uk is designed adequately,
the control performance can be improved. On the other hand, even though
Uk is inadequate, the use of the above adaptive NN controller still results in a
stable tracking.
Remark 6.14. In Theorem 6.4, it requires that the observer (6.29) turns on
time t* in advance. This is because the high-gain observer may exhibit a
peaking phenomenon and the estimated state errors might be very large in
the initial transient period. If the observer is turned on time t* before the
controller is put into operation, Lemma 6.4 guarantees that the state estimation
1r:t 1 _y(k) is bounded by the constant €hk which only depends on Yk, € and ~i'
and therefore the peaking of the controller can be avoided. Another method to
overcome the peaking problem is to introduce an estimate saturation or control
input saturation [95] [2] [82]. Thus during the short transient period when the
state estimates exhibit peaking, the saturation prevents the peaking from being
transmitted to the plant.
6.4.3 Application to CSTR

Continuously Stirred Tank Reactor System
The continuously stirred tank reactor (CSTR) system given by Lightbody and
Irwin [116] is shown in Figure 6.3. This system consists of a constant volume
reactor cooled by a single coolant stream flowing in a cocurrent fashion. An
irreversible, exothermic reaction, A --+ B, occurs in the tank. The objective is
to control the effluent concentration, C a by manipulating the coolant flow rate,
qe. The process is described by the following differential equations
2..{Cao - Ca) - aoCae-~ (6.128)
V
~ (Tf - Ta) + a1Cae- R.t + a3qe [1 - e-?c] (Tel - Ta) (6.129)
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.
Cae Tr
Ca Ta
v
Figure 6.3: Continuously stirred tank reactor system
Define the state variables, input and output as
x = [Xl,X2f = [Ca,Ta]T, u = qc, Y = C a

System (6.128)-{6.129) can be written in the form of system (6.1)
(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
Table 6.1: Parameters of CSTR System
Parameter Description Nominal value

q process flowrate 100l/min
GaO concentration of component A 1 mol/l
Tf feed temperature 350K
Tef inlet coolant temperature 350K
V volume of tank 100l
ha heat transfer coefficient 7 x 105 J/min. K
ao preexponential factor 7.2 x 10 1O min- 1
E/R activation energy 1 x 104 K
(-~H) heat of reaction 2 x 104 cal/mol
PI, Pc liquid densities 1 x 103 g/l
Gp,Gpe heat capacities 1 cal/g· K
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
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.
Adaptive State Feedback Control Using MNN

The proposed adaptive state feedback controller based on MNNs is applied to
the CSTR system. The signals e s and v are generated by
with '\1 = 3.0 and c = 0.1. The controller is taken as
(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
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 ," I, 'I

I I
I
I, I
102
,
II:'
100 I I " 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
-10 " ''

"
-12 "i\
i
-14 i
-16
Tinemifl Time min
10 15 ~ ~ ~ 35 -180 5 10 15 20 2S 3() 35
(c) NN weights IIWII ("-") (d) The bounding control Ub ("-")

and IIVII F ("- -") and NN output ("- -")
Figure 6.4: Responses of state feedback controller using MNNs

Adaptive Output Feedback Control Using MNN
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.
0.' 3 112 rmn

mo~
...
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
(a) Output tracking Yd ("_._.") (b) Control input

PI ("- -") and NN ("-") PI ("- -") and NN ("-")
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
(d) State estimation error iJ -

30 -40 50 60 70 80
7r2/ E
and IIVIIF ("- -")
Figure 6.5: Responses of PI control and adaptive output feedback NN control

using MNNs
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
Xi = Ii(Xi) + gi(Xi)Xi+I, 1 ::; i ::; n - 1

xn = !n(Xn) + gn(xn)u, n 2: 2 (7.1)
Y = Xl
where Xi = [Xl,'" ,XiJT E Ri, i = 1,··· ,n, u E R, Y E R are the state

variables, system input and output, respectively; IiO and gi(-), i = 1,··· , n
are unknown smooth functions. The possible controller singularity problem
usually met in adaptive control is avoided without using projection. However,
the integral Lyapunov function based design is usually too complicated to use.
In this chapter, different control schemes are presented for different classes
of triangular nonlinear systems based on backstepping. The proposed control
methods can not only guarantee the boundedness of closed-loop system sig-
nals, but also render that the system output y follows a desired trajectory Yd

184 Chapter 7. Triangular Nonlinear Systems
generated from the following smooth, bounded reference model
Xdi = fdi(Xd}, 1 :S i :S rn (7.2)

Yd = Xdl, rn 2: n
where Xd = [Xdl' Xd2," . ,XdmV E R m are the states, Yd E R is the system

output, fdiC), i = 1,2,," ,rn are known smooth nonlinear functions.
In Section 7.2, we first present an adaptive neural control scheme for system
(7.1) where the affine term gn(xn} is assumed to be independent of state X n. By
utilizing this special property, the controller singularity problem is overcome
completely, and the stability of the resulting adaptive system is guaranteed
without the requirement for the integral Lyapunov function [192].
In Section 7.3, a direct adaptive NN control approach is proposed for a class
of partially known nonlinear systems in the strict-feedback form, in which a pri-
ori information of the system under study can be incorporated into controller
design, and the overparameterization problem [92] caused by the parametric
uncertainties in adaptive backstepping is avoided without using tuning func-
tions [103]. With the help of NN approximation, this scheme much simplifies
the backstepping design, since it reduces the large amount of algebraic manip-
ulations at each step in the backstepping design procedure.
In Section 7.4, adaptive neural control of nonlinear pure-feedback systems
is studied. Note that in the literature, while the nonlinear strict-feedback sys-
tems have been much investigated via backstepping design, the broader class of
nonlinear pure-feedback systems have been less explored. Only a few results are
available in the literature for the control of pure-feedback systems with para-
metric uncertainties [92,104,133,167]. The pure-feedback systems represent a
more general class of triangular systems which have no affine appearance of
the variables to be used as virtual controls. As indicated in [104], it was quite
restrictive to find the explicit virtual controls to stabilize the pure-feedback
systems by using integral backstepping. For the control of uncertain pure-
feedback systems with unknown nonlinear functions, in this chapter, direct
adaptive neural control schemes are presented, where the idea of backstepping
is still employed, i.e., some of the state variables are considered as "virtual con-
trols", and intermediate control laws are designed in the constructive design
procedures. Implicit function theorem (Lemma 2.8) is first exploited to assert
the existence of the continuous desired virtual controls. Then, NN approxima-
tors are employed to approximate the continuous desired virtual controls aT
and desired practical control u·. With the help of NN approximation, there
is no need to solve the implicit functions for the explicit virtual controls and
the practical controller to cancel the unknown functions in backstepping design.
With mild assumptions on the partial derivatives of the unknown functions, the
developed adaptive NN control scheme achieves semi-global uniform ultimate
boundedness of all the signals in the closed loop.
7.2 Special Systems in Strict-Feedback Form 185
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.
7.2 Special Systems in Strict-Feedback Form

In this section, motivated by the fact that better results can be obtained when
more properties of the studied systems are exploited [49,191J, a direct adaptive
NN control scheme is presented for a class of nonlinear system in the following
strict-feedback form
Xi = gi(Xi)Xi+l + h(Xi), 1 SiS n - 1
xn = gn(Xn-l)U + fn(xn), n ~ 2 (7.3)
Y= Xl
where Xi = [Xl,'" ,XiV E R i , i = 1"" ,n, U E R, y E R are state vari-

ables, system input and output, respectively; h(-) and gi(-), i = 1" .. ,n are
unknown smooth functions. Note that for the control of the uncertain strict-
feedback system (7.1), one of the main difficulties comes from the uncertain
affine terms gi(-) (i = 1"" ,n). When gic-) (i = 1"" ,n) are known ex-
actly, while fiO (i = 1"" ,n) are composed of dominant uncertain nonlin-
earities and/or parametric uncertainties, both robust adaptive backstepping
designs [84,104,143, 184J and neural control scheme provided in [147J are ap-
plicable. When gi(') (i = 1"" ,n) are unknown, iffeedback linearization type
controllers Qi = gJ) ( - ii 0 + Vi) are considered, where ii (-) and 9i (-) are the
estimates of fiO and gi(')' respectively, and Vi is a new control to be defined,
difficulties arise when 9i(-) -+ 0, which is referred to the controller singUlarity
problem. By combining adaptive neural design with backstepping method-
ology, a smooth adaptive neural controller has been presented for uncertain
strict-feedback system (7.1) in Chapter 5, where integral Lyapunov functions
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
in the closed-loop remain bounded.

Since gi(')' i = 1, ... ,n are smooth functions, they are therefore bounded
within the compact set O. Accordingly, we can make the following assumption
as commonly being done in the literature.
Assumption 7.1. The signs of gi (.) are known, and there exist constants gil ~
giO > 0 such that gil ~ Igi(·)1 ~ giO, 'l:fxn E 0 eRn.
The above assumption implies that smooth functions gi (.) are strictly either
positive or negative. Without losing generality, we shall assume gil ~ gi(Xi) ~
giO > 0, 'l:fxn E 0 C Rn.
The derivatives of giO are given by
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,
hnn(Z) = WT 8(Z) (7.4)
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
h(Z) = W*T 8(Z) + E, 'l:fZ E Oz (7.5)
where W* are ideal constant weights, and E is the approximation error.

Assumption 7.3. There exist ideal constant weights W* such that kl ::; E*
with constant E* > 0 for all Z E Oz.
7.2.1 Direct Adaptive NN Control

In this section, by restricting our attention to uncertain strict-feedback system
(7.3) where gn(-) is unknown and independent of state X n , a singularity-free
direct adaptive NN control scheme is proposed by combining direct adaptive
NN design with backstepping [57J. At each recursive step i (1 SiS n) of
the backstepping design, the desired virtual control at and the desired real
control u'" = a~ are first shown to exist which possess some desired stabilizing
properties. The desired virtual control at (i = 1,··· , n) contains uncertainties
fi(-) and gi{') (i = 1,··· , n), and thus cannot be implemented in practice.
To solve this problem, the virtual control ai and the practical control u are
constructed by using RBF NNs WiT 8 i {Zi) to parameterize the unknown parts
in the desired virtual control at and the desired control u*. Then, adaptive
techniques are used to update the weights of neural networks for closed-loop
stability. By using the virtual control ai, the ith-order subsystem is stabilized
with respect to a Lyapunov function Vi, The control law u is designed in the
last step to stabilize the whole closed-loop system with respect to an overall
Lyapunov function Vn , which is the sum of all the sub-Lyapunov functions.
The detailed design procedure is described in the following steps. For clar-
ity and conciseness of presentation, Step 1 and 2 are described with detailed
explanations, while Step i and Step n are simplified, with redundant equations
and explanations being omitted.
Step 1
Define Zl = Xl - Xdl. Its derivative is
(7.6)
By viewing X2 as a virtual control input, i.e., if we choose at ~ X2 as the control

input for the zl-subsystem in the above equation, and consider the Lyapunov
function candidate VZ1 = -2 1
91
zr,
whose derivative is
. . 2 .
. zl Z l gl z 1
VZ1 = - - - - 2
2 = zl[h{X1) + 91{X1)a1.. - . J
Xd1 -
91
-22 Zl
2
91 gl 91
then there exists a desired feedback control
(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.
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
where Zl ~ [Xl,xdl]T C R2. By employing an RBF NN WT81 (Zl) to approx-

imate hl(Zl), ai can be expressed as
(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)
the Zl equation (7.6) becomes
(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)
where WI = WI - Wi- Throughout this chapter, we shall define () = () - (-)*.

Consider the following Lyapunov function candidate
1 2 1 - T -1 -
VI = 2g l (Xl)Zl + 2Wl r l WI (7.12)
where r 1 = rT > 0 is an adaptation gain matrix.

The derivative of V1 along (7.11) is
Consider the following adaptation law
(7.14)
where a1 > 0 is a small constant. The a-modification term a1 W1 is introduced

to improve the robustness in the presence ofthe NN approximation error E1 [74J.
Without such a modification term, the NN weight estimates W1 might drift to
very large values, which will result in a variation of a high-gain control scheme.
Let C1 = ClO + C11, with ClO and Cu > O. Then, equation (7.13) becomes
(7.15)
By completion of squares, we have
(7.16)
Because - (ClO + !712 )zr S -

2g I
(ClO - gl: )zr, by choosing
2g lO
ClO such that
we have the following inequality
(7.17)
where the coupling term ZIZ2 will be canceled in the next step.
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
a; = -z] - C2Z2 - ~(h(X2)

g2(X2)
- 0: 1 ) (7.18)
where C2 is a positive constant to be specified later. From (7.9), it can be seen

that al is a function of Xl, Xd and W1 . Thus, Ql is given by
(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)
denote the unknown part of a 2 in (7.18), with Z2 ~ [xI,~, ¢dT c R4.

(Please see Remark 7.2 for the definition of Z2). By employing an RBF NN
wi 8 2(Z2) to approximate h 2(Z2), a; can be expressed as
(7.21)
Define the error variable Z3 = X3 - a2 and choose the virtual control

(7.22)
Then, we have
(7.23)
Remark 7.2. Though the unknown function h 2 (Z2) (7.20) is a function of

X2, Xd and WI, the large number of neural weight estimates W1 are not recom-
mended to be taken as inputs to the NN because of the curse of dimensionality
of RBF NN [66]. By defining intermediate variables aaaXll and ¢l, which are
available through the computation of X2, Xd and Wl , the NN approximation
wi 8 2(Z2) of the unknown function h 2(Z2) can be computed by using the
minimal number of NN inputs Z2 = [xI, ~, ¢1 jT. The introductions of inter-

mediate variables help to avoid curse of dimensionality, and make the proposed
neural control scheme computationally implementable (please see Remark 7.4
for detailed discussion). The same idea of choosing the input variables of NNs
is also used in the following design steps.
1 2 1 - T ~1-
V2 = Vi + 292 (_)
X2
z2 + - W 2 r 2 W 2
2
(7.24)
where r 2 = rI > 0 is an adaptation gain matrix. The derivative of V2 is

. . Z2 Z2 t12Z~ - T ~1 ~
V2 2" + W" r" W 2
= VI + -92- - -92 --
·
= VI - ZIZ2 + Z2 Z3 -
2
c2 Z " -
-
92 2
-22 z"
92-
+ Z2 E2
-T -T I~
-W2 S2(Z2)Z2 + W2 r 2 W 2 (7.25)
(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
ir.2 < Z"Z _ '"'

2
c* z2 _ ' " ' ak k
2 IIW 112 + '"'
2
ak
IIW*112
k
2~
+ '"'
*2 (7.27)
- 3 ~ kO k ~ 2 ~ 2 ~ 4Cki
k=l k=l k=l k=l
• 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)
, . 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)
denote the unknown part in at (7.28), with

Z•. ~ [-'!' ... oai-l A.. IT
X., oai-l
II
UXI
, 'll , '1'0-1 C
UXi-I
R2i (7.31)
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
Zi = h(Xi) + 9i(Xi)(Zi+1 + ai) - Qi-I

-T
= 9i[Zi+1 - Zi-I - CiZi - Wi 8 i (Zi) + €il (7.34)

1 2 1- T -1-
Vi = Vi-I + 29i(Xi) zi + '2 Wi r i Wi (7.35)
(7.36)
where (J'i > 0 is a small constant. Let Ci = Cia + Ci1, where Cia and Cil > O.
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
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)
denote the unknown part in u" (7.38), with
[ -T 8an-l aan-l ]T 2n
Zn = Xn , -8--'"'' -a--,¢n-l
Xl Xn-l
c R (7.41)
By employing an RBF NN Wn T Sn(Zn) to approximate hn(Zn), u" can be

expressed
(7.42)
Choosing the practical control law as

(7.43)
we have
(7.44)
Consider the overall Lyapunov function candidate

1 2 1 - T -1-
Vn =Vn- 1 + 2 ( )Zn+-2Wnrn Wn (7.45)
gn Xn-l
Wn = Wn = r n[Sn(Zn)zn - Un Urn] (7.46)

where Un > 0 is a small constant. Let Cn = Cno + Cn 1, where CnO and Cn1 > O.
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
V.n < ~ C"

-L..tkOk-L..t
z2 ~ ukll W
2
kl1 2 + 8
k=l k=l
n - T
::; - L 2 7 L7
n -1-
z~ - W k ~k Wk +8
k=l 9kO k=l
[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.
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)
1
L -z~
n
~ < 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
lead to curse of dimensionality [66]. By introducing some intermediate vari-

ables aaCkiXl- 1 , ••• , aaCki - 1 and ¢i-l as inputs to RBF NN VVTSi(Zi), where the
X1.-1
intermediate variables are available through the computation of system states
Xi and neural weight estimates Wl, ... , Wi-l, the NN approximation can be
implemented by using the minimal number of NN input variables. Hence, the
differentiation requirement can be satisfied by using RBF NNs to approximate
all the uncertainties in the desired virtual controls and practical control, while
intermediate variables are introduced to make the NN approximation compu-
tationally feasible.
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.
7.2.2 Simulation studies

In practical application of the proposed scheme, 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 lat-
tice on Rn can uniformly approximate sufficiently smooth functions on closed,
bounded subsets. Furthermore, given only crude estimates of the smoothness
of the function being approximated, it is feasible to select the centers and
variances of a finite number of Gaussian nodes, so that the resulting NNs are
capable of uniformly approximating the required function to a chosen tolerance
everywhere on a pre-specified subset. Accordingly, in the following simulation
studies, we will select the centers and widths on a regular lattice in the respec-
tive compact sets.
The proposed adaptive NN controller is applied to the following two uncer-
tain strict-feedback systems
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
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
Xd2 = -Xdl + ,6(1 - X~1}Xd2 (7.58)

Yd = Xdl
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)
(a) Output Y ("-") follows Yd ("- -") (b) State X2(t}
"
-, 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
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)
and NN weights WI and W2 are updated by (7.14) and (7.26) correspondingly.
-30 8 10 1Z 14 " 20 -3o'--~--'--'---'.-'"::,,--',,:-, ---,':--'"::,,-,=-,---:'".

T".(SGaIdsJ In,,,{SecaOs)
(a) Output y ("-") follows Yd ("- -") (b) State X2(t)
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 ("- -")
Figure 7.2: Responses of adaptive NN controller for plant (7.58)

7.3 Partially Known Nonlinear Systems 201
Neural networks W[ 8 1 (Zl) contains 25 nodes (i.e., h = 25), with centers

ILl (l = 1"" , h) evenly spaced in [-6,6] x [-6,6]' and widths rll = 3 (l =
1", . ,h). Neural networks wi 8 2(Z2) contains 108 nodes (i.e., b = 108),
with centers ILl (l = 1, .. , , b) evenly spaced in [-4,4] x [-4,4] x [-4, 4] x [-6,6],
and widths 7]1 = 4 (l = 1, ... , h). The design parameters of the above controller
are Cl = 2.5, C2 = 2.5, f l = f2 = diag{2.0}, ()1 = ()2 = 0.2. The initial weights
WI (0) = 0.0, W 2 (0) = 0.0.
Figure 7.1 (a)-(d) shows the simulation results of applying controller (7.59)
to system (7.57) for tracking desired signal Yd with f3 = 0.2. The initial con-
ditions [Xl(0),X2(0)]T = [1.2,1.0jT and [Xdl(0),Xd2(0)]T = [1.5,0.8]T. From
Figure 7.1 (a), we can see that fairly good tracking performance is obtained.
The boundedness of state X2, control signal u and NN weight estimates WI, W2
are shown in Figure 7.1 (b)-(d) respectively.
Keeping all design parameters unchanged, Figure 7.2 (a)-(d) shows the sim-
ulation results of applying the same controller (7.59) to system (7.58) for track-
ing a different desired signal Yd with f3 = 0.7, from a different initial conditions
[Xl(0),X2(O)]T = [1.5,O.6]T, and [Xdl(O),Xd2(O)jT = [1.1,1.3jT. From Figure
7.2 (a), it can be seen that good tracking performance is obtained. The bound-
edness of system state X2, control signal u and NN weights WI, W2 are shown
in Figure 7.2 (b)-(d) respectively.
These simulation results demonstrate that the adaptive NN controller (7.43)
can achieve good control performance for uncertain strict-feedback systems
(7.3). Furthermore, the controller is independent of the complexities of the
system nonlinearities. With the compact sets for NN approximation being
made large enough, the adaptive NN controller (7.43) can be applied to un-
certain strict-feedback systems (7.3) without repeating the complex controller
design procedure for different system nonlinearities.
7.3 Partially Known Nonlinear Systems

In this section, a direct adaptive NN control scheme is presented for partially
known nonlinear systems in the following strict-feedback form [178]
Xi = giXi+l + !i(Xi) + Or Fi(Xi) + c,oi(Xi), 1 ::; i ::; n-1

xn = gnu + !i(Xn) + O~Fn(xn) + c,on(xn), n ;::: 2 (7.61)
y=Xl
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
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.
The control objective is to design a direct adaptive NN controller for system

(7.61) under the assumption that full system state xn is measurable, such that
(i) all the signals in the closed-loop system remain bounded, and (ii) the output
y follows a desired trajectory Yd in (7.2).
7.3.1 Adaptive Neural Control Design

In this section, direct adaptive NN design is presented using backstepping. It
should be noted that the employment of NN approximators avoids the over-
parametrization problem [92J in adaptive backstepping design without using
tuning functions [103], as will be detailed in the following derivation. The
design procedure consists of n steps as follows.
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
ar = -ClZl - ~ ['PI (Xl) + efFl(xd + h(xl) -

gl
!dl(Xd)]
T 1
= -ClZl - 'Til Fry1 - -'PI (Xl) (7.63)
gl
where Cl is a positive constant to be specified later, 1]1 := [~, .l...]T is an

g1 g1
unknown constant vector, and Fry1 = [Fr, (11 - !dl)jT is a known function
vector. Note that because of the introduction of 1]1, we only need to estimate
.l...
g1
rather than gl.
For uniformity of presentation, let hl(Zl) = .l...<Pl(Xt)
g1
denote the unknown
function in (7.63), where Zl = [xlf E 0 1 c R. By employing a RBF NN
Wr8 l (Zl) to approximate hl(Zl), a! can be expressed as
(7.64)
where W{ denotes the vector of ideal constant weights, and lEI I S E! is the ap-
proximation error with constant E! > O. Accordingly, the controller singularity
problem caused by the estimation of gl is avoided with the employment of the
neural network WI T 8 1(Zl) and the introduction of 1]1.
Since 1]1 and wt are unknown, ai cannot be realized in practice. Define
Z2 = Xi - al and let
(7.65)
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)
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~
-WI Sl(Z1)Z1 + W 1 rW1 W 1

-T -T 1 ~
(7.68)
Consider the following adaptation laws
ry,l = 17.1 = r 1'/1 [Fl'/l Zl - 0'1'/1 ih]
(7.69)
W1 = WI = rW, [Sl(Zl)Zl - O'W, W1 ]
where > 0 are small constants. Let

0'1'/1' O'w1 C1 = ClO + Cll, where ClO and
Cll > 0. Then, equation (7.68) becomes
(7.70)
-T -T - - 2 -
= + Wt) ::; -O'Wl + O'W1 II W 1 1111 wt II
A
-O'W1 W 1 W1 -O'Wl W 1 (W1 II W 1 11

- 2 2
< _ O'w1 11W1 11 + O'w111Wili
- 2 2
-O'l'/liifTJI = -O'l'/liif( i h + 1]n ::; -0'1'/1// i hl/ 2 + 0'I'/ll/ii11///1]rl/
< _ 0'1'/1 I/ii1//2 + 0'1'/11/1]il/ 2
- 2 2
E 2 E*2
-Cllzi + ZlE1 ::; -cllzi + Zl/ E1/ ::; _1_ ::; _1_ (7.71)
4C11 4Cll
Therefore, we have the following inequality
. < ~ ~
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
Step 2
The derivative of Z2 = X2 - al is
(7.73)
When viewing X3 as a virtual control input to stabilize the (Zl' z2)-subsystem,

there exists a desired feedback control
(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)
which is computable. Then, a 2 is given by
(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)
denote the unknown function in (7.77), with Za = [Xa, ~]T E fh C R3. By

employing a RBF NN WaTSa(Za) to approximate ha(Z2), a:; can be expressed
as
(7.79)
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)
where Z2 = [X2' ~]T E fh C R3 is the same as in (7.78). In this case,

the parametric uncertainties 91 and lh in "12 = [~, .1..., ~, ll1.]T will need to be
· t ed agam
est Ima . .m t he contro11er d '
eSlgn, wh'lC h'IS known
[12 92 as
92 overparame
92 t'
nza t'IOn.
Note that due to the existence of the unknown nonlinearities <PI (Xl) and
Cf'a(X2) in system (7.61), neural network W 2T S 2(Z2) has to be used to approxi-
mate the unknown nonlinear function h 2(Z2) (7.78) or h 2(Z2) (7.81). By define
a:; as in (7.77) rather than in (7.80), the parametric uncertainties 91 and (h
have been combined into the unknown function h 2(Z2) as in (7.78). Using
neural network W 2T S 2(Z2) to approximating h2(Za) (7.78) instead of h;(Z2)
(7.81), there is no need to estimate 91 and (h repeatedly in the controller,
or to use tuning functions to deal with the partial derivative terms ~91X2
and ~Or FI (xt). Therefore, with the help of neural approximation, the over-
parametrization problem caused by the parametric uncertainties 91 and 01 in
adaptive backstepping design is avoided without using tuning functions [103].
It is noticed that neural network can also be used to approximate the over-
:2 Or
all nonlinear function [Cf'2(X2) + F2(X2) + h(X2) - a1] rather than h2(Za)
(7.78). Since 92 is a constant though unknown, we introduce "12 = [~, V :2
and F'12 = [Fi, (f2 - ¢l)V such that F'12 =T/r :2 [Or
F2(X2) + h(X2) - ¢1],
which is separated from the overall nonlinear function. Thus, neural network
W 2TS2 (Z2) is only used to approximate the unknown nonlinear function h 2(Z2)
(7.78), which might have lower nonlinearity, less complexity and smaller magni-
tudes than the overall nonlinear function. By doing this, smaller approximation
errors might be produced, and better control performance might be achieved.
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)
where r'12 = r~2 > 0, 1\v2 = rW2> O. The derivative of V2 is
(7.85)
ry.2 = iJ.2 = r'12 [F'12 Z2 -

a'12 r,2]
(7.86)
W2 = W2 = rW2[82(Z2)Z2 - aW2W2]
where a'12' aW2 > 0 are small constants. Let C2 = C20 + C2l, where C20 and
C21> O. Then, equation (7.85) becomes
(7.87)
-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)
Thus, we have the following inequality
(7.89)
Step i (3::; i ::; n -1)
The derivative of Zi = Xi - Qi-l is
(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
where Ci is a positive constant to be specified later, and ai-l is a function of

Xi-I, Xd, ill,'" , ili-l and WI,'" , Wi-I' Thus, ai-l can be expressed as
(7.92)
where
which is computable. Then, at is given by

a~t -- -z·'t- 1 - C'~'
1,"'1.
- 'l1Tp,
'Ii 'f1i
~ 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
where Zi = [Xi, 880<,-"

Xl
... , 880<~-1]T
X1.-1
E SG i C R 2i - 1. By employing a RBF NN
WiTS.i(Zi) to approximate hi(Zi), at can be expressed as
(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
Zi = gi(ZH1 + ai) -/- IPi(Xi) -/- Or Fi(Xi) -/- fi(Xi) - Qi-1

= gdzH1 - Zi-I - CiZi - 'T-T - T
Ji Fryi - Wi Si(Zi) -/- Ei] (7.98)

1 2 1 -T -1- 1 - T -1-
Vi = Vi-I -/- -Zi
2gi -/- -'TJi
2
rry• 'TJi -/- -Wi
2
r wWi
•
(7.99)
where f 1). = f~ > 0, fw, = f~, > O. The derivative of Vi is
- T - T ~
-Wi Si(Zi)Zi + Wi fw,1 Wi (7.100)
~.i = f,.i = f1).[F1),Zi -(1),i]i] (7.101)

Wi = Wi = fw, [Si (Zi) Zi - OW, Wi]
where (1)" ow. > 0 are small constants. Let Ci = CiO + Cil, where CiQ and
> O. Then, equation (7.100) becomes
Cil
(7.102)

-T -.. -
-T~
-ow, Wi Wi = -ow, Wi (Wi + Wi):::; -ow,IIWili 2
+ Ow. II Will
-
II Wi.. II
< _ Ow.l! Wi 112 + ow.llWtll2
- 2 2
-(1),r,[i]i = -(1),r,[(r,i +17;):::; -(1),IIr,iIl 2 +(1).IIr,illll17ill
< _ (1).IIr,i 112 + (1), l117i 112
- 2 2
*2
-CilZi2 + Zi€i :::; -CilZi2 + Zi I€i I :::; €.
_1_ (7.103)
4Cil
(7.104)
Step n
This is the final step. The derivative of Zn = Xn - Qn-l is
(7.105)
To stabilize the whole system (ZI,'" , zn), there exists a desired feedback con-
trol
where en is a positive constant to be specified later, and cYn-1 is a function of

Xn-l, Xd, r,1,'" , r,n-I and WI,'" , Wn- I . Thus, cYn-1 can be expressed as
where
<Pn-I
(7.108)
which is computable. Then, u* is given by
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
where Zn = [xn, O';!n-l

UXI
, ... , ~"'n-llT
UXn-l
E rl n C R 2n - 1. By employing a RBF NN
Wn TSn(Zn) to approximate hn(Zn), u* can be expressed as
(7.111)
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
in = gnu + 'Pn(xn) + ()~Fn(xn) + fn(xn) - cl n-1

= gn[-Zn-1 - CnZn - r,~Fl1n - WJ Sn(Zn) + IOn] (7.113)

1 2 1 -T -1- 1 - T -1 -
Vn = Vn- 1 + 2gn Zn + 2T1nrl1n TIn + 2Wn rWnWn (7.114)
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)
-h.n = ~.n = r I1n [FI1.. Zn - a l1 .. 7}nl

(7.116)
W n = Wn = rw.. [Sn(Zn)zn - aWn Wnl
where al1n , aWn > 0 are small constants. Let Cn = Cno + Cn1, where Cno and
> o. Then, equation (7.115) becomes
Cn 1
(7.117)
-awnWn Wn = -awnWn (Wn + W~)::; -aWn II Wn II +aw.. IIWnIlIlW:1I

-T' -T - - 2 -
< awn llWn l 2 + aWn1lW~1I2

2 2
-al1n r,~ r,n = -al1 .. r,~ (r,n + TI~) ::; -al1 .. lIr,nll 2 + U I1 .. lIr,nllllTlnll
< _ U l1n l17Jn II 2 + U I1 .. IITlnll 2
- 2 2
(7.118)
(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
then from (7.119) we have the following inequality
:s; -,Vn + 8 (7.120)
Let p:= bIT > 0, then (7.114) satisfies
(7.121)
Therefore Zi, iii and Wi are uniformly bounded. Since Zl Xl - a;dl, Zi =

Xi - Qi-l and Xd is bounded, Xi is bounded. Using (7.112), we conclude that
control u is also bounded. Thus, all the signals in the closed-loop system remain
bounded.
n 1
~ _z2 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) =
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.
7.3.2 Numerical Simulation

To verify the effectiveness of the proposed scheme, the developed adaptive NN
controller is applied to the following system
Xl =0.5Xl + 0.1xi + sinxl + 1.2x2

X2 = 0.3Xl + X2 + 0.2x~ + XIX2 + 1.7u
Y = Xl (7.126)
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)
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)
(7.129)
and T]i, Wi (i = 1,2) are updated by
~i = iii = r 1'/i [F1'/i Zi - a1'/i T]i]

. .
Wi = Wi = rWi [Si(Zi)Zi - aWi Wi] (7.130)
In practical applications of the proposed scheme, the selection of the centers

and widths of RBF has a great influence on the performance of the designed
controller. Since Gaussian RBF NNs arranged on a regular lattice on R n can
uniformly approximate sufficiently smooth functions on closed, bounded sub-
sets, in the following simulation studies, we will select the centers and widths
on a regular lattice in the respective compact sets. On the other hand, though
increasing Ci and li, and decreasing ai might result in a better tracking perfor-
mance, it is noticed that increasing Ci will lead to a high gain control scheme,
and 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 error,
and the increase in NN node number li is limited by computational feasibility.
Therefore, the design parameters should be adjusted carefully for achieving
suitable control performance in practical applications.
Neural networks W[SI(ZI) contains 5 nodes (i.e., h = 5), with centers J.tll
(l = 1,··· , h) evenly spaced in [-6,6]' and widths all = 2 (l = 1,··· , ll).
Neural networks wi S2(Z2) contains 36 nodes (i.e., b = 36), with centers J.t21
(l = 1,··· , l2) evenly spaced in [-6,6J x [-6, 6J x [-6,6]' and widths a21 = 3
(l = 1,··· , l2). The design parameters of the above controller are CI = 2.5,
C2 = 2.5, r l = r 2 = diag{1.0}, a1'/l = ff1'/2 = 0.1, aWl = aW2 = 0.1. The initial
weights WI (0) = 0.0, W2(0) = 0.0, the initial conditions [Xl (0), X2(0)jT =
[1.1, 0.5jT and [Xdl (0), Xd2(0)]T = [1.3,0.7]T.
In the following, simulation studies are conducted for two cases: (i) no a
priori information is available, i.e., FI(XI) = 0, h(XI) = 0, and F2(X2) = 0,
h (X2) = O. The unknown nonlinear functions are represented by CPI (Xl) =
0.5XI + O.lx~ + sin Xl and CP2(X2) = 0.3Xl + X2 + 0.2x~ + XIX2, and gl = 1.2
and g2 = 1.7 are assumed unknown to the controller; (ii) h(XI) = O.lx~,
Fl(Xl) = Xl, f(X2) = 0 and F2(X2) = [Xl,X2jT represent the known system
7.4 Pure-feedback Nonlinear Systems 217
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.
7.4 Pure-feedback Nonlinear Systems

While the nonlinear strict-feedback systems have been much investigated via
backstepping design, only a few results are available in the literature for the con-
trol of pure-feedback nonlinear systems [92,104,133,167]. The pure-feedback
system represents a more general class of triangular systems which have no
affine appearance of the variables to be used as virtual controls. In prac-
tice, there are many systems falling into this category, such as mechanical sys-
tems [41], aircraft flight control system [71], biochemical process [104], Duffing
oscillator [33], etc. As indicated in [104], it was quite restrictive to find the ex-
plicit virtual controls to stabilize the pure-feedback systems by using integrator
backstepping. In [92,104], while excellent results are given for global stabiliza-
tion of parametric strict-feedback systems, only local stability is achieved in a
well defined region around origin for parametric pure-feedback systems. By im-
posing additional restrictions on the nonlinearities, global stability is obtained
for a special case of the parametric pure-feedback systems in [167]. Note that
in [92,104,167]' the nonlinearities are known smooth functions, and the un-
known parameters occur linearly.
In this section, adaptive NN control schemes are proposed for the following
uncertain nonlinear pure-feedback systems [56]
~1 : (7.131)
-,
-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
(c) u: case (i) ("- -") and case (ii) ("_")

Figure 7.3: Responses of adaptive NN controller
" "
" "
""
,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
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
where Xi = [Xl, ... , XiV E Ri, i = 1, ... , n, u E R, y E R are state variables,

system input and output, respectively; /i(Xi,Xi+l) (i = 1,··· ,n-2), fjO and
gj (.) (j = n - 1, n) are unknown smooth functions. Here the difference between
I;l and I;2 only lies in gnO, i.e., gnO = gn(Xn-l) in I;l, while gn(-) = gn(xn)
in I;2.
Due to the difficulties for controlling non-affine systems, the systems con-
sidered in this section are affine in control u. Moreover, they are also affine in
Xn in the Xn-l equations. The two classes of pure-feedback systems I;l and
I;2 are motivated by previous work on the control of both pure-feedback sys-
tems [104] and uncertain strict-feedback systems [192]. As mentioned before,
this is mainly due to the fact that it is very difficult to find virtual controls
ai in terms of Xi in backstepping design procedure [104]. To overcome this
difficulty, direct adaptive NN control schemes are presented for I;l and I;2 in
this section. The idea of integrator backstepping is still employed, Le., some of
the state variables are considered as "virtual controls", and intermediate con-
trollaws are designed in the constructive design procedures. In the proposed
schemes, implicit function theorem (Lemma 2.8) is firstly exploited to assert
the existence of the continuous desired virtual controls at for i = 1, ... , n - 2.
Here the desired virtual controls at are not functions of Xi+!, but functions
of Xi. Secondly, NN approximators are used to approximate the continuous
desired virtual controls at and desired practical control u*. With the help
of NN approximation, there is no need to solve the implicit functions for the
explicit virtual controls and the practical controller to cancel the unknown
functions in backstepping design. With mild assumptions on the partial deriv-
atives of the unknown functions /i(Xi, Xi+l) (i = 1,··· , n-2) as well as on gj(-)
(j = n - 1, n), the developed adaptive NN control scheme achieves semi-global
uniform ultimate boundedness of all the signals in the closed-loop. Moreover,
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 control objective is to design direct adaptive NN controllers for systems
I;l and I;2 such that (i) all the signals in the closed-loop remain semi-globally
uniformly ultimately bounded, and (ii) the output y follows a desired trajectory
Yd in (7.2).
Assume that the states of the reference model remain bounded, i.e., Xd E Od,
Vt ;::: o.
7.4.1 Direct Adaptive NN Control for ~1
In the control of strict-feedback system, it is required that the affine term

giO "I 0 [192J. Similarly, for the control of pure-feedback system ~1 and
~2, define gi(Xi, Xi+l) := afi~Xi'Xi+t), i = 1,' .. ,n - 2. Note that the partial
X,+l
derivatives gi(Xi, Xi+l) are still unknown nonlinear functions. The following
assumption is made for gi('), i = 1", . ,n.
Assumption 7.4. The signs of gi(Xi, Xi+l), i = 1, ... ,n - 2, gn-l(Xn-l) and
gn(xn-t} are known, and there exist constants gil > giO > 0 such that (i)
IgiOI > giO > 0, VXn ERn, and (ii) IgiOI ~ gil < 00, VXn E SlXn eRn where
Slxn is a compact region, i = 1, ... ,n.
The above assumption implies that partial derivatives 9i('), i = 1",' , n
are strictly either positive or negative. Without losing generality, it is assumed
that gi 0 > giO > 0, VXn E Rn.
Accordingly, the derivatives of gi(-), i = 1, ... , n are given by
(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
the ith-order subsystem is stabilized with respect to a Lyapunov function ~

by the design of a stabilizing function ni, where an RBF NN is employed to
approximate the unknown part in intermediate desired feedback control ni·
The control law u is designed in the last step.
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)
where kl > 0 is a constant. It is clear that VI is a function of Xl and Xd.

Considering the fact that avr/ aX2 = 0, the following inequality holds
a[h (Xl, X2) + VI] 0
n > glO >
uX2
According to Lemma 2.8, by viewing X2 as a virtual control input, for every

value of Xl and VI, there exists a smooth ideal control input X2 = ni(XI, VI)
such that
(7.136)
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
Since Xd is assumed to be bounded, similar to Assumption 7.5, it is reason-

able to assume that lillAl I is bounded by some constant within some com-
pact regions. For uniformity of presentation, we assume that lillA101 ::; gld,
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)
By employing an RBF NN WTSl(Zl) to approximate aiexl,vl), where Zl =

[Xl, Xdl, ZlJ T E 0 1 C R 3 , at can be expressed as
(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)
where Cl is a positive constant to be specified later. Then, the dynamics of Zl

is governed by
(7.141 )
where WI = WI - Wr
Through out this section, define () = () - (-)*.
(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)
(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)
By completion of squares, the following inequalities hold
(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)
where k2 > 0 is a constant. Considering the fact that 8V2/0X3 = 0, the

following inequality holds
0[h(x2, X3) + V2J 0

n
UX3
> 920 >
According to Lemma 2.8, by viewing X3 as a virtual control input, for every
value of X2 and V2, there exists a smooth ideal control input X3 = a;(x2' V2)
such that
(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
and a 2, and from (7.149)(7.151), it can be seen that a~ is a function of X2, Z2

and aI, i.e., a function of Xl, X2, Xd and WI, the deriv~tion of 92).,2 is
Similar to Step 1, it is reasonable to assume that 192).,21 is bounded by some

constant within some compact regions. For uniformity of presentation, we
assume that 192).,2 (.) 1 ::; g2d, 't/xn E nXn , Xd E nd and WI E nWl' where g2d
is the same as in Assumption 7.5 for 1.921. In other words, Assumption 7 ..5 on
.92(X2, X3) is valid for .92).,2' Subsequently, such an assumption is also made for
.9i).,j in the following steps.
Combining (7.148}-(7.151) yields
(7.152)
Since al is a function of Xl, xd and WI, al is given by
(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,
neural network weights WI are not recommended to used as inputs to the NN

controller under construction. This is because of the curse of dimensionality
of RBF NN which may result in the number of NN inputs to be too large
for wi S2(Z2). By defining the intermediate variable ~ and <PI which are
computable, the NN approximation wi S2(Z2) for 0(2{X2, 112) can be computed
by using the minimal number of NN inputs Z2 = [X2, ~, <PI, Z2V. Note also
that the variable vector Xd does not appear in Z2, because it has been combined
into the intermediate variable <Pl.
Since W; is unknown, 0(2 cannot be realized in practice. Let W2 be the
estimate of Wi. Define Z3 = X3 - 0(2 and let
(7.156)
where C2 is a positive constant to be specified later. Then, the dynamics of Z2

is governed by
(7.157)
(7.158)
The derivative of "1/2 is
(7.159)
(7.160)
where a2 > 0 is a small constant. Let C2 = C20 + C21, where C20 and C21 > o.
(7.161)
(7.162)
Because - (C20 + -g2)..2

2 - ) Z22 ~ - ( C20 - 2 - ) Z2,
-g2d 2 bY cl '
100smg C20 such t 1lat
2g 2 )..2 29 20
* := 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)
where k i > 0 is a constant. Considering the fact that 8vi/8xHl = 0, the

following inequality holds
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)
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)
Since ai-l is a function of Xi-l, Xd and Wl , ... , Wi - l , O:i-l is given by
(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)
where Ci is a positive constant to be specified later. Then, equation (7.168)

becomes
-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)
Vi = Vi-I + - -1Z i2 + -Wi

1 - T -1-
r i Wi (7.174)
2gi)'i 2
The derivative of Vi is
. . 2 .
V=V + ZiZi _ gi>"Zi + wTr~lw.
• .-1 '.
g.>.., 2gi>"i
2 ..'
= \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.
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)
-T" -T - - 2 -
-aiWi Wi = + Wt) ~ -ailiWill
-aiWi (Wi +ailiWillllwtll
aillWdl2 ai II wt 112
< - 2 + 2
(7.178)
B ecause - (Cm + -,gi>"i

-2- ) Zi2 ~ - ( CiO - gid
- 2 ) Zi'
2 b Y Cl '
100Sll1g Cm suc h t 1lat
2gi >"i 2gio
* :=
CiO CiO -
gid
-2 >0
2gio
the derivative of Vi satisfies
i
~aj IIW*112
j i
~ Ej
*2
+ L.,.. + L.,.. -4C '1 (7.179)
j=1
2 j=1 J
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)
Then, the dynamics of Zn-l is governed by

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
Wn - 1 = Wn - 1 = r n - 1 [-Sn-1(Zn-1)Zn-1 - a n -1 Wn-l] (7.189)
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)

-T -T -
-an-l W n - 1 W n - 1
~
= -an-l Wn_1(Wn - 1 + W n..- 1 )
S -an-l IIWn-l 112 +an-lIlWn-11lIlW:_11I
(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
by choosing C(n-l)O such that
".- g(n-l)d 0
c(n-l)O . - C(n-1)O - 2 2 > (7.192)
g(n-l)O
the derivative of Vn - I satisfies
(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 Cn is a positive constant to be specified later.

Since an-l is a function of Xn-l, Xd and WI,'" ,Wn - I , an-l is given by
n-2
",oan-l (_ ) oan - 1 ( ) ()
a n -l = 6 -o--Ik Xk+l + -0-- !n-l + gn-lxn + ¢n-l 7.196
k=1 Xk Xn-I
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.
Since W,,: is unknown, u* cannot be realized in practice. Let Wn be the

estimate of W,,:, and
(7.200)
where Cn is a positive constant to be specified later. Then, equation (7.194)

becomes
(7.201)

_ 1 2 1 - T -1-
Vn - Vn- 1 + -Zn
2gn
+ -Wn
2
rn Wn (7.202)
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.
'r _ .
~n - Vn- 1 - Zn-1 Zn - (CnO + -gn2 )Zn2 - 2
Cn1Zn - Zncn - (TnWn Wn
- T '
(7.205)
2gn
-(Tn W,;CWn + W~) ::; -(TnIIWnI12 + (TnIIWnIIIlW":1I

a n llWn l12 an11W~112
< - 2 + 2
(7.206)
Because -(cnO + in? )z~ ::; -{CnO - 9~d )z~, by choosing CnO such that
gn 2gno
C*nD := CnD - -gnd

2- >
0
29 no
the following inequality hold
n-2 k n n Ilw II" n Ilw"112 n *2

Vn < _ '"" _J_. Z2 _ '"" C"OZ2 _ ' " " (Yj j ~ + '"" (Yj j + '"" l
~ g.>. J ~ J J ~ 2 ~ 2 ~ 4C1
j=l J J j=l j=l j=l j=l J
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)
Let p := oh > 0, then (7.202) satisfies
(7.209)
Therefore and Wi are uniformly ultimately bounded. Since Zl = Xl - Xd1,

Zi
Zi = Xi - and Xd is bounded, it can be seen that Xi is bounded. Using
O:i-1
(7.200), it is concluded that control u is also bounded. Thus, all the signals in
the closed-loop system remain bounded.
Combining equation (7.202) and inequality (7.209) yields
(7.210)
Let g* = maxl~i~n{gi1}' Then, the following inequality holds
(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.
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.
7.4.2 Direct Adaptive NN Control for 1:2

In this section, the design procedure is very similar to that of Section 7.4.1,
except that integral Lyapunov function is employed in controller design to avoid
the possible singularity problem caused by gn(Xn) in the last equation of E 2 .
For system E 2 , all the assumptions on gi(-), i = 1"" ,n - 1 are the same.
The following assumption is made for gn(xn).
Assumption 7.6. The sign of gn(Xn) is known, and there exist a constant
gnO > 0 and a known smooth function 9n(Xn) such that 9n(Xn) 2: Ign(xn)1 2:
gnO, "i/xn ERn. Without losing generality, it is assumed that 9n(Xn) 2: gn(xn) 2:
gnO, "i/xn ERn.
The direct adaptive controller is given by
(7.214)
where
Zl = Xl - Xdl, Zi+l = Xi+! - ai, 1::; i ::; n - 1

, T
al = -ClZl + WI 8 1 (Zl)
, T
ai = -Zi-l - CiZi
Si(Zi) + Wi
Zl = [Xl,Xdl,ZlfEf!ICR 3
- Oai-l
Zi = [Xi, -8--'
8ai-1
... , -8--' <!Ji-l. Zi
IT E f!i C R
2i+1·
,2 ::; z ::; n - 2
Xl Xi-l
_ [- 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-
mates to Wt, and the adaptation laws being given by
(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
VZn = lzn (Jg;;1'1;ax(X n -1,a + an-1)da

= Z~ 11 8g;;~ax(Xn-I' 8zn + a n-l)d8 (7.218)
Noting that 1 ::; g;;~ax(xn-l,8zn + an-I) :S 9n(Xn-I,8zn + an-d/gno (As-

sumption 7.6), the following inequality holds
(7.219)
Therefore, VZn is positive definite and radically unbounded with respective to

Zn. Choose
(7.220)
as a Lyapunov function candidate. Its time derivative becomes
(7.221)
Since an-I is a function of Xn-I, Xd and WI,'" ,Wn - I , O:n-I is given by

n-'}
~ oan-I _
O:n-I ~ -O--!k(Xk+I) + oan-I )
-J:l--(fn-I + gn-IXn + ¢n-I
()
7.222
k=l Xk UXn-I
where
(7.223)
is computable.
Using the fact that
we obtain
(7.224)
where
(7.225)
with
To stabilize the whole system (Zl,'" ,Zn), there exists a desired feedback
control
(7.226)
where Cn is a positive constant to be specified later.

By employing an RBF NN Wn T Sn(Zn) to approximate -hn(Zn), u· can
be expressed
(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
U = gn(~n) [ - Zn-l - CnZn + WJ Sn(Zn)] (7.228)
The derivative of Vn becomes

. . 2
Vn = Vn- 1 - Zn-lZn - CnZn - Znfn
-T -T
+Wn Sn(Zn)Zn + Wn r;:;- l~W n (7.229)
lirn = W n = r n[-Sn(Zn)zn - an Wnl (7.230)
where an > 0 is a small constant. Let Cn = CnO + Cnl> where Cnl > 0 and
(7.231)
with 'Y being a positive constant. Then, equation (7.229) becomes

. . 2 2 -T~
Vn = Vn - 1 - Zn-1Zn - CnOZn - CnlZn - Znfn - anWn Wn (7.232)
(7.233)
Then, the derivative of Vn satisfies
n-2 k n-1 n IIW- 112

v:n < - "L -L
.
z2 _
J
" c * Z2 _ C z2 _ ,,<Tj
L JO J nO n L 2
j
j=1 gJ)' j=1 j=1
+" 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
Let p := Jh > 0, then (7.220) satisfies
(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. <>
7.4.3 Simulation studies

To verify the effectiveness of the proposed approach, the developed adaptive
NN controller is applied to the following nonlinear system:
(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)
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)
where Z2 = X2 - aI, Z3 = X3 - a2 and Z3 = [Xl, X2, X3'~'~' 4>2]T E R6

with
Zl = Xl - Yd, al = -ClZl + W[Sl(Zd, Zl = [Xl, Xdl, Zl]T E R3

aal . aal . aal ~
4>1 = - a Xdl + - a Xd2 + -.-W1
Xdl Xd2 aWl
• T aa1 T 4
a2 = -Zl - C2 Z2 + W 2 8 2(Z2), Z2 = [Xl, X2, - a ,4>1] E R
Xl
aa2 . aa2 . aa2 ~ aa2 ~
4>2 = --Xdl + --Xd2 + -.-W 1 + - . - W2 (7.240)
aXdl aXd2 aWl aW2
and NN weights WI, W 2 and W3 are updated by (7.144), (7.189) and (7.204)
correspondingly.
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.
,~
2 "
1
o 2 4 e 6 W U '4 ~ 1& H -"'--~-~~--~'--~1O---:'12:----;'~'~1~'-,~,----:"
..
Time (Secon::lll) Time (Se.:ondll
.• , ,
,~
" .;\ ~ :: ~ 'I~~~~:~
1\
\,
J,/I
'" r~ " ,:;}:I\~i:1 f, /'; .'1: I~I ,. :' :I t\ /,11,
/\t*·i.
2
It • ,,\, - - ~':r:{\!:, ~, ~
I
~J:Y:'r:~::~~'~
..;r......
-., " ~ \11,11 ,1

II',
I, ~ \'
1/
~ ~~ ~
"II
"
-"
o 2 " e 6 W
TlIfM(SeamcIsl
U 14 18 D H -". • 10
TIIIII/Slccru:II1
12 . " "
Figure 7.5: Xl ("-") and Xdl ("- -") Figure 7.6: X2 ( " - " ) and X3 ("- -")
';E:j'
."
-.,
~'~-2~~'--~'--~'--~1O~~12~~1~'~1~'--'~'~"·
Time (SecCllldil
.~ o 1 " II e 10
TIIIIII!Sc0ftd5I
12 !4 HI l' 20
Figure 7.8: IIWIII ("-"), IIW211 ("- -")

Figure 7.7: Control input u
and IIW3 11 ("-.-.")
7.5 MIMO Nonlinear Systems

In this section, we consider adaptive neural control of the broader class of
multi-input/multi-output (MIMO) nonlinear systems. For MIMO nonlinear
systems, the control problem is very complicated due to the couplings among
various inputs and outputs. Moreover, when there exist uncertain parame-
ters and/or unknown nonlinear functions in the input coupling matrL'<, it is in
general very difficult to deal with the problem of controlling MIMO nonlinear
systems. For some MIMO nonlinear systems with parametric uncertainties in
the input coupling matrix, several adaptive control schemes have been proposed
based on feedback linearization techniques (e.g., [23,133,138,165]). In these
schemes, to remove the couplings of system inputs, an estimate of "decoupling
matrix" is usually needed and required to be invertible during the parameter

adaptation. Therefore, additional efforts have to be made to avoid the possible
singularity problem when calculating the inverse of the estimated "decoupling
matrix", e.g., by using projection algorithm [165] to keep the estimated para-
meters inside a feasible set in which the singularity problem does not happen.
The disadvantage of using projection is that, it usually requires a priori knowl-
edge for the feasible parameter set and no systematic procedure is available for
constructing such a set for a general plant [74J. As an alternative approach,
an adaptive multilayer neural network controller was proposed in 1112J for a
special class of MIMO nonlinear plants, rigid robot systems, using several nice
properties of the input decoupling matrix of robotic systems. An energy-type
Lyapunov function is chosen to develop the stable neural controller and adap-
tation laws. More recently, in [51], a class of MIMO nonlinear systems with a
triangular structure in control inputs was considered. By using this triangular
property, integral Lyapunov functions are introduced to construct a Lyapunov-
based controJ structure, which does not try to cancel the "decoupling matrix"
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.
In this section, adaptive neural control schemes are presented for a class of
uncertain MIMO nonlinear systems El in block-triangular forms. The MIMO
system El is composed of subsystems which are interconnected. Compared
with the MIMO system considered in [51J, the block-triangular MIMO system
in this section is more general in system state interconnections, which makes
it difficult to conclude the stability of the whole system by stability analysis of
individual subsystem separately. However, due to the block-triangular struc-
ture properties, it is feasible to design for each subsystem a full state feedback
controller, and the stability of the whole system can be concluded by stability
analysis of all the state variables in a nested iterative manner. By exploiting
the special properties of the affine terms of the MIMO system Et, the devel-
oped scheme avoids the controller singularity problem completely without using
projection algorithms. The stability of the resulting adaptive systems is guar-
anteed without the requirement for the integral Lyapunov function [51], which
is regarded as difficult to use. By employing NNs to approximate all the uncer-
tain nonlinear functions in the controller design, the developed scheme achieves
semi-global uniform ultimate boundedness of all the signals in the closed-loop
of the MIMO systems. The outputs of the system are proven to converge to a
small neighborhood of the desired trajectories. The control performance of the
closed-loop system is guaranteed by suitably choosing the design parameters.
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.

Consider the following uncertain MIMO nonlinear systems El in the block-
triangular forms
Xlh = ft,il (Xl,(il-ell)'" . ,Xm,(il-elm))
+ gl,il (Xl,(h -ell)' ... ,Xm ,(il-elm))Xl,il +1
1 ::; il ::; PI - 1
X1,Pl = h,Pl (X) + gl,Pl (Xl,Pl-1, ... ,Xm ,Pm-l)Ul
(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
where X = [XT,2' Xf,2]T with Xj,2 = [Xj,l, Xj,2]T, j = 1,2, and

X1,1 = h,l(X1,l,X2,3) + gl,1(X1,1,X2,3)X1,2,
X1,2 = h,2(X) + gl,2(Xl,l, X2,3)U1
X2,1 = h,l(X2,1) + g2,I(X2,l)X2,2,
X2,2 = h,2(X2,2) + g2,2(X2,2)X2,3, (7.244)
X2,3 = h,3(x1,1,x2,3) + g2,3(X1,1,X2,3)X2,4,
X2,4 = h,4(X, Ul) + 92,4(Xl,I, X2,3)U2
Yj = Xj,I, j = 1,2
where Xj,i' = [Xj,I,'" ,Xj,ij]T, j = 1,2, il = 1,2, i2 = 1"" , 4, and X =

[XT,2' Xi4]7'. System ES, shows the case when the orders of the subsystems are
the same (PI = P2 = 2), while system ES2 demonstrates the situation when
the orders of the subsystems are different (P1 = 2, P2 = 4). These two cases
show how the state variables of one subsystem are embedded into the other
subsystem according to the values of i j - (2jl as stated above in Remark 7.15.
Compared with the MIMO system considered in [51], where the system in-
terconnections are only limited to hpj (X), the above block-triangular MIMO
system (7.241) has the most system interconnections that the approach, which
we are going to present, can handle due to the introduction of Xl,(ij-ejL)'
The system interconnections are represented by terms fJ,ij (Xl,(ij-ejd, ... ,
Xm,(ij-ej",)) and 9j,i)Xl,(ij-ejd, ... , Xm,(irej",)), i j = 1,··· ,Pj - I, j =
I,··· ,711, as well as by terms fJ,Pj (X,Ul,'" ,Uj-l) and 9j,Pj(X1,Pl-l, ... ,
Xm ,Pm-l). Note that hpj (X, Ul,'" , Uj-1) (j = 1,··· ,711) are very general
in the sense that they include not only the system states X, but also the
system inputs ul,'" , Uj-l, where Ul,'" ,Uj-l are the control signals of the
(1, ... ,j - l)th subsystems directly passing through to the jth subsystem.
Notice that the affine terms 9j,Pj(Xl,Pl-l,'" ,Xj,p",-l), j = 1,··· ,711 are
assumed to be independent of the states Xl,Pl"" ,Xj,Pj' By utilizing this
special structure property, which can be found in many practical systems, the
controller singularity problem is avoided, and the stability of the resulting adap-
tive system is guaranteed without the requirement for the integral Lyapunov
function [51].
The control objective is to design adaptive neural controller for system
(7.241) such that (i) all the signals in the closed-loop remain semi-globally
uniformly ultimately bounded, and (ii) the output Yj follows the desired tra-
jectories Ydj generated from the following smooth, bounded reference model
Xdji = fdji(Xd), 1::; i ::; nj
(7.245)
Ydj = Xd:jl, 1 ::; j ::; 711
where nj ~ Pj, j = 1, ... , rn, Xd = [Xdll, ... ,Xd1n" Xd21, ... , Xd2n2' ... , Xdml,
"', Xdmn",]T E Rn'+"+n", are the states, Ydj E R, 1::; j ::; 711, is the system
outputs, fdji(') are known smooth nonlinear functions.
Note that in the following derivation of the adaptive neural controller, NN

approximation is only guaranteed within a compact set. Accordingly, the sta-
bility results obtained in this work are semi-global in the sense that, as long
as the input variables of the NNs remain within some compact sets, where the
compact sets can be made as large as desired, there exists controller(s) with
sufficiently large number of NN nodes such that all the signals in the closed-loop
remain bounded.
For the control of the uncertain MIMO system ~1 (7.241), we make the
following assumption as commonly being done in the literature.
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.
Assumption 7.S. There exist constants gJd,i. ]

> 0, ij = 1, .. , , Pj, j = 1, ... , Tn
such that 19j,iJ) I ::; gj,ij in the compact set nj,ij "it 2: O.
For uncertain MIMO nonlinear system (7.241), we employ the idea of back-
stepping to design controllers for all the subsystems of (7.241). Note that
because all subsystems in.system (7.241) are interconnected, it is difficult to
conclude the stability of the whole system by stability analysis of individual
subsystem separately. However, due to the block-triangular structure prop-
erty, it is feasible to design for each subsystem a full state feedback controller,
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
Define Zj,l = Xj,l - Xdjl. Its derivative is
Zj,l = !j,l(Xl,(l-l!jl)'··· ,Xm,(l-I!jm))
+gj,l(Xl,(l-l!jl)'··· ,Xj,(m-l!jm))Xj,2 - Xdjl (7.248)
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)
where is a positive design constant to be specified later. gj,l (Xj,l) and

Cj,l
!j,l(Xj,l)are unknown smooth functions of Xj,l.
Let hj,dZj,l) = gj\ (fj,l - Xdjl) denote the unknown part of ai,l' with
(7.250)
By employing an RBF NN WDSj,l(Zj,l) to approximate hj,l(Zj,d, ai,l can

be expressed as
(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)
Then, we have
(7.253)
(7.254)
The derivative of vj) is

. . 2
V· Zj) Zj,l _ gj,l Zj,l + W- T r-. 1 W~ .' 1
j,l = g. 2g2 ),1 ),1 J,
J,l j,l
2 9j,1 2
= Zj,l Z j,2 - Cj,l Z j,l - 2g2 Zj,l + Zj,lEj,l
),1
(7.255)
Consider the adaptation law for Wj,l as

. .
ltV ), 1 = W 1 = rw [8
), l(Z
J,l) l)Z'l
,], ), - 0"'1Hr·
] , ] , 1] (7.256)
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)

*
-T -
-O"j,l W j ,l W j ,l = -T -
-O"j,l W j ,l (Wj,l + Wj,l)
::; -O"j,dIWj,lI12 + O"j,lllWj,lllllWj~lll
< _ O"j,lllWj ,lll2 + O"j,l IIWj,l 112
- 2 2
(7.258)
. 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
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
Step i·) (2 <

- i'J <
- p'J - 1)
Define Zj,ij = Xj,i j - aj,ij-l. Its derivative is
Zj,i j = fJ,ij(Xl,(ij-l!jt),··· ,xm,(ij-l!jm»)
+gj,ij (Xl,(ij-l!jd,'" , X m ,(i j -l!jm»)Xj,ij +l - aj,ij-l (7.260)

where ejl = Pj - PI, I = 1, ... ,711, and if i j - ejl :::; 0, then the corresponding
variable vector Xj,(i j -l!jl) does not exist.
By viewing Xj,ij+l as a virtual control to stabilize the (Zj,I,'" , Zj,ij)-sub-
subsystem of the jth subsystem, there exists a desired feedback control
(7.261 )
where Cj,ij > 0 is a design constant to be specified later, aj,ij-l is a function of

Xl,(ij-l-i!jt}, ... , X m ,(ij-l-i!jm)' Xd and Wj ,l, ... , Wj ,ij-l' Therefore, aj,ij-l
can be expressed as
(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)
denote the unknown part of aJ~ , i J., where

T
x-T
[1,(ij-l!jt)'
... , x-T
m,(ij-l!jm)'
( Oaj,ij-l
0- .
)
" ...
x1,('j-l-l!jl)
( 0-
oaj i_l)T
'J , <Pj,ij-l
]T E rlj,ij (7.264)
X m ,(ij-l-l!jm)
By employing an RBF NN WEj Sj,ij (Zj,i,) to approximate hj,i j (Zj,iJ, a;,ij

can be expressed as
(7.265)
where WJ*, •..3 denotes the ideal constant weights, and IEJ' ' i J I :::; EJ* i'J is the approx-
I
imation error with constant EJ\ , J > O.

Since WJ*i.3 is unknown, aJ*i.
I , J
cannot be realized in practice. Let us define
Zj,ij+1 = Xj,ij+1 - aj,ij and
(7.266)
Then, we have
ij,ij = hij + gj,ij(Zj,ij+l + aj,ij) - aj,ij-1

- T -.
= gj,ij[Zj,ij +1 - Zj,i j -1 - Cj,ijZj,ij - Wj,V~j,ij(Zj,iJ + Ej,ij](7.267)
Remark 7.17. The principle for designing the neural network Wj~j Sj,ij (Zj,ij)
is to use as few neurons as possible to approximate the unknown function
Though hj,ij (Zj,i,) is a function of X1,(ij-l1j1)' " ' , Xm,(ij -l1jm)' Xd
hj,ij (Zj,i,).
and W j ,l, "', W j ,i j -1, the weight estimates W j ,l, "', W j ,ij -1 are not recom-
mended to be taken as inputs to the NN because of the curse of dimensionality
of RBF NN (see, e.g. [66]). By defining intermediate variables (aa\~~;'~i1=:jt!) T,
"', ( aCl.ji·_1
.' J
a-Xm.(lj-l-"jm) )T and ¢j,i'-l,
J
1
wlich 1 tle
are availa ble t h rougl 1 computa-
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.
(7.268)
The derivative of \!j,ij is
(7.269)
Consider the adaptation law for Wj,ij as
(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
The derivative of Zj,Pj = Xj,Pj - OJ,Pj -1 is

To stabilize the jth subsystem (Zj,l,' .. ,Zj,Pj)' there exists a desired feedback
control
(7.273)
where Cj,p; > 0 is a design constant to be specified later, Ci.j,Pi -1 is a function

of X1,(pt- I ), " ' , xm,(p",-I), Xd and W j ,l,'" ,Wj,P;-l' Therefore, Qj,p;-I can
be expressed as
(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)
By employing an RBF NN Wlp; Sj,p; (Zj,p;) to approximate hj,pj (Zj,Pj)' uj

can be expressed
(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
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,
The derivative of Vj,Pj is
.
Vj,Pj -V.
- J,pj-1
+ z·J,Pj i·J,Pj
gj,Pj
(7.280)
Consider the, adaptation law for Wj,Pj as
(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
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
~ -1'j Vj,Pj + 6j (7.286)
Let V = 2::;~1 Vj,Pj. Its derivative is

m m
V= L ~,Pj < L( -1'j Vj,Pj + 6j )

j=l j=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
of the interconnection between the subsystems in system (7.241), we cannot

conclude the stability of the whole closed-loop system by stability analysis of
individual subsystem separately. However, due to the structure property of
system (7.241), we can prove the stability of the states X of the whole MIMO
system in a nested iterative manner, as will be shown in the proof of the
following theorem.
Theorem 7.5. {55JConsider the closed-loop system consisting of the plant

(7.241), the reference model (7.245), the controllers (7.277) and the NN weight
updating laws (7.256)(7.270)(7.281) for all the subsystems (j = 1, ... ,m). As-
sume there exist sufficiently large compact sets S1 j ,i; such that Zj,i; E S1 j ,i;, i j =
1, ... ,Pj, j = 1, ... ,m for all t 2: O. Then, for bounded initial conditions,
(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.
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
Ur ~ {(j, i j ) I Xj,ij not proven bounded yet} (7.290)
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.
Using (7.277), we conclude that Ul is bounded because it is a function of

bounded variables X, Xd and Wl,l, ... , Wl,Pl' Similarly, we have that controls
Uj, j = 2, ... ,Tn are also bounded. Thus, all the signals in the closed-loop
system remain bounded.
To provide some estimates of the regions of attraction of equation (7.287),
we consider the following two conditions.
1) if
(7.291)
where Osl is given in (7.288), then according to Theorem 2.14 in [156],

all the states X and the neural weights Wj = [WD,'" , WDi]T (j =
1"" ,rn) will remain in Osl, i.e.,
(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)
Let 9J' = max1<i-<p'{§J' i'}' Then, we have

- J_:J 1 :J
(7.297)
that is,
Pj
L Z;'k < 2gjwj + 2gj \Ij,Pj (O)exp( -"Ijt) (7.298)
k=l
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.
a) System ESl (7.243) for case 1)

After proving the boundedness of Xl,l and X2,I, we can proceed to prove
the boundedness of aj,lt j = 1,2, because aj,I (7.252) is a function of
Xl,l, X2,1. Xd and Wj,l. Since the states Xl,l and X2,l have been proven
bounded, we conclude that aj,I, j = 1,2 remain bounded. This leads to
the boundedness of Xl,2 and X2,2 by definition. Thus, the states X of the
overall system (7.244) are proven bounded.
b) System ES2 (7.244) for case 2)

As stated in case 2), after proving the boundedness of Xl,l and X2,lt
we cannot proceed to prove the boundedness of al,l because as can
be seen from (7.252) and (7.250), al,I is also a function of X2,3, where
X2,3 = [X2,l' X2,2, X2,3]T, and X2,2, X2,3 have not been proven bounded yet.
According to (7.289), we know that q = 2 (i q = 2 in this step), which
means that we can prove the boundedness of a2,l of the second subsystem
in ES2 first. (From (7.252), it is clear that 0:2,1 is a function OfX2,I, Xd,
W2 ,l, which have been proven bounded already). This will lead to the
boundedness of X2,2 by definition.
By using (7.289), we find that (q,i q ) = (2,3)' which means that we can
prove the boundedness of Cl!2,2 and X2,3. From (7.266) and (7.264), we
can see that 0:2,2 is bounded because it is a function of X2,l, X2,2, Xd,
W2 ,l and W2 ,2, which have all been proven bounded. The boundedness
of a2,2 will lead to the boundedness of X2,3 by definition. However, we
cannot prove the boundedness of 0:1,1 and Xl,2 in this step, because we
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
B2 : Xj,i, = hij (X1,(ij -{l,d' ... ,Xm,(i j -(ljm») (7.300)

+ eIijFj,ij (X1,(ij-{ljl)"" , Xm,(ij-i!jm»)
+ Cj,ij (XI,(i j -(ljl)' .. , , Xm,(ij -i!jm») + gj,ij Xj,ij+l,
1 :s: i j :s: Pj - 1
Xj,Pj = fj,pj (X, UI, ... , Uj-l) + eIpj Fj,pj (X, UI, ... , Uj-d
+ Cj,Pj (X, UI,'" ,Uj-1) + gj,pjUj
Yj = Xj,l, 1:S: j :s: m
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

262 Chapter 8. Conclusion
nonlinearly parametrized plants. Through combining neural networks, inte-

gral Lyapunov function, and adaptive backstepping design, stable adaptive NN
control was then investigated for a class of strict-feedback nonlinear systems.
Finally, adaptive NN control design has been extended to multivariable nonlin-
ear systems with a triangular control structure. In all these proposed designs,
transient behaviours of the resulting adaptive systems were explicitly provided,
and the relationships between control performance and design parameters were
also discussed for guiding the control system design.
For adaptive control of non-affine nonlinear systems, this book addressed
three aspects in Chapter 6: (i) state feedback NN control for general non-affine
systems, which guarantees regional stability of the closed-loop neural systems,
(ii) output feedback NN control for non-affine systems using high-order ob-
server, which ensures that the tracking error converges to an adjustable neigh-
borhood of zero, and (iii) application study for the CSTR plant demonstrate
the effectiveness of the developed control schemes.
Several classes of triangular nonlinear systems have also been thoroughly
studied in Chapter 7. Direct adaptive NN control approaches have been pro-
posed for (i) special nonlinear system in strict-feedback form with property of
9n(xn-d being independent of state Xn; (ii) partially known nonlinear system
in strict-feedback form; (iii) pure-feedback nonlinear systems; and (iv) MIMO
nonlinear systems in block-triangular forms. All the developed adaptive neural
control schemes in this chapter achieve semi-global uniform ultimate bounded-
ness 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.
8.2 Design Flexibility

Though Lyapunov design is sufficient, it is a difficult problem to find a Lya-
punov function satisfying the requirements of Lyapunov design. Lyapunov
functions are additive, like energy, i.e., Lyapunov functions for combinations
of subsystems may be derived by adding the Lyapunov functions of the sub-
systems. This point can be seen clearly in the adaptive control design and
backstepping design. In general, Lyapunov design is usually a trial-and-error
process and there is a lack of systematic methods. Different choices of Lyapunov
functions may result in different control structures and control performance. As
shown in the book, Lyapunov design is not unique, and is problem dependent.
It has been demonstrated that superior control laws can be obtained by exploit-
ing the properties of the system under study [46J. The choices for functions Qi
in the iterative backstepping design are by no means unique, nor is the final
choice for the control law for systems in strict-feedback form. In actual applica-
tions, different types of Lyapunov functions, quadratic, energy-based quadratic
8.3 Further Research 263
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.
8.3 Further Research

Although this book solved several important problems in the adaptive neural
network control, many open questions remain for future research. In the fol-
lowing, some suggestions are made for further studies:
1. Persistent excitation (PE) for ne'ural control systems
Due to the limited space and time, the book did not discuss the effects
of PE conditions for the adaptive NN systems. Though the results of
this book show that stability and tracking performance are guaranteed
without such a condition, PE condition is actually very important for
the parameter convergence and the improvement of system robustness
and performance. Works [38,40,64] have studied PE conditions of lin-
early parametrized networks and provided several conditions for system
identification and adaptive NN control. For multilayer NNs, no theoreti-
cal result is available in the literature to guarantee/check PE conditions.
The research on PE condition and its effects on adaptive system is an
important topic in adaptive neural control field.
2. Enlarging the class of nonlinear systems handled by neural control tech-
niques
In the development of adaptive nonlinear control, two types of control
schemes have been considered [91]: the uncertainty-constrained schemes
[18,174] which impose restrictions (matching conditions) on the loca-
tion of unknown parameters, and the nonlinearity-constrained schemes
[133,165,175] which impose restrictions on the type of nonlinearities of
the original systems. Motivated by such a design idea, the authors be-
lieve that similar conditions can also be exploited to enlarge the class of
nonlinear systems for which stable adaptive neural control techniques are
applicable.
264 Chapter 8. Conclusion
3. Adaptive neural control for MIMO nonlinear system

At the present stage, few results are available in the literature for adap-
tive control of MIMO nonlinear systems. This book only addressed the
NN control problem for some special classes of MIMO systems. It is
believed that pursuing different ways to utilize the structure properties
of different MIMO systems would be beneficial to derive a stable neural
control scheme for a larger class of multi variable nonlinear systems.
4. Adaptive NN control for non-affine nonlinear systems

Because of the complexity of the studied problem, the work on non-
affine nonlinear systems presented in the book are only some preliminary
results. Research efforts can be directed to study the global adaptive
control scheme, further relax the assumptions made in the book, develop
other stable output feedback NN approach, and investigate adaptive NN
scheme for multivariable non-affine nonlinear systems. The development
of stable adaptive NN control for non-affine systems is an important and
challenging problem in the future.
5. Control of Discrete-t'ime and Sampled-Data Systems
All of the approaches developed in the book are based on continuous-
time system representation because compared with discrete-time system
control, design and analysis of the continuous-time adaptive systems are
more tractable. At the present stage, it appears to us that many re-
sults presented in the books cannot be obtained directly for discrete- time
neural control systems. In particular, it is very hard to find a similar
"integral" Lyapunov function for discrete-time nonlinear systems which
satisfies the required properties in the proof, and all the elegant methods
for strict-feedback systems in continuous-time are not directly applicable
to discrete-time strict-feedback systems due to the noncausal problem in
the controller design procedure via backstepping. Only preliminary re-
sults are available for discrete-time nonlinear systems [53, 54J. Sampled-
data systems are hybrid dynamical systems that contain both discrete
and continuous signals. They represent an important class of nonlinear
systems that adaptive NN control should be investigated.
References
[IJ J. S. Albus. A new approach to manipulator control: the cerebellar model

articulation controller (cmac). Journal of Dynamic Systems, 97:220-227,
1975.
[2] B. Aloliwi and H. K. Khalil. Robust adaptive output feedback con-
trol of'nonlinear systems without persistence of excitation. Automatica,
33:202.5-2032, 1997.
[3J A. M. Annaswamy, F. P. Skantze, and A. P. Loh. Adaptive control of con-
tinuous time systems with convex/concave parametrization. Automatica,
34:33-49, 1998.
[4] T. M. Apostol. Mathematical Analysis. Addison-Wesley, Reading, Mass,
1957.
[5J K. J. Astrom and B. Wittenmark. Adaptive Control. Addison-Wesley,
Reading, Mass, 1989.
[6J A. G. Balestrino, De Maria, and A. S. I. Zinober. Nonlinear adaptive
model-following control. Antomatica, 20:.5.59-568, 1994.
[7J A. R. Barron. Approximation and estimation bounds for artificial neural
networks. In Proc. 4th Ann. Workshop on Comp'utational Learning The-
ory, pages 243-249, San Mateo: Morgen Kanffman, August 1991.
[8J A. R. Barron. Universal approximation bounds for superpostion for a
sigmoidal function. IEEE Trans. on Information Theory, 39:930-945,
1993.
[9J A. R. Barron. When do neural nets avoid the curse of dimensionality?
NATO advanced study institute - Prom statistics to neural networks, June
21-July 2,1993.
[10] S. Behtash. Robust output tracking for nonlinear system. Int. J. Control,
51:1990,1381-1407.
266 References
[l1J L. Blum, F. Cucker, M. Shub, and S. Smale. Complexity and Real Com-
putation. Springer-Verlag, New York, 1998.
[12J J. D. Boskovic. Stable adaptive control of a class of fed-batch fermenta-

tion processes. In Proc. of the 8th Yale Workshop on Adaptive and Learn-
ing Systems, pages 263-268, Yale University, New Haven, CT, March
1994.
[13J J. D. Boskovic. Observer-based adaptive control of a class of bioreac-

tor processes. In Proc. American Control Conference, pages 1171-1176,
Washington, June 1995.
[14] J. D. Boskovic. Stable adaptive control of a class of first-order nonlinearly

parameterized plants. IEEE Trans. Automatic Control, 40:347-350, 1995.
[15] J. D. Boskovic. Stable adaptive control of a class of nonlinearly-

parametrized bioreactor processes. In Proc. American Control Confer-
ence, pages 263-268, Washington, June 1995.
[16] J. D. Boskovic. Adaptive control of a class of nonlinearly parameterized

plants. IEEE Trans. Automatic Control, 43:930-933, 1998.
[17J M. Brown and C. Harris. Neurofuzzy Adaptive Modelling and Control.

Prentice-Hall International, London, 1994.
[18J G. Campion and G. Bastin. Indirect adaptive state feedback control of

linearly parametrized nonlinear systems. Int. J. Adaptive Control Signal
Process, 4:345-358, 1990.
[19J R. H. Cannon. Dynamics of Physical Systems. McGraw-Hill, New York,

1967.
[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.
[21J F. C. Chen. Back-propagation neural networks for nonlinear self-tuning

adaptive control. IEEE Control Systems Magazine, 8:44-48, 1988.
[22] F. C. Chen and H. K. Khalil. Adaptive control of nonlinear systems using

neural networks. Int. J. Control, 55:1299-1317, 1992.
[23J F. C. Chen and H. K. Khalil. Adaptive control of a class of nonlinear

discrete-time systems using neural networks. IEEE Trans. Automatic
Control, 40:791-801, 1995.
References 267
[24] F. C. Chen and C. C. Liu. Adaptively controlling nonlinear continuous-

time systems using multilayer neural networks. IEEE Trans. Automatic
Control, 39:1306-1310, 1994.
[25] T. P. Chen and H. Chen. Approximation capability to functions of several
variables, nonlinear functionals, and operators by radial basis function
neural networks. IEEE Trans. Neural Networks, 6:904-910, 1995.
[26] V. Cherkassky, D. Ghering, and F. Mulier. Comparison of adaptive meth-
ods for function estimation from samples. IEEE Trans. Neural Networks,
7:969-984, 1996.
[27] L. O. Chua, M. Komuro, and T. Matsumoto. The double scroll family
parts i and ii. IEEE Trans. Circuits Syst. Part I, 33:1073-1118, 1986.
[28] M. A. Cohen and S. Grossberg. Absolute stability of global pattern

formation and parallel memory storage by competitive neural networks.
IEEE Trans. Systems, Man and Cybernetics, 13:815-826, 1983.
[29] S. Commuri and F. L. Lewis. Cmac neural networks for control of non-
linear dynamical systems: Structure, stability and passivity. Automatica,
33:635-641, 1997.
[30] G. Cybenko. Approximation by superpositions of a sigmoidal function.

Mathematics of Control, Signals and Systems, 2:303-314, 1989.
[31] A. Datta and P. A. Ioannou. Performance analysis and improvement
in model reference adaptive control. IEEE Trans. Automatic Control,
39:2370-2387, 1994.
[32] J. Dieudonne. Foundations of Modern Analysis. Academic Press, New

York, 1969.
[33] X. Dong, G. Chen, and L. Chen. Adaptive control of the uncertain duffing
oscillator. Int. J. Bifurcation and Chaos, 7:1651-1658,1997.
[34] D. Driankov and R. Palm. Advances in Fuzzy Control. Physica- Verlag,

New York, 1998.
[35] B. Egardt. Stability of Adaptive Control. Springer-Verlag, New York,
1979.
[36] B. Egardt. Stability analysis of continuous-time adaptive control systems.
SIAM J. Control Optimization, 18:540-558, 1980.
[37] F. Esfandiari and H. K. Khalil. Output feedback stabilization of fully
linearizable system. Int. J. Control, 56:1007-1037,1992.
268 References
[38] J. Farrell. Stability and approximator convergence in nonparametric

nonlinear adaptive control. IEEE Trans. Neural Networks, 9:1008-1020,
1998.
[39] J. Farrell. Persistence of excitation conditions in passive learning control.

Automatica, 33:699-703, 1999.
[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.
[41] A. Ferrara and L. Giacomini. Control of a class of mechanical systems

with uncertainties via a constructive adaptive/second order vsc approach.
Trans. ASME J. Dynamic Systems, Measurement and Control, 122:33-
39,2000.
[42] R. Fierro and F. L. Lewis. Control of a nonholonomic mobile robot using

neural 'networks. IEEE Trans. Neural Networks, 9:589-600, 1998.
[43] A. L. Fradkov. Speed-gradient scheme and its application in adaptive

control. Automation and Remote Control, 40:1333-1342, 1979.
[44] K. 1. Funahashi. On the approximate realization of continuous mappings
by neural networks. Neural Networks, 2:183-192, 1989.
[45] S. S. Ge. Robust adaptive nn feedback linearization control of nonlinear
systems. Int. J. Systems Science, 27:1327-1338, 1996.
[46] S. S. Ge. Lyapunov design. Submitted to Encyclopedia of Life Support
Systems (EOLSS), Article No. 6.43.21.13, EOLSS Publishers, UK, 2001.
[47] S. S. Ge, C. C. Hang, and T. Zhang. Adaptive neural network control
of nonlinear systems by state and output feedback. IEEE Trans. on
Systems, Man, and Cybernetics, Part B, 29:818-828, 1999.
[48] S. S. Ge, C. C. Hang, and T. Zhang. A direct adaptive controller for
dynamic systems with a class of nonlinear parameterizations. A utomatica,
35:741-747, 1999.
[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.
[50] S. S. Ge, C. C. Hang, and T. Zhang. Nonlinear adaptive control using

neural networks and its application to cstr systems. J. Process Control,
9:313-323, 1999.
References 269
[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.
[52] S. S. Ge, T. H. Lee, and C. J. Harris. Adaptive Neural Network Control

of Robotic Manipulators. World Scientific, London, Dec. 1998.
[53] S. S. Ge, G. Y. Li, and T. H. Lee. Adaptive nn control for a class of

discrete-time nonlinear systems. Submitted to Int. J. Control, 2001.
[54] S. S. Ge, G. Y. Li, and T. H. Lee. Adaptive nn control for a class of strict-
feedback discrete-time nonlinear systems. Submitted to Automatica, 2001.
[55] S. S. Ge and C. Wang. Adaptive neural control of uncertain mimo non-
linear systems. Submitted to IEEE Trans. Neural Networks, 2001.
[56] S. S. Ge and C. Wang. Adaptive nn control of uncertain nonlinear pure-

feedback system. To appear in Automatica, 2001.
[57] S. S. Ge and C. Wang. Direct adaptive nn control of a class of nonlinear
systems. Submitted to IEEE Trans. Neural Networks, 2001.
[58] F. Girosi and T. Poggio. Networks and the best approximation property.
Artificial Intelligence Lab. Memo, Oct., 1989.
[59] C. J. Goh. Model reference control of non-linear systems via implicit

function emulation. Int. J. Control, 60:91-1i5, 1994.
[60] C. J. Goh and T. H. Lee. Direct adaptive control of nonlinear systems

via implicit function emulation. Control-Theory and Advance Technology,
10:539-552, 1994.
[61] G. C. Goodwin and D. Q. Mayne. A parameter estimation perspective of

continuous time model reference adaptive control. Automatica, 23:57-70,
1987.
[62] G. C. Goodwin, P. J. Ramadge, and P. E. Caines. Discrete-time multi-

variable adaptive control. IEEE Trans. Automatic Control, 25:449-456,
1980.
[63] G. C. Goodwin and K. S. Sin. Adaptive Filtering Prediction and control.

Prentice-Hall, Englewood Cliffs NJ, 1984.
[64] D. Gorinevsky. On the persistency of excitation in radial basis function

network identification of nonlinear systems. IEEE Trans. Neural Net-
works, 6:1237-1244, 1995.
270 References
[65] M. M. Gupta and D. H. Rao. Nettro-Control Systems: Theory and Ap-

plications. IEEE Press, Piscataway, NJ, 1994.
[66] S. Haykin. Neural Networks - A Comprehensive Foundation. Macmillan

College, USA, 1994.
[67] M. W. Hirsch and S. Smale. Differential Equations,Dynamical Systems,

and Linear Algebra. Acadamic Press, San Diego, CA, 1974.
[68] J. J. Hopfield. Neurons with graded response have collective computa-

tional properties like those of two-state neurons. Proc. Nat. Acad. Sci,
81:3088-3092,1984.
[69] K. Hornik, M. Stinchcombe, and H. White. Multilayer feedforward net-

works are universal approximator. Neural Networks, 2:359-366, 1989.
[70] K. J. Hunt, D. Sbarbaro, R. Zbikowski, and P. J. Gawthrop. Neural
networks for control system-a survey. Attiomatica, 28:1083-1112, 1992.
[71J L. R. Hunt and C. Meyer. Stable inversion for nonlinear systems. Attto-
matica, 33:1549-1554, 1997.
[72] F. Ikhouane and M.Krstic. Adaptive backstepping with parameter pro-
jection: robustness and asymptotic performance. Automat'ica, 34:429-
435, 1998.
[73J P. A. Ioannou and P. V. Kokotovic. Adaptive Systems with Redttced
Models. Springer-Verlag, New York, 1983.
[74] P. A. Ioannou and J. Sun. Robust Adaptive Control. Englewood Cliffs,

NJ: Prentice-Hall, 1995.
[75] P. A. Ioannou and K. Tsakalis. A robust direct adaptive controller. IEEE

Trans. Automatic Control, 31:1033-1034,1986.
[76] B. Irie and S. Miyake. Capabilities of three-layered peceptrons. In Prof

IEEE Int. Conf Neural Networks, pages 636-641, Atlanta, GA, 1988.
[77] A. Isidori. Nonlinear Control System. Springer-Verlag, Berlin, 2nd edition

1989, 3rd edition, 1995.
[78] S. Jagannathan, S. Commuri, and F. L, Lewis. Feedback linearization

using cmac neural networks. Automatica, 34:547-557, 1998.
[79] S. Jagannathan and F. L. Lewis. Discrete-time neural net controller for

a class of nonlinear dynamical systems. IEEE Trans. Automatic Control,
41: 1693-1699, 1996.
References 271
[80] S. Jagannathan and F. L. Lewis. Identification of nonlinear dynamical

systems using multilayered neural networks. Automatica, 32:1707-1712,
1996.
[81] S. Jain and F. Khorrami. Decentralized adaptive control of a class of

large-scale interconnected nonlinear systems. IEEE Trans. Automatic
Control, 42:136-154, 1997.
[821 M. Jankovic. Adaptive output feedback control of non-linear feedback

line ariz able system. Int. J. Control, 10:1-18, 1996.
[83] M. Jankovic. Adaptive nonlinear output feedback tracking with a partial

high-gain observer and backstepping. IEEE Trans. Automatic Control,
42:106-113,1997.
[84] Z. P. Jiang and L. Praly. Design of robust adaptive controllers for nonlin-
ear systems with dynamic uncertainties. Automatica, 34:825-840, 1998.
[85] L. Jin, P. N. Nikiforuk, and M. M. Gupta. Direct adaptive output tracking
control using multilayered neural networks. lEE Proc. Control Theory
Applications, 140:393-398, 1993.
[86] L. Jin, P. N. Nikiforuk, and M. M. Gupta. Fast neural learning and
control of discrete-time nonlinear systems. IEEE Trans. Syst, Man, and
Cybernetic, 25:478-488, 1995.
[87] T. A. Johansen and P. A. Ioannou. Robust adaptive control of minimum
phase non-linear systems. Int. J. Adaptive Control Signal Process, 10:61-
78,1996.
[88] A. Judisky, Q. Zhang, B. Delyon, P. Y. Golorennec, and A. Benveniste.

Wavelets in identification: Wavelets, splines, neurones, fuzzies, how good
for identification? IRIBA, Intern Publication no. 849, 1994.
[891 J. Kallkuhl, K. J. Hunt, R. Zbikowski, and Eds A. J. Dzielinshi. Applica-
tions of Neural Adaptive Control Technology. Would Scientific, Singapore,
1997.
[90J r. Kanellakopoulos, P. V. Kokotovic, and R. Marino. Robus,tness of adap-

tive nonlinear control under an extended matching condition. In Proc.
IFAC Symp. Nonlinear Contr. Syst. Design, pages 192-197, Capri, Italy,
March 1989.
[91J r. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse. Systematic design

of adaptive controller for feedback linearizable systems. IEEE Trans.
Automatic Control, 36:1241-1253,1991.
272 References
[92J I. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse. Systematic design

of adaptive controller for feedback line ariz able systems. IEEE Trans.
Automatic Control, pages 1241-1253, 1991.
[93J I. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse. Adaptive output
feedback control of systems with output nonlinearities. IEEE Trans.
Automatic Control, 37:1166-1182, 1992.
[94J H. K. Khalil. Nonlinear Systems. Mecmillan, New York, 1992.
[95J H. K. Khalil. Adaptive output feedback control of nonlinear system repre-
sented by input-output models. IEEE Trans. Automatic Control, 41:177-
188, 1996.
[96] H. K. Khalil. Nonlinear systems. Prentice Hall, New Jersey, USA, 1996.
[97] T. Khanna. Foundations of Neural Networks. Addison Wesley, Reading,
MA, 1~90.
[98] Y. H. Kim, F. L. Lewis, and C. T. Abdallah. A dynamic recurrent
neural-network-based adaptive observer for a class of nonlinear systems.
Automatica, 33:1539-1543, 1997.
[99J E. B. Kosmatopoulos. Universal stabilization using control lyapunov
functions, adaptive derivative feedback and neural network approxima-
tors. In Proc. IEEE Conf. Decision and Contro~ pages 2444-2449, Kobe,
Japan, December 1996.
[100] E. B. Kosmatopoulos, M. M. Polycarpou, M. A. Christodoulou, and P. A.
Ioannou. High-order neural network structures for identification of dy-
namical systems. IEEE Trans. Neural Networks, 6:422-431, 1995.
[101J G. Kreisselmeier and B. D. O. Anderson. Robust model reference adap-
tive control. IEEE Trans. Automatic Control, 31:127-133, 1986.
[102J M. Krstic. Invariant manifolds and asymptotic properties of adaptive
nonlinear stabilizers. IEEE Trans. Automatic Control, 41:817-829, 1996.
[103] M. Krstic, I. Kanellakopoulos, and P. Kokotovic. Adaptive nonlinear
control without overparameterization. Systems Control Letters, 19:177-
185,1992.
[104J M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic. Nonlinear and Adap-

tive Control Design. Wiley and Sons, NY, 1995.
[105J M. Krstic, P. V. Kokotovic, and I. Kanellakopoulos. Transient perfor-
mance improvement with a new class of adaptive controllers. Systems
Control Letters, 21:451-461, 1993.
References 273
[106] C. Kwan, F. L. Lewis, and D. M. Dawson. Robust neural-network control

of rigid-link electrically driven robots. IEEE Trans. Neural Networks,
9:581-588, 1998.
[107] V. Lakshmikantham and S. Leela. Differential and Integral Inequalities.
Academic Press, New York, 1969.
[1081 Y. D. Landau. Adaptive Control. Marcel Dekker, New York, 1979.
[109] A. U. Levin and K. S. Narendra. Control of nonlinear dynamical systems
using neural networks-part ii: observability, identification, and control.
IEEE Trans. Neural Networks, 7:30-42, 1996.
[110] F. L. Lewis, K. Liu, and A. Yesildirek. Neural net robot controller with
guaranteed tracking performance. IEEE Trans. Neural Networks, 3:703-
715, 1995.
[1111 F. L. ~ewis and T. Parisini. Guest editorial: Neural network feedback
with guaranteed stability. Int. J. Control, 70:337-339, 1998.
[112] F. L. Lewis, A. Yesildirek, and K. Liu. Multilayer neural-net robot con-
troller with guaranteed tracking performance. IEEE Trans. Neural Net-
works, 7:388-398, 1996.
[113] F. L. Lewis, A. Yesildirek, and K. Liu. Robust backstepping control of
induction motors using neural networks. IEEE Trans. Neural Networks,
11:1178-1187,2000.
[114] Z. H. Li and M. Krstic. Optimal design of adaptive tracking controllers
for non-linear systems. Automatica, 33:1459-1473, 1997.
[115] Z. H. Li and M. Krstic. Geometric/asymptotic properties of adaptive
nonlinear systems with partial excitation. IEEE Trans. Automatic Con-
trol, 39:419-425, 1998.
[116] G. Lightbody and G. W. Irwin. Direct neural model reference adaptive
control. lEE Proc. Control Theory Applications, 142:31-43, 1995.
[117] Y. Liguni, H. Sakai, and H. Tokumaru. A nonlinear regulater design in
the presence of system uncertainties using multilayer neural networks.
Neural networks, 2:410-417, 1991.
[118] C. T. Lin. Neural Fuzzy Control Systems with Structure and Parameter
Learning. Springer-Verlag, London Limited, 1994.
[119] J. S. Lin and 1. Kanellakopoulos. Nonlinearities enhance parameter con-
vergence in output-feedback systems. IEEE Trans. Automatic Control,
43:204-222, 1998.
274 References
[120J J. S. Lin and 1. Kanellakopoulos. Nonlinearities enhance parameter con-

vergence in strict feedback systems. IEEE Trans. Automatic Control,
44:89-94, 1999.
[121J C. C. Liu and F. C. Chen. Adaptive control of non-linear continuous-
time systems using neural networks-general relative degree and mimo
cases. Int. J. Control, 58:317-335, 1993.
[122J G. P. Liu, V. Kadirkamanathan, and S. A. Billings. Variable neural net-
works for nonlinear adaptive control. In Proc. 13th IFA C World Congress,
pages 1407-1412, San Francisco, CA, July 1996.
[123J G. P. Liu, V. Kadirkamanathan, and S. A. Billings. Identification of
nonlinear dynamical systems by growing rbf networks. Proceeding: 3rd
ECC, Sep. 1995.
[124] A. P. Loh, A. M. Annaswamy, and F. P. Skanze. Adaptive control of

dynamic systems with nonlinear parametrization. In Proc. 4th Europ.
Contr. Conf., pages 407-412, Brussels, August 1997.
[125J Y. S. Lu and J. S. Chen. Design of a perturbation estimator using the
theory of variable-structure systems and its application to magnetic lev-
itation systems. IEEE Trans. Ind. Electron, 42:281-289, 1995.
[126J R. Marino and P. Tomei. Global adaptive output-feedback control of non-

linear systems, part i: Linear parameterization. IEEE Trans. Automatic
Control, 38:17-32, 1993.
[127J R. Marino and P. Tomei. Global adaptive output-feedback control of
nonlinear systems, part ii: Nonlinear parameterization. IEEE Trans.
Atttomatic Control, 38:17-48, 1993.
[128J R. Marino and P. Tomei. Nonlinear Adaptive Design: Geometric, Adap-
tive, and Robust. Printice Hall, London, 1995.
[129J R. Marino and P. Tomei. Robust adaptive state-feedback tracking for
nonlinear systems. IEEE Trans. Automatic Control, 43:84-89, 1998.
[130J W. S. McCulloch and W. Pitts. A logical calculus of the ideas immanent
in nervous activity. Bull. Math. biophys, 5: 115-133, 1943.
[131J C. A. Micchelli. Interpolation of scattered data: Distancematrices and
conditionally positive definite functions. Constructive Approx'imation,
2:11-22, 1986.
[132J A. S. Morse. Global stability of parameter-adaptive control systems.

IEEE Trans. Automatic Control, 25:433-439, 1980,
References 275
[133J K. Nam and A. Arapostations. A model-reference adaptive control

scheme for pure-feedback nonlinear systems. IEEE Trans. Automatic
Control, 33:803-811, 1988.
[134J K. S. Narendra and A. M. Annaswamy. A new adaptive law for robust
adaptation without persistent excitation. IEEE Trans. Automatic Con-
trol, 32:134-145,1987.
[135J K. S. Narendra and A. M. Annaswamy. Stable Adaptive System. Prentice-
Hall, Englewood Cliffs, NJ, 1989.
[136J K. S. Narendra and J. Balakrishnan. Improving transient response of
adaptive control systems using multiple models and switching. In Proc.
32nd IEEE Can! Decision and Control, pages 1067-1072, San Antonio,
Texas, December 1993.
[137J K. S. Narendra, Y. H. Lin, and L. S. Valavani. Stable adaptive con-
troller design, part ii: Proof of stability. IEEE Trans. Automatic Control,
25:440-448, 1980.
[138J K. S. Narendra and S. Mukhopadhyay. Adaptive control of nonlinear

multivariable system using neural networks. Neural networks, 7:737-752,
1994.
[139J K. S. Narendra and S. Mukhopadhyay. Adaptive control using neural

networks and approximate models. IEEE Trans. Neural Networks, 8:475-
485, 1997.
[140J K. S. Narendra and K. Parthasarathy. Identification and control of dy-

namic systems using neural networks. IEEE Trans. Neural Networks,
1:4-27, 1990.
[141] G. Nurnberger. Approximation by Spline Functions. Springer-Verlag,

New York, 1989.
[142] R. Ortega. Some remarks on adaptive neuro-fuzzy systems. Int. .1. of
Adaptive contr. and Signal Processing, 10:79-83, 1996.
[143J Z. Pan and T. Basar. Adaptive controller design for tracking and distur-
bance attenuation in parametric strict-feedback nonlinear systems. IEEE
Trans. Automatic Control, 43:1066-1083, 1998.
[144] P. Paretto and J. J. Niez. Long term memory storage capacity of multi-
connected neural networks. Bioi. Cybern, 54:53-63, 1986.
[145J B. B. Peterson and K. S. Narendra. Bounded error adaptive control.
IEEE Trans. A1ttomatic Control, 27:1161-1168,1982.
276 References
[146] T. Poggio and F. Girosi. Networks for approximation and learning. Proc.
of IEEE, 78:1481-1497, 1990.
[147] M. M. Polycarpou. Stable adaptive neural control scheme for nonlinear

systems. IEEE Trans. Automatic Control, 41:447-451, 1996.
[148] M. M. Polycarpou and P. A. Ioannou. Identification and control using

neural network models: design and stability analysis. Technical Report
91-09-01, Department of Electrical Engineering Systems, University of
Southern California, 1991.
[149] M. M. Polycarpou and P. A. Ioannou. On the existence and uniqueness of

solutions in adaptive control systems. IEEE Trans. Automatic Control,
38:474-479, 1993.
[150] M. M. Polycarpou and P. A. Ioannou. A robust adaptive nonlinear control

design. Automatica, 32:423-427, 1996.
[151J M. M. Polycarpou and M. J. Mears. Stable adaptive tracking of uncertain

systems using nonlinearly parametrized on-line approximators. Int. J.
Control, 70:363-384, 1998.
[152] M. M. Polycarpou, M. J. Mears, and S. E. Weaver. Adaptive wavelet

control of nonlinear systems. In Proc. 36th IEEE Conf. Decision and
Control, pages 3890-3895, San Diego, CA, December 1997.
[153] J. B. Pomet and L. Praly. Adaptive nonlinear regulation: estimation

from the lyapunov equation. IEEE Trans. Automatic Control, 37:729-
740, 1992.
[154] L. Praly. Robust model reference adaptive controllers - part i: stability

analysis. In Proc. 23th IEEE Can! Decision and Control, pages 1009-
1014, Las Vegas, December 1984.
[155] D. Psaltis, A. Sideris, and A. Yamamura. A multilayered neural network

controller. IEEE Control Sys. Magazine, 8:17-21, 1988.
[156J Z. Qu. Robust control of nonlinear uncertain systems. John Wiley and
Sons, New York, 1998.
[157] J. R. Rice. The Approximation of Functions. Addison-Wesley, Reading,

MA,1964.
[158] O. E. Rossler. An equation for continuous chaos. Physica Letts., 57 A:397-

398,1976.
References 277
[159J D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal

representations by error propagation. Parallel Distributed Processing,
1:318-362, 1986.
[160] N. Sadegh. A perception network for functional identification and control
of nonlinear systems. IEEE Trans. Neural Networks, 4:982-988, 1993.
[161] R. M. Sanner and J. E. Slotine. Gaussian networks for direct adaptive
control. IEEE Trans. Neural Networks, 3:837-863, 1992.
[162] R. M. Sanner and J. E. Slotine. Structurally dynamic wavelet networks
for adaptive control of robotic systems. Int. J. Control, 70:405-421,1998.
[163] R. M. Sanner and J. J. E. Slotine. Stable adaptive control and recursive
identification using radial gaussian networks. In Proc. 30th IEEE Con!
Decision and Control, Brighton, England, December 1991.
[164] S. S. S.astry and M. Bodson. Adaptive Control: Stability, Convergence,
and Robustness. Prentice-Hall, Englewood Cliff, NJ, 1989.
[165] S. S. Sastry and A. Isidori. Adaptive control of linearizable systems.
IEEE Trans. Automatic Control, 34:1123-1131,1989.
[166] R. Sepulchre, M. Jankovic, and P. V. Kokotovic. Constructive Nonlinear
Control. Springer-Verlag, London, 1997.
[167] D. Seto, A. M. Annaswamy, and J. Baillieul. Adaptive control of nonlinear
systems with a triangular structure. IEEE Trans. Automatic Control,
39:1411-1428,1994.
[168J L. Shi and S. K. Singh. Decentralized adaptive controller design for large-
scale systems with higher order interconnections. IEEE Trans. Automatic
Control, 37:1106-1118,1992.
[169] J. E. Slotine and W. Li. Applied Nonlinear Control. Prentice-Hall, Inc.,
Englewood Cliff, NJ, 1991.
[170J J. T. Spooner and K. M. Passino. Adaptive control of a class of decen-
tralized nonlinear systems. IEEE Trans. Automatic Control, 41:280-284,
1996.
[171] J. T. Spooner and K. M. Passino. Stable adaptive control using fuzzy
systems and neural networks. IEEE Trans. Fuzzy Systems, 4:339-359,
1996.
[172] J. Sun. A modified model reference adaptive control scheme for im-
proved transient performance. IEEE Trans. Automatic Control, 34:630-
635, 1994.
278 References
[173] J. A. K. Suykens, J. P. L. Vandewalle, and B L.R. De Moor. Artificial

neural networks for modelling and control of non-linear systems. Kluwer
Academic, USA, 1996.
[174] D. G. Taylor, P. V. Kokotovic, R. Marino, and 1. Kanellakopoulos. Adap-
tive regulation of nonlinear systems with unmodeled dynamics. IEEE
[175] A. R. Teel, P. V. Kokotovic R. R. Kadiyala, and S. S. Sastry. Indirect
techniques for adaptive input-output linearization of non-linear systems.
Int. J. control, 53:193-222, 1991.
[176] J. Tsinias and N. Kalouptsidis. Invertability of nonlinear analytic single-
input systems. IEEE Trans. Automatic Control, 28:931-933, 1983.
[177] M. Vidyasagar. Nonlinear systems analysis, volume 2nd ed. Prentice-
Hall, Inc., Englewood Cliff, NJ, 1993.
[178] C. Wang and S. S. Ge. Adaptive nn control of partially known nonlinear
strict-feedback systems. Submitted to Asian Journal of Control, 200l.
[179] L. X. Wang. Adaptive Fuzzy Systems and Control: Design and Analysis.
Prentice-Hall, Englewood Cliff, NJ, 1994.
[180] L. X. Wang. Modeling and control of hierarchical systems with fuzzy
systems. Automatica, 33:1041-1053, 1997.
[181] C. Y. Wen. Robust adaptive control of proper systems. IEEE Trans.
Automatic Control, 43:1579-1584, 1998.
[182] C. Y. Wen and Y. C. Soh. Decentralized adaptive control using integrator
backstepping. Automatica, 33:1719-1724, U197.
[183] B. Wid row and F. W. Smith. Pattern-recognizing control systems. In
Proc. of Computer and Information Sciences, 1964.
[184] B. Yao and M. Tomizuka. Adaptive robust control of siso nonlinear
systems in a semi-strict feedback form. Automatica, 33:893-900, 1997.
[185] B. E. Ydstie. Transient performance and robustness of direct adaptive
control. IEEE Trans. Automatic Control, 37:1091-1105,1992.
[186] A. Yesidirek and F. L. Lewis. Feedback linearization using neural net-
works. Automatica, 31:1659-1664, 1995.
[187] D. S. Yoo and M. J. Chung. A variable structure control with simple
adaptation laws for upper bounds on the norm of the uncertainties. IEEE
References 279
[188] S. H. Yu and A. M. Annaswamy. Adaptive control of nonlinear dynamic

systems using O-adaptive neural networks. Automatica, 33:1975-1995,
1997.
[189] S. H. Yu and A. M. Annaswamy. Stable neural controllers for nonlinear

dynamic systems. A utomatica, 34:641-650, 1998.
[190] Q. Zhang and A. Benvensite. Wavelet network. IEEE Trans. Neural

Networks, 3:889-898, 1992.
[191] T. Zhang, S. S. Ge, and C. C. Hang. Design and performance analysis of
a direct adaptive controller for nonlinear systems. Alttomatica, 35:1809-
1817, 1999.
[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.
[193] Y. P. Zhang. Robustness and Performance of Nonlinear Adaptive Control

Systems. PhD thesis, Faculty of Electrical Engineering, University of
Southern California, CA, USA, 1996.
[194] Y. P. Zhang and P. A. Ioannou. Robustness and performance of a mod-
ified adaptive backstepping controller. Int. J. Adapt. Control Signal
Process, 12:247-265, 1998.
[195] Y. P. Zhang and P. A. Ioannou. Robustness of nonlinear control systems
with respect to un modeled dynamics. IEEE Trans. AutomaticControl,
44: 119-124, 1999.
Index
C k ,13 Differentiation of integrals, 22

L 1 ,14
L 2 ,14 Equilibrium point, 16
L2 transient bound, 103 Extended matching condition, 107
Lee!) 14
Loo tracking error, 56, 65, 103, 134 Feedback linearization, 47
a-modificatic;m, 41, .52 Filtered error, 49, 82
el-modification,41 First-order system, 108
Function
Adaptive backstepping design, 105, continuous, 13
183 differentiable, 14
positive, 15
Backstepping, 107, 183 strictly positive, 15
Barbalat's lemma, 15 Function approximation, 27
Block-triangular forms, 243 Fuzzy systems, 2, 29
Brunovskyfornl, 47, 48, 82
Gaussian function, 29
Cauchy-Schwartz inequality, 21 General nonlinear system, 139
Chattering, 77 Growth condition, 107
Class IC function, 17
Common tuning algorithm, 31, 41 Holder's inequality, 22
Comparison Theorem, 22 High-gain control, 137
Continuous function, 13 High-gain observer, 146
Controllability, 6 HONN,30
Controller singularity problem, 185 Hurwitz, 19, 142
Couplings, 242
CSTR,176 Implicit desired feedback control, 142
Implicit Function Theorem, 23,143,
Desired feedback control, 49 219
DFCl,85 Input-output stability, 25
DFC2,86 Integral formula, 20
DFC3,88 Integral Lyapunov Function, 81, 83
Diffeomorphism, 141 Integration by Parts, 21
Differential inequality, 25 Inverted pendulum, 68
Index 282
Jacobian matrL'C, 14 Polynomial, 2, 29

Projection algorithm, 34
Lie derivate, 140 Proper transfer function, 26
Higher-order, 140 Pure-feedback nonlinear system :E 1,
LIP parametrization, 88 220
Lipschitz, 13 Pure-feedback nonlinear system :E 2 ,
LPNN, 29, 147 235
Lyapunov function, 18
Lyapunov theorem, 18 RBF,29
Relative degree, 141
Matching condition, 107 Residual term, 40
MatrL'C
indefinite, 15 Schwartz inequality, 103
negative semi-definite, 15 Sets, 13
negative definite, 15 bounded, 13
positive semi-definite, 15 closed, 13
positive definite, 15 compact, 13
properties, 15 convex, 13
skew-symmetric, 15 open, 13
symmetric, 15 SGUUB, 17, 184
Mean square tracking error, 134 Singularity, 47, 81
Mean Value Theorem, 20, 170 Sliding surface, 19
MIMO nonlinear systems, 127, 242 Spline, 2, 29
MNN, 35, 160 Stability, 16
Strict-feedback nonlinear systems,
Non-affine nonlinear systems, 140, 107
219
Non-linearly parameterized network, Taylor series expansion, 148, 154
35 Transient performance, 56, 65
Nonlinear parameterization, 88 Triangular form, 81
Norm Triangular nonlinear systems, 183
Frobenius, 12
function, 14 Uncertain strict-feedback system, 186
matrix, 12
Variable structure control, 73
vector, 12
Wavelet networks, 2, 29
Order difference, 244, 260
Weighted control Lyapunov function,
Overparameterization problem, 201
83
Parameter drift, 5 Weighting function, 83
Partially known nonlinear systems,
201
Pendulum of variable length, 57

(The Springer International Series on Asian Studies in Computer and Information Science 13) Shuzhi S. Ge, Chang C. Hang, Tong H. Lee, Tao Zhang (auth.) - Stable Adaptive Neural Network Control-Springe.pdf

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(The Springer International Series on Asian Studies in Computer and Information Science 13) Shuzhi S. Ge, Chang C. Hang, Tong H. Lee, Tao Zhang (auth.) - Stable Adaptive Neural Network Control-Springe.pdf

Uploaded by

Copyright:

Available Formats

STABLE ADAPTIVE NEURAL

FULLY TUNED RADIAL BASIS FUNCTION NEURAL NETWORKS FOR FLIGHT

NONLINEAR CONTROL SYSTEMS AND POWER SYSTEM DYNAMICS by

DATA MANAGEMENT AND INTERNET COMPUTING FOR IMAGE/PATTERN

COMMON WAVEFORM ANALYSIS: A New and Practical Generalization of Fourier

DOMAIN MODELING-BASED SOFTWARE ENGINEERING: A Formal Approach, by

AUTOMATED BIOMETRICS: Technologies and Systems, by David D. Zhang;

FUZZY LOGIC AND SOFT COMPUTING, by Guoqing Chen, Mingsheng Ying

INTELLIGENT BUILDING SYSTEMS, by Albert Ting-pat So, Wai Lok Chan;

PERFORMANCE EV ALUA TION, PREDICTION AND VISUALIZATION OF

ROBUST MODEL-BASED FAULT DIAGNOSIS FOR DYNAMIC SYSTEMS by Jie

Department 0/Electrical Engineering

SPRINGER SCIENCE+BUSINESS MEDIA, LLC

Stable adaptive neural network control/by S. S. Ge ... ret al.J.

Copyright © 2002 by Springer Science+Business· Media New York

Printed on acidjree paper.

Jinlan, Sao Chin, Ruth, Yuan

our loved ones

I am pleased to acknowledge the assistance to the editorial work by

3 Neural Networks and Function Approximation 27

4 SISO Nonlinear Systems 47

5 ILF for Adaptive Control 81

6 Non-affine Nonlinear Systems 139

6.5 Conclusion ......... . 182

7 Triangular Nonlinear Systems 183

proven through Lyapunov stability analysis. Transient performance of adaptive

Shuzhi S. Ge, Chang C. Hang, Tong H. Lee and Tao Zhang

S. S. Ge et al., Stable Adaptive Neural Network Control

work control only. For other linear-in-the-parameters function approximators

1.2 Adaptive Control

to early 1990s. Several important results were presented based on feedback

1.3 Neural Network Control

processing, modelling and system control. The approximation abilities of neural

investigated in the literature. In an effort to solve the problem, several re-

1.4 Instability Mechanisms in Adaptive Neural

designs by introducing projection algorithm [148,171,179] and modified

i; = f(x) + g(x)u (1.1)

3. Transient Behaviour Problem

4. Initial conditions and reference signals

5. Choice of design parameters

The third objective of this book is to solve the above-mentioned control

The fourth objective is to enlarge the class of nonlinear systems handled by

The final objective is to develop neural adaptive approaches for non-affine

1.5 Outline of the Book

2.2 Mathematical Preliminaries

S. S. Ge et al., Stable Adaptive Neural Network Control

2.2.1 Norms for Vectors and Signals

for 1 :S p < 00 (2.1)

Induced Norm of Matrices

For p = 1,2,00, we have the corresponding induced norms as follows:

IIAIII = max L laijl (column sum)

IIAII2 = m~x JAi(AT A)

IIAIIe<> = m~x L laijl (row sum)

II A IIJ,.= L a~ = tr(A T A) (2,6)

N(X,E) = {z E R n Illz -xii < E} (2.7)

such that N(x, E) C S.

IIx - yll < 8 ===> IIf(x) - f(y)11 < E (2.8)

A function f is continuous on a set of S if it is continuous at every point

j(x) = lim l(x + t5x) - l(x) (2.9)

II flip = (10 00 If(t)IP dt) lip, for p E [1, (0) (2.11)

1111100 = sup If(t)l, for p = 00 (2.12)

By letting p = 1,2,00, the corresponding normed spaces are called L 1 , L 2 ,

Loo ~ {I: R+ --4 R 11111100 = sup

Lemma 2.2 (Barbalat's Lemma). Let f(t) be a differentiable function, if