Advanced Textbooks in Control and Signal Processing

Nhan T. Nguyen

Model-
Reference
Adaptive
Control
A Primer
Advanced Textbooks in Control and Signal
Processing

Series editors
Michael J. Grimble, Glasgow, UK
Michael A. Johnson, Oxford, UK
Linda Bushnell, Seattle, WA, USA
More information about this series at http://www.springer.com/series/4045
Nhan T. Nguyen

Model-Reference Adaptive
Control
A Primer

123
Nhan T. Nguyen
Intelligent Systems
NASA Ames Research Center
Moffett Field, CA
USA

ISSN 1439-2232 ISSN 2510-3814 (electronic)

Advanced Textbooks in Control and Signal Processing
ISBN 978-3-319-56392-3 ISBN 978-3-319-56393-0 (eBook)
https://doi.org/10.1007/978-3-319-56393-0

Library of Congress Control Number: 2017953807

MATLAB® and Simulink® are registered trademarks of The MathWorks, Inc., 1 Apple Hill Drive,
Natick, MA 01760-2098, USA, http://www.mathworks.com.

© Springer International Publishing AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To my wife Rosa and my children Daniel and
Natalia
Series Editors’ Foreword

The Advanced Textbooks in Control and Signal Processing series is designed as a

vehicle for the systematic textbook presentation of both fundamental and innovative
topics in the control and signal processing disciplines. It is hoped that prospective
authors will welcome the opportunity to publish a more rounded and structured
presentation of some of the newer emerging control and signal processing tech-
nologies in this textbook series. However, it is useful to note that there will always
be a place in the series for contemporary presentations of foundational material in
these important engineering areas.
The science of adaptive control design and implementation, as presented in this
textbook, is motivated by some very practical problems found in real-world control
applications. Probably the most important issue is maintaining the desired perfor-
mance specifications and closed-loop system stability in the presence of changing
process dynamics. The way a real system can change is an interesting topic in itself.
A system may operate for long periods with unchanging parameters or character-
istics and then change both significantly and rapidly when, for example, a new
operating regime begins. Alternatively, a system may be changing all the time,
sometimes slowly, as in the case of long-term component wear and degradation,
and sometimes more rapidly as a product evolves or new operating regimes are
entered.
When system change is predictable, the most widely applied forms of adaptive
control are probably the controller gain schedule and the controller schedule per se.
These technologies have the advantages of application transparency and, in
safety-critical situations, verifiability. This approach has links to the methods of
multimodel control design.
Another approach to adaptive control is that of online updates and improvements
to linear parameter varying control designs when the underlying system is non-
linear, partially uncertain, or slowly changing. A technique that found much favor
was that of self-tuning control. Although this technique has a very transparent
architecture, it was soon realized that, even when plants were linear, these types of
adaptive control problems are nonlinear systems that need nonlinear systems
analysis for their study.

vii
viii Series Editors’ Foreword

The other theme in this textbook is the use of a model-reference control

framework. In the literature of deterministic control designs, such schemes are more
likely to be termed “model-following” control systems. This class of techniques is
an interesting field in its own right. Quite a few contributions focus on the use of
classical PID controllers in the model-following framework to achieve additional
robustness and performance gains in the presence of process uncertainty.
A fundamental purpose of the model in these schemes is to instil the desired control
system performance specifications such as overshoot, rise time, zero steady-state
error, and prescribed stability margins. An important art in this method is to specify
the model so that the closed-loop process can actually achieve the desired perfor-
mance specifications.
Combining adaptive control with model-following design yields the class of
methods known as model-reference adaptive control (MRAC). To understand such
methods requires a good grounding in nonlinear systems methods and Lyapunov
stability theory. This is where the text Model-Reference Adaptive Control: A Primer
by Nhan T. Nguyen begins. Dr. Nguyen has distilled his many years of experience
into this textbook for which the material has been tested on lecture courses given by
the author. The material is presented in sharply focussed sections with examples,
chapter exercises, and supporting references. Dr. Nguyen is a leading research
scientist at the NASA Ames Research Center at Moffett Field, California, USA, and
his extensive research has included an exploration of the potential of MRAC in
aerospace applications. This experience is reflected in the textbook through the
discussions he presents on the certification and validation of adaptive control
methods in safety-critical systems and high-risk human aerospace applications.
Outcomes from Dr. Nguyen’s experiences are also evident in the final chapter of the
textbook that contains the details of a number of case studies mostly drawn from the
aerospace field. The Series Editors are pleased to welcome this application-
orientated textbook to the Advanced Textbooks in Control and Signal Processing
series and believe it will find a wide readership in the control and aerospace
communities.

January 2017 M. J. Grimble

M. A. Johnson
Industrial Control Centre
Glasgow, Scotland, UK
Preface

This textbook is developed from a series of lecture notes for graduate courses that the
author has taught. The textbook includes new research materials that the author
developed during his research at NASA Ames Research Center. The intended readers
of this book are masters-level graduate students and beginning doctoral students. The
author notes that there exist many excellent advanced adaptive control textbooks that
provide in-depth rigorous mathematical control theory aimed at doctoral and post-
doctoral researchers. This textbook is aimed at applied control theory and applica-
tions intended to provide the readers a sufficient working understanding of adaptive
control theory through applications. Examples and problem sets are used to reinforce
the materials and aid the understanding through applications of adaptive control
techniques. A solutions manual for many of the problems, available free of charge to
instructors who adopt this textbook for use in teaching their courses, can be
downloaded from http://www.springer.com/book/9783319563923. During the
course of teaching, the author feels that such a textbook could help beginning
graduate students to better understand and appreciate the richness of adaptive control
without feeling being overwhelmed by the mathematics of real analysis required in
adaptive control theory. It is with this objective in mind that the author decided to
compose this manuscript.
Adaptive control is a well-researched subject, and as such, the subject is enriched
by a voluminous body of research literature. Many new advancements in adaptive
control theory are still being developed in the present time. The textbook does not
attempt to incorporate all of the latest developments in adaptive control theory, for
this would require an extensive endeavor which would be exceedingly beyond the
scope of this textbook. Rather, the book tries to provide a foundation in the
model-reference adaptive control with basic well-accepted adaptive control tech-
niques. Parameter estimation by least-squares techniques and uncertainty approxi-
mation by neural networks are covered. The second half of the book is devoted to
the subject of robust adaptive control. Robustness issues with model-reference
adaptive control are discussed. Standard methods to improve robustness are covered
in the remainder of the second half of the textbook. These include well-established
methods, such as the dead zone, projection, r modification, and e modification.

ix
x Preface

More recent robust adaptive control methods are also covered. These include the
author’s work in optimal control modification and bi-objective optimal control
modification, as well as two recent well-known adaptive control methods: the
adaptive loop recovery and L1 adaptive control. The author well recognizes that this
is by no means a complete exposition of all the latest advancements in robust
adaptive control theory. The readers will be referred to a list of references provided
in the bibliography for other adaptive control methods. Finally, this book includes a
chapter on applications. In particular, the applications on adaptive flight control are
exhibited from the author’s body of research in this field.

Acknowledgements

The author wishes to acknowledge NASA Aeronautics Research Mission

Directorate for providing the author with the opportunity to conduct research in
adaptive control theory in the course of conducting aviation safety research. The
author is appreciative of all the support from his colleagues at NASA Ames
Research Center and NASA Armstrong (formerly Dryden) Flight Research Center
during the course of research and flight testing of adaptive flight controllers on
NASA research aircraft. In particular, the author wants to thank Joseph Totah and
Kalmanje Krishnakumar of NASA Ames Research Center for their organizational
support to allow the pursuit of this research. The author is appreciative of the
friendly collaboration with John Bosworth, Curtis Hanson, and in particular John
Burken of NASA Armstrong Flight Research Center on flight test validation of
adaptive control. The author would like to acknowledge Prof. Anthony Calise, Prof.
Naira Hovakimyan, and Prof. Kumpati Narendra for allowing their work to be
included in the manuscript. The author also would like to acknowledge Stephen
Jacklin at NASA Ames Research Center for his thorough and insightful input on the
certification of adaptive control. Finally, the author would be remiss without
acknowledging the support of the author’s family that allows him the personal
freedom to pursue the research that goes into this manuscript.

Santa Clara, CA, USA Nhan T. Nguyen

November 2016
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 2
1.2 Verification and Validation Challenges for Adaptive Flight
Control Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 6
1.2.1 Verification by Simulation of Adaptive Flight
Control Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Adaptive Control Metrics . . . . . . . . . . . . . . . . . . . . . 8
1.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Nonlinear Systems . . . . . . . . . . . . . . . ......... . . . . . . . . . . . . . . 17
2.1 Equilibrium and Linearization . ......... . . . . . . . . . . . . . . 21
2.2 Local Stability and Phase Plane Analysis . . . . . . . . . . . . . . . . 24
2.3 Other Nonlinear Behaviors . . . ......... . . . . . . . . . . . . . . 27
2.4 Summary . . . . . . . . . . . . . . . . ......... . . . . . . . . . . . . . . 28
2.5 Exercises . . . . . . . . . . . . . . . . ......... . . . . . . . . . . . . . . 29
Reference . . . . . . . . . . . . . . . . . . . . . . ......... . . . . . . . . . . . . . . 30
3 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1 Vector and Matrix Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1 Vector Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.2 Matrix Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Compact Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3 Existence and Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.1 Cauchy Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.2 Global Lipschitz Condition . . . . . . . . . . . . . . . . . . . . 39
3.3.3 Local Lipschitz Condition . . . . . . . . . . . . . . . . . . . . . 40

xi
xii Contents

3.4 Positive Definite, Symmetric and Anti-Symmetric Matrices . . . 41

3.4.1 Positive-Definite Matrix and Function . . . . . . . . . . . . 41
3.4.2 Anti-Symmetric Matrix . . . . . . . . . . . . . . . . . . . . . . . 43
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4 Lyapunov Stability Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.1 Stability Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.1.1 Stability Definition . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.2 Asymptotic Stability . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.3 Exponential Stability . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Lyapunov’s Direct Method . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.2 Lyapunov Theorem for Local Stability . . . . . . . . . . . . 57
4.2.3 Lyapunov Theorem for Exponential Stability . . . . . . . 60
4.2.4 Radially Unbounded Functions . . . . . . . . . . . . . . . . . 61
4.2.5 Barbashin–Krasovskii Theorem for Global
Asymptotic Stability . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.6 LaSalle’s Invariant Set Theorem . . . . . . . . . . . . . . . . 63
4.2.7 Differential Lyapunov Equation . . . . . . . . . . . . . . . . . 67
4.3 Stability of Non-Autonomous Systems . . . . . . . . . . . . . . . . . . 69
4.3.1 Uniform Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3.2 Uniform Boundedness . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.3 Barbalat’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5 Model-Reference Adaptive Control . . . . . . . . . . . . . . . . . . . . . .... 83
5.1 Composition of a Model-Reference Adaptive
Control System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.1.1 Uncertain Plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.1.2 Reference Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.1.3 Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1.4 Adaptive Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2 Direct MRAC for First-Order SISO Systems . . . . . . . . . . . . . 90
5.2.1 Case I: a and b Unknown but Sign of b Known . . . . . 90
5.2.2 Case II: a and b Known . . . . . . . . . . . . . . . . . . . . . . 93
5.3 Indirect MRAC for First-Order SISO Systems . . . . . . . . . . . . 95
5.4 Direct MRAC for Second-Order SISO Systems . . . . . . . . . . . 99
5.4.1 Case I: A and B Unknown but Sign of b known . . . . . 99
5.4.2 Case II: A and B Known . . . . . . . . . . . . . . . . . . . . . . 105
5.5 Indirect MRAC for Second-Order SISO Systems . . . . . . . . . . 106
Contents xiii

5.6 Direct MRAC for MIMO Systems . . . . . . . . . . . . . . . . . . . . . 110

5.6.1 Case I: A and K Unknown, but B and Sign
of K Known . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.6.2 Case II: A, B, K ¼ I Known . . . . . . . . . . . . . . . . . . . 117
5.7 Indirect MRAC for MIMO Systems . . . . . . . . . . . . . . . . . . . . 118
5.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6 Least-Squares Parameter Identification . . . . . . . . . . . . . . . . . . . . . . 125
6.1 Least-Squares Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.2 Convex Optimization and Least-Squares Gradient Method . . . . 128
6.3 Persistent Excitation and Parameter Convergence . . . . . . . . . . 130
6.4 Recursive Least-Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.5 Indirect Adaptive Control with Least-Squares Parameter
Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.6 Estimation of Time Derivative Signals . . . . . . . . . . . . . . . . . . 146
6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7 Function Approximation and Adaptive Control with
Unstructured Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.1 Polynomial Approximation by Least-Squares . . . . . . . . . . . . . 152
7.2 Neural Network Approximation . . . . . . . . . . . . . . . . . . . . . . . 154
7.3 Adaptive Control with Unstructured Uncertainty . . . . . . . . . . . 159
7.3.1 Recursive Least-Squares Direct Adaptive Control
with Matrix Inversion . . . . . . . . . . . . . . . . . . . . . . . . 160
7.3.2 Modified Recursive Least-Squares Direct Adaptive
Control without Matrix Inversion . . . . . . . . . . . . . . . . 164
7.3.3 Least-Squares Gradient Direct Adaptive Control . . . . . 168
7.3.4 Least-Squares Gradient Direct Adaptive Control
with Neural Network Approximation . . . . . . . . . . . . . 173
7.3.5 MRAC with Neural Network Approximation . . . . . . . 175
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8 Robustness Issues with Adaptive Control . . . . . . . . . . . . . . . . . . . . 185
8.1 Parameter Drift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
8.2 Non-minimum Phase Systems . . . . . . . . . . . . . . . . . . . . . . . . 190
8.3 Time-Delay Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.4 Unmodeled Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.5 Fast Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
xiv Contents

8.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
9 Robust Adaptive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
9.1 Dead-Zone Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
9.2 Projection Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
9.3 r Modification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
9.4 e Modification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
9.5 Optimal Control Modification . . . . . . . . . . . . . . . . . . . . . . . . 234
9.5.1 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
9.5.2 Derivation of Optimal Control Modification . . . . . . . . 240
9.5.3 Lyapunov Stability Analysis . . . . . . . . . . . . . . . . . . . 247
9.5.4 Linear Asymptotic Property . . . . . . . . . . . . . . . . . . . . 254
9.6 Adaptive Loop Recovery Modification . . . . . . . . . . . . . . . . . . 261
9.7 L1 Adaptive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
9.8 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
9.9 Covariance Adjustment of Adaptation Rate . . . . . . . . . . . . . . . 280
9.10 Optimal Control Modification for Systems with Control Input
Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
9.11 Bi-Objective Optimal Control Modification for Systems
with Control Input Uncertainty . . . . . . . . . . . . . . . . . . . . . . . 291
9.12 Adaptive Control for Singularly Perturbed Systems
with First-Order Slow Actuator Dynamics . . . . . . . . . . . . . . . 307
9.13 Optimal Control Modification for Linear Uncertain Systems
with Unmodeled Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 314
9.14 Adaptive Control of Non-Minimum Phase Plants
with Relative Degree 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
9.14.1 Minimum Phase Plant . . . . . . . . . . . . . . . . . . . . . . . . 323
9.14.2 Non-Minimum Phase Plant . . . . . . . . . . . . . . . . . . . . 325
9.15 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
9.16 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
10 Aerospace Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
10.1 Inverted Pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
10.2 Double Pendulum in Robotic Applications . . . . . . . . . . . . . . . 354
10.3 Adaptive Control of Aircraft Longitudinal Dynamics . . . . . . . 359
10.4 Recursive Least-Squares and Neural Network Pitch Attitude
Adaptive Flight Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
10.5 Adaptive Control of Flexible Aircraft . . . . . . . . . . . . . . . . . . . 374
Contents xv

10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression

Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
10.7 Adaptive Flight Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
10.8 Hybrid Adaptive Flight Control . . . . . . . . . . . . . . . . . . . . . . . 414
10.9 Adaptive Flight Control for F-18 Aircraft with Optimal
Control Modification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
10.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
10.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Suggested Exam Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
About the Author

Dr. Nhan Nguyen is a senior research scientist at NASA Ames Research Center
with over 31 years of experience. He is the technical group lead of the Advanced
Control and Evolvable Systems research group in the Intelligent Systems Division
with 22 researchers conducting aerospace flight control research. He is currently
leading a project team to conduct research and development of advanced aeroser-
voelastic control technologies for flexible aircraft. He served as the Project Scientist
in the NASA Integrated Resilient Aircraft Control (IRAC) Project from 2007 to
2010 with an annual budget of $21 million that funded advanced research in
adaptive control conducted by NASA, industry, and academia. His prior position
was the Deputy Principal Investigator of the IRAC project from 2005 to 2006. His
work on optimal control modification adaptive control culminated in a successful
flight validation experiment on NASA F-18 research aircraft in 2010. He has
received several prestigious awards including the NASA Exceptional Scientific
Achievement Medal, NASA Exceptional Achievement Medal, two-time NASA
Ames Honors for Excellence as Engineer, and more than 50 other NASA career
awards. He has four US patents and two other patents in pending. He has over 15
NASA invention disclosures. He has published over 200 peer-reviewed technical
publications. He is the chair of the American Institute of Aeronautics and
Astronautics (AIAA) Intelligent Systems Technical Committee with 50 profes-
sional members. He has received two AIAA Distinguished Services Awards.

xvii
Chapter 1
Introduction

Abstract This chapter provides a brief summary of the recent advancements made in
model-reference adaptive control theory. Adaptive control is a promising technology
that can improve the performance of control systems in the presence of uncertainty
due to a variety of factors such as degradation and modeling uncertainty. During the
past decade, the adaptive control research community has produced several advance-
ments in adaptive control theory along with many novel adaptive control methods
as a result of the increased government research funding. Many of these new adap-
tive control methods have added new capabilities in terms of improved performance
and robustness that further increase the viability of model-reference adaptive control
as a future technology. Flight test validation of adaptive control on full-scale air-
craft and unmanned arial vehicles has increased the confidence in model-reference
adaptive control as a possible new flight control technology for aerospace vehicles
in the near future. In spite of the five decades of research in adaptive control, the
fact still remains that currently no adaptive control system has ever been deployed
on any safety-critical or human-rated production systems. Many technical problems
remain unresolved. As a nonlinear control method, the lack of well-accepted met-
rics for adaptive control system design presents a major hurdle for certification. The
development of certifiable adaptive control systems is viewed as a major techni-
cal challenge for the adaptive control research community to address in the current
research.

Adaptive control is a well-researched topic in control theory that spans several

decades. Many adaptive control applications can be found in aerospace and other
settings, but very few or none have been deployed on safety-critical production
systems. In this chapter, the learning objectives are:
• To understand a brief history of adaptive control research and the extensive
on-going research in this topic;

© Springer International Publishing AG 2018 1

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_1
2 1 Introduction

• To recognize the fact that, in spite of the many advancements in the field, adoption
of adaptive control technologies is not widespread and its applications are limited
in scope; and
• To recognize that verification, validation, and certification are viewed as significant
technical barriers that need to be further researched in order to increase the con-
fidence in and the trustworthiness of adaptive control technologies, which would
enable future adoption.

1.1 Overview

Adaptive control is a study of a class of nonlinear control methods that deal with
systems with uncertainty. The uncertainty can be due to unforeseen internal changes
in system dynamics or external disturbances. An adaptive control system can be
broadly described as any control system that has the ability to adjust control design
parameters such as control gains online based on inputs received by the plant in
order to accommodate system uncertainty as illustrated in Fig. 1.1. The adjustable
parameters are called adaptive parameters, and the adjusting mechanism, which is
described by a set of mathematical equations, is called an adaptive law. In most
cases, a typical adaptive law is nonlinear. This nonlinearity makes adaptive control
inherently difficult to design and analyze since many traditional design and analysis
methods for linear time-invariant (LTI) control systems such as bode plot, phase and
gain margins, eigenvalue analysis are not applicable.
Adaptive control has a long history dated back to the early 1950s when there was
an extensive research interest in designing advanced autopilots for high-performance
aircraft that operated over a wide range of flight conditions [1]. After a significant
research and development effort, gain-scheduling control was gaining acceptance
over adaptive control due to the inherent difficulty associated with nonlinearity of
adaptive control. On the other hand, gain-scheduling control can take advantage of
classical control techniques in the design selection of suitable control gains that are
scheduled as a function of aircraft operating conditions.
The 1960s saw the emergence of the modern control theory and Lyapunov stabil-
ity theory that also contributed to adaptive control theory. Whitaker et al. developed
model-reference adaptive control with the sensitivity method and MIT rule. During
this development phase, there was a lack of stability proofs and good understand-
ing of the properties of adaptive control. In 1967, an adaptive flight controller was
flight-tested by NASA on one of the three X-15 experimental hypersonic aircrafts
[2, 3]. Several successful test flights were made before an ill-fated test flight occurred
that resulted in a fatal crash of one of these vehicles. This event and the technical
difficulties led to a diminished interest in adaptive control.
1.1 Overview 3

Fig. 1.1 Classes of automatic control for dynamical systems

In the 1970s, the Lyapunov stability theory was established as a foundation for
model-reference adaptive control. This represented a breakthrough in adaptive con-
trol as it led to the development of improved model-reference adaptive control
schemes with the accompanied Lyapunov stability theory. The excitements soon
gave way to a realization in the 1980s that, in spite of the Lyapunov stability theory,
adaptive control can exhibit unstable behaviors in the presence of small disturbances
or unmodeled dynamics [4]. This led to the recognition that model-reference adap-
tive control was sensitive to system modeling accuracy and the resultant mismatch
between a real system and a plant model of the system. This lack of robustness
led to the development of the σ modification [5] and e modification to improve the
stability properties of adaptive control [6]. These robust modification schemes rep-
resent a new class of “robust adaptive control.”
In the 1990s and up to the present time, research in adaptive control remains
highly active. Neural networks were introduced as a mechanism for adaptation
[7–12]. This led to the development of a class of adaptive control called “intelli-
gent control” or “neural network adaptive control” which utilizes neural networks to
approximate model uncertainties, but the basic framework remains to be the same as
model-reference adaptive control [13–21].
The following decade saw a resurgence of interest in adaptive control research.
This also coincided with a period of increased research funding made possible by
NASA [22] and other US government agencies. In particular, NASA has been an
active key player in adaptive control technology development as illustrated on the
time line in Fig. 1.2. During this period, the adaptive control research community
produced several advancements in adaptive control theory along with many novel
adaptive control methods as a result of the increased research funding. The list of
these new advancements is long, and it would not do justice to include all of these
new methods in this limited space. The reader is referred to a subset of this list from
which those novel adaptive control methods that were funded by NASA during this
decade can be found. These methods include: adaptive control based on retrospective
cost optimization by Santillo and Bernstein [23, 24], adaptive control with reference
model modification by Stepanyan and Krishnakumar [25, 26], adaptive loop recov-
4 1 Introduction

ery by Calise and Yucelen [27, 28], bounded linear stability analysis metric-driven
adaptive control by Nguyen [29, 30], combined/composite model-reference adaptive
control by Lavretsky [31, 32], derivative-free model-reference adaptive control by
Yucelen and Calise [33, 34], hybrid adaptive control by Nguyen [35–37], K modifica-
tion by Kim et al. [38], Kalman filter modification by Yucelen and Calise [39, 40], L1
adaptive control by Hovakimyan and Cao [41–43], least-squares adaptive control by
Nguyen [44, 45], concurrent learning least-squares adaptive control by Chowdhary
and Johnson [46, 47], modified state observer adaptive control by Balaskrishnan [48],
multivariable model-reference adaptive control by Guo and Tao [49], optimal control
modification [50, 51] and multiobjective optimal control modification by Nguyen
[52, 53], Q modification by Volyanskyy et al. [54, 55], and parameter-dependent
Riccati equation adaptive control by Kim et al. [56]. Many of these new adaptive
control methods have added new capabilities in terms of improved performance and
robustness that further increase the viability of model-reference adaptive control as
a future technology.

Fig. 1.2 Adaptive control research time line

In terms of flight validation, an intelligent flight control system based on the

sigma-pi neural network adaptive control by Calise and Rysdyk [13] was developed
and flight-tested by NASA on a F-15 aircraft at NASA Armstrong (former Dryden)
Flight Research Center in the early 2000s to demonstrate the capability of neural
network adaptive control [57, 58]. In 2010, another flight test program was conducted
on a F/A-18 aircraft at NASA Armstrong Flight Research Center to demonstrate a
new simplified adaptive flight controller based on the optimal control modification
[59–62]. A pilot-in-the-loop high-fidelity flight simulation study of several adaptive
control methods was conducted at NASA Ames Research Center in 2009 [63, 64].
A flight experiment of the L1 adaptive control was conducted on NASA AirSTAR
flight test vehicle at NASA Langley Research Center in 2009 [65]. Another flight
experiment of the L1 adaptive control was conducted on a UAV (unmanned arial
vehicle) at the Naval Postgraduate School in 2009 [66]. Model-reference adaptive
1.1 Overview 5

control was also evaluated for fight testing on a Beechcraft Bonanza fly-by-wire
testbed in 2010 [67]. These flight experiments and many others that followed have
increased the confidence in model-reference adaptive control as a possible new flight
control technology for aerospace vehicles in the near future. They also revealed that
further flight validation for adaptive control is still needed to mature this technology.
Research in adaptive control continues at the present time. It would be well beyond
the scope of this book to include all current research activities. The interested reader is
referred to the references for more information on numerous applications for aircraft
[68–79], spacecraft [21, 80–83], UAVs [15, 47, 65, 66, 84], aerospace structures
[85, 86], robotic systems [18, 87], munition systems [88], fluid systems [89], and
many others.
The current research in adaptive control generally lacks the ability to deal with
integrated effects that exist in many different system designs and operations. These
include but are not limited to: complex uncertainties due to unmodeled dynamics
[4, 90], gross changes in system dynamics resulting from unintended operations and
structural damage [36, 76, 91], unknown component failures and anomalies [80,
92–94], high degree of design complexity [59], novel actuators and sensors [79, 95],
multiphysics interactions [85, 86, 96–98], and many other effects.
Adaptive control plays a major role in the aerospace domain. In aerospace vehicles
with damaged structures or failed flight control surfaces or operating in off-nominal
flight events, the vehicles can exhibit numerous coupled effects such as aerodynam-
ics, vehicle dynamics, structural dynamics, and propulsion. These coupled effects
can impose a wide range of uncertainties on the performance of a flight control sys-
tem. Thus, even though an adaptive control system may be shown to be stable in a
nominal flight condition, it may fail to provide enough stability in the presence of
these uncertainties [99, 100]. For example, conventional aircraft flight control sys-
tems incorporate aeroservoelastic (ASE) notch filters to prevent control signals from
exciting aeroelastic wing modes. However, if changes in aircraft dynamics are signif-
icant, the frequencies of the aeroelastic modes may change enough to render the ASE
notch filters ineffective. This could cause the control signals to excite wing modes
which could lead to problems for a pilot to control an aircraft. Another problem
faced by adaptive control is the ability of adaptive control systems to accommodate
slow or degraded flight control actuators such an impaired flight control surface or
engine as a flight control effector [79, 101]. The dissimilar actuator rates due to slow
actuator dynamics can cause problems with adaptive control and can potentially lead
to pilot-induced oscillations (PIO) [102].
To adequately deal with these coupled effects, integrated approaches in adaptive
control research need to be developed. These integrated approaches will be required to
develop new fundamental multidisciplinary methods in adaptive control and system
modeling. Unmodeled dynamics are a significant source of uncertainty and a causal
factor of instability of adaptive control systems in high-gain adaptation situations.
Future research in adaptive control should leverage fundamental understanding of the
structures of these secondary dynamics by incorporating multidisciplinary methods
into the design of an adaptive control system. With a better understanding of the
6 1 Introduction

system uncertainties, more effective adaptive control methods may be developed to

achieve improved robustness in the presence of uncertainties.

1.2 Verification and Validation Challenges for Adaptive

Flight Control Systems

In spite of the five decades of research in adaptive control, the fact still remains that
currently no adaptive control system has ever been deployed on any safety-critical
or human-rated production systems such as passenger transport aircraft [99, 103–
106]. It should be noted that adaptive control has been successfully developed and
deployed in munition production systems [88]. The problem lies in the difficulty
with the certification of adaptive control systems since existing certification methods
for linear time-invariant (LTI) control systems cannot readily be used for nonlinear
adaptive control systems. Research to address the notion of metrics for adaptive
control began to appear during the first decade of the twenty-first century [14, 107–
115]. The goal of this research is to develop performance and stability metrics for
adaptive control similar to overshoot, settling time, phase and gain margins for LTI
control systems. These metrics, if accepted, could pave a path toward certification
that would potentially lead to the adoption of adaptive control as a future control
technology for safety-critical and human-rated production systems.
Development of certifiable adaptive control systems represents a major challenge
to overcome. Adaptive control systems with learning algorithms will never become
part of the future unless it can be proven that they are highly safe and reliable.
Rigorous methods for adaptive control software verification and validation must
therefore be developed to ensure that adaptive control system software failures will
not occur to verify that the adaptive control system functions as required, to eliminate
unintended functionality, and to demonstrate that certification requirements imposed
by regulatory bodies such as the Federal Aviation Administration (FAA) can be
satisfied [104, 105].
The ability of an adaptive control system to modify a pre-designed flight control
system is a strength and a weakness at the same time. On the one hand, the abil-
ity of adaptive control systems to accommodate system degradation is considered
as a major advantage since traditional gain-scheduling control methods tend to be
less capable of handling off-nominal flight conditions outside their design operating
envelopes. On the other hand, serious problems with adaptive control can still ex-
ist with regard to unmodeled dynamics and high-gain adaptation. Adaptive control
systems are sensitive to these potential problems as well as many others including
actuator dynamics and exogenous disturbances. To be certifiable, an adaptive flight
control system must be able to prove that these effects as well as other factors such
as time delay, system constraints, and measurement noise are handled in a globally
satisfactory manner.
1.2 Verification and Validation Challenges for Adaptive Flight Control Systems 7

1.2.1 Verification by Simulation of Adaptive Flight Control

Systems

Simulation is an integral part of the verification of adaptive control systems [104, 105,
116, 117]. Many aspects of adaptive control systems, in particular convergence and
stability, can only be analyzed with simulations that provide enough detail and fidelity
to model significant nonlinear dynamics. For example, stall upset of an aircraft cannot
be expressed as a linear model since this effect is highly nonlinear and unsteady.
Simulation provides a fairly rapid way to accomplish the following tasks:
• Evaluation and comparison of different adaptive control algorithms;
• Tuning control gains and weight update laws;
• Determination of progress of adaptation at each step;
• Evaluation of the effects of process and measurement noise on convergence of
adaptive parameters;
• Determination of stability boundaries;
• Validation with actual flight computer hardware;
• Conducting piloted evaluation of adaptive control in a flight simulator;
• Simulating ad hoc techniques of improving the adaptation process, such as adding
persistent excitation to improve identification and convergence, or stopping the
adaptation process after the tracking error converges to within a specified tolerance
or after a specified number of iterations.
Simulations differ primarily in the fidelity with which the plant is modeled. Higher
fidelity simulations require more complicated mathematical models of the adaptive
control system and also a greater use of actual (and expensive) controller hardware.
The behavior of simple linear models is compared to that of higher fidelity nonlinear
models when they are available to ensure that analyses performed using the linear
models still apply. In order to be cost effective, the lowest fidelity testbed is usually
used as much as possible.
The lowest fidelity simulations are usually run on a desktop computer. This simu-
lation typically includes the control laws and a linear or nonlinear plant of the system
dynamics. A linear model is most often used in early control law design and analysis
or to calculate linear gain and phase margins. Changes to the plant model can be
simulated by changing the system transfer function from one matrix to another with
varying frequency. By varying the amount of change, the stability boundaries of the
system can be determined. Concomitant with this process is an evaluation of the sys-
tem tuning parameters that are used in the adaptive control algorithms. The desktop
simulation environment provides a quick way to compare different adaptive control
algorithms and controller architectures. Only the most promising designs need to be
simulated using higher fidelity simulations.
Higher fidelity simulation testbeds use actual flight hardware (or even aircraft) in
the simulation of the control loop and are often run in dedicated computing environ-
ments with a cockpit and out-the-window graphics [117, 118]. These simulations
may include a cockpit to interface with the pilot and can either be fixed-base or
8 1 Introduction

motion-base. Motion-base simulators additionally provide the pilot with simulations

of physical (motion and visual) cues of the actual flight [63]. Typically, they contain
software models of nonlinear aircraft dynamics, actuator models, and sensor models.
Using the actual aircraft flight computer is a particularly important advantage of this
simulation, since all computers tend to handle exceptions differently and may have
differences in their computational processes. Either an actual or iron-bird aircraft
may be used in a higher fidelity simulation to provide realistic actuator dynamics,
sensor noise, actual flight wiring, and some structural interactions. These testbeds
allow for a complete check out of all interfaces to the flight hardware, timing tests,
and various failure modes and effects analysis (FMEA) testing, which is not possible
in a low fidelity simulation.

1.2.2 Adaptive Control Metrics

In spite of the many advancements made in adaptive control research and the poten-
tial benefits of adaptive control systems, the absence of verification and validation
methods for adaptive control systems remains a major hurdle to the implementation
of adaptive control in safety-critical or human-rated production systems. This hurdle
can be traced to the lack of performance and stability metrics for adaptive control.
The development of verifiable metrics for adaptive control is an important aspect
of adaptive control research in order to mature adaptive control technology for use
in future safety-critical and human-rated production systems. Stability metrics of
adaptive control are an important consideration for assessing system robustness to
unmodeled dynamics, time delay, high-gain learning, and exogenous disturbances.
Therefore, it is imperative to define a proper set of stability and performance metrics
for adaptive systems as a first step in the development of reliable verification and
validation methods that could lead to the certification of adaptive control software.
Another potential benefit of metrics for adaptive control is the consideration for
metrics-driven adaptive control. Metrics-driven adaptive control is the notion that
adaptation in certain cases should be driven by the need to trade off between stability
and performance to maintain operational safety of a control system [30]. Research in
this area has resulted in some initial analysis methods for computing stability metrics
online that could be used to adjust the adaptive parameters of an adaptive control
system to improve stability margins of the closed-loop system [29].
Generally, the classical gain and phase margins for LTI systems are not applicable
to adaptive control systems which are nonlinear. Some results began to appear in the
literature on the definition and evaluation of candidate metrics for adaptive control
systems [14, 109, 110, 112, 115]. The use of a parameter sensitivity as a metric for
the output of neural networks has been investigated [103]. Stability metrics based
on the Lyapunov analysis and passivity theory have been studied [108]. Metrics for
evaluation of stability and robustness of adaptive control systems can also be obtained
from optimization frameworks [107, 119]. The notion of time-delay margin as a
stability metric has been studied by many authors [120, 121].
1.2 Verification and Validation Challenges for Adaptive Flight Control Systems 9

There are several unaddressed issues related to the implementation of adaptive

control technology in aerospace vehicles to accommodate uncertainty in adverse
events. These issues include but are not limited to: (1) achievable stability metrics of
adaptive control as related to uncertainty bounds; (2) adaptation in the presence of
static and dynamic saturation of actuators; (3) cross-coupling between longitudinal
and lateral-directional axes due to failures, damage, and different rates of adaptation
in each axis; (4) online reconfiguration and control reallocation using nontraditional
control effectors such as engines; and (5) timescale separation in actuator systems
with different time latencies such as conventional control surfaces and engines.

1.3 Summary

Adaptive control is a well-researched subject over the past several decades. Adap-
tive control is a promising technology that could provide improved performance of
control systems in the presence of uncertainty due to a variety of factors such as degra-
dation and modeling uncertainty. During the past decade, advancements in adaptive
control theory along with many novel adaptive control methods have been made by
researchers in the field. Many of these new adaptive control methods have added new
capabilities in terms of improved performance and robustness that further increase
the viability of model-reference adaptive control as a future technology. Flight test
validation of adaptive control on full-scale aircraft and unmanned arial vehicles has
increased the confidence in model-reference adaptive control as a possible new flight
control technology for aerospace vehicles in the near future.
In spite of this, adaptive control is not well-accepted for applications in safety-
critical or human-rated production systems. Many technical problems remain unre-
solved. As a nonlinear control method, the lack of well-accepted metrics for adaptive
control system design as opposed to linear control systems presents a major hurdle
for certification.

References

1. Åström, K. J., & Wittenmark, B. (2008). Adaptive control: Dover Publications Inc.
2. Jenkins, D. R. (2000). Hypersonics Before the Shuttle: A Concise History of the X-15 Research
Airplane, NASA SP-2000-4518.
3. Staff of the Flight Research Center. (1971). Experience with the X-15 Adaptive Flight Control
System, NASA TN D-6208.
4. Rohrs, C. E., Valavani, L., Athans, M., & Stein, G. (1985). Robustness of continuous-time
adaptive control algorithms in the presence of unmodeled dynamics. IEEE Transactions on
Automatic Control, AC-30(9), 881–889.
5. Ioannou, P., & Kokotovic, P. (1984). Instability analysis and improvement of robustness of
adaptive control. Automatica, 20(5), 583–594.
6. Narendra, K. S. & Annaswamy, A. M. (1987). A new adaptive law for robust adaptation
without persistent excitation. IEEE Transactions on Automatic Control, AC-32(2), 134–145.
10 1 Introduction

7. Cybenko, G. (1989). Approximation by superpositions of a sigmoidal function. Mathematics

of Control Signals Systems, 2, 303–314.
8. Lee, T., & Jeng, J. (1998). The Chebyshev-Polynomials-based unified model neural networks
for function approximation. IEEE Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics, 28(6), 925–935.
9. Micchelli, C. A. (1986). Interpolation of scattered data: distance matrices and conditionally
positive definite functions. Constructive Approximation, 2, 11–12.
10. Moody, J. (1992). The effective number of parameters: An analysis of generalization and
regularization in nonlinear learning systems. Advances in Neural Information Processing
Systems (Vol. 4). San Mateo: Morgan Kaufmann Publishers.
11. Suykens, J., Vandewalle, J., & deMoor, B. (1996). Artificial neural networks for modeling
and control of non-linear systems: Dordrecht: Kluwer Academic Publisher.
12. Wang, X., Huang, Y., & Nguyen, N. (2010). Robustness quantification of recurrent neural
network using unscented transform. Elsevier Journal of Neural Computing, 74(1–3),
13. Calise, A. J., & Rysdyk, R. T. (1998). Nonlinear adaptive flight control using neural networks.
IEEE Control System Magazine, 18(6), 1425.
14. Ishihara, A., Ben-Menahem, S., & Nguyen, N. (2009). Protection ellipsoids for stability
analysis of feedforward neural-net controllers. In International Joint Conference on Neural
Networks.
15. Johnson, E. N., Calise, A. J., El-Shirbiny, H. A., & Rysdyk, R. T. (2000). Feedback lineariza-
tion with neural network augmentation applied to X-33 attitude control. In AIAA Guidance,
Navigation, and Control Conference, AIAA-2000-4157, August, 2000.
16. Kim, B. S., & Calise, A. J. (1997). Nonlinear flight control using neural networks. Journal of
Guidance, Control, and Dynamics, 20(1), 26–33.
17. Lam, Q., Nguyen, N., & Oppenheimer, M. (2012). Intelligent adaptive flight control using
optimal control modification and neural network as control augmentation layer and robustness
enhancer. In AIAA Infotech@Aerospace conference, AIAA-2012-2519, June, 2012.
18. Lewis, F. W., Jagannathan, S., & Yesildirak, A. (1998). Neural network control of robot
manipulators and non-linear systems. Boca Raton: CRC.
19. Rysdyk, R. T., Nardi, F., & Calise, A. J. (1999). Robust adaptive nonlinear flight control
applications using neural networks. In American Control Conference.
20. Steinberg, M. L. (1999). A comparison of intelligent, adaptive, and nonlinear flight control
laws. In AIAA Guidance, Navigation, and Control Conference, AIAA-1999-4044, August,
1999.
21. Zou, A., Kumar, K., & Hou, Z. (2010). Attitude control of spacecraft using chebyshev neural
networks. IEEE Transactions on Neural Networks, 21(9), 1457–1471.
22. Krishnakumar, K., Nguyen, N., & Kaneshige, J. (2010, December). Integrated resilient aircraft
control. In Encyclopedia of Aerospace Engineering (Vol. 8). New Jersey: Wiley. ISBN: 978-
0-470-75440-5.
23. Santillo, M. A. & Bernstein, D. S. (2008). A retrospective correction filter for discrete-time
adaptive control of non-minimum phase systems. In IEEE Conference on Decision and Con-
trol, December, 2008.
24. Santillo, M. A., & Bernstein, D. S. (2010). Adaptive control based on retrospective cost
optimization. AIAA Journal of Guidance, Control, and Dynamics, 33(2), 289–304.
25. Stepanyan V. & Krishnakumar, K. (2010). MRAC revisited: Guaranteed performance with
reference model modification. In American Control Conference.
26. Stepanyan, V., & Krishnakumar, K. (2012). Adaptive control with reference model modifica-
tion. AIAA Journal of Guidance, Control, and Dynamics, 35(4), 1370–1374.
27. Calise, A. J., Yucelen, T., Muse, J., & Yang, B. (2009). A loop recovery method for adaptive
control. In AIAA Guidance, Navigation, and Control Conference, AIAA-2009-5967, August
2009.
28. Calise, A. J., & Yucelen, T. (2012). Adaptive loop transfer recovery. AIAA Journal of Guidance,
Control, and Dynamics, 35(3), 807–815.
References 11

29. Nguyen, N., Bakhtiari-Nejad, M., & Huang, Y. (2007). Hybrid adaptive flight control with
bounded linear stability analysis. In AIAA Guidance, Navigation, and Control Conference,
AIAA-2007-6422, August 2007.
30. Bakhtiari-Nejad, M., Nguyen, N., & Krishnakumar, K. (2009). Adjustment of adaptive gain
with bounded linear stability analysis to improve time-delay margin for metrics-driven adap-
tive control. In AIAA Infotech@Aerospace Conference, AIAA-2009-1801, April 2009.
31. Lavretsky, E. (2009). Combined/composite model reference adaptive control. In AIAA Guid-
ance, Navigation, and Control Conference, AIAA-2009-6065, August 2009.
32. Lavretsky, E. (2009). Combined/composite model reference adaptive control. IEEE Transac-
tions on Automatic Control, 54(11), 2692–2697.
33. Yucelen, T., & Calise, A. J. (2010). Derivative-free model reference adaptive control. In AIAA
Guidance, Navigation, and [11, 5] Control Conference, AIAA-2009-5858, August 2010.
34. Yucelen, T., & Calise, A. J. (2011). Derivative-free model reference adaptive control. AIAA
Journal of Guidance, Control, and Dynamics, 34(4), 933–950.
35. Nguyen, N., Krishnakumar, K., Kaneshige, J., & Nespeca, P. (2006). Dynamics and adaptive
control for stability recovery of damaged asymmetric aircraft. In AIAA Guidance, Navigation,
and Control Conference, AIAA-2006-6049, August 2006.
36. Nguyen, N., Krishnakumar, K., Kaneshige, J., & Nespeca, P. (2008). Flight dynamics model-
ing and hybrid adaptive control of damaged asymmetric aircraft. AIAA Journal of Guidance,
Control, and Dynamics, 31(3), 751–764.
37. Nguyen, N. (2011). Hybrid adaptive flight control with model inversion adaptation. In Ad-
vances in Flight Control Systems. Croatia: Intech Publishing. ISBN 978-953-307-218-0.
38. Kim, K., Yucelen, T., & Calise, A. J. (2010). K-modification in adaptive control. In AIAA
Infotech@Aerospace Conference, AIAA-2010-3321, April 2010.
39. Yucelen, T., & Calise, A. J. (2009). A Kalman filter optimization approach to direct adaptive
control. In AIAA Guidance, Navigation, and Control Conference, AIAA-2010-7769, August
2009.
40. Yucelen, T., & Calise, A. J. (2010). A Kalman filter modification in adaptive control. AIAA
Journal of Guidance, Control, and Dynamics, 33(2), 426–439.
41. Cao, C., & Hovakimyan, N. (2007). Guaranteed transient performance with L1 adaptive
controller for systems with unknown time-varying parameters and bounded disturbances:
Part I. In American Control Conference, July 2007.
42. Cao, C., & Hovakimyan, N. (2008). Design and analysis of a novel L1 adaptive control
architecture with guaranteed transient performance. IEEE Transactions on Automatic Control,
53(2), 586–591.
43. Hovakimyan, N., & Cao, C. (2010). L1 Adaptive control theory: Guaranteed robustness with
fast adaptation. Society for Industrial and Applied Mathematics.
44. Nguyen, N., Burken, J., & Ishihara, A. (2011). Least-squares adaptive control using Chebyshev
orthogonal polynomials. In AIAA Infotech@Aerospace Conference, AIAA-2011-1402, March
2011.
45. Nguyen, N. (2013). Least-squares model reference adaptive control with Chebyshev orthog-
onal polynomial approximation. AIAA Journal of Aerospace Information Systems, 10(6),
268–286.
46. Chowdhary, G., & Johnson, E. (2010). Least squares based modification for adaptive control.
In IEEE Conference on Decision and Control, December 2010.
47. Chowdhary, G., & Johnson, E. (2011). Theory and flight-test validation of a concurrent-
learning adaptive controller. AIAA Journal of Guidance, Control, and Dynamics, 34(2), 592–
607.
48. Balakrishnan, S. N., Unnikrishnan, N., Nguyen, N., & Krishnakumar, K. (2009). Neuroad-
aptive model following controller design for non-affine and non- square aircraft systems. In
AIAA Guidance, Navigation, and Control Conference, AIAA-2009-5737, August 2009.
49. Guo, J., & Tao, G. (20011). A Multivariable MRAC design for aircraft systems under failure
and damage conditions. In American Control Conference, June 2011.
12 1 Introduction

50. Nguyen, N., Krishnakumar, K., & Boskovic, J. (2008). An optimal control modification to
model-reference adaptive control for fast adaptation. In AIAA Guidance, Navigation, and
Control Conference, AIAA 2008-7283, August 2008.
51. Nguyen, N. (2012). Optimal control modification for robust adaptive control with large adap-
tive gain. Systems and Control Letters, 61(2012), 485–494.
52. Nguyen, N. (2014). Multi-objective optimal control modification adaptive control method for
systems with input and unmatched uncertainties. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2014-0454, January 2014.
53. Nguyen, N., & Balakrishnan, S. N. (2014). Bi-objective optimal control modification adaptive
control for systems with input uncertainty. IEEE/CAA Journal of Automatica Sinica, 1(4),
423–434.
54. Volyanskyy, K. Y., & Calise, A. J. (2006). A novel q-modification term for adaptive control.
In American Control Conference, June 2006.
55. Volyanskyy, K. Y., Haddad, W. M., & Calise, A. J. (2009). A new neuroadaptive control
architecture for nonlinear uncertain dynamical systems: Beyond σ - and e-modifications. IEEE
Transactions on Neural Networks, 20, 1707–1723.
56. Kim, K., Yucelen, T., Calise, A.J., & Nguyen, N. (2011). Adaptive output feedback control
for an aeroelastic generic transport model: A parameter dependent Riccati equation approach.
In AIAA Guidance, Navigation, and Control Conference, AIAA-2011-6456, August 2011.
57. Bosworth, J., & Williams-Hayes, P. S. (2007). Flight test results from the NF-15B IFCS project
with adaptation to a simulated stabilator failure. In AIAA Infotech@Aerospace Conference,
AIAA-2007-2818, May 2007.
58. Williams-Hayes, P. S. Flight test implementation of a second generation intelligent flight
control system. In NASA TM-2005-213669.
59. Hanson, C., Johnson, M., Schaefer, J., Nguyen, N., & Burken, J. (2011). Handling qualities
evaluations of low complexity model reference adaptive controllers for reduced pitch and roll
damping scenarios. In AIAA Guidance, Navigation, and Control Conference, AIAA-2011-
6607, August 2011.
60. Hanson, C., Schaefer, J., Johnson, M., & Nguyen, N. Design of low complexity model refer-
ence adaptive controllers. In NASA-TM-215972.
61. Nguyen, N., Hanson, C., Burken, J., & Schaefer, J. (2016). Normalized optimal control modi-
fication and flight experiments on NASA F/A-18 aircraft. AIAA Journal of Guidance, Control,
and Dynamics.
62. Schaefer, J., Hanson , C., Johnson, M., & Nguyen, N. (2011). Handling qualities of model
reference adaptive controllers with varying complexity for pitch-roll coupled failures. In IAA
Guidance, Navigation, and Control Conference, AIAA-2011-6453, August 2011.
63. Campbell, S., Kaneshige, J., Nguyen, N., & Krishnakumar, K. (2010). An adaptive control
simulation study using pilot handling qualities evaluations. In AIAA Guidance, Navigation,
and Control Conference, AIAA-2010-8013, August 2010.
64. Campbell, S., Kaneshige, J., Nguyen, N., & Krishnakumar, K. (2010). Implementation and
evaluation of multiple adaptive control technologies for a generic transport aircraft simulation.
In AIAA Infotech@Aerospace Conference, AIAA-2010-3322, April 2010.
65. Gregory, I. M, Cao, C., Xargay, E., Hovakimyan, N., & Zou, X. (2009). L1 adaptive control
design for NASA AirSTAR flight test vehicle. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2009-5738, August 2009.
66. Kitsios, I., Dobrokhodov, V., Kaminer, I., Jones, K., Xargay, E., Hovakimyan, N., Cao, C.,
Lizarraga, M., Gregory, I., Nguyen, N., & Krishnakumar, K. (2009). Experimental validation
of a metrics driven L1 adaptive control in the presence of generalized unmodeled dynamics.
In AIAA Guidance, Navigation, and Control Conference, AIAA-2009-6188, August 2009.
67. Steck, J., Lemon, K., Hinson, B., Kimball, D., & Nguyen, N. (2010). Model reference adap-
tive fight control adapted for general aviation: Controller gain simulation and preliminary
flight testing on a bonanza fly-by-wire Testbed. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2010-8278, August 2010.
References 13

68. Bhattacharyya, S., Krishnakumar, K., & Nguyen, N. (2012). Adaptive autopilot designs for
improved tracking and stability. In AIAA Infotech@Aerospace Conference, AIAA-2012-2494,
June 2012.
69. Burken, J., Nguyen, N., & Griffin, B. (2010). Adaptive flight control design with optimal
control modification for F- 18 aircraft model. In AIAA Infotech@Aerospace Conference,
AIAA-2010-3364, April 2010.
70. Campbell, S., Nguyen, N., Kaneshige, J., & Krishnakumar, K. (2009). Parameter estimation
for a hybrid adaptive flight controller. In AIAA Infotech@Aerospace Conference, AIAA-2009-
1803, April 2009.
71. Chen, S., Yang, Y., Balakrishnan, S. N., Nguyen, N., & Krishnakumar, K. (2009). SNAC
convergence and use in adaptive autopilot design. In International Joint Conference on Neural
Networks, June 2009.
72. Eberhart, R. L., & Ward, D. G. (1999). Indirect adaptive flight control system interactions.
International Journal of Robust and Nonlinear Control, 9, 1013–1031.
73. Hinson, B., Steck, J., Rokhsaz, K., & Nguyen, N. (2011). Adaptive control of an elastic general
aviation aircraft. In AIAA Guidance, Navigation, and Control Conference, AIAA-2011-6560,
August 2011.
74. Lemon, K., Steck, J., Hinson, B., Rokhsaz, K., & Nguyen, N. (2011). Application of a six
degree of freedom adaptive controller to a general aviation aircraft. In AIAA Guidance, Nav-
igation, and Control Conference, AIAA-2011-6562, August 2011.
75. Nguyen, N., Tuzcu, I., Yucelen, T., & Calise, A. (2011). Longitudinal dynamics and adaptive
control application for an aeroelastic generic transport model. In AIAA Atmospheric Flight
Mechanics Conference, AIAA-2011-6319, August 2011.
76. Rajagopal, K., Balakrishnan, S.N., Nguyen, N., & Krishnakumar, K. (2010). Robust adap-
tive control of a structurally damaged aircraft. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2010-8012, August 2010.
77. Reed, S., Steck, J., & Nguyen, N. (2011). Demonstration of the optimal control modification
for general aviation: Design and simulation. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2011-6254, August 2011.
78. Reed, S., Steck, J., & Nguyen, N. (2014). Demonstration of the optimal control modification
for 6-DOF control of a general aviation aircraft. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2014-1292, January 2014.
79. Stepanyan, V., , Nguyen, N., & Krishnakumar, K. (2009). Adaptive control of a transport
aircraft using differential thrust. In AIAA Guidance, Navigation, and Control Conference,
AIAA-2009-5741, August 2009.
80. Boskovic, J., Jackson, J., Mehra, R., & Nguyen, N. (2009). Multiple-model adaptive fault-
tolerant control of a planetary lander. AIAA Journal of Guidance, Control, and Dynamics,
32(6), 1812–1826.
81. Lam, Q., & Nguyen, N. (2013). Pointing control accuracy and robustness enhancement of
an optical payload system using a direct adaptive control combined with an optimal control
modification. In AIAA Infotech@Aerospace Conference, AIAA-2013-5041, August 2013.
82. Lefevre B., & Jha, R. (2009). Hybrid adaptive launch vehicle ascent flight control. In AIAA
Guidance, Navigation, and Control Conference, AIAA-2009-5958, August 2009.
83. Swei, S., & Nguyen, N. (2016). Adaptive estimation of disturbance torque for orbiting space-
craft using recursive least squares method. In AIAA Infotech@Aerospace Conference, AIAA-
2016-0399, January 2016.
84. Hovakimyan, N., Kim, N., Calise, A. J., Prasad, J. V. R., & Corban, E. J. (2001). Adaptive
output feedback for high-bandwidth control of an unmanned helicopter. In AIAA Guidance,
Navigation and Control Conference, AIAA-2001-4181, August 2001.
85. Nguyen, N., Swei, S., & Ting, E. (2015). Adaptive linear quadratic gaussian optimal control
modification for flutter suppression of adaptive wing. In AIAA Infotech@Aerospace Confer-
ence, AIAA 2015-0118, January 2015.
86. Kumar, M., Rajagopal, K., Balakrishnan, S. N., & Nguyen, N. (2014). Reinforcement learning
based controller synthesis for flexible aircraft wings. IEEE/CAA Journal of Automatica Sinica,
1(4), 435–448.
14 1 Introduction

87. Ishihara, A., Al-Ali, K., Kulkarni, N., & Nguyen, N. (2009). Modeling Error Driven Robot
Control. In AIAA Infotech@Aerospace Conference, AIAA-2009-1974, April 2009.
88. Sharma, M., Calise, A., & Corban, J. E. (2000). Application of an adaptive autopilot design
to a family of guided munitions. In AIAA Guidance, Navigation, and Control Conference,
AIAA-2000-3969, August 2000.
89. Nguyen, N., Bright, M., & Culley, D. (2007). Adaptive feedback optimal control of flow
separation on stators by air injection. AIAA Journal, 45(6),
90. Nguyen, N. (2013). Adaptive control for linear uncertain systems with unmodeled dynam-
ics revisited via optimal control modification. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2013-4988, August 2013.
91. Lemaignan, B. (2005). Flying with no flight controls: handling qualities analyses of the
baghdad event. In AIAA Atmospheric Flight Mechanics Conference, AIAA-2005-5907, August
2005.
92. Nguyen, N. (2012). Bi-objective optimal control modification adaptive control for systems
with input uncertainty. In AIAA Guidance, Navigation, and Control Conference, AIAA-2012-
4615, August 2012.
93. Nguyen, N. (2010). Optimal control modification adaptive law for time-scale separated sys-
tems. In American Control Conference, June 2010.
94. Stepanyan, V., & Nguyen, N. (2009). Control of systems with slow actuators using time scale
separation. In AIAA Guidance, Navigation, and Control Conference, AIAA-2009-6272, August
2009.
95. Nguyen, N., & Stepanyan, V. (2010). Flight-propulsion response requirements for directional
stability and control. In AIAA Infotech@Aerospace Conference, AIAA-2010-3471, April 2010.
96. Tuzcu, I., & Nguyen, N. (2009). Aeroelastic modeling and adaptive control of generic transport
model. In AIAA Atmospheric Flight Mechanics, AIAA-2010-7503, August 2010.
97. Tuzcu, I., & Nguyen, N. (2010). Modeling and control of generic transport model. In 51st
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference,
AIAA-2010-2622, April 2010.
98. Yucelen, T., Kim, K., Calise, A., & Nguyen, N. (2011). Derivative-free output feedback
adaptive control of an aeroelastic generic transport model. In AIAA Guidance, Navigation,
and Control Conference, AIAA-2011-6454, August 2011.
99. Nguyen, N., & Jacklin, S. (2010). Stability, convergence, and verification and validation
challenges of neural net adaptive flight control. In Applications of Neural Networks in High
Assurance Systems Studies in Computational Intelligence (Vol. 268, pp. 77–110). Berlin:
Springer.
100. Nguyen, N., & Jacklin, S. (2007). Neural net adaptive flight control stability, verification
and validation challenges, and future research. In International Joint Conference on Neural
Networks, August 2007.
101. Nguyen, N., Ishihara, A., Stepanyan, V., & Boskovic, J. (2009). Optimal control modification
for robust adaptation of singularly perturbed systems with slow actuators. In AIAA Guidance,
Navigation, and Control Conference, AIAA-2009-5615, August 2009.
102. Gilbreath, G. P. (2001). Prediction of Pilot-Induced Oscillations (PIO) due to Actuator Rate
Limiting Using the Open-Loop Onset Point (OLOP) Criterion. In M.S. Thesis, Air Force
Institute of Technology, Wright-Patterson Air Force Base, Ohio, 2001.
103. Schumann, J., & Liu, Y. (2007). Tools and methods for the verification and validation of
adaptive aircraft control systems. In IEEE Aerospace Conference, March 2007.
104. Jacklin, S. (2009). Closing the certification gaps in adaptive flight control software. In AIAA
Guidance, Navigation, and Control Conference, AIAA-2008-6988, August 2008.
105. Jacklin, S. A., Schumann, J. M., Gupta, P. P., Richard, R., Guenther, K., & Soares, F. (2005).
Development of advanced verification and validation procedures and tools for the certification
of learning systems in aerospace applications. In AIAA Infotech@aerospace Conference,
September 2005.
106. Jacklin, S. A. (2015). Survey of verification and validation techniques for small satellite
software development. In Space Tech Expo Conference, May 2015.
References 15

107. Crespo, L. G., Kenny, S. P., & Giesy, D. P. (2008). Figures of merit for control verification.
In AIAA Guidance, Navigation, and Control Conference, AIAA 2008-6339, August 2008.
108. Hodel, A. S., Whorton, M., & Zhu, J. J. (2008). Stability metrics for simulation and flight-
software assessment and monitoring of adaptive control assist compensators. In AIAA Guid-
ance, Navigation, and Control Conference, AIAA 2008-7005, August 2008.
109. Ishihara, A., Ben-Menahem, S., & Nguyen, N. (2009). Time delay margin computation via
the Razumikhin method for an adaptive control system. In AIAA Guidance, Navigation, and
Control Conference, AIAA-2009-5969, August 2009.
110. Ishihara, A., Ben-Menahem, S., & Nguyen, N. (2009). Time delay margin estimation for direct
adaptive control of a linear system. In IASTED International Conference on Identification,
Control and Applications, August 2009.
111. Ishihara, A., Nguyen, N., & Stepanyan, V. (2010). Time delay margin estimation for adaptive
outer- loop longitudinal aircraft control. In AIAA Infotech@Aerospace Conference, AIAA-
2010-3455, April 2010.
112. Nguyen, N., & Summers, E. (2011). On time delay margin estimation for adaptive control and
robust modification adaptive laws. In AIAA Guidance, Navigation, and Control Conference,
AIAA-2011-6438, August 2011.
113. Nguyen, N., & Boskovic, J. (2008). Bounded linear stability margin analysis of nonlinear
hybrid adaptive control. In American Control Conference, June 2008.
114. Rajagopal, K., Balakrishnan, S. N., Nguyen, N., & Krishnakumar, K. (2013). Time delay mar-
gin analysis of modified state observer based adaptive control. In AIAA Guidance, Navigation,
and Control Conference, AIAA-2013-4755, August 2013.
115. Stepanyan, V., Krishnakumar, K., Nguyen, N., & Van Eykeren, L. (2009). Stability and per-
formance metrics for adaptive flight control. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2009-5965, August 2009.
116. Belcastro, C., & Belcastro, C. (2003). On the validation of safety critical aircraft systems, part I:
Analytical and simulation methods. In AIAA Guidance, Navigation, and Control Conference,
AIAA-2003-5559, August 2003.
117. Belcastro, C., & Belcastro, C. (2003). On the validation of safety critical aircraft systems,
Part II: Analytical and simulation methods. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2003-5560, August 2003.
118. Duke, E. L., Brumbaugh, R. W., & Disbrow, D. (1989). A rapid prototyping facility for flight
research in advanced systems concepts. In IEEE Computer, May 1989.
119. Matsutani, M., Jang, J., Annaswamy, A., Crespo, L. G., Kenny, S. P. (2008). An adaptive
control technology for safety of a GTM-like aircraft. In NASA CR-2008-1, December 2008.
120. Annaswamy, A., Jang, J., & Lavretsky, E. (2008). Stability margins for adaptive controllers
in the presence of time-delay. In AIAA Guidance, Navigation, and Control Conference, AIAA
2008-6659, August 2008.
121. Li, D., Patel, V. V., Cao, C., Hovakimyan, N., & Wise, K. (2007). Optimization of the time-
delay margin of L1 adaptive controller via the design of the underlying filter. In AIAA Guid-
ance, Navigation, and Control Conference, AIAA 2007-6646, August 2007.
Chapter 2
Nonlinear Systems

Abstract A brief overview of nonlinear systems is presented in this chapter. Non-

linear systems are inherently more complex to study than linear systems. Nonlinear
systems possess many complex behaviors that are not observed in linear systems.
Multiple equilibrium points, limit cycle, finite escape time, and chaos are illustrative
of some of the complex behaviors of nonlinear systems. Global stability of a nonlin-
ear system over its entire solution domain is difficult to analyze. Linearization can
provide information on the local stability of a region about an equilibrium point. The
phase plane analysis of a nonlinear system is related to that of its linearized systems
because the local behaviors of the nonlinear system can be approximated by the
behaviors of its linearized systems in the vicinity of the equilibrium points. Because
nonlinear systems can have multiple equilibrium points, one important fact to note
is that the trajectories of a nonlinear system can exhibit unpredictable behaviors.

Nonlinear systems are abundant in nature. In fact, most real-world systems are inher-
ently nonlinear. Linear systems in some regards are viewed as idealization of non-
linear systems in subspaces of nonlinear solutions. Nonlinear systems possess many
complex behaviors that are not observed in linear systems. Multiple equilibrium
points, limit cycle, finite escape time are illustrative of some of the complex behav-
iors of nonlinear systems. The learning objectives of this chapter are the following:
• To develop an understanding of equilibrium concepts and linearization; and
• To be able to assess local stability of equilibrium points.

Consider a typical control block diagram as shown in Fig. 2.1.

Fig. 2.1 Typical control block diagram

© Springer International Publishing AG 2018 17

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_2
18 2 Nonlinear Systems

The objective of a controller is to enable the output of a system, called a plant, to

track or follow a command input. A system generally can be described by a nonlinear
plant
ẋ = f (x, u, t) (2.1)
 
subject to an initial condition x (t0 ) = x0 , where x (t) = x1 (t) x2 (t) . . . xn (t) ∈
 
Rn is called a state vector, u (t) = u 1 (t) u 2 (t) . . . u m (t) ∈ Rm is a control
 
vector, t ∈ R+ is time, and f () = f 1 () f 2 () . . . f n () ∈ Rn is a nonlinear
transition function.
The output, or response, of the system can generally be written as

y = h (x, u, t) (2.2)
 
where y (t) = y1 (t) y2 (t) . . . yl (t) ∈ Rl is the system output and h () =
 
h 1 () h 2 () . . . h l () ∈ Rl is a nonlinear output function.
Equations (2.1) and (2.2) together form the system state-space representation.
When m = l = 1, the system is said to be a single-input, single-output (SISO) sys-
tem. When m = 1 and l = 1, the system is said to be a multiple-input, multiple-output
(MIMO) system. The system can also be single-input, multiple-output (SIMO); or
multiple-input, single-output (MISO).
When the system is explicitly dependent on time t as in Eqs. (2.1) and (2.2), the
system is said to be non-autonomous or time-varying. When the explicit dependency
on time is not present, then the system is said to be autonomous or time-invariant.
A special case of nonlinear systems is a linear affine-in-control form which is
given by
ẋ = f (x, t) + g (x, t) u (2.3)

y = h (x, t) (2.4)

In contrast to nonlinear systems, a linear time-varying (LTV) state-space system

is expressed by
ẋ = A (t) x + B (t) u (2.5)

y = C (t) x + D (t) u (2.6)

where the matrices A (t), B (t), C (t), and D (t) are functions of t.
When these matrices are constant matrices, then the following system is linear
time-invariant (LTI):
ẋ = Ax + Bu (2.7)

y = C x + Du (2.8)
2 Nonlinear Systems 19

LTI systems are much easier to analyze since there are a wealth of design and
analysis tools for such systems. As a result, many control system designs are still
based on a LTI representation.
The general solution for a LTI state-space system is given by
 t
A(t−t0 )
x =e x0 + e A(t−τ ) Bu (τ ) dτ (2.9)
t0

 t
y = Ce A(t−t0 ) x0 + C e A(t−τ ) Bu (τ ) dτ + Du (t) (2.10)
t0

where e At is a matrix exponential function which is defined by the inverse Laplace

transform  
e At = L −1 (s I − A)−1 (2.11)

Example 2.1:
• The following second-order spring-mass-damper system

m ẍ + c ẋ + k (x) x = bu

where m, c, and b are constant, k (x) is a nonlinear spring function, and u (x, ẋ) =
f (x, ẋ) is a nonlinear state feedback control is classified as an autonomous and
nonlinear time-invariant system.
• The following spring-mass-damper system

m ẍ + c ẋ + k (x) x = bu

where u (x, ẋ, t) = f (x, ẋ, t) is a nonlinear state feedback command-following

control is classified as an non-autonomous and nonlinear time-varying system.
• The following spring-mass-damper system

m ẍ + c ẋ + k (t) x = bu

where k (t) is a time-varying spring function is classified as a non-autonomous

and nonlinear time-varying system if u (x, ẋ, t) is a nonlinear state feedback
command-following control or as a non-autonomous and linear time-varying sys-
tem if u (x, ẋ, t) is a linear state feedback command-following control.

In the block diagram in Fig. 2.1, the controller block represents a feedback control
action such as a proportional-integral-derivative (PID) control. The output of the
controller is a command to an actuator or servo system that actually controls the
plant. A typical actuator system employed in many control systems is a servomotor
that translates an actuator command signal into either a rotational or translational
motion. The actuator system can be a source of nonlinearity due to such behaviors as
20 2 Nonlinear Systems

amplitude saturation and rate limiting when the actuator command exceeds certain
limits. These nonlinear sources can make even a LTI control system to become
nonlinear. Generally, actuator saturation should be avoided in order for a linear
system to retain its linearity. Thus, actuators are sized appropriately to provide a
bandwidth much larger than the control bandwidth of the system and the control
gains are designed so as not to result in an actuator amplitude saturation or rate
limiting. When properly designed, the actuator command will follow closely the
actuator output. In an ideal case, the effect of the actuator system may be neglected
during an initial design process.
Example 2.2: Consider a second-order plant

X (s) b
G (s) = = 2
U (s) s + 2ζ ωn s + ωn2

whose actuator is subject to amplitude saturation and rate limiting described by

⎧
⎪
⎨u min u c (t) < u min
u (t) = sat (u c (t)) = u c (t) u min ≤ u c (t) ≤ u max
⎪
⎩
u max u c (t) > u max
⎧
⎪
⎨u (t − Δt) + λmin Δt u̇ c (t) < λmin
u (t) = rate (u c (t)) = u c (t) λmin ≤ u̇ c (t) ≤ λmax
⎪
⎩
u (t − Δt) + λmax Δt u̇ c (t) >≤ λmax

where sat () is called a saturation function, rate () is called a rate limiter function,
and u̇ c (t) is the actuator slew rate which is computed as

u c (t) − u c (t − Δt)
u̇ c (t) =
Δt
The saturation and rate limiter functions are nonlinear discontinuous functions
since the input does not map linearly with the output when the amplitude or rate
limit is exceeded.
The actuator command u c (t) is the output of the feedback controller. In this
example, the goal is to enable x (t) to track a sinusoidal input

r (t) = a sin ωt

The controller is designed to be a proportional-integral-derivative type

ki
Uc (s) = k p + + kd s E (s)
s

where E (s) = R (s) − X (s).

2 Nonlinear Systems 21

Note that the system closed-loop transfer function for an ideal case with no actu-
ator amplitude saturation is

X (s) bkd s 2 + bk p s + bki

= 3  
R (s) s + (2ζ ωn + bkd ) s 2 + ωn2 + bk p s + bki

Even though the controller is linear, the response with the actuator saturation
becomes quite nonlinear.
  2.2 shows the closed-loop response for G (s) =
Figure
5
s 2 +5s+6
, U (s) = 15 + 8
s
+ 5s E (s) with the amplitude limit at ±1 for r (t) = sin t.

0.8

0.6

0.4

0.2
r, x

−0.2

−0.4

−0.6

−0.8
r x
−1
0 2 4 6 8 10
t

Fig. 2.2 Response due to amplitude saturation

2.1 Equilibrium and Linearization

All nonlinear autonomous systems possess a set of equilibrium points about which
the systems exhibit certain stability properties. Denoting x ∗ as an equilibrium point,
then for an autonomous system, the equilibrium points are the real roots of
 
f x∗ = 0 (2.12)

A system that is initially at an equilibrium point will remain there at all times
since an equilibrium is a constant solution of an autonomous nonlinear system.
Unlike linear systems which can have only one equilibrium point, nonlinear sys-
tems can have multiple isolated equilibrium points. Under certain conditions, the
behavior of a nonlinear system can be studied by linearization which can reveal
some important local stability properties of the system at these equilibrium points.
Linearization generally cannot predict the behavior of a nonlinear system far from
an equilibrium point about which the linearization is performed.
22 2 Nonlinear Systems

Consider an autonomous system

ẋ = f (x) (2.13)

The solution can be expressed as

x = x ∗ + x̃ (2.14)

where x̃ (t) is a small perturbation in the solution of x (t) about the equilibrium x ∗ .
Then, Eq. (2.13) can be written as
 
ẋ ∗ + x̃˙ = f x ∗ + x̃ (2.15)

But, by the definition of an equilibrium point, ẋ ∗ = 0, and applying the Taylor

series expansion yields the linearization of the nonlinear system (2.13) about the
equilibrium point x ∗ as
       
x̃˙ = f x ∗ + x̃ = f x ∗ + J x ∗ x̃ + · · · ≈ J x ∗ x̃ (2.16)

where J (x ∗ ) is the Jacobian matrix of f (x) evaluated at the equilibrium point x ∗

computed by
⎡ ∂ f1 ∂ f1 ⎤
∂ x1 ∂ x2
· · · ∂∂xf1n
⎢ ∂ f2 ∂ f2 ⎥
 ∗ ∂ f ⎢ ∂ x ∂ x · · · ∂∂ xf2 ⎥
⎢ 1 2 2 ⎥
J x = =⎢ ⎥ (2.17)
∂ x x=x ∗ ⎢ ... ... . . . ... ⎥
⎣ ⎦
∂ fn ∂ fn
∂ x1 ∂ x2
· · · ∂∂ xfnn ∗ x=x

Example 2.3 Consider a rotating pendulum without friction in Fig. 2.3. The equa-
tion of motion is given by
g
θ̈ + sin θ − ω2 sin θ cos θ = 0
l

This can be cast in a state-space form by setting x1 (t) = θ (t) and x2 (t) = θ̇ (t) as
   
ẋ1 x2
=
ẋ2 − gl sin x1 + ω2 sin x1 cos x1

The equilibrium points can be found by setting ẋ1 (t) = 0 and ẋ2 (t) = 0. This gives
the following solution:
2.1 Equilibrium and Linearization 23
⎧   
⎨cos−1 lωg 2 , ω ≥ gl
x1∗ = 
⎩0, π , ω < gl

x2∗ = 0

Fig. 2.3 Rotating pendulum

Physically, this means that one of the equilibrium points occurs when the angular
speed ω exceeds a certain value and the pendulum will be suspended by an angle
as the centrifugal force exerted on the pendulum is in balance with its weight. The
other two equilibrium points occur when the angular speed ω is low enough and the
pendulum is at either the bottom or the top in the vertical plane. Thus, this nonlinear
system has three equilibrium points.
The equation of motion can be linearized by computing its Jacobian matrix as
 
J= 0 1
− gl cos x1 + ω2 cos2 x1 − sin2 x1 0

Evaluating the Jacobian matrix at the equilibrium points yields

  g    0 1

−1
J cos , 0 = g2
lω2 l 2 ω2
− ω2 0
 
0 1
J (0, 0) =
− gl + ω2 0
 
0 1
J (π, 0) = g
l
+ ω2 0
24 2 Nonlinear Systems

The solutions of both

 the nonlinear and linearized equations of motion are in good
agreement for ω < gl as shown in Fig. 2.4. At a higher value of θ (t), clearly the
accuracy of the linearized solution tends to suffer as shown in Fig. 2.4. This is one of
the drawbacks of the linearization in that the small perturbation assumption typically
breaks down when the solution is not close enough to an equilibrium. The validity
of a linearized solution therefore must be verified to ensure that the predicted linear
behavior is a good approximation of the nonlinear behavior in the proximity to an
equilibrium point.

40
0.5 Nonlinear EOM
ω = 0.8(g/l)
20 Linear EOM
θ, deg

−20

−40
0 5 10 15 20
t, sec

80
Nonlinear EOM
ω = 1.2(g/l)0.5
Linear EOM
60
θ, deg

40

20
0 5 10 15 20
t, sec

g
Fig. 2.4 Comparison between nonlinear and linear solutions, l =1

2.2 Local Stability and Phase Plane Analysis

Global stability of a nonlinear system over its entire solution domain is difficult to
analyze. Linearization can provide information on the local stability in a region about
an equilibrium point. For a LTI system, there is only one equilibrium point which
is the origin. For a linearized system, the origin corresponds to an equilibrium point
of the nonlinear system about which the linearization is performed. Recall that a
LTI system is absolutely stable if the eigenvalues of the transition matrix A all have
negative real part. That is,

(λ (A)) < 0 (2.18)

A is then said to be a Hurwitz matrix. Notationally, it can also be written as

λ (A) ∈ C− where C− is a space of complex numbers having negative real part.
Phase portraits are plots of trajectories of the solution which can be useful for
studying the behaviors of second-order nonlinear systems. Phase portraits can also
be used to study the behaviors of linearized systems from which local stability of
the nonlinear systems can be learned. For a second-order LTI system, there are two
2.2 Local Stability and Phase Plane Analysis 25

eigenvalues, denoted by λ1 and λ2 . The trajectories in the vicinity of the equilibrium

point can exhibit different characteristics depending on the values of λ1 and λ2 .
1. Stable or unstable node occurs when both λ1 and λ2 are real and have the same
sign. When λ1 < 0 and λ2 < 0, the node is a stable node and all trajectories
converge to the node. When λ1 > 0 and λ2 > 0, the node is an unstable node
and the trajectories diverge from the node. The behaviors of a stable node and an
unstable node are illustrated in Fig. 2.5.

Unstable Node Stable Node

x x
2 2

x1 x1

Fig. 2.5 Phase portrait of a node

2. Saddle point occurs when both λ1 and λ2 are real and have opposite signs. Because
one of the eigenvalue corresponds to an unstable pole, half of the trajectories
diverge from the center of a saddle point as shown in Fig. 2.6. The system is
always on the verge of instability.

Saddle Point
x2

x
1

Fig. 2.6 Phase portrait of a saddle point

3. Stable or unstable focus occurs when both λ1 and λ2 are a complex conjugate
pair. When Re (λ1 ) < 0 and Re (λ2 ) < 0, the focus is a stable focus and the
trajectories spiral toward the focus as illustrated in Fig. 2.7. When Re (λ1 ) > 0
and Re (λ2 ) > 0, the focus is an unstable focus and the trajectories spiral outward
from the focus as shown in Fig. 2.7.
26 2 Nonlinear Systems

Unstable Focus Stable Focus

x x
2 2

x x
1 1

Fig. 2.7 Phase portrait of a focus

4. Center occurs when both λ1 and λ2 are purely imaginary. All trajectories encircle
the center point at the origin with concentric-level curves as illustrated in Fig. 2.8.

Center
x2

x
1

Fig. 2.8 Phase portrait of a center

The phase plane analysis of a nonlinear system is related to that of its linearized
system because the local behaviors of the nonlinear system can be approximated
by the behaviors of its linearized systems in the vicinity of the equilibrium points.
Because a nonlinear system can have multiple equilibrium points, one important fact
to note is that the trajectories of the nonlinear solution can exhibit unpredictable
behaviors. For example, if the solution starts with an initial condition near a stable
equilibrium point, it does not necessarily guarantee that the ensuing trajectories will
converge to a stable node or focus if there exists an unstable equilibrium point in the
solution set.
Example 2.4: From Example 2.3, the eigenvalues of the linearized systems are

   g  g2
∗ −1
λ1,2 J x1 = cos = ±i ω 2−
lω2 l 2 ω2

  ∗  g
λ1,2 J x1 = 0 = ±i − ω2
l
2.2 Local Stability and Phase Plane Analysis 27

  ∗  g
λ1,2 J x1 = π = ± + ω2
l

The first two equilibrium points are centers, and the last equilibrium point is a
saddle point. The phase portrait is shown in Fig. 2.9.

1
2

0
x

−1

−2

−3

−4
−4 −3 −2 −1 0 1 2 3 4
x
1


g g
Fig. 2.9 Phase portrait of θ̈ + l sin θ − ω2 sin θ cos θ = 0, l = 1, ω = 0.8 gl

2.3 Other Nonlinear Behaviors

Nonlinear systems differ from linear systems in that they can exhibit many complex
behaviors [1]. One feature which has been discussed is the existence of multiple
isolated equilibrium points. There are many other complex behaviors that are unique
to nonlinear systems, such as finite escape time, limit cycle, chaos.
1. Finite escape time: An unstable linear system becomes unbounded as time
approaches to infinity. For nonlinear systems, an interesting phenomenon exists.
It is possible for a nonlinear system to become unbounded in a finite interval of
time. The solution then exhibits a finite escape time.

Example 2.5: The nonlinear system

ẋ = x 2

subject to x (0) = 1 has a solution

1
x (t) = −
t −1

which is defined only for t ∈ [0, 1). The solution has a finite escape time at t = 1.
28 2 Nonlinear Systems

2. Limit cycle: For some nonlinear systems, a limit cycle is a periodic nonlinear
solution represented by a closed trajectory in the phase plane such that all tra-
jectories in its vicinity either converge to it or diverge from it. A limit cycle can
be classified as stable, unstable, or neutrally stable depending on the behaviors
of the trajectories of the solution in its vicinity. A stable limit cycle and a stable
equilibrium point are the only two types of “regular” attractors.
Example 2.6: The Van der Pol oscillator described by
 
ẍ − μ 1 − x 2 ẋ + x = 0

has a stable limit cycle as shown in Fig. 2.10.

1
2

0
x

−1

−2

−3

−4

−5
−5 0 5
x1

Fig. 2.10 Stable limit cycle of Van der Pol oscillator, μ = 0.5

2.4 Summary

Nonlinear systems can have multiple isolated equilibrium points. Under certain con-
ditions, the behavior of a nonlinear system can be studied by linearization which
can reveal some important local stability properties of the nonlinear system at these
equilibrium points. Linearization generally cannot predict the behavior of a nonlin-
ear system far from an equilibrium point about which the linearization is performed.
Nonlinear systems differ from linear systems in that they can exhibit many complex
behaviors, such as finite escape time, limit cycle, chaos. Global stability of a nonlin-
ear system over its entire solution domain is difficult to analyze. Linearization can
provide information on the local stability in a region about an equilibrium point. The
phase plane analysis of a nonlinear system is related to that of its linearized system
because the local behaviors of the nonlinear system can be approximated by the
behaviors of its linearized systems in the vicinity of the equilibrium points. Because
a nonlinear system can have multiple equilibrium points, one important fact to note
is that the trajectories of the nonlinear system can exhibit unpredictable behaviors.
2.5 Exercises 29

2.5 Exercises

1. Consider the following PID control system with the actuator amplitude saturation
and rate limiting:

5
G (s) =
s 2 + 5s + 6

with k p = 15, ki = 8, and kd = 5 as the control gains. The actuator has both
amplitude and rate limits between −1 and 1.
a. Compute the characteristic roots of the ideal closed-loop system without con-
sideration for actuator amplitude saturation and rate limiting.
b. Construct a Simulink model for a sinusoidal input r (t) = sin t. Plot the input,
the ideal output without the actuator amplitude saturation and rate limiting,
and the actual output for a simulation time t = 10 sec. Also plot the actuator
command signal u c (t) and the control input signal to the plant u (t).
c. Comment on the effect of rate limiting.
2. Given
θ̈ + cθ̇ + 2 sin θ − 1 = 0

a. Find all the equilibrium points of the system for −π ≤ θ (t) ≤ π .

b. Linearize the system and compute the eigenvalues about all the equilibrium
points.
c. Classify the types of the equilibrium points on a phase plane and plot the phase
portrait of the nonlinear system.
3. Repeat Exercise 2 for
   
ẋ1 −x1 + x1 x2
=
ẋ2 x2 − x1 x2

4. Analytically determine the solution of the following nonlinear system:

ẋ = |x| x 2

with a general initial condition x (0) = x0 .

a. Let x0 = 1. Does the solution have a finite escape time? If so, determine it.
b. Repeat part (a) with x0 = −1.
c. Comment on the effect of initial condition on the stability of the system.
30 2 Nonlinear Systems

Reference

1. Khalil, H. K. (2001). Nonlinear systems. Upper Saddle River: Prentice-Hall.

Chapter 3
Mathematical Preliminaries

Abstract This chapter presents some basic mathematical fundamentals for adaptive
control theory. Vector and matrix norms are defined. The existence and uniqueness of
a solution of a nonlinear differential equation are stated by the Cauchy theorem and
the Lipschitz condition. Positive-valued functions are an important class of functions
in adaptive control theory. Positive definiteness of a real-valued function is defined.
The properties of a positive-definite matrix are given.

To be able to develop basic working knowledge of adaptive control, some mathe-

matical fundamental concepts need to be understood. In this chapter, the learning
objectives are to develop a basic understanding of:

• Norms as metric measures of vectors and matrices;

• Existence and uniqueness of ordinary differential equations in connection with the
Lipschitz condition; and
• Positive-definite functions which are an important class of functions that are fre-
quently used to establish stability of adaptive control.

3.1 Vector and Matrix Norms

3.1.1 Vector Norms

 
The state vector x = x1 x2 . . . xn belongs to a set of real numbers in an n-
dimensional Euclidean space Rn . The Euclidean space is a metric space endowed
with a notion of “distance” measure or norm. A real-valued quantity x defined on
R is said to be a norm of x if it satisfies the following conditions:

1. Positivity
x ≥ 0 ∀x ∈ Rn (3.1)

© Springer International Publishing AG 2018 31

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_3
32 3 Mathematical Preliminaries

2. Positive definiteness
x = 0 (3.2)

if and only if x = 0.
3. Homogeneity
αx = |α| x (3.3)

for any scalar α.

4. Triangle inequality
x + y ≤ x + y (3.4)


x is said to be a semi-norm if it satisfies all the above conditions except the positive
definiteness condition.
A p-norm of a vector x ∈ Cn , where Cn is a complex number space that general-
izes Rn , is defined as  n 1/ p

x p = |xi | p
(3.5)
i=1

for p = 1, 2, . . . , ∞.
In special cases when p = 1, 2, ∞, the 1-, 2-, and infinity norms of x are defined
as
n
x1 = |xi | (3.6)
i=1


 n

x2 = |xi |2 (3.7)
i=1

x∞ = max |xi | (3.8)

1≤i≤n

The 2-norm is also called the Euclidean norm which is a “distance” measure of
the vector x from the origin.
The inner product in the Euclidean space Rn is defined as


n
x, y
= xi yi (3.9)
i=1

In R2 and R3 spaces, the inner product is also known as the dot product. The inner
product is also a measure in the Euclidean space since

x, x
= x22 (3.10)
3.1 Vector and Matrix Norms 33

A Euclidean space with a dot product is called an inner product space.

Let x and y be any two vectors, then the Cauchy-Schwartz inequality for the inner
product is given by
x, y
≤ x2 y2 (3.11)

Example 3.1 Verify that x2 , where x ∈ Rn , satisfies the norm conditions.
Firstly, x2 can be expressed as

x2 = x12 + x22 + · · · + xn2

It is obvious that x2 ≥ 0 and x2 = 0 if and only if xi = 0 ∀i = 1, 2, . . . , n.

Thus, x2 satisfies the positivity and positive definiteness conditions.
Since

αx2 = (αx1 )2 + (αx2 )2 + · · · + (αxn )2 = |α| x12 + x22 + · · · + xn2 = |α| x2

then x2 satisfies the homogeneity condition.

Let y ∈ Rn . Then,

x + y22 = (x1 + y1 )2 + (x2 + y2 )2 + · · · + (xn + yn )2

= x12 + x22 + · · · + xn2 + 2x1 y1 + 2x2 y2 + · · · + 2xn yn + y12 + y22 + · · · + yn2
= x22 + 2 x, y
+ y22

Using the Cauchy-Schwartz inequality, one can write

 2
x + y22 ≤ x22 + 2 x2 y2 + y22 = x2 + y2

Taking the square root of both sides yields the triangle inequality

x + y2 ≤ x2 + y2

Thus, x2 satisfies all the norm conditions.



When x (t) is a function of t, then the L p norm is defined as

 1/ p
∞ 
n
x p = |x (t)| p dt (3.12)
0 i=1

provided the integral exists, in which case x ∈ L p , the space with the L p norm.
34 3 Mathematical Preliminaries

In particular, the L1 , L2 , and L∞ norms are defined as

∞ 
n
x1 = |xi (t)| dt (3.13)
t0 i=1



 ∞ 
n
x2 = |xi (t)|2 dt (3.14)
t0 i=1

x∞ = sup max |xi (t)| (3.15)

t≥t0 1≤i≤n

where the notation sup denotes the supremum over all t that yields the largest value
of the argument.

Example 3.2 Let  

2e−t
x=
−e−2t

for all t ≥ 0.

5
|x |+|x |
1 2
2 2
4.5 x1+x2

4 max(x1,x2)
∞
Integrand of ||x||1, ||x| , ||x||

3.5
2

2.5

1.5

0.5

0
0 1 2 3 4 5
t

Fig. 3.1 Integrands of L p norms

The integrands of the L1 , L2 , and L∞ norms of x (t) are shown in Fig. 3.1.
The L1 , L2 , and L∞ norms of x (t) can be evaluated analytically as
∞  −t  5
x1 = 2e + e−2t dt =
0 2
 
∞   9 3
x2 = 4e−2t + e−4t dt = =
0 4 2
3.1 Vector and Matrix Norms 35

x∞ = 2

Example 3.3 The function x (t) = sin t does not have a L1 or L2 norm because the
integral of sin t for t ∈ [0, ∞) does not exist. However, the L∞ norm of x (t) exists
and is equal to 1. Thus, x (t) ∈ L∞ .

3.1.2 Matrix Norms

Let A ∈ Cm × Cn be an m-by n complex-valued matrix, then the induced p-norm

of A is defined as
 Ax p
A p = sup (3.16)
x=0 x p

In special cases when p = 1, 2, ∞, then the 1-, 2-, and infinity norms of A are
defined as
m
 
A1 = max ai j  (3.17)
1≤ j≤n
i=1


A2 = λmax (A∗ A) (3.18)


n
 
A∞ = max ai j  (3.19)
1≤i≤m
j=1

where ai j is an element of the matrix A and A∗ is the complex conjugate transpose

of A. If A ∈ Rm × Rn is a real-valued matrix, then A∗ = A .
Another type of matrix norm is the Frobenius norm, which is not an induced norm
and is defined as


  m  n
 2
 A F = trace (A A) =
∗ ai j  (3.20)
i=1 j=1

where the trace operator is the sum of the diagonal elements of a square matrix, that
is,
n
trace (A) = aii (3.21)
i=1

The matrix norm has the following properties:

ρ (A) ≤ A (3.22)

36 3 Mathematical Preliminaries

A + B ≤ A + B (3.23)

AB ≤ A B (3.24)

where A and B are any matrices of appropriate dimensions and ρ is the spectral
radius operator of a square matrix which is defined as

ρ (A) = max |λi | (3.25)

1≤i≤n

where λi is the ith eigenvalue of A.

Example 3.4 The norms of the following matrix A are to be computed:

 
1+i 0
A=
2 −1

The 1-norm of A is computed as


2
  √ 
A1 = max ai j  = max (|1 + i| + 2, |−1|) = max 2 + 2, 1 = 3.4142
1≤ j≤2
i=1

The 2-norm of A is computed as

 
∗ 1−i 2
A =
0 −1
    
1−i 2 1+i 0 6 −2
A∗ A = =
0 −1 2 −1 −2 1
 
λ1,2 A∗ A = 0.2984, 6.7016

A2 = λmax (A∗ A) = 2.5887

The infinity norm of A is computed as


2
  √ 
 A∞ = max ai j  = max (|1 + i| , 2 + |−1|) = max 2, 3 = 3
1≤i≤2
j=1

The Frobenius norm of A is computed as

 √
A F = trace (A∗ A) = 6 + 1 = 2.6458
3.1 Vector and Matrix Norms 37

The spectral radius of A is computed as

λ1,2 (A) = −1, 1 + i

√
ρ (A) = max (|−1| , |1 + i|) = 2 = 1.4142

Thus,
ρ (A) ≤ A

3.2 Compact Set

A compact set is a mathematical representation of a region that contains a collection

of mathematical objects. A compact set in a Euclidean space Rn is a closed and
bounded set. In studying the Lyapunov stability theory, it is convenient to use the
notion of a compact set to represent the collection of all the trajectories of a closed-
loop adaptive control system.
Example 3.5 The set S ⊂ R2 given by
 
S = x ∈ R2 : |x1 | ≤ a, |x2 | ≤ b; a > 0, b > 0

is a compact set that represents a closed rectangular region in R2 as shown in Fig. 3.2.

Fig. 3.2 A compact set

On the other hand, the complementary set S c such that S ∪ S c = R2 (the union
of S and S c is R2 )
 
S c = x ∈ R2 : |x1 | > a, |x2 | > b

is an open set and therefore is not a compact set.

38 3 Mathematical Preliminaries

3.3 Existence and Uniqueness

A nonlinear differential equation may or may not have a unique solution that is
dependent continuously on an initial condition. Certain requirements are imposed
on a nonlinear differential equation to ensure the existence of a unique solution. The
existence and uniqueness are fundamental requirements for any nonlinear differential
equations.

3.3.1 Cauchy Theorem

The existence of a solution of a nonlinear differential equation can be stated by the

Cauchy theorem as follows:
Theorem 3.1 Given
ẋ = f (x, t) (3.26)

with an initial condition x (t0 ) = x0 , if f (x, t) is at least piecewise continuous in

the closed region such that

|t − t0 | ≤ T, x − x0  ≤ R (3.27)

where T > 0 and R > 0 are some positive constants, then there exists t1 > t0 such
that at least one solution of Eq. (3.26) exists and is continuous over the time interval
[t0 , t1 ].

Geometrically, this simply means that continuity of f (x, t) in a closed region
that contains the initial condition ensures that there exists at least one continuous
solution that lies
 therein.
 This is illustrated in Fig. 3.3. The gradient of a continuous
solution of x t; t0, x0 starting from an initial condition x (t0 ) = x0 is the value of
the function f (x, t) at any point on the trajectory of x (t). For the gradient to exist,
f (x, t) must be a continuous function.

Fig. 3.3 Continuity of f (x, t) in closed region

3.3 Existence and Uniqueness 39

Example 3.6 Consider the differential equation

1
ẋ =
x −1

with x (0) = 1. It is obvious that f (x) is discontinuous for x = 1 since f (x) is

defined for x (t) > 1 and x (t) < 1 for all t. So there is no solution for the equation
with this initial condition. Now, suppose x (0) = −1. Then, f (x) is defined in a
region x (t) < 1 and t ≥ 0 which includes x0 = −1. Therefore, a solution exists for
all x (t) ≤ −1 and t ≥ 0.
Note that the existence of a solution does not imply that the solution is unique.
The requirement of uniqueness is imposed by the Lipschitz condition which is stated
in the following theorem.

3.3.2 Global Lipschitz Condition

The uniqueness of a solution of a nonlinear differential equation can be stated by the

following theorem:
Theorem 3.2 If f (x, t) is piecewise continuous in t and there exists a positive
constant L such that

 f (x2 , t) − f (x1 , t) ≤ L x2 − x1  (3.28)

for all x1 , x2 ∈ Rn and t ∈ [t0 , t1 ], then Eq. (3.26) has a unique solution for all
t ∈ [t0 , t1 ].

Inequality (3.28) is called the Lipschitz condition and the constant L is called a
Lipschitz constant. When x1 and x2 are any two points in the Euclidean space Rn ,
then f (x, t) is said to be globally Lipschitz. Geometrically, the Lipschitz condition
essentially imposes the continuity and boundedness of the partial derivative or the
Jacobian of f (x, t) with respect to x (t) as illustrated in Fig. 3.4.

Fig. 3.4 Lipschitz condition

40 3 Mathematical Preliminaries

Example 3.7 The differential equation

2
ẋ = 3x 3

subject to x (0) = 0, not only has a solution x (t) = t 3 but also has a zero solution
x (t) = 0. Thus, this equation does not have a unique solution.
2
The function f (x) = 3x 3 does not satisfy the global Lipschitz condition since
the derivative of f (x)

f (x) = 2x − 3
1

is not bounded at x = 0.

Example 3.8

• The function f (x) = x is globally Lipschitz since it has continuous and bounded
derivative for all x (t) ∈ R. The Lipschitz constant L is equal to 1.
• The function f (x) = sin x is globally Lipschitz since its derivative is continuous
and bounded with a Lipschitz constant L = 1 for all x (t) ∈ R.

• The function f (x) = x 2 is not globally Lipschitz since its derivative, f (x) = 2x,
is unbounded for x (t) ∈ R as x (t) → ±∞.

Note that the global Lipschitz condition is very restrictive since a function with this
condition must have a bounded derivative for all values of x (t) ∈ Rn as x (t) → ±∞.
Many nonlinear physical models fail to satisfy this condition since physically there
are usually upper and lower limits placed on variables of most if not all physical
models. These limits thus does not require x (t) to be in a domain that occupies
the entire Euclidean space Rn , but rather in a subset of it. This leads to a notion of
local Lipschitz condition, which is less restrictive and can be easily satisfied by any
nonlinear differential equations.

3.3.3 Local Lipschitz Condition

The uniqueness of a solution of a nonlinear differential equation in a finite neighbor-

hood can be stated by the following theorem:
Theorem 3.3 If f (x, t) is piecewise continuous in t and there exists a positive
constant L such that

 f (x2 , t) − f (x1 , t) ≤ L x2 − x1  (3.29)

3.3 Existence and Uniqueness 41

for some x1 and x2 in a finite neighborhood of x0 and t ∈ [t0 , t1 ], then Eq. (3.26) has
a unique solution in a finite time interval t ∈ [t0 , t0 + δ], where t0 + δ < t1 .

The local Lipschitz condition now only imposes the continuity and boundedness
of the partial derivative of f (x, t) with respect to x (t) over a finite neighborhood of
x0 as a subset of Rn instead of over all x (t) ∈ Rn . Therefore, it is a less restrictive
condition that can easily be met. In most cases, the local Lipschitz condition is a
reasonable assumption for many continuous nonlinear functions.
Example 3.9 The differential equation
2
ẋ = 3x 3

subject to x (0) = 1 has a unique solution x (t) = (t + 1)3 for a finite neighborhood
of x0 , for example, x (t) ∈ [1, 2]. f (x) satisfies the local Lipschitz condition since
the derivative of f (x) is now continuous and bounded for all x (t) ∈ [1, 2].

Example 3.10 The function f (x) = x 2 is locally Lipschitz for any finite neighbor-
hood of x (t) ∈ [x1 , x2 ], where x1 and x2 are finite-valued.

3.4 Positive Definite, Symmetric and Anti-Symmetric

Matrices

3.4.1 Positive-Definite Matrix and Function

The Lyapunov stability theory for nonlinear systems often invokes a class of functions
called positive-definite functions. One such function is a quadratic function

V (x) = x  P x (3.30)

where V (x) ∈ R is a scalar function, x ∈ Rn is a vector, and P ∈ Rn × Rn is an

n-by n matrix.
Definition 3.1 The quadratic scalar function V (x) is said to be positive definite if
V (x) > 0 or positive semi-definite if V (x) ≥ 0 for all x = 0. Conversely, V (x) is
said to be negative definite if V (x) < 0 and negative semi-definite if V (x) ≤ 0 for
all x = 0.

Definition 3.2 The matrix P is said to be a positive-definite matrix, denoted as

P > 0, if there exists a function V (x) such that

V (x) = x  P x > 0 (3.31)

42 3 Mathematical Preliminaries

Thus, P > 0 implies V (x) > 0.


A positive (negative) (semi)-definite matrix is defined the same way as a positive
(negative) (semi-)definite function.
A positive-definite matrix P has the following properties:
• P is a symmetric matrix, that is,

P = P (3.32)

• All eigenvalues of P are real and positive, that is, λ (P) > 0.
• A scalar multiplication with a constant causes the product to have the same sign
definiteness as the constant, that is, α P > 0 if α > 0 and α P < 0 if α < 0.
• Let A be any arbitrary matrix of appropriate dimension, then

A P A > 0 (3.33)

• Let Q be a positive-definite matrix of the same dimension as P, then

P+Q>0 (3.34)

• The quadratic function V (x) is bounded from below and above by

λmin (P) x22 ≤ V (x) ≤ λmax (P) x22 (3.35)

which can be illustrated geometrically in Fig. 3.5, where λmin (.) and λmax (.)
denote the minimum and maximum eigenvalues of a matrix, respectively.
The above properties apply for a negative definite matrix by reversing the inequality
sign.

Fig. 3.5 Boundedness of positive-definite quadratic function

3.4 Positive Definite, Symmetric and Anti-Symmetric Matrices 43

Example 3.11
• The identity matrix I is a positive-definite matrix.
• The matrix  
53
P=
32

is positive definite since λ1.2 (P) = 0.1459, 6.8541 > 0.

• Let  
1 −1 0
A=
0 2 1

be an arbitrary matrix, then

 
12
A P A = >0
25
 
since λ1,2 A P A = 0.1716, 5.8284 > 0.
• The quadratic function V (x) is expressed as

V (x) = x  P x = 5x12 + 6x1 x2 + 2x22

which, at a first glance, does not seem to be positive definite. By completing the
square, it can be shown that V (x) > 0 as follows:
 2  2  
3 3 3 √ 
V (x) = − √ x1 + √ x1 + 2 √ x1
5x12 2x2 + 2x22
2 2 2
 2
1 3 √
= x12 + √ x1 + 2x2 > 0
2 2

3.4.2 Anti-Symmetric Matrix

Definition 3.3 A matrix Q ∈ Rn × Rn is said to be an anti-symmetric (or skew-

symmetric) matrix if
Q = −Q  (3.36)

The elements of Q are then

qi j = −q ji (3.37)

qii = 0 (3.38)
44 3 Mathematical Preliminaries

An anti-symmetric matrix Q has the following properties:

• The eigenvalues of Q are purely imaginary if the matrix dimension n is even, and
includes a zero if n is odd.
• det (Q) = 0 or Q is singular if n is odd.
• Let A be any arbitrary matrix of appropriate dimension, then the product A Q A
is also an anti-symmetric matrix.
• Let V (x) be a quadratic scalar function that is formed with Q. Then,

V (x) = x  Qx = 0 (3.39)

Example 3.12 The matrix ⎡ ⎤

0 −4 −1
Q = ⎣ 4 0 −1 ⎦
1 1 0

is an anti-symmetric matrix. Its eigenvalues are λ1.2,3 (Q) = 0, ±4.2426i. Q is sin-

gular since one of its eigenvalues is zero or equivalently det (Q) = 0.
The quadratic function given by

V (x) = x  Qx = 4x1 x2 + x1 x3 + x2 x3 − 4x1 x2 − x1 x3 − x2 x3

is identically zero. 

Any arbitrary square matrix A ∈ Rn × Rn can be decomposed into a symmetric

part and anti-symmetric part as follows:

A=M+N (3.40)

where M is called the symmetric part of A and N is called the anti-symmetric part
of A such that
1 
M= A + A (3.41)
2
1 
N= A − A (3.42)
2
If V (x) is a quadratic function formed with A, then

1  
V (x) = x  Ax = x A + A x (3.43)
2
3.5 Summary 45

3.5 Summary

Adaptive control theory requires a basic understanding of mathematics of matrix

algebra. More advanced adaptive control theory requires a solid foundation in real
analysis which is beyond the scope of this book. Various types of vector and matrix
norms are defined. The existence and uniqueness of a solution of a nonlinear differ-
ential equation require the continuity condition and the Lipschitz condition. Positive-
valued functions are an important class of functions in adaptive control theory. The
positive definiteness of a real-valued function is defined. The properties of a positive-
definite matrix are given.

3.6 Exercises

1. Verify that the 1-norm of x ∈ Rn


n
x1 = |xi |
i=1

satisfies the norm conditions.

2. Compute analytically the 1-, 2-, infinity, and Frobenius norms of
⎡ ⎤
1 0 −2
A=⎣ 4 0 2 ⎦
−1 3 2

and verify the answers with MATLAB using the function “norm.”

Note: MATLAB may be used to compute the eigenvalues.

3. Decompose A into its symmetric part P and anti-symmetric part Q. Write the
quadratic function V (x) = x  P x. Is V (x) positive (semi-)definite, negative
(semi-)definite, or neither?
4. Given a set C ⊂ R2
 
C = x ∈ R2 : x12 + 4x22 − 1 < 0

Is C a compact set? Write the set notation for the complementary set C c . Plot
and illustrate the region in R2 that represents C .
5. For each of the following equations, determine if f (x) is locally Lipschitz at
x = x0 or globally Lipschitz:
46 3 Mathematical Preliminaries
√
a. ẋ = x 2 + 1, x0 = 0.
−x 3 , x0 = 1.
b. ẋ = √
c. ẋ = x 3 + 1, x0 = 0.
Chapter 4
Lyapunov Stability Theory

Abstract Stability of nonlinear systems are discussed in this chapter. Lyapunov

stability, asymptotic stability, and exponential stability of an equilibrium point of a
nonlinear system are defined. The Lyapunov’s direct method is introduced as an indis-
pensable tool for analyzing stability of nonlinear systems. The Barbashin–Krasovskii
theorem provides a method for global stability analysis. The LaSalle’s invariant set
theorem provides a method for analyzing autonomous systems with invariant sets.
Stability of non-autonomous systems involves the concepts of uniform stability, uni-
form boundedness, and uniform ultimate boundedness. The Barbalat’s lemma is an
important mathematical tool for analyzing asymptotic stability of adaptive control
systems in connection with the concept of uniform continuity of a real-valued func-
tion.

Stability is an important consideration of any dynamical systems with feedback con-

trol. Stability for LTI systems can be analyzed by many well-established methods
such as eigenvalue analysis, root locus, phase and gain margins, etc. For nonlinear
systems, Lyapunov stability theory provides a powerful technique for stability analy-
sis of such systems. The Lyapunov stability theory is central to the study of nonlinear
adaptive control [1–4]. In this chapter, the learning objectives are to develop a basic
understanding of:

© Springer International Publishing AG 2018 47

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_4
48 4 Lyapunov Stability Theory

• Various stability concepts for autonomous and non-autonomous systems, such as

local stability, asymptotic stability, exponential stability, uniform stability, and
uniform boundedness;
• Lyapunov’s direct method and LaSalle’s invariant set theorem for analyzing sta-
bility of nonlinear systems; and
• Uniform continuity concept and Barbalat’s lemma for analyzing stability of non-
autonomous systems.

4.1 Stability Concepts

Consider an autonomous system

ẋ = f (x) (4.1)

with an initial condition x (t0 ) = x0, where f (x) is locally Lipschitz in some
subset D of Rn and the solution x t; t0, x0 exists and is unique in a region
B R = {x (t) ∈ Rn : x < R} ⊂ D of an equilibrium point x ∗ . The region B R
can be thought of as a hypersphere in Rn with the origin at x = x ∗ . Colloquially, it
is often referred to in the literature as a ball B R . Since x ∗ is a constant vector, for
convenience, the autonomous system can be transformed by shifting the equilibrium
point to the origin at x = 0. Let y (t) = x (t) − x ∗ , then
 
ẏ = f y + x ∗  g (y) (4.2)

whose equilibrium is the origin y ∗ = 0.

Thus, for convenience, the equilibrium point for autonomous systems described
by Eq. (4.1) is understood to be x ∗ = 0.
Example 4.1 The system    
ẋ1 −x1 + x1 x2
=
ẋ2 x2 − x1 x2

has an equilibrium at x1∗ = 1 and x2∗ = 1.

The system can be transformed by letting y1 (t) = x1 (t)−1 and y2 (t) = x2 (t)−1
which yields
     
ẏ1 − (y1 + 1) + (y1 + 1) (y2 + 1) y1 y2 + y2
= =
ẏ2 (y2 + 1) − (y1 + 1) (y2 + 1) −y1 y2 − y1
4.1 Stability Concepts 49

4.1.1 Stability Definition

Definition 4.1 The equilibrium point x ∗ = 0 of a system starting at an initial con-

dition x (t0 ) = x0 is said to be stable (in the sense of Lyapunov) if, for any R > 0,
there exists some r (R) > 0 such that

x0  < r ⇒ x < R, ∀t ≥ t0 (4.3)

Otherwise, the equilibrium point is unstable.

Stability concept essentially implies that, given a system with an initial condition
close to the origin, the trajectory of the system can be kept arbitrarily close to it.
Figure 4.1 illustrates the stability concept.

Fig. 4.1 Stability concept

Note that instability for linear systems means that the solution grows exponentially
as t → ∞ due to unstable poles in the right half plane, resulting in unbounded
signals. For nonlinear systems, instability of an equilibrium does not always lead to
unbounded signals. For example, the Van der Pol oscillator in Example 2.6 has a
stable limit cycle that encloses an unstable equilibrium point at the origin. So, the
equilibrium point in theory is unstable and the system cannot stay arbitrarily close to
it. If we choose any arbitrary circle B R to be completely inside the limit cycle, then
no matter how close the initial condition is to the origin, the trajectory of the system
will eventually escape the circle B R as illustrated in Fig. 4.2. However, the trajectory
tends to the limit cycle and remains there as t → ∞.
50 4 Lyapunov Stability Theory

1.5

0.5
Br
2

0
x

−0.5

−1
B
−1.5 R

−2

−2 −1 0 1 2
x
1

Fig. 4.2 Unstable origin of Van der Pol Oscillator

4.1.2 Asymptotic Stability

The stability concept in the Lyapunov sense does not explicitly imply that the tra-
jectory of a nonlinear system will eventually converge to the origin. For example, an
ideal spring-mass system without friction will display a sinusoidal motion forever if
it is subject to a disturbance. So, the system is stable in the Lyapunov sense but does
not converge to the origin. Asymptotic stability is a stronger stability concept than
the Lyapunov stability concept and is defined as follows:

Definition 4.2 The equilibrium point x ∗ = 0 is said to be asymptotically stable if

there exists some r > 0 such that

x0  < r ⇒ lim x = 0 (4.4)

t→∞

All trajectories starting within B R will eventually converge to the origin. The
origin is then said to be attractive. For a second-order system, both stable focus and
stable node are attractive. The largest such region is called a region of attraction,
defined as  
R A = x (t) ∈ D : lim x (t) = 0 (4.5)
t→∞

It is noted that the asymptotic stability concept in the definition above is a local
concept for any initial condition that lies within the ball B R . If an equilibrium point
of a system is asymptotically stable for all initial conditions x0 ∈ Rn , then the
4.1 Stability Concepts 51

equilibrium point is said to be asymptotically stable in the large. This notion is

equivalent to global asymptotic stability.
Example 4.2 The equilibrium point of the system

ẋ = −x 2

with x (0) = x0 > 0 is asymptotically stable since the solution

x0
x (t) =
x0 t + 1

tends to zero as t → ∞. The region of attraction is


x0
R A = x (t) ∈ R+ : x (t) = , x0 > 0
x0 t + 1

Note that the equilibrium point is unstable if x0 < 0 and has a finite escape time
at t = −1/x0 . So, the equilibrium is asymptotically stable for all x (t) ∈ R+ but not
asymptotically stable in the large.

4.1.3 Exponential Stability

The rate of convergence of a solution of a nonlinear differential equation can be

estimated by comparing its solution to an exponential decay function [2, 4]. This
gives rise to a notion of exponential stability which is defined as follows:

Definition 4.3 The equilibrium point x ∗ = 0 is said to be exponentially stable if

there exist two strictly positive constants α and β such that

x ≤ α x0  e−β(t−t0 ) , ∀x ∈ B R , t ≥ t0 (4.6)

This definition gives a local version of the exponential stability concept for some
initial condition x0 close to the origin. If the origin is exponentially stable for all
initial conditions x0 ∈ Rn , then the equilibrium point is said to be exponentially
stable in the large. The constant β is called the rate of convergence.
It is noted that exponential stability implies asymptotic stability, but the converse
is not true.
Example 4.3 The differential equation
 
ẋ = −x 1 + sin2 x

subject to x (0) = 1 is bounded from below and above by

52 4 Lyapunov Stability Theory

−2 |x| ≤ |ẋ| ≤ − |x|

if x (t) > 0.
The solution is bounded from below and above as shown in Fig. 4.3 by

e−2t ≤ |x (t)| ≤ e−t

Therefore, the equilibrium is exponentially stable and the rate of convergence

is 1.

1
. 2
x=−x(1+sin x)
0.9
.
x=−x
.
x=−2x
0.8

0.7

0.6

0.5
x

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4
t

Fig. 4.3 Exponential stability

4.2 Lyapunov’s Direct Method

4.2.1 Motivation

Consider a spring-mass-damper system with friction as shown in Fig. 4.4.

The equation of motion without external forces is described by

m ẍ + c ẋ + kx = 0 (4.7)

where m is the mass, c > 0 is the viscous friction coefficient, and k is the spring
constant.
4.2 Lyapunov’s Direct Method 53

Fig. 4.4 Spring-mass system

The system in the state-space form is expressed as

   
ẋ1 x2
= (4.8)
ẋ2 − mc x2 − k
x
m 1

where x1 (t) = x (t) and x2 (t) = ẋ (t). The system has an equilibrium point at
(0, 0), that is, it is at rest with zero displacement and velocity
The spring-mass-damper system possesses two types of energy: (1) kinetic energy
and (2) potential energy. The kinetic energy of any moving point mass is given by

1 2 1 1
T = mv = m ẋ 2 = mx22 (4.9)
2 2 2
The potential energy for a spring is given by

1 2 1
U= kx = kx12 (4.10)
2 2
The energy of the spring-mass-damper system is the sum of the kinetic energy
and potential energy. Thus, the energy function of the system is defined as

1 2 1 2
E = T +U = mx + kx (4.11)
2 2 2 1
Note that the energy function is a quadratic positive-definite function.
The friction force also does work on the mass. This type of work is called a
non-conservative work which is usually due to a dissipative force, as opposed to
a conservative work of which potential energy is one form. The work function is
defined in general as

W = F.d x (4.12)

where F is a force acting on a mass that displaces it by an infinitesimal distance d x,

and the integral is evaluated over a path that the mass traverses.
54 4 Lyapunov Stability Theory

For a viscous friction force, the work done is evaluated as

  
W = c ẋd x = c ẋ 2 dt = cx22 dt (4.13)

The total energy of the system is the sum of the energy and the work done. Thus,

1 2 1 2
E+W = mx + kx + cx22 dt (4.14)
2 2 2 1

According to the first law of thermodynamics, the total energy of a closed system
is neither created or destroyed. In other words, the total energy is conserved and is
equal to a constant. Thus,
E + W = const (4.15)

or equivalently
Ė + Ẇ = 0 (4.16)

This energy conservation law can be easily verified for the spring-mass-damper
system as
 
d c k
Ė + Ẇ = mx2 ẋ2 +kx1 ẋ1 + = mx2 − x2 − x1 +kx1 x2 +cx22 = 0
cx22 dt
dt m m
(4.17)
The time derivative of the energy function is evaluated as

c k
Ė = mx2 ẋ2 + kx1 ẋ1 = mx2 − x2 − x1 + kx1 x2 = −cx22 ≤ 0 (4.18)
m m

for c > 0.
The reason Ė is only negative semi-definite is because Ė can be zero for any
x1 = 0.
Thus, for a dissipative system, the time derivative of the positive definite energy
function is a negative semi-definite function. That is,

Ė ≤ 0 (4.19)

The equilibrium point is then stable. Thus, stability of a dynamical system can
be studied by examining the time derivative of the energy function. The Lyapunov
stability theory is motivated by the concept of energy. In fact, the energy function
is a Lyapunov function. Whereas the energy function is unique for a given physical
system, a Lyapunov function can be any positive-definite function that satisfies the
negative (semi-)definiteness of its time derivative.
Lyapunov’s direct method is a powerful tool for assessing stability of an equilib-
rium of a nonlinear system directly without solving the system’s dynamical equation.
4.2 Lyapunov’s Direct Method 55

The motivation of the method is based on the energy concept of a mechanical system.
From the spring-mass-damper example, the following observations are made:
• The energy function is positive definite.
• The time rate of the energy function is negative semi-definite in which case the
equilibrium is stable.
Aleksandr Mikhailovich Lyapunov (1857–1918) recognized that stability of a system
can be proven without developing a true knowledge of the system energy using a
class of positive-definite functions, known as Lyapunov functions, provided they can
be found.
Definition 4.4 A function V (x) is said to be a Lyapunov function if it satisfies the
following conditions:
• V (x) is positive definite; i.e.,
V (x) > 0 (4.20)

and has a continuous first partial derivative.

• V̇ (x) is at least negative semi-definite; i.e.,

∂V ∂V
V̇ (x) = ẋ = f (x) ≤ 0 (4.21)
∂x ∂x
or
V̇ (x) < 0 (4.22)


Geometrically, a Lyapunov function may be illustrated by a bowl-shaped surface
as shown in Fig. 4.5. The Lyapunov function is a level curve on the bowl starting

Fig. 4.5 Illustration of Lyapunov function

56 4 Lyapunov Stability Theory

at the top and progressing downward toward the bottom of the bowl. The value of
the Lyapunov function thus decreases toward zero at the bottom of the bowl which
represents a stable equilibrium.
Example 4.4 For the spring-mass-damper system, the energy function is clearly a
Lyapunov function. Suppose one chooses another Lyapunov candidate function

V (x) = x12 + x22 > 0

Then,
 
c k k c
V̇ (x) = 2x1 ẋ1 +2x2 ẋ2 = 2x1 x2 +2x2 − x2 − x1 = 2x1 x2 1 − −2 x22
m m m m

Note that V̇ (x) is not negative (semi-)definite because of the term with the product
x1 x2 . So, this candidate function is not a Lyapunov function.
In many systems, finding a Lyapunov function is not trivial. Unfortunately, there
is no straightforward way to obtain a Lyapunov function. Perhaps, the most obvious
Lyapunov function for any system is the energy function, but it is not always easy to
identify such a function for a nonlinear system unless one knows the physics of the
system.

Example 4.5 Consider a pendulum with viscous friction whose motion is described
by
ml 2 θ̈ + cθ̇ + mgl sin θ = 0

which is expressed in a state-space form as

   
ẋ1 x2
=
ẋ2 − mlc 2 x2 − gl sin x1

with x1 (t) = θ (t) and x2 (t) = θ̇ (t).

It is not obvious what a Lyapunov candidate function would look like for this
system. One can try
V (x) = x12 + x22

but that would not yield V̇ (x) ≤ 0 since

 c g 
V̇ (x) = 2x1 ẋ1 + 2x2 ẋ2 = 2x1 x2 + 2x2 − 2 x2 − sin x1  0
ml l
The kinetic energy and potential energy for the pendulum are known to be

1 2 2
T = ml x2
2
4.2 Lyapunov’s Direct Method 57

U = mgl (1 − cos x1 )

So, the energy function is

1 2 2
E = T +U = ml x2 + mgl (1 − cos x1 ) > 0
2

Using this as a Lyapunov function, then evaluating V̇ (x) gives

 c g 
V̇ (x) = ml 2 x2 ẋ2 + mgl sin x1 ẋ1 = ml 2 x2 − 2 x2 − sin x1 + mgl sin x1 x2 = −cx22 ≤ 0
ml l

Thus, it is not surprising that the energy function can always be used for a
Lyapunov function, provided such a function can be found.


In summary, it can be seen that the Lyapunov’s direct method is a powerful tech-
nique for studying stability of an equilibrium point of a nonlinear system.

4.2.2 Lyapunov Theorem for Local Stability

Theorem 4.1 Let x ∗ = 0 be an equilibrium point and if there exists a Lyapunov

function V (x) > 0 for all x (t) ∈ B R such that V̇ (x) ≤ 0 for all x (t) ∈ B R , then
the equilibrium is locally stable in a Lyapunov sense. Moreover, if V̇ (x) < 0 for all
x (t) ∈ B R , then the equilibrium is locally asymptotically stable.


It is important to note that the Lyapunov’s direct method only gives a sufficient
condition for stability. Failure of a Lyapunov candidate function to satisfy the stability
condition does not imply that the equilibrium is unstable. It simply means that a good
Lyapunov candidate function may not have been identified. An exception to this rule
is the energy function which provides both the necessary and sufficient conditions
for stability.
Example 4.6 For the pendulum in Example 4.5, V̇ (x) is negative semi-definite, so
the equilibrium is locally stable.
Example 4.7 Consider the system
   
ẋ1 x13 + x1 x22 − x1
=
ẋ2 x23 + x12 x2 − x2

Choose a Lyapunov candidate function

V (x) = x12 + x22

58 4 Lyapunov Stability Theory

Then, V̇ (x) is evaluated as

   
V̇ (x) = 2x1 ẋ1 + 2x2 ẋ2 = 2x1 x13 + x1 x22 − x1 + 2x2 x23 + x12 x2 − x2
  
= 2 x12 + x22 x12 + x22 − 1

Observing V̇ (x), one can conclude that V̇ (x) < 0 for all x ∈ B R where
 
B R = x (t) ∈ D ⊂ R2 : x12 + x22 < 1

The equilibrium is asymptotically stable. The region of attraction is B R within

which all the trajectories converge to the equilibrium.
Example 4.8 For the spring-mass-damper system, consider a Lyapunov candidate
function
V (x) = x P x > 0
 
where x (t) = x1 (t) x2 (t) and P = P > 0 to be determined such that V̇ (x) <
0 for asymptotic stability.
Expressing V (x) as
  
  p11 p12 x1
V (x) = x1 x2 = p11 x12 + 2 p12 x1 x2 + p22 x22
p12 p22 x2

where pi j are elements of P, then evaluating V̇ (x) yields

V̇ (x) = 2 p11 x1 ẋ1 + 2 p12 (x1 ẋ2 + ẋ1 x2 ) + 2 p22 x2 ẋ2

 
c k c k
= 2 p11 x1 x2 + 2 p12 x1 − x2 − x1 + 2 p12 x22 + 2 p22 x2 − x2 − x1
m m m m
  
k c k c
= −2 p12 x12 + 2 p11 − p12 − p22 x1 x2 + 2 p12 − p22 x2
m m m m 2

Since V̇ (x) < 0, one can choose

V̇ (x) = −2x12 − 2x22

Equating terms then yields

k m
p12 = 1 ⇒ p12 =
m k
c m m m 
p12 − p22 = −1 ⇒ p22 = ( p12 + 1) = +1
m c c k
c k c k c m k
p11 − p12 − p22 = 0 ⇒ p11 = p12 + p22 = + +
m m m m k c c
4.2 Lyapunov’s Direct Method 59

The matrix c 
+ m
+ k m
 mk 
P= k c c
m
k
m
c k
+1

can be verified to be positive definite for m > 0, c > 0, and k > 0.

The system is then asymptotically stable. It is noted that since the system is linear,
stability is always referred to in a global context.
Another approach to be considered is as follows:
The system can be expressed as

ẋ = Ax

where  
0 1
A=
− mk − mc

is a Hurwitz matrix with negative real part eigenvalues.

Proceeding to evaluate V̇ (x) yields

V̇ (x) = ẋ P x + x P ẋ

which upon substitution yields

 
V̇ (x) = x A P x + x P Ax = x A P + P A x < 0

This inequality is satisfied if and only if

A P + P A < 0

which is called a linear matrix inequality (LMI) that can be solved for P.
Alternatively, one can write the LMI as a linear matrix equation

A P + P A = −Q

where Q = Q > 0 is a positive-definite matrix. This equation is known as the

algebraic Lyapunov equation.
Thus, setting Q = 2I , where I is an identity matrix, which in this case is a 2-by-2
matrix, then yields
V̇ (x) = −2x x = −2x12 − 2x22

The matrix P is then solved from

A P + P A = −2I
60 4 Lyapunov Stability Theory

There are numerical methods that can be used to solve the Lyapunov equation.
For a low matrix dimension less than 4, the equation can be solved analytically. For
this example, one can set up
       
0 − mk p11 p12 p11 p12 0 1 −2 0
+ =
1 − mc p12 p22 p12 p22 − mk − mc 0 −2

then expand
   
−2 p12 mk p11 −
 p12 m − pc 22
c
m
k
−2 0
=
p11 − p12 m − p22 m 2 p12 − p22 m
c k
0 −2

and equate terms to solve for pi j .

4.2.3 Lyapunov Theorem for Exponential Stability

Theorem 4.2 Let x ∗ = 0 be an equilibrium point and if there exists a Lyapunov

function V (x) > 0 for all x (t) ∈ B R such that V̇ (x) < 0 for all x (t) ∈ B R and
there also exist two positive constants η and β such that

V (x) ≤ η x2 (4.23)

and
V̇ (x) ≤ −βV (x) (4.24)

then the equilibrium is locally exponentially stable.

Example 4.9 Consider

 
ẋ = −x 1 + sin2 x

subject to x (0) = 1.
Choose a Lyapunov candidate function

V (x) = x 2 = x2

So,
V (0) = x02 = 1

V̇ (x) is computed as
 
V̇ (x) = 2x ẋ = −2x 2 1 + sin2 x ≤ −2 x2 = −2V (x) < 0
4.2 Lyapunov’s Direct Method 61

Integrating V̇ (x) yields

V (t) ≤ V (0) e−2t

or
x 2 ≤ e−2t

which is equivalent to
|x| ≤ e−t

4.2.4 Radially Unbounded Functions

Definition 4.5 A continuous, positive-valued function ϕ (x) ∈ R+ is said to belong

to class K ; i.e., ϕ (x) ∈ K , if
• ϕ (0) = 0
• ϕ (x) is strictly increasing for all x (t) ≤ R or x (t) < ∞
Definition 4.6 A continuous, positive-valued function ϕ (x) ∈ R+ is said to belong
to class K R; i.e., ϕ (x) ∈ K R, if
• ϕ (0) = 0
• ϕ (x) is strictly increasing for all x (t) < ∞
• lim x→∞ ϕ (x) = ∞
Definition 4.7 A continuous, positive-valued function V (x) ∈ R+ with V (0) = 0
is said to be a radially unbounded function if there exists a function ϕ (x) ∈ K R
such that V (x) ≥ ϕ (x) for all x (t) ∈ Rn . Thus, V (x) must be infinitely large
when x tends to infinity. That is,

V (x) → ∞ as x → ∞ (4.25)

4.2.5 Barbashin–Krasovskii Theorem for Global Asymptotic

Stability

The asymptotic stability concept in the Lyapunov sense of an equilibrium point is a

local concept such that there exists V (x) > 0 for all x ∈ B R , where B R is a finite
region in D ⊂ Rn for which the function f (x) is locally Lipschitz, then V̇ (x) < 0.
There exists a region of attraction R A ⊂ D ⊂ Rn in which all trajectories will
converge to the origin. On the other hand, asymptotic stability in the large is a global
concept that requires the region of attraction to extend to the entire Euclidean space
Rn . As a result, V (x) > 0 must be defined for all x ∈ Rn .
There is an additional requirement imposed on V (x) for stability in the large. That
is, V (x) is required to be a radially unbounded function. The condition of the radial
62 4 Lyapunov Stability Theory

unboundedness of V (x) ensures that all trajectories in the large when x → ∞
will be attracted to the origin. The global stability Lyapunov condition can be stated
by the following Barbashin–Krasovskii theorem [2, 4]:

Theorem 4.3 The equilibrium point x ∗ = 0 is asymptotically stable in the large if

there exists a radially unbounded Lyapunov function V (x) > 0 for all x (t) ∈ Rn
such that V̇ (x) < 0 for all x (t) ∈ Rn .

Example 4.10 Consider a scalar linear system

ẋ = −ax

where a > 0, whose equilibrium point at the origin is asymptotically stable in the
large. Suppose a Lyapunov candidate function is chosen as

x2
V1 (x) = >0
1 + x2

V1 (x) ∈ K but V1 (x) ∈ / K R since V1 (0) = 0 and V1 (x) is strictly increasing

for all x (t) < ∞, but lim x→∞ V1 (x) = 1. This means that one cannot analyze
global asymptotic stability of the origin of this system using this Lyapunov candidate
function since

2x ẋ 2x 3 ẋ −2ax 2
V̇1 (x) = −   =  2 → 0
1 + x2 1 + x2
2
1 + x2

as x → ∞, which implies that the origin is not asymptotically stable as x → ∞.
Therefore, the origin is stable but not asymptotically stable in the large, which is a
contradiction.
Now, suppose another Lyapunov candidate function is chosen as

V2 (x) = x 2 > 0

V2 (x) is a radially unbounded function since there exists a function ϕ (x) =

αx 2 ∈ K R, where α < 1, such that V2 (x) ≥ ϕ (x) for all x (t) < ∞. Global
asymptotic stability of the origin can be analyzed using this radially unbounded
Lyapunov candidate function. Evaluating V̇2 (x) yields

V̇2 (x) = 2x ẋ = −2ax 2 < 0

which implies that the origin is indeed asymptotically stable in the large.
4.2 Lyapunov’s Direct Method 63

4.2.6 LaSalle’s Invariant Set Theorem

Asymptotic stability requires that V̇ (x) < 0. Yet, for the spring-mass-damper sys-
tem, if the energy function is chosen as a Lyapunov function, then V̇ (x) ≤ 0, even
though the solution of the system is clearly asymptotic in the presence of a dissipa-
tive force due to the viscous friction. The LaSalle’s invariant set theorem can resolve
this apparent contradiction when an asymptotically stable equilibrium point of an
autonomous system only satisfies the Lyapunov condition V̇ (x) ≤ 0.

Definition 4.8 For an autonomous system, a set M is said to be invariant if every

trajectory that starts from a point in M will remain in M for all future time [2, 4].
That is,
x (0) ∈ M ⇒ x (t) ∈ M , ∀t ≥ t0 (4.26)

Example 4.11
• An equilibrium point is an invariant set because by definition x (t) = x ∗ is a
constant solution of an autonomous system so that

x (t) = x (t0 ) = x ∗ ∈ M , ∀t ≥ t0

• A region of attraction R A is an invariant set since all trajectories in R A will remain

in R A for all future time and converge to the origin as t → ∞
 
R A = x (t) ∈ M : lim x (t) = 0
t→∞

• The limit cycle of the Van der Pol oscillator is an invariant set since any point on
the limit cycle will remain on it for all future time.
• The entire Euclidean space Rn is a trivial invariant set since all trajectories must
be in some subspace that belongs in Rn .

The LaSalle’s invariant set theorem is stated as follows:

Theorem 4.4 Given an autonomous system, let V (x) > 0 be a positive-definite

function with a continuous first partial derivative such that V̇ (x) ≤ 0 in some finite
region B R ⊂ D. Let R be a set of all points where V̇ (x) = 0. Let M be the largest
invariant set in R. Then, every solution x (t) starting in B R approaches M as t → ∞.

Example 4.12 Consider

   
ẋ1 −x13 − x1 x22 + x1
=
ẋ2 −x23 − x12 x2 + x2

Choose a Lyapunov candidate function

64 4 Lyapunov Stability Theory

V (x) = x12 + x22

Then, V̇ (x) is evaluated as

   
V̇ (x) = 2x1 −x13 − x1 x22 + x1 + 2x2 −x23 − x12 x2 + x2
   
= −2x12 x12 + x22 − 1 − 2x22 x12 + x22 − 1
  
= −2 x12 + x22 x12 + x22 − 1

V̇ (x) < 0 in a set S where

 
S = x (t) ∈ B R : x12 + x22 − 1 > 0

but V̇ (x) ≥ 0 in the complementary set

 
S c = x (t) ∈ B R : x12 + x22 − 1 ≤ 0

which represents a circular region that includes the origin. Therefore, the origin is
unstable.
Let R be a set of all points where V̇ (x) = 0. Then,
 
R = x (t) ∈ B R : g (x) = x12 + x22 − 1 = V (x) − 1 = 0

in fact represents a bounded solution x (t) since all trajectories either inside or outside
of S c will move toward R and remain in R as illustrated in Fig. 4.6. Thus, R is an
invariant set. One can also verify this by taking the time derivative of the function
g (x) that represents the trajectories in the set R

Fig. 4.6 Bounded solution set

4.2 Lyapunov’s Direct Method 65

d
ġ (x) = (V − 1) = V̇ (x) = 0, ∀x (t) ∈ R
dt
The Lyapunov function can be solved analytically by noting that

V̇ = −2V (V − 1)

which leads to
dV
= −2dt
V (V − 1)

Using the partial fraction, this can be expressed as


1 1
− d V = −2dt
V −1 V

which yields the following general solution:

V0
V (t) =
V0 − (V0 − 1) e−2t

As t → ∞, V (t) tends to a constant solution as shown in Fig. 4.7 with

lim V (t) = 1
t→∞

1.8

1.6

1.4

1.2 Invariant Set

V(t)

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4
t

Fig. 4.7 Trajectories of Lyapunov function tending to an invariant set

66 4 Lyapunov Stability Theory

which is in fact the set R. Therefore, the set R is a bounding set of all possible
solutions x (t). Thus, x (t) ∈ L∞ ; i.e., x (t) is bounded.
Example 4.13 Consider the spring-mass-damper system. Choose the Lyapunov func-
tion to be the energy function

1 2 1 2
V (x) = mx + kx
2 2 2 1
Then,
V̇ (x) = mx2 ẋ2 + kx1 ẋ1 = −cx22 ≤ 0

Since V̇ (x) is negative semi-definite, then the origin is only stable in the Lyapunov
sense, but not asymptotically stable as one would expect for a spring-mass-damper
system with friction. Let R be a set of all points where V̇ (x) = 0. Then,
 
R = x (t) ∈ R2 : V̇ (x) = 0 ⇒ x2 = 0

is a collection of all points that lie on the x1 -axis. It follows that any point on this
axis must satisfy
m ẋ2 + kx1 = 0

or
k
ẋ2 = ẍ1 = − x1
m

If x1 (t) = 0, then ẍ1 (t) = 0 with sgn (ẍ1 ) = −sgn (x1 ) in R, where sgn () is the
sign function which returns 1 if the argument is positive, 0 if the argument is zero,
or -1 if the argument is negative. This means that a point on this axis cannot remain
in R because the acceleration ẍ1 causes the point to move toward the origin, unless
it is already at the origin.
Another way to find an invariant set is to evaluate the derivative of the function
that describes R and set it equal to zero. Hence,

ẋ2 = 0

which is satisfied if and only if x1 (t) = 0.

Thus, the invariant set M ⊂ R is a set that contains only the origin. Then,
according to the LaSalle’s invariant set theorem, all trajectories will converge to the
origin as t → ∞. The origin then is asymptotically stable.

This example brings up an interesting observation that can be stated in the
following corollary of the LaSalle’s invariant set theorem:
Corollary 4.1 Let V (x) > 0 be a positive definite function with a continuous
first
 partial derivative such  that V̇ (x) ≤ 0 in some finite region B R ⊂ D. Let R =
x (t) ∈ B R : V̇ (x) = 0 and suppose that no solution can stay in R except the trivial
4.2 Lyapunov’s Direct Method 67

solution x = 0. Then, the origin is asymptotically stable. Moreover, if V (x) > 0 is
a positive definite radially unbounded function and R = x (t) ∈ Rn : V̇ (x) = 0 ,
then the origin is asymptotically stable in the large.

4.2.7 Differential Lyapunov Equation

The Lyapunov equation has a connection to the optimal control theory. In particular,
the Lyapunov equation can be viewed as a special case of the Riccati equation for a
Linear Quadratic Regulator (LQR) optimal control problem. Consider the following
LTI system:
ẋ = Ax (4.27)

subject to x (t0 ) = x0 , where x (t) ∈ Rn and A ∈ Rn × Rn .

It is of interest to find a condition that minimizes the following quadratic cost
function:  t f
min J = x Qxdt (4.28)
t0

where Q > 0 ∈ Rn × Rn is a positive definite matrix.

The solution can be established by the Pontryagin’s maximum principle in the
optimal control theory [5, 6]. The Hamiltonian function of this system is defined as

H = x Qx + λ (Ax + Bu) (4.29)

where λ (t) ∈ Rn is called an adjoint or co-state vector.

The adjoint equation is given by

∂ H
λ̇ = − = −Qx − A λ (4.30)
∂x
subject to the transversality (terminal time) condition
 
λ tf = 0 (4.31)

Choose a solution of λ (t) in the form of

λ (t) = P (t) x (4.32)

where P (t) ∈ Rn × Rn is a time-varying matrix.

Then, the adjoint equation is evaluated with the system dynamics as

Ṗ x + P Ax = −Qx − A P x (4.33)
68 4 Lyapunov Stability Theory

By factoring out x (t), the differential Lyapunov equation is obtained as

Ṗ + P A + A P + Q = 0 (4.34)
 
subject to P t f = 0.
Contrasting this with the differential Riccati equation

Ṗ + P A + A P − P B R −1 B P + Q = 0 (4.35)

the differential Lyapunov equation is a special case of the differential Riccati equation
for R → ∞.
Note that the differential Lyapunov equation is defined backward in time with the
transversality condition given at the final time. By transforming into a time-to-go
variable, τ = t f − t, then

dP
= P A + A P + Q (4.36)
dτ
subject to P (0) = 0 in the time-to-go coordinate.
If A is Hurwitz and let t f → ∞ which corresponds to an infinite time horizon
solution, then the time-varying solution of the differential Lyapunov equation tends
to a constant solution of the algebraic Lyapunov equation

P A + A P + Q = 0 (4.37)

The constant solution of P is given by

 τ

P = lim e A τ Qe Aτ dτ (4.38)
τ →∞ 0

which is positive definite for Q > 0 and requires that A be Hurwitz since the solution
must be a stable solution such that

lim e Aτ = 0 (4.39)
τ →∞

Example 4.14 Numerically compute P, given

   
0 1 10
A= , Q=
−4 −4 01

The differential Lyapunov equation in time-to-go can be solved using any numer-
ical technique for solving differential equations such as the Euler or Runge–Kutta
method. For example, the Euler method for solving the Lyapunov equation is as
follows:  
Pi+1 = Pi +
τ Pi A + A Pi + Q
4.2 Lyapunov’s Direct Method 69

1.2

0.8

0.6
P

0.4

0.2

0
p11 p12=p21 p22
−0.2
0 2 4 6 8 10
τ=tf−t

Fig. 4.8 Numerical solution of Lyapunov equation

where i denotes the value of P (τ ) at a time τi = i

τ .
The equation can be integrated until the solution converges to within a specified
error. Figure 4.8 illustrates the solution convergence of the differential Lyapunov
equation. The result is obtained as
 
1.125 0.125
P=
0.125 0.15625

4.3 Stability of Non-Autonomous Systems

Most of the concepts for Lyapunov stability for autonomous systems can be applied
to non-autonomous systems with some additional considerations [2, 4].
Consider a non-autonomous system

ẋ = f (x, t) (4.40)

subject to x (t0 ) = x0 , where f (x, t) is locally Lipschitz in D ×[0, ∞) and D ⊂ Rn .

The notion of the origin as an equilibrium point now takes on a different meaning
in that the equilibrium point x ∗ must be time-invariant for all t ≥ t0 and satisfies
 
f x ∗, t = 0 (4.41)
70 4 Lyapunov Stability Theory

Otherwise, the system does not have a “true” equilibrium point in the Lyapunov
sense.
Example 4.15
• The system
ẋ = g (t) h (x)

has an equilibrium point x ∗ where

 
h x∗ = 0

• The system
ẋ = g (t) h (x) + d (t)

does not have a true equilibrium point since h (x) would have been a function of
t to satisfy ẋ (t) = 0, which is a contradiction unless g (t) = αd (t) for some
constant α.

4.3.1 Uniform Stability

The Lyapunov stability definition for a non-autonomous system is defined as follows:

Definition 4.9 The equilibrium point x ∗ = 0 is said to be stable (in the sense of
Lyapunov) if, for any R > 0, there exists some r (R, t0 ) > 0 such that

x0  < r ⇒ x < R, ∀t ≥ t0 (4.42)

Note that the difference in this definition as compared to that for an autonomous
system is the ball of radius r that encloses x0 now may depend on the initial time t0 .
Thus, the stability of the origin may also be dependent on the initial time.
The concept of uniform stability is an additional consideration for non-autonomous
systems. Uniform stability implies that the radius r = r (R) is not dependent on
the initial time, and so are the stability properties of the equilibrium point. For
autonomous systems, stability is independent of the initial time. This property is
highly desirable since it eliminates the need for examining the effect of the initial
time on the stability of a non-autonomous system.

Definition 4.10 Given a Lyapunov function V (x, t) for a non-autonomous system

that satisfies the following conditions:

V (0, t) = 0 (4.43)
4.3 Stability of Non-Autonomous Systems 71

and
0 < V (x, t) ≤ W (x) (4.44)

where W (x) > 0 is a positive-definite function, then V (x, t) is said to be a positive

definite, decrescent function.

Example 4.16 The Lyapunov candidate function

  
V (x, t) = 1 + sin2 t x12 + x22

is bounded from above by

 
0 < V (x, t) ≤ 2 x12 + x22 = W (x)

Since W (x) > 0, then V (x, t) is a positive definite, decrescent function.


The Lyapunov’s direct method for a non-autonomous system is stated in the
following theorem:

Theorem 4.5 If there exists a positive definite, decrescent Lyapunov function

V (x, t) for all x (t) ∈ B R and t ≥ 0 such that

∂V ∂V
V̇ (x, t) = f (x, t) + ≤0 (4.45)
∂x ∂t
then the origin is said to be uniformly stable in the Lyapunov sense. Moreover,
if V̇ (x, t) < 0, then the origin is said to be uniformly asymptotically stable, and
additionally, if the region B R is extended to the entire Euclidean space Rn , then the
origin is said to be uniformly asymptotically stable in the large.

4.3.2 Uniform Boundedness

When a non-autonomous system does not have an equilibrium point, stability of such
a system is defined by the notion of boundedness [2, 4].

Definition 4.11 The solution of a non-autonomous system is said to be uniformly

bounded if, for any R > 0, there exists some r (R) > 0 independent of the initial
time t0 such that
x0  < r ⇒ x ≤ R, ∀t ≥ t0 (4.46)

Moreover, the solution is said to be uniformly ultimately bounded if, for any
R > 0, there exists some r > 0 independent of R and the initial time t0 such that

x0  < r ⇒ x ≤ R, ∀t ≥ t0 + T (4.47)

72 4 Lyapunov Stability Theory

where T = T (r ) is some time interval after the initial time t0 .



The uniform ultimate boundedness concept simply means that the solution may
not be uniformly bounded initially according to Definition 4.11 but eventually
becomes uniformly ultimately bounded after some time has passed. The constant
R is called a bound if the solution is uniformly bounded or an ultimate bound if the
solution is uniformly ultimately bounded.
The Lyapunov’s direct method can be applied to a non-autonomous system
according to the following theorem:
Theorem 4.6 Given a Lyapunov function V (x, t) for all x ≥ R and t ∈ [0, ∞),
then the solution of the non-autonomous system (4.40) is said to be uniformly
bounded if there exist functions ϕ1 (x) ∈ K R and ϕ2 (x) ∈ K R such that [2]
• ϕ1 (x) ≤ V (x, t) ≤ ϕ2 (x)
• V̇ (x, t) ≤ 0
for all x ≥ R and t ∈ [0, ∞). In addition, if there exists a function ϕ3 (x) ∈ K R
such that
• V̇ (x, t) ≤ −ϕ3 (x)
for all x ≥ R and t ∈ [0, ∞), then the solution is said to be uniformly ultimately
bounded.
Example 4.17 Consider
ẋ = −x + 2 sin t

subject to x (0) = x0 .
The system does not have an equilibrium. The solution is

x = (x0 + 1) e−t + sin t − cos t

√
If x0  < r and recognizing that e−t ≤ 1 and |sin t − cos t| ≤ 2, then
√ √
x ≤ x0 + 1 + 2<r +1+ 2=R
√
Thus, one can choose r (R) = R − 1 − 2 according to Definition 4.11. √
So, the
solution is uniformly bounded. Suppose x0 = 1, then the bound is R = 2 + 2.
Moreover, as t → ∞, then the solution tends to

x → sin t − cos t

so that √
x ≤ 2=R

independent of r . √
The solution then is also uniformly ultimately bounded with an
ultimate bound of 2 as shown in Fig. 4.9.
4.3 Stability of Non-Autonomous Systems 73

3
R (Bound for t ≥0)
Lyapunov Ultimate Bound
2

1
r R (Ultimate Bound for t ≥T)

0
x

−1

−2

−3

0 5 10 15 20
t

Fig. 4.9 Uniform ultimate boundedness

The Lyapunov’s direct method is now applied to determine the uniform bound-
edness of a solution. Consider a Lyapunov candidate function for this system

V (x) = x 2 > 0

Note that one can always find ϕ1 (x) ∈ K R and ϕ2 (x) ∈ K R for this
Lyapunov candidate function such that ϕ1 (x) ≤ V (x) ≤ ϕ2 (x), for example,
ϕ1 (x) = ax 2 with a < 1 and ϕ2 (x) = bx 2 with b > 1.
Then,
V̇ (x) = 2x ẋ = 2x (−x + 2 sin t) ≤ −2x 2 + 4 x

We see that 
V̇ (x) ≤ −2V (x) + 4 V (x)
√
Let W (t) = V (t) = x. Then,

V̇ √
Ẇ = √ = − V + 2 ≤ −W + 2
2 V

The solution of W (t) is

W ≤ (x0  − 2) e−t + 2

Thus,
lim x = lim W ≤ 2 = R
t→∞ t→∞
74 4 Lyapunov Stability Theory

Choose
ϕ3 (x) = 2 x2 − 4 x

Note that ϕ3 (x) ∈ K R (verify!). Then, it follows that

V̇ (x) ≤ −ϕ3 (x)

Then, V̇ (x) ≤ 0 if −2x 2 + 4 x ≤ 0 or x ≥ 2. Therefore, according to

Theorem 4.6, the solution x (t) is uniformly ultimately bounded with a Lyapunov
ultimate bound of 2. It is noted that the Lyapunov ultimate bound is always more
conservative or greater than or equal to the ultimate bound derived from the actual
solution as illustrated in Fig. 4.9.
There is another way of showing that x (t) is uniformly ultimately bounded. By
completing the square, V̇ (x) can also be expressed as

V̇ (x) ≤ −2 (x − 1)2 + 2

Then, V̇ (x) ≤ 0 if −2 (x − 1)2 + 2 ≤ 0 or x ≥ 2. Since V̇ (x) ≤ 0 outside

the compact set x ≤ 2, but V̇ (x) > 0 inside it, therefore the solution x (t) is
uniformly ultimately bounded. Any trajectory that starts outside the compact set will
reach the ultimate bound x = 2 because V̇ (x) ≤ 0 outside the compact set. Any
trajectory that starts inside the compact set will move away from the origin because
V̇ (x) > 0 inside the compact set, but will eventually be attracted to the ultimate
bound x = 2.

4.3.3 Barbalat’s Lemma

The LaSalle’s invariant set theorem can be used to show asymptotic stability of
an equilibrium point for an autonomous system when V̇ (x) is only negative semi-
definite. This theorem cannot be used for non-autonomous systems. Therefore, it can
be much more difficult to show asymptotic stability for a non-autonomous system
than for an autonomous system. Barbalat’s lemma is a mathematical tool that can be
used to address this situation to some extent [4].
Firstly, the concept of uniform continuity needs to be introduced. A mathematical
formal definition of uniform continuity of a function is given as follows:
Definition 4.12 The function f (t) ∈ R is uniformly continuous on a set D if, for
any  > 0, there exists some δ () > 0 such that

|t2 − t1 | < δ ⇒ | f (t2 ) − f (t1 )| < , ∀t1 , t2 (4.48)


The following statements are equivalent to the definition of uniform continuity:
4.3 Stability of Non-Autonomous Systems 75

• Suppose a function f (t) is continuous on a closed finite interval t ∈ [t1 , t2 ]. Then,

f (t) is uniformly continuous on t ∈ [t1 , t2 ].
• Suppose a function
 f (t) is differentiable on a set D, and there exists a constant
M > 0 such that  f˙ (t) < M for all t. Then, f (t) is uniformly continuous on D.
In the simplest term, uniform continuity of a differentiable function f (t) requires
its derivative f˙ (t) to exist and be bounded.
Example 4.18
• The function f (t) = t 2 for all t ∈ [0, ∞) is continuous but is not uniformly
continuous since f˙ (t) is not bounded for all t ∈ [0, ∞).
• The function f (t) = t 2 for t ∈ [0, 1] is uniformly continuous since f (t) is
continuous for t ∈ [0, 1].√
• The function f (t) = t for all t ∈ [0, ∞) does not have a bounded derivative
f˙ (t) = 2√1
t
for all t ∈ [0, ∞), and since the interval is semi-open and infinite,
one cannot readily conclude that f (t) is not uniformly continuous on t ∈ [0, ∞).
However, this function is actually uniformly continuous on t ∈ [0, ∞). To see
this, we note that the interval can be divided into two subintervals t ∈ [0, a] and
t ∈ [a, ∞) where a > 0. Then, f (t) is uniformly continuous on t ∈ [0, a]
since f (t) is continuous on t ∈ [0, a], and furthermore f (t) is also uniformly
continuous on t ∈ [a, ∞) since f˙ (t) = 2√ 1
t
is bounded on t ∈ [a, ∞). Therefore,
in totality, f (t) is uniformly continuous on t ∈ [0, ∞).
• Consider a stable LTI system
ẋ = Ax + Bu

with x(t0 ) = x0 and a continuous bounded input u (t). The system is exponentially
stable with the solution
 t
x = e−A(t−t0 ) x0 + e A(t−τ ) Bu (τ ) dτ
t0

Thus, x (t) is a continuous bounded signal with a bounded derivative ẋ (t) for all
t ∈ [0, ∞). Therefore, x (t) is uniformly continuous. Any output signal

y = C x + Bu

is also uniformly continuous if u (t) has a bounded derivative. The system is then
said to be bounded-input-bounded-output (BIBO) stable.

The Barbalat’s lemma is now stated as follows:

Lemma 4.1 If the limit of a differentiable function f (t) as t → ∞ exists and is

finite, and if f˙ (t) is uniformly continuous, then limt→∞ f˙ (t) = 0.
76 4 Lyapunov Stability Theory

Example 4.19
• The function f (t) = e−t has a finite limit as t → ∞. To determine the uniform
2

continuity of the first derivative f˙ (t) = −2te−t , we need to determine whether or

2

not the second derivative f¨ (t) is bounded for all t ∈ [0, ∞). The second derivative
f¨ (t) = −2e−t + 4t 2 e−t , in fact, is bounded because the exponential term e−t
2 2 2

decreases at a much faster rate than the power term t 2 . Therefore, limt→∞ f˙ (t) =
−2te−t = 0 using the L’Hospital rule.
2
0. In fact, one can verify lim 
t→∞
• The function f (t) = 1t sin t 2 which tends to zero as t → ∞ but whose derivative
   
f˙ (t) = − t12 sin t 2 + 2 cos t 2 does not have a limit as t → ∞. Thus, it can be
seen that even if the limit of a differentiable function f (t) exists and is finite as
t → ∞, it does not necessarily imply that limt→∞ f˙ (t) = 0 since f˙ (t) may not
be uniformly continuous, that is, f˙ (t) has a bounded derivative
  or equivalently
f¨ (t) is bounded. Therefore, the function f (t) = 1t sin t 2 does not satisfy the
Barbalat’s lemma.
• The function f (t) = sin (ln t) whose derivative f˙ (t) = 1t cos (ln t) tends to zero
but f (t) does not have a finite limit as t → ∞. Thus, limt→∞ f˙ (t) = 0 does
not necessarily imply that the limit of a differentiable function f (t) exists and is
finite. Therefore, the converse of the Barbalat’s lemma is not true.

The Barbalat’s lemma is now extended to the Lyapunov’s direct method to examine
the asymptotic stability of a non-autonomous system by the following Lyapunov-like
lemma [4]:

Lemma 4.2 If a positive-definite function V (x, t) has a finite limit as t → ∞, and

if V̇ (x, t) is negative semi-definite and uniformly continuous for all t ∈ [0, ∞), then
V̇ (x, t) → 0 as t → ∞.

Example 4.20 Consider a simple adaptive control system

 
ẋ = −ax + b u + θ ∗ w (t)

where a > 0, w (t) ∈ L∞ is a bounded time-varying disturbance, and θ ∗ is an

unknown constant parameter.
To cancel out the effect of the time-varying disturbance, an adaptive controller is
designed as
u = −θ (t) w (t)

where θ (t) is an adaptive parameter that estimates θ ∗ .

If θ (t) → θ ∗ as t → ∞, then the adaptive controller perfectly cancels out the
disturbance and the closed-loop system tends to an ideal reference model

ẋm = −axm

where xm (t) is the desired response of x (t).

4.3 Stability of Non-Autonomous Systems 77

The adaptive parameter is computed as

θ̇ = −bew (t)

where e (t) = xm (t) − x (t) is called a tracking error, described by the tracking error
equation
ė = ẋm − ẋ = −ae + bθ̃ w (t)

where θ̃ (t) = θ (t) − θ ∗ is the parameter estimation error.

The combined system is non-autonomous due to w (t). Both the variables e (t)
and θ (t) are influenced by the tracking error equation ė (t) and the adaptive law
θ̇ (t). To show that the system is stable, choose the following Lyapunov candidate
function that includes both the variables e (t) and θ̃ (t):

V (e, θ ) = e2 + θ̃ 2

Then,
   
V̇ e, θ̃ = 2e −ae + bθ̃ w (t) + 2θ̃ [−bew (t)] = −2ae2 ≤ 0

Since V̇ (e, θ ) is negative semi-definite, e (t) ∈ L∞ and θ (t) ∈ L∞ , i.e., they

are bounded, but the LaSalle’s invariant set theorem cannot be used to show that
the tracking error e (t)
 converges
 to zero. This is where the Barbalat’s lemma comes
in handy. Firstly, V e, θ̃ must be shown to have a finite limit as t → ∞. Since
 
V̇ e, θ̃ ≤ 0, then

     ∞  
V e (t → ∞) , θ̃ (t → ∞) − V e (t0 ) , θ̃ (t0 ) = V̇ e, θ̃ dt
t0
 ∞
= −2a e2 (t) dt = −2a e22
t0

   
V e (t → ∞) , θ̃ (t → ∞) = V e (t0 ) , θ̃ (t0 ) − 2a e22
= e2 (t0 ) + θ̃ 2 (t0 ) − 2 e22 < ∞
 
So, V e, θ̃ has a finite limit as t → ∞. Since e2 exists, therefore e (t) ∈
L2 ∩ L∞ .  
Next, V̇ e, θ̃ must be shown to be uniformly continuous. This can be done by
   
examining the derivative of V̇ e, θ̃ to see if it is bounded. V̈ e, θ̃ is computed as
78 4 Lyapunov Stability Theory
   
V̈ e, θ̃ = −4ae −ae + bθ̃ w (t)
 
Since e (t) ∈ L2 ∩ L∞ and θ̃ (t) ∈ L∞ by the virtue of V̇ e, θ̃ ≤ 0, and
   
w (t) ∈ L∞ by assumption, then V̈ e, θ̃ ∈ L∞ . Therefore, V̇ e, θ̃ is uniformly
 
continuous. It follows from the Barbalat’s lemma that V̇ e, θ̃ → 0 and hence
e (t) → 0 as t → ∞. Note that one cannot conclude that the system is asymptotically
stable since only e (t) → 0 as t → ∞, but θ̃ (t) is only bounded.

4.4 Summary

The Lyapunov stability theory is the foundation of nonlinear systems and adap-
tive control theory. Various stability concepts for autonomous and non-autonomous
systems are introduced. The Lyapunov’s direct method is an indispensable tool
for analyzing stability of nonlinear systems. Barbashin–Krasovskii theorem pro-
vides a method for global stability analysis. LaSalle’s invariant set theorem provides
another complementary tool for analyzing systems with invariant sets. Stability of
non-autonomous systems involves the concepts of uniform stability, uniform bound-
edness, and uniform ultimate boundedness. Barbalat’s lemma is an important math-
ematical tool for analyzing stability of adaptive control systems in connection with
the concept of uniform continuity of a real-valued function.

4.5 Exercises

1. Given      
ẋ1 x1 x12 + x22 − 1 − x2
=
ẋ2 x1 + x2 x12 + x22 − 1

a. Determine all the equilibrium points of the system and linearize the system
about the equilibrium points to classify the types of equilibrium points.
b. Use the Lyapunov candidate function

V (x) = x12 + x22

to determine the types of Lyapunov stability of the equilibrium points and

their corresponding regions of attraction, if any.
2. Given 
1
ẋ = x −1 + sin x
2
4.5 Exercises 79

subject to x (0) = 1.
a. Determine the upper and lower bound solutions.
b. Use the Lyapunov candidate function

V (x) = x 2

to determine the type of Lyapunov stability and the upper bound of V (x)
as an explicit function of time.
3. Use the Lyapunov candidate function

V (x) = x12 + x22

to study stability of the origin of the system

    
ẋ1 (x2 − x1 ) x12 + x22 
=
ẋ2 (x1 + x2 ) x12 + x22

4. Given
ẋ = Ax

a. Calculate analytically P that solves

A P + P A = −2I

where  
0 1
A=
−4 4

and verify the result using the MATLAB function “lyap.”

b. Determine if P is positive or negative (semi-)definite. What can be said
about stability of the origin of this system.
5. Given      
ẋ1 x1 1 − x12 − x22 + x2 
=
ẋ2 −x1 + x2 1 − x12 − x22

a. Use the Lyapunov candidate function

V (x) = x12 + x22

to determine the type of Lyapunov stability of the origin.

b. Find an invariant set.
c. Solve for V (t) as an explicit function of time and plot the trajectories of
V (t) for V (0) = 0.01, 0.5, 1, 1.5, 2.
80 4 Lyapunov Stability Theory

6. Given ⎡ ⎤
0 1 0
A = ⎣ −1 −1 −2 ⎦
1 0 −1

Determine whether or not A is Hurwitz. If so, compute P using the Euler method
to integrate the differential Lyapunov equation

dP
= P A + A P + I
dτ
subject to P (0) = 0, where τ is time-to-go. Plot all six elements of P on the
same plot and verify the result at the final time-to-go with the MATLAB function
“lyap.”
7. Use the Lyapunov’s direct method to determine an ultimate bound of the solution
x (t) for the following equation:

ẋ = −x + cos t sin t

subject to x (0) = 1. Plot the solution x (t) for 0 ≤ t ≤ 20.

8. Given a non-autonomous system

ẋ = (−2 + sin t) x − cos t

a. Show that the system is uniformly ultimately bounded by the Lyapunov

theorem for non-autonomous systems. Also determine the ultimate bound
of x.
b. Plot the solution by numerically integrating the differential equation and
show that it satisfies the ultimate bound.
9. Given  
ẋ = − 1 + sin2 t x + cos t

a. Use the Lyapunov candidate function

V (x) = x 2

√ the upper bound of V̇ (x) as a function of V (x).

to determine
b. Let W = V . Solve for the inequality W (t) as an explicit function of time
and determine the ultimate bound of the system.
c. Show that the system is uniformly ultimately bounded.
10. For the following functions:
 2
a. f (t) = sin e−t
b. f (t) = e− sin
2
t
4.5 Exercises 81

Plot f (t) for t ∈ [0, 5]. Determine whether or not the limit of f (t) exists as
t → ∞ and f˙ (t) is uniformly continuous. If so, use the Barbalat’s lemma to
show that f˙ (t) → 0 as t → ∞ and verify by taking the limit of f˙ (t) as t → ∞.
11. Consider the following adaptive control system:

ė = −e + θ x

θ̇ = −xe

where e (t) = xm (t) − x (t) is defined as the tracking error between a given
explicit reference time signal xm (t) which is assumed to be bounded, i.e.,
xm (t) ∈ L∞ , and the state variable x (t). Show that the adaptive system is
stable and that e (t) → 0 as t → ∞.

References

1. Åström, K. J. (2008). and Wittenmark B. Adaptive Control: Dover Publications Inc.

2. Khalil, H. K. (2001). Nonlinear systems. Upper Saddle River: Prentice-Hall.
3. Ioannu, P. A., & Sun, J. (1996). Robust Adaptive Control. Upper Saddle River: Prentice-Hall,
Inc.
4. Slotine, J.-J., & Li, W. (1991). Applied nonlinear control. Upper Saddle River: Prentice-Hall,
Inc.
5. Anderson, B., & Moore, J. (1971). Linear optimal control. Upper Saddle River: Prentice-Hall,
Inc.
6. Bryson, A. E., & Ho, Y. C. (1979). Applied optimal control: optimization, estimation, and
control. New Jersey: Wiley Inc.
Chapter 5
Model-Reference Adaptive Control

Abstract This chapter presents the fundamental theory of model-reference adaptive

control. Various types of uncertainty are defined. The composition of a model-
reference adaptive control system is presented. Adaptive control theory for first-
order single-input single-output (SISO) systems, second-order SISO systems, and
multiple-input multiple-output (MIMO) systems is presented. Both direct and indi-
rect adaptive control methods are discussed. The direct adaptive control methods
adjust the control gains online directly, whereas the indirect adaptive control meth-
ods estimate unknown system parameters for use in the update of the control gains.
Asymptotic tracking is the fundamental property of model-reference adaptive con-
trol which guarantees that the tracking error tends to zero in the limit. On the other
hand, adaptive parameters are only bounded in the model-reference adaptive control
setting.

When designing a controller for a system, a control designer typically would like
to know how the system behaves physically. This knowledge is usually captured
in the form of a mathematical model. For many real-world applications, modeling
of physical systems can never be perfect as systems may have parameter variations
due to nonlinearity, parameter uncertainty due to modeling inaccuracy or imprecise
measurements, uncertainty in exogenous disturbances coming from the operating
environment, or other sources of uncertainty. The role of a modeling specialist is
to reduce the system uncertainty as much as practicable. The control designer then
uses a mathematical model of the system to design a controller which may incorpo-
rate performance measures and stability margins to account for any residual system
uncertainty that cannot be completely accounted for.
In situations when the system uncertainty may become significant beyond a level
of desired tolerance that can adversely affect the performance of a controller, adaptive
control can play an important role in reducing the effects of the system uncertainty on
the controller performance. Situations that may warrant the use of adaptive control
could include unintended consequences of off-nominal modes of operation such
as system failures or highly uncertain operating conditions, and complex system
behaviors that can result in an increase in the complexity and hence cost of the
modeling efforts. In this chapter, the learning objectives are as follows:

© Springer International Publishing AG 2018 83

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_5
84 5 Model-Reference Adaptive Control

• To develop a basic understanding of system uncertainty and the composition of a

typical model-reference adaptive control system and its functionality;
• To be able to apply various model-reference adaptive control techniques for direct
and indirect adaptation for first-order, second-order, and MIMO systems;
• To be able to perform a Lyapunov stability proof of model-reference adaptive
control using the Lyapunov’s direct method and Barbalat’s lemma; and
• To recognize that model-reference adaptive control achieves asymptotic tracking
but only boundedness of adaptive parameters.

A typical model-reference adaptive control (MRAC) system block diagram is
shown in Fig. 5.1.
There are generally two classes of adaptive control: (1) direct adaptive control and
(2) indirect adaptive control [1, 2]. Adaptive control architectures that combine both
types of adaptive control are also frequently used and are referred to as composite [2,
3], combined, or hybrid direct–indirect adaptive control [4]. A typical direct adaptive
controller may be expressed as
u = k x (t) x + kr (t) r (5.1)

where k x (t) and kr (t) are adjustable control gains. The mechanism to adjust these
control gains is via an adaptive law. Thus, a direct adaptive control in effect adjusts a
feedback control mechanism of a control system directly to cancel out any unwanted
system uncertainty so that the performance of the control system can be regained in
the presence of significant system uncertainty.
In contrast, an indirect adaptive controller achieves the same objective by adjusting
the control gains in an indirect way, which may be expressed as
u = k x ( p (t)) x + kr ( p (t)) r (5.2)

where p (t) are system parameters that are estimated online to update the control
gains.

Fig. 5.1 A model-reference adaptive control system

5 Model-Reference Adaptive Control 85

Example 5.1 Consider a second-order LTI system

ẍ + 2ζ ωn ẋ + ωn2 x = u

A controller is designed to enable the output x (t) to track a constant command

r (t) and meet performance specifications for a closed-loop damping of ζm and a
bandwidth frequency of ωm . A PD (proportional-derivative) controller is designed
as
u = k p e + kd ė + kr r

where e (t) = r (t) − x (t), to achieve the following closed-loop characteristic:

 
ẍ + (2ζ ωn + kd ) ẋ + ωn2 + k p x = (kr + k p )r

Choose
ωn2 + k p = ωm2 ⇒ k p = ωm2 − ωn2

2ζ ωn + kd = 2ζm ωm ⇒ kd = 2 (ζm ωm − ζ ωn )

kr = ωn2

Then, the output x (t) tracks the command r (t) as t → ∞.

Suppose the open-loop system natural frequency ωn suddenly changes to a new
value ωn∗ which may be unknown. The closed-loop system can now have a drastically
different performance if the original control gains are used. To maintain the same
performance specifications, the ideal control gains must be

k ∗p = ωm2 − ωn∗2 = k p + ωn2 − ωn∗2

   
kd∗ = 2 ζm ωm − ζ ωn∗ = kd + 2ζ ωn − ωn∗

kr∗ = ωn∗2 = kr − ωn2 + ωn∗2

A direct adaptive controller seeks to adjust the original control gains k p , kd ,

and kr toward the ideal control gains k ∗p , kd∗ , and kr∗ , respectively, directly without
knowing the value of the uncertain parameter ωn∗ . On the other hand, an indirect
adaptive controller seeks to adjust the control gains by estimating online the uncertain
parameter ωn∗ and use the estimate of ωn∗ to re-compute the control gains as if the
parameter estimate is the true value. This approach is often referred to as the Certainty
Equivalence Principle.
86 5 Model-Reference Adaptive Control

5.1 Composition of a Model-Reference Adaptive Control

System

5.1.1 Uncertain Plant

Adaptive control can deal with either linear or nonlinear plants with various types of
uncertainty which can be structured uncertainty, unstructured uncertainty, or unmod-
eled dynamics.
1. Structured uncertainty is a source of uncertainty with uncertain parameters but
known functional characteristics. It is also often referred to as parametric uncer-
tainty.
Example 5.2 A linear spring-mass-damper system with an uncertain spring constant

m ẍ + c ẋ + k ∗ x = u

where k ∗ is an uncertain parameter, is a system with structured or parametric uncer-

tainty. The function x (t) associated with k ∗ is a known characteristic that appears in
the structured uncertainty.
2. Unstructured uncertainty is a source of uncertainty for which neither parameters
or functional characteristics are certain.
Example 5.3 A spring-mass-damper system with an uncertain spring characteristic
 
m ẍ + c ẋ + f ∗ x, k ∗ = u

where f ∗ () is an uncertain function, is a system with unstructured uncertainty.

3. Unmodeled dynamics is a source of uncertainty that represents system internal
or external dynamics that are not included in a plant model because they may be
unmeasurable, unobservable, or assumed incorrectly to be negligible.
Example 5.4 The following linear spring-mass-damper system has unmodeled
dynamics:
m ẍ + c ẋ + k ∗ x = c1 y + u

ẏ = c2 x + c3 y

where y (t) is an internal system state that is not modeled, and ci , i = 1, 2, 3 are
parameters.
4. Matched uncertainty is a type of structured uncertainty that can be matched by
the control input for a class of MIMO linear affine-in-control systems of the form
 
ẋ = f (x) + B u + Θ ∗ Φ (x) (5.3)
5.1 Composition of a Model-Reference Adaptive Control System 87

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, B ∈ Rn × Rm

is a control input matrix, Θ ∗ ∈ R p × Rm is a matrix of uncertain parameters, and
Φ (x) ∈ R p is a known bounded regressor function.
The quantity Θ ∗ Φ (x) is called a parametric matched uncertainty since it appears
in the range space of the control input matrix B. Recall from linear algebra that a
range or column space of a matrix B consists of all possible products Bu. When a
parametric uncertainty is matched, the control input can cancel out the uncertainty
completely when the adaptation is perfect.

Example 5.5 The parametric uncertainty in the following LTI system:

        
ẋ1 −1 1 x1 1   x1
= + u + δ 1 δ2
ẋ2 −1 −2 x2 1 x2

is a matched uncertainty where δ1 and δ2 are uncertain parameters.

5. Unmatched uncertainty is a type of uncertainty that cannot be matched by the

control input for a class of MIMO linear affine-in-control systems of the form

ẋ = f (x) + Bu + Θ ∗ Φ (x) (5.4)

A parametric uncertainty cannot be matched if the control input matrix B ∈

Rn × Rm is a non-square “tall” matrix, i.e., n > m, or if B ∈ Rn × Rn is a rank-
deficient square matrix such that its inverse does not exist. In such a case, the
control input cannot completely cancel out the uncertainty by adaptive control.
Otherwise, the uncertainty may be cast as a matched uncertainty by the following
pseudo-inverse transformation:

 −1 ∗ 
ẋ = f (x) + B u + B  B B  Θ Φ (x) (5.5)

 −1
where B  B B  is the right pseudo-inverse of a full-rank non-square “wide”
 −1
matrix B ∈ R × R with n < m and rank (B) = n, or where B  B B 
n m
=
B −1 for a full-rank square matrix B.

Example 5.6 The parametric uncertainty in the following LTI system:

        
ẋ1 −1 1 x1 1 δ11 x1 + δ12 x2
= + u+
ẋ2 −1 −2 x2 1 δ21 x1 + δ22 x2

is an unmatched uncertainty since B is a “tall” matrix. Intuitively, a tall matrix B

implies the number of control inputs is less than the number state variables. Therefore,
it would be difficult for the control inputs to cancel out the uncertainty in all the state
variables.
88 5 Model-Reference Adaptive Control

Example 5.7 The parametric uncertainty in the following LTI system:

⎤ ⎡
      
u1   
ẋ1 −1 1 x1 1 −1 2 ⎣ ⎦ δ δ x1
= + u 2 + 11 12
ẋ2 −1 −2 x2 1 2 −1 δ21 δ22 x2
u3

is actually a matched uncertainty since B is a full-rank “wide” matrix whose pseudo-

inverse exists ⎡ ⎤
  11
−1 1
B B B = ⎣0 1⎦
3 10

So, the system can be cast as

⎛⎡ ⎤ ⎡ ⎤ ⎞
       u δ + δ21 δ12 + δ22  
ẋ1 −1 1 x1 1 −1 2 ⎝⎣ 1 ⎦ 1 ⎣ 11 ⎦ x 1 ⎠
= + u2 + δ21 δ22
ẋ2 −1 −2 x2 1 2 −1 3 x2
u3 δ11 δ12

6. Control input uncertainty is a type of uncertainty that exists in the control input
matrix for a class of MIMO linear affine-in-control systems of the form

ẋ = f (x) + BΛu (5.6)

where Λ is a positive diagonal matrix whose diagonal elements represent the

control input effectiveness uncertainty which can be in the amplitude or in the
sign or both. When the uncertainty is in the amplitude, a control saturation can
occur and may worsen the performance of a controller. When the uncertainty is
in the sign, a control reversal can occur and potentially can cause instability.
An alternative form of a control input uncertainty is given by

ẋ = f (x) + (B + ΔB) u (5.7)

which is less common in adaptive control.

5.1.2 Reference Model

A reference model is used to specify a desired response of an adaptive control system

to a command input. It is essentially a command shaping filter to achieve a desired
command following. Since adaptive control is formulated as a command following
or tracking control, the adaptation is operated on the tracking error between the
reference model and the system output. A reference model must be designed properly
for an adaptive control system to be able to follow. Typically, a reference model is
formulated as a LTI model, but a nonlinear reference model can be used although
a nonlinear design always brings up many complex issues. A LTI reference model
5.1 Composition of a Model-Reference Adaptive Control System 89

should capture all important performance specifications such as rise time and settling
time, as well as robustness specifications such as phase and gain stability margins.
Example 5.8 For Example 5.1, the reference model for an adaptive control system
could be selected to be a second-order system as

ẍm + 2ζm ωm ẋm + ωm2 xm = ωm2 r

where xm (t) is a model-reference signal that only depends on the reference command
input r (t). 
The tracking error is defined as

e = xm − x (5.8)

The objective of an adaptive control system is to adapt to system uncertainty so as

to keep the tracking error as small as possible. In an ideal case when e (t) → 0, then
the system state follows the model-reference signal perfectly, i.e., x (t) → xm (t).

5.1.3 Controller

A controller must be designed to provide overall system performance and stability

for a nominal plant without uncertainty. Thus, it can be thought of as a baseline or
nominal controller. The type of controllers is dictated by the objective of a control
design. A controller can be linear or nonlinear but as always nonlinear controllers
are much more difficult to design, analyze, and ultimately certify for operation in
real systems. The controller can be a nominal controller augmented with an adaptive
controller or a fully adaptive controller. The adaptive augmentation control design
is more prevalent and generally should be more robust than a fully adaptive control
design.

5.1.4 Adaptive Law

An adaptive law is a mathematical relationship that expresses explicitly how adaptive

parameters should be adjusted to keep the tracking error as small as possible. An
adaptive law can be either linear time-varying or nonlinear. In any case, stability of
an adaptive control system usually must be analyzed using Lyapunov stability theory.
Many different adaptive laws have been developed, and each has its own advantages
as well as disadvantages. Ultimately, designing an adaptive control system comes
down to a trade-off between performance and robustness. This trade-off can be made
by a suitable selection of an adaptive law and a set of tuning parameters that are built
into an adaptive law.
90 5 Model-Reference Adaptive Control

5.2 Direct MRAC for First-Order SISO Systems

Consider a first-order nonlinear SISO system

ẋ = ax + b [u + f (x)] (5.9)

subject to x (0) = x0 , where f (x) is a structured matched uncertainty that can be

linearly parametrized as


p
f (x) = θi∗ φi (x) = Θ ∗ Φ (x) (5.10)
i=1

 
where Θ ∗ = θ1 θ2 . . . θ p ∈ R p is an unknown constant vector, and Φ (x) =
 
φ1 (x) φ2 (x) . . . φ p (x) ∈ R p is a vector of known bounded basis functions.

5.2.1 Case I: a and b Unknown but Sign of b Known

A reference model is specified as

ẋm = am xm + bm r (5.11)

subject to xm (0) = xm 0 , where am < 0 and r (t) ∈ L∞ is a piecewise continuous

bounded reference command signal, so that xm (t) is a uniformly bounded model-
reference signal.
Firstly, define an ideal controller that perfectly cancels out the uncertainty and
enables x (t) to follow xm (t) as

u ∗ = k x∗ x + kr∗r (t) − Θ ∗ Φ (x) (5.12)

where the superscript ∗ denotes ideal constant values which are unknown.
Upon substituting into the plant model, we get the ideal closed-loop plant
 
ẋ = a + bk ∗ x + bkr∗r (5.13)

Comparing the ideal closed-loop plant to the reference model, the ideal gains k x∗
and kr∗ can be determined by the following model matching conditions:

a + bk x∗ = am (5.14)

bkr∗ = bm (5.15)
5.2 Direct MRAC for First-Order SISO Systems 91

It turns out that the solutions for k x∗ and kr∗ always exist since there are two
independent equations with two unknowns.
The actual adaptive controller is an estimate of the ideal controller with a goal
that in the limit the adaptive controller approaches the ideal controller. Let

u = k x (t) x + kr (t) r − Θ  (t) Φ (x) (5.16)

be the adaptive controller, where k x (t), kr (t), and Θ (t) are the estimates of k x∗ , kr∗ ,
and Θ ∗ , respectively.
The adaptive controller is a direct adaptive controller since k x (t), kr (t), and Θ (t)
are estimated directly without the knowledge of the unknown system parameters a,
b, and Θ ∗ .
Now, define the estimation errors as

k̃ x (t) = k x (t) − k x∗ (5.17)

k̃r (t) = kr (t) − kr∗ (5.18)

Θ̃ (t) = Θ (t) − Θ ∗ (5.19)

Substituting these into the plant model gives

⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟
ẋ = ⎝a + bk x∗ +bk̃ x ⎠ x + ⎝ bkr∗ +bk̃r ⎠ r − bΘ̃  Φ (x) (5.20)
   
am bm

Let e (t) = xm (t) − x (t) be the tracking error. Then, the closed-loop tracking
error equation is established as

ė = ẋm − ẋ = am e − bk̃ x x − bk̃r r + bΘ̃  Φ (x) (5.21)

Note that the tracking error equation is non-autonomous due to r (t).

Now, the task of defining the adaptive laws to adjust k x (t), kr (t), and Θ (t) is
considered next. This can be accomplished by conducting a Lyapunov stability proof
as follows:

Proof Choose a Lyapunov candidate function

 
  k̃ 2 k̃ 2
V e, k̃ x , k̃r , Θ̃ = e2 + |b| x + r + Θ̃  Γ −1 Θ̃ >0 (5.22)
γx γr

where γx > 0 and γr > 0 are called the adaptation (or learning) rates for k x (t) and
kr (t), and Γ = Γ  > 0 ∈ R p × R p is a positive-definite adaptation rate matrix for
Θ (t).
92 5 Model-Reference Adaptive Control
 
V̇ e, k̃ x , k̃r , Θ̃ is evaluated as

 
  2k̃ x k̃˙ x 2k̃r ˜k˙r  −1 ˙
V̇ e, k̃ x , k̃r , Θ̃ = 2eė + |b| + + 2Θ̃ Γ Θ̃
γx γr
   
˙
k̃ ˙
k̃
x r
= 2am e2 + 2k̃ x −ebx + |b| + 2k̃r −ebr + |b|
γx γr


+ 2Θ̃  ebΦ (x) + |b| Γ −1 Θ̃˙ (5.23)
 
Since b = |b| sgnb, then V̇ e, k̃ x , k̃r , Θ̃ ≤ 0 if

k̃˙ x
− exsgnb + =0 (5.24)
γx

k̃˙r
− er sgnb + =0 (5.25)
γr

eΦ (x) sgnb + Γ −1 Θ̃˙ = 0 (5.26)

Because k x∗ , kr∗ , and Θ ∗ are constant, therefore from Eqs. (5.17)–(5.19) k̃˙ x = k̇ x ,
k̃˙r = k̇r , and Θ̃˙ = Θ̇. Thus, the following adaptive laws are obtained:

k̇ x = γx xesgnb (5.27)

k̇r = γr r esgnb (5.28)

Θ̇ = −Γ Φ (x) esgnb (5.29)

Then,  
V̇ e, k̃ x , k̃r , Θ̃ = 2am e2 ≤ 0 (5.30)
 
Since V̇ e, k̃ x , k̃r , Θ̃ ≤ 0, then e (t), k x (t), kr (t), and Θ (t) are bounded. Then,
   
lim V e, k̃ x , k̃r , Θ̃ = V e0 , k̃ x0 , k̃r0 , Θ̃0 + 2am e 22 (5.31)
t→∞

 = e0 , k̃ x (0)
where e (0)  = k̃ x0 , k̃r (0) = k̃r0 , and Θ̃ (0) = Θ̃0 .
So, V e, k̃ x , k̃r , Θ̃ has a finite limit as t → ∞. Since e 2 exists, therefore
e (t) ∈ L2 ∩ L∞ , but ė ∈ L∞ .
5.2 Direct MRAC for First-Order SISO Systems 93
 
V̇ e, k̃ x , k̃r , Θ̃ can be shown to be uniformly continuous by examining its deriv-
 
ative to see whether it is bounded. V̈ e, k̃ x , k̃r , Θ̃ is computed as
 

V̈ e, k̃ x , k̃r , Θ̃ = 4am eė = 4am e am e − bk̃ x x − bk̃r r + bΘ̃  Φ (x) (5.32)
 
Since e (t), k x (t), kr (t), and Θ (t) are bounded by the virtue that V̇ e, k̃ x , k̃r , Θ̃
≤ 0, x (t) is bounded because e (t) and xm (t) are bounded, r (t) is a bounded ref-
erence
 command  signal, and Φ (x) is bounded because  x (t) is bounded; therefore,
V̈ e, k̃ x , k̃r , Θ̃ is bounded. Thus, V̇ e, k̃ x , k̃r , Θ̃ is uniformly continuous. It fol-
 
lows from the Barbalat’s lemma that V̇ e, k̃ x , k̃r , Θ̃ → 0, hence e (t) → 0 as
t → ∞. The tracking error is asymptotically stable, but the whole adaptive control
system is not asymptotically stable since k x (t), kr (t) and Θ (t) can only be shown
to be bounded.

5.2.2 Case II: a and b Known

If a and b are known, then the gains k x and kr need not be estimated since they are
known and can be found from the model matching conditions

am − a
kx = (5.33)
b
bm
kr = (5.34)
b
Then, the adaptive controller is given by

u = k x x + kr r − Θ  (t) Φ (x) (5.35)

where only Θ (t) needs to be adjusted.

The tracking error equation is then obtained as

ė = am e + bΘ̃  Φ (x) (5.36)

The adaptive law can be found to be

Θ̇ = −Γ Φ (x) eb (5.37)

Proof To show that the adaptive law is stable, choose a Lyapunov candidate function
94 5 Model-Reference Adaptive Control
 
V e, Θ̃ = e2 + Θ̃  Γ −1 Θ̃ > 0 (5.38)

Then,
 
V̇ e, Θ̃ = 2eė + 2Θ̃  Γ −1 Θ̃˙ = 2am e2 + 2ebΘ̃  Φ (x) − 2Θ̃  ebΦ (x) = 2am e2 ≤ 0
(5.39)

The Barbalat’s lemma can be used to show that the tracking error is asymptotically
stable, i.e., e (t) → 0 as t → ∞.

Example 5.9 Let a = 1, b = 1, am = −1, bm = 1, r (t) = sin t, and f (x) =

θ ∗ x (t), where θ ∗ is an unknown constant but for simulation purposes is taken to be
θ ∗ = 0.1. Then, the control gains are computed as

am − a
kx = = −2
b
bm
kr = =1
b
The adaptive controller is given by

u = −2x + r − θ x

θ̇ = −γ xeb

where γ = 1 is chosen as the adaptation rate for θ (t).

Note that the controller is nonlinear even though the plant is linear. Figure 5.2
illustrates a block diagram of the adaptive controller.

Fig. 5.2 Adaptive control block diagram

5.2 Direct MRAC for First-Order SISO Systems 95

0.02 1

0 0.5

x, xm
e −0.02 0

−0.04 −0.5
x xm
−0.06 −1
0 10 20 30 40 50 0 10 20 30 40 50
t t

2 0.1

0 0.05
u

θ
−1 θ
*
θ
−2 0
0 10 20 30 40 50 0 10 20 30 40 50
t t

Fig. 5.3 Adaptive control system response

The results are shown in Fig. 5.3. It can be seen that e (t) → 0 and x (t) → xm (t)
as t → ∞. The estimate θ (t) also converges to the correct value of the uncertain
parameter θ ∗ , although the convergence is quite gradual. The convergence rate can
be increased by increasing the adaptation rate γ , but a large value of γ can lead to
an increase in the sensitivity of the control system to noise and unmodeled dynamics
that can lead to instability. In other words, a larger value of γ results in a better
tracking performance but at the same time degrades the system robustness. In a
practical design, the adaptation rate must be chosen carefully to maintain a sufficient
robustness while achieving a desired level of performance.

5.3 Indirect MRAC for First-Order SISO Systems

Consider the system in Sect. 5.2.1 with a and b unknown, but sign of b is known.
From the model matching conditions, if a and b can be estimated, then the gain k x
and kr can be obtained. Therefore, the objective of indirect adaptive control is to
estimate system parameters which are then used to update the control gains. Hence,
indirect adaptive control is essentially a parameter identification method.
Let
am − â (t)
k x (t) = (5.40)
b̂ (t)

bm
kr (t) = (5.41)
b̂ (t)
96 5 Model-Reference Adaptive Control

Let ã (t) = â (t) − a and b̃ (t) = b̂ (t) − b be the estimation errors. Now, the plant
model is expressed as
 


ẋ = ax + b̂ − b̃ u + Θ ∗ Φ (x) (5.42)

Then, substituting Eqs. (5.16), (5.40), and (5.41) into Eq. (5.42) yields
 
am − â bm 
∗
ẋ = ax + b̂ x+ r − Θ Φ (x) + Θ Φ (x)
b̂ b̂
 
am − â bm 
∗
− b̃ x+ r − Θ Φ (x) + Θ Φ (x)
b̂ b̂

 am − â bm
= (am − ã) x + bm r − bΘ̃ Φ (x) − b̃ x+ r (5.43)
b̂ b̂

Let
ū = k x (t) x + kr (t) r (5.44)

Then, the tracking error equation is established as


ė = ẋm − ẋ = am e + ãx + b̃ū + bΘ̃ Φ (x) (5.45)

The Lyapunov’s direct method is now used to find the adaptive laws as follows:
Proof Choose a Lyapunov candidate function

  ã 2 b̃2
V e, ã, b̃, Θ̃ = e2 + + + |b| Θ̃  Γ −1 Θ̃ > 0 (5.46)
γa γb

where γa >0 and γb >0 are the adaptation rates for â (t) and b̂ (t), respectively.
Then, V̇ e, ã, b̃, Θ̃ is evaluated as

  2ã ã˙ 2b̃b̃˙

V̇ e, ã, b̃, Θ̃ = 2eė + + + 2 |b| Θ̃  Γ −1 Θ̃˙
γa γb
   
ã˙ b̃˙
= 2am e + 2ã xe +
2
+ 2b̃ ūe +
γa γb


+ 2 |b| Θ̃ Φ (x) esgnb + Γ −1 Θ̃˙

(5.47)

˙ Thus, the adaptive laws are

Since a and b are constant, then ã˙ = â˙ and b̃˙ = b̂.
obtained as
â˙ = −γa xe (5.48)
5.3 Indirect MRAC for First-Order SISO Systems 97

b̂˙ = −γb ūe (5.49)

Θ̇ = −Γ Φ (x) esgnb (5.50)

Then,  
V̇ e, ã, b̃, Θ̃ = 2am e2 ≤ 0 (5.51)
 
Since V̇ e, ã, b̃, Θ̃ ≤ 0, then e (t), â (t), b̂ (t), and Θ (t) are bounded. Then,
   
lim V e, k̃ x , k̃r , Θ̃ = V e0 , ã0 , b̃0 , Θ̃0 + 2am e 22 (5.52)
t→∞

 Θ̃0 are defined previously, ã (0) = a0 , and b̃ (0) = b0 .

where e0 and
So, V e, ã, b̃, Θ̃ has a finite limit as t → ∞. Since e 2 exists, therefore
e (t) ∈ L2 ∩ L∞, but ė ∈ L∞ .
V̇ e, ã, b̃, Θ̃ can be shown to be uniformly continuous by examining its deriv-
 
ative to see whether it is bounded. V̈ e, ã, b̃, Θ̃ is computed as
 


V̈ e, ã, b̃, Θ̃ = 4am eė = 4am e am e + ãx + b̃ū + bΘ̃ Φ (x) (5.53)
 
Since e (t), â (t), b̂ (t), and Θ (t) are bounded by the virtue that V̇ e, ã, b̃, Θ̃ ≤
0, x (t) is bounded because e (t) and xm (t) are bounded, ū (t) is bounded because x (t)
is bounded and r (t) is a bounded reference  command
 signal, and Φ (x) is bounded 
because x (t) is bounded; therefore, V̈ e, ã, b̃, Θ̃ is bounded. Thus, V̇ e, ã, b̃, Θ̃
 
is uniformly continuous. It follows from the Barbalat’s lemma that V̇ e, ã, b̃, Θ̃ →
0, hence e (t) → 0 as t → ∞. The tracking error is asymptotically stable.


It should be noted that the possibility of b̂ (t) = 0 does exist, and in such a case,
the control gains will “blow up.” Thus, indirect MRAC may not be as robust as direct
MRAC. To prevent this from occurring, the adaptive law for b̂ (t) must be modified
so that the adaptation can take place in a closed subset of R that does not include
b̂ (t) = 0. One technique for modification is the projection method which assumes a
priori knowledge of b [2].
Suppose a lower bound on b is known, i.e., 0 < b0 ≤ |b|, then the adaptive law
can be modified by the projection method as
⎧ " " " " " "
" "
⎨ " " " " d "b̂"
−γb ūe if "b̂" > b0 or if "b̂" = b0 and ≥0
b̂˙ = dt (5.54)
⎩
0 otherwise
98 5 Model-Reference Adaptive Control

The projection method is essentially a constrained optimization which will be

discussed in detail in Chap. 9. In the simplest term, the projection method allows the
adaptation to take place in such a manner that an adaptive parameter will not violate
its a priori bound. A simple explanation of the modified adaptive law is as follows:
" " one knows that b is bounded from below by b0 , i.e., |b| ≥ b0 , then as
Suppose
" "
long as "b̂" > b0 , the unmodified adaptive law can be used normally. Now, suppose
" " " "
" " " " " "
" " d "b̂" d "b̂"
"b̂" = b0 , there are two cases to consider: dt < 0 and dt ≥ 0.
" "
" " " " " "
d "b̂" " " " "
1. If dt
< 0, then " b̂ " is decreasing and "b̂" < b0 at some time t + Δt, which would
" "
" " " "
" " d "b̂"
violate the constraint "b̂" ≥ b0 . Therefore, to satisfy the constraint on b, dt = 0.
" "
" " " " " "
d "b̂" " " " "
2. On the other hand, if dt ≥ 0, then "b̂" is non-decreasing and "b̂" ≥ b0 , so that
the unmodified adaptive law can be used normally.
" "
" "
The modified adaptive law thus guarantees that "b̂" will always be greater than or
equal to b0 .
 
Proof Because of the modification, V̇ e, ã, b̃, Θ̃ will no longer be the same and
" "
" " " "
" " d "b̂"
is now dependent on the conditions on "b̂" and dt . Thus,

 
  b̃˙
V̇ e, ã, b̃, Θ̃ = 2am e + 2b̃ ūe +
2
γb
⎧ " " " " " "
" "
⎪
⎨2am e2 ≤ 0 " " " " d "b̂"
if "b̂" ≥ b0 or if "b̂" = b0 and dt
≥0
= " " " "
" " (5.55)
⎪
⎩a e2 + 2b̃ūe if ""b̂"" = b and d "b̂" < 0
m 0 dt

" "
" " " "
" " d "b̂"
Consider the second case when "b̂" = b0 and dt < 0 for which the sign definite-
" "
  " "
d "b̂"
ness of V̇ e, ã, b̃, Θ̃ is still undefined. The condition dt < 0 gives
" "
" "
d "b̂"
˙
= b̂sgnb = −γb ūesgnb < 0 ⇒ ūesgnb > 0 (5.56)
dt
" "
" "
Since "b̂" = b0 , then
 
" " 
" "
2b̃ūe = 2 b̂ − b ūe = 2 "b̂" sgnb − |b| sgnb ūe = 2 (b0 − |b|) ūesgnb (5.57)
5.3 Indirect MRAC for First-Order SISO Systems 99

Since |b| ≥ b0 , which implies |b| = b0 + δ > 0 where δ ≥ 0, then ūesgnb > 0
implies
2b̃ūe = 2 (b0 − b0 − δ) ūesgnb = −2δ ūesgnb ≤ 0 (5.58)

Therefore,
 
V̇ e, ã, b̃, Θ̃ = 2am e2 + 2b̃ūe = 2am e2 − 2δ ūesgnb ≤ 2am e2 ≤ 0 (5.59)

Using the usual Barbalat’s lemma, one can conclude that e (t) → 0 as t → ∞.

5.4 Direct MRAC for Second-Order SISO Systems

Consider a second-order nonlinear SISO system

ÿ + 2ζ ωn ẏ + ωn2 y = b [u + f (y, ẏ)] (5.60)

where ζ and ωn are unknown and f (y, ẏ) = Θ ∗ Φ (y, ẏ) is defined in a manner
similar to Eq. (5.10).
 
Let x1 (t) = y (t), x2 = ẏ (t), and x (t) = x1 (t) x2 (t) ∈ R2 . The state-space
form of the system is  
ẋ = Ax + B u + Θ ∗ Φ (x) (5.61)

where    
0 1 0
A= , B = (5.62)
−ωn2 −2ζ ωn b
 
Let xm 1 (t) = ym (t), xm 2 = ẏm (t), and xm (t) = xm 1 (t) xm 2 (t) ∈ R2 . A
reference model is given by
ẋm = Am xm + Bm r (5.63)

where r (t) ∈ R is a bounded command signal, Am ∈ R2 ×R2 is Hurwitz and known,

and Bm ∈ R2 is also known.

5.4.1 Case I: A and B Unknown but Sign of b known

Firstly, the ideal controller is defined as

u ∗ = K x∗ x + kr∗r − Θ ∗ Φ (x) (5.64)

where K x∗ ∈ R2 and kr∗ ∈ R are constant but unknown ideal gains.

100 5 Model-Reference Adaptive Control

Comparing the ideal closed-loop plant to the reference model, the model matching
conditions are
A + B K x∗ = Am (5.65)

Bkr∗ = Bm (5.66)

Note that in general, one cannot always assume that K x∗ and kr∗ exist because A,
Am , B, and Bm may have different structures that do not allow the solutions for K x∗
and kr∗ to be determined. In most cases, if A and B are known, then K x∗ and kr∗ can
be designed by any standard non-adaptive control techniques to stabilize the closed-
loop system and enable it to follow a command. Then, Am and Bm can be computed
from A, B, K x∗ , and kr∗ .

Example 5.10 A second-order SISO system and a reference model are specified as
       
0 1 0 0 1 0
A= , B= , Am = , Bm =
−1 −1 1 −16 −2 2

Then, utilizing the pseudo-inverse, K x∗ and kr∗ can be solved as

   
 −1   0 1 0 1  
K x∗ = B B 
B (Am − A) = 0 1 − = −15 −1
−16 −2 −1 −1
 −1 
kr∗ = B  B B Bm = 2

Now, suppose  
1 1
Am =
−16 −2

Then, the solution of K x∗ is the same (verify!), but the model matching condition
is not satisfied since  
0 1
A + BK =  = Am
−16 −2

Thus, it is important to state an explicit assumption that there exist constant but
unknown K x∗ and kr∗ such that the model matching conditions are satisfied. For a
second-order SISO system, the model matching conditions are satisfied if Am and
Bm have the same structures as those of A and B, respectively.
A full-state feedback adaptive controller is designed as

u = K x (t) x + kr (t) r − Θ  Φ (x) (5.67)

where K x (t) ∈ R2 and kr (t) ∈ R.

5.4 Direct MRAC for Second-Order SISO Systems 101

Let K̃ x (t) = K x (t) − K x∗ , k̃r (t) = kr (t) − kr∗ , and Θ̃ (t) = Θ (t) − Θ ∗ be the
estimation errors, then the closed-loop plant becomes
⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟
ẋ = ⎝ A + B K x∗ +B K̃ x ⎠ x + ⎝ Bkr∗ +B k̃r ⎠ r − B Θ̃  Φ (x) (5.68)
   
Am Bm

The closed-loop tracking error equation is now obtained as

ė = ẋm − ẋ = Am e − B K̃ x x − B k̃r r + B Θ̃  Φ (x) (5.69)

where e (t) = xm (t) − x (t) ∈ R2 .

Proof To find the adaptive laws, choose a Lyapunov candidate function

 
  k̃ 2
V e, K̃ x , k̃r , Θ̃ = e Pe + |b| K̃ x Γx−1 K̃ x + r + Θ̃  ΓΘ−1 Θ̃ > 0 (5.70)
γr

where Γx = Γx > 0 ∈ R2 × R2 is a positive-definite adaptation rate matrix for

K x (t) and P = P  > 0 ∈ R2 × R2 that solves the following Lyapunov equation:

P Am + A
m P = −Q (5.71)

where Q = Q  > 0 ∈ R2× R2 .

Then, V̇ e, K̃ x , k̃r , Θ̃ is evaluated as

 
  2k̃r k̃˙r
 
V̇ e, K̃ x , k̃r , Θ̃ = ė Pe + e P ė + |b| 2 K̃ x Γx−1 K̃˙ x +  −1 ˙
+ 2Θ̃ ΓΘ Θ̃
γr
(5.72)
Substituting the tracking error equation yields
   

V̇ e, K̃ x , k̃r , Θ̃ = e P Am + A  
m P e + 2e P B − K̃ x x − k̃r r + Θ̃ Φ (x)
 
−1 ˙  2k̃r k̃˙r  −1 ˙
+ |b| 2 K̃ x Γx K̃ x + + 2Θ̃ ΓΘ Θ̃ (5.73)
γr

Let pi j , i = 1, 2, j = 1, 2, be the elements of P and notice that

2e P B = 2e P̄b ∈ R (5.74)

 
where P̄ = p12 p22 .
102 5 Model-Reference Adaptive Control
 
Then, V̇ e, K̃ x , k̃r , Θ̃ can be expressed as
 

V̇ e, K̃ x , k̃r , Θ̃ = −e Qe + 2 |b| sgn (b) − K̃ x x − k̃r r + Θ̃  Φ (x) e P̄
 
−1 ˙  2k̃r k̃˙r  −1 ˙
+ |b| 2 K̃ x Γx K̃ x + + 2Θ̃ ΓΘ Θ̃ (5.75)
γr

or
   
V̇ e, K̃ x , k̃r , Θ̃ = −e Qe + 2 |b| K̃ x −xe P̄sgnb + Γx−1 K̃˙ x
 
 k̃˙r
+ 2 |b| k̃r −r e P̄sgnb +
γr


+ 2 |b| Θ̃  Φ (x) e P̄sgnb + ΓΘ−1 Θ̃˙ (5.76)

Thus, the following adaptive laws are obtained:

K̇ x = Γx xe P̄sgnb (5.77)

k̇r = γr r e P̄sgnb (5.78)

Θ̇ = −ΓΘ Φ (x) e P̄sgnb (5.79)

It follows that
 
V̇ e, K̃ x , k̃r , Θ̃ = −e Qe ≤ −λmin (Q) e 22 ≤ 0 (5.80)
 
Since V̇ e, K̃ x , k̃r , Θ̃ ≤ 0, therefore e (t), K x (t), kr (t), and Θ (t) are bounded.
Then,
   
lim V e, K̃ x , k̃r , Θ̃ = V e0 , K̃ x0 , k̃r0 , Θ̃0 − λmin (Q) e 22 (5.81)
t→∞

 
So, V e, K̃ x , k̃r , Θ̃ has a finite limit as t → ∞. Since e 2 exists, therefore
e (t) ∈ L2 ∩ L∞ , but  ė ∈ L∞ .
V̇ e, K̃ x , k̃r , Θ̃ can be shown to be uniformly continuous by examining its
derivative to see whether it is bounded, where
   
V̈ e, K̃ x , k̃r , Θ̃ = −ė Qe − e Q ė = −e Q A + A Q e


− 2e Q Am e − B K̃ x x − B k̃r r + B Θ̃  Φ (x) (5.82)
5.4 Direct MRAC for Second-Order SISO Systems 103
 
Since e (t), K x (t), kr (t), and Θ (t) are bounded by the virtue of V̇ e, K̃ x , k̃r , Θ̃
≤ 0, x (t) is bounded because e (t) and xm (t) are bounded, r (t) is a bounded
reference
 command  signal, and Φ (x) is bounded
  x (t) is bounded, therefore
because
V̈ e, K̃ x , k̃r , Θ̃ is bounded. Thus, V̇ e, K̃ x , k̃r , Θ̃ is uniformly continuous. It
 
follows from the Barbalat’s lemma that V̇ e, K̃ x , k̃r , Θ̃ → 0, hence e (t) → 0 as
t → ∞. Therefore, the tracking error is asymptotically stable.

Example 5.11 Design an adaptive controller for a second-order system

 
ÿ + 2ζ ωn ẏ + ωn2 y = b u + Θ ∗ Φ (y)
 
where ζ > 0, ωn > 0, b > 0, and Θ ∗ = θ1∗ θ2∗ are unknown, and
 
1
Φ (y) =
y2

The reference model is given by

ÿm + 2ζm ωm ẏm + ωm2 ym = bm r

where ζm = 0.5, ωm = 2, bm = 4, and r (t) = sin 2t. For simulation purposes,

the unknown
 parameters
 may be assumed to be ζ = −0.5, ωn = 1, b = 1, and
Θ ∗ = 0.5 −0.1 .
√
Note that the open-loop plant is unstable with the eigenvalues λ (A) = 1±2 3 on
the right half plane, and the ideal control gains K x∗ and kr∗ exist and are equal to
 −1   
K x∗ = B  B B (Am − A) = −3 −3

bm
kr∗ = =4
b
Let Q = I , then the solution of the Lyapunov equation (5.71) is
$ % $ %
3 1 1
P= 2 8
1 5
⇒ P̄ = 8
5
8 16 16

   
Let Γx = diag γx1 , γx2 and ΓΘ = diag γθ1 , γθ2 . Then, the adaptive laws are
$ %    $ %
x1   1
k̇ x1 γ 0
K̇ x = 
= Γx xe P̄ sgn (b) = x1 e1 e2 8
k̇ x2    0 γx2 x2 5
16
1
  
1 5 γx1 x 1
= e1 + e2
8 16 γx2 x 2
104 5 Model-Reference Adaptive Control

1 5
k̇r = γr r e P̄sgnb = e1 + e2 γr r
8 16

     $ %
θ̇1  γ θ1 0 1   18
Θ̇ = = −ΓΘ Φ (x) e P̄sgnb = − e1 e2
θ̇2 0 γ θ2 x12 5
16
  
1 5 γ θ1
=− e1 + e2
8 16 γ 2
θ2 x 1

0.1 0.5

0.05
1

0 0
e

−0.05

−0.1 −0.5
0 20 40 60 0 20 40 60
t t

2 x 4 x
x m x m
1 1 2 2
1 2
1

2
x , xm

2 m
x ,x

0 0
1

−1 −2

−2 −4
0 20 40 60 0 20 40 60
t t

Fig. 5.4 Adaptive control system response

For simulations, all adaptation rates are chosen to be 100 and all initial conditions
are set to zero. The results are shown in Figs. 5.4 and 5.5.
It is noted that the plant follows the reference model very well only after a short
time, but the adaptive gains K x (t) and kr (t) and the adaptive parameter Θ (t) are
converging much more slowly. When they converge, some of them do not converge
to their true values, such as k x2 (t) and kr (t). This is one of the properties of MRAC
whereby there is no assurance on the convergence of adaptive parameters to their
true values.
5.4 Direct MRAC for Second-Order SISO Systems 105

2
K ,k,Θ
r

0
x

−2

−4
kx kx kr θ1 θ2
1 2

−6
0 50 100 150 200
t

Fig. 5.5 Adaptive gains and adaptive parameters

5.4.2 Case II: A and B Known

If A and B are known, then it is assumed that there exist K x and kr that satisfy the
model matching conditions
A + B K x = Am (5.83)

Bkr = Bm (5.84)

For a second-order system, if Am and Bm have the same structures as those of A

and B, respectively, then K x and kr can be determined by using the pseudo-inverse
method.
Let the adaptive controller be

u = K x x + kr r − Θ  Φ (x) (5.85)

Then, the closed-loop plant is

ẋ = (A + B K x ) x + Bkr r − B Θ̃  Φ (x) (5.86)

and the tracking error equation is

ė = Am e + B Θ̃  Φ (x) (5.87)
106 5 Model-Reference Adaptive Control

Proof Choose a Lyapunov candidate function

 
V e, Θ̃ = e Pe + Θ̃  Γ −1 Θ̃  (5.88)
 
Then, V̇ e, Θ̃ is evaluated as
 
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) + 2Θ̃  Γ −1 Θ̃˙ (5.89)

Since e P B ∈ R is a scalar value, then

 
V̇ e, Θ̃ = −e Qe + 2Θ̃  Φ (x) e P B + 2Θ̃  Γ −1 Θ̃˙


= −e Qe + 2Θ̃  Φ (x) e P B + Γ −1 Θ̃˙ (5.90)

Thus, the following adaptive law is obtained:

Θ̇ = −Γ Φ (x) e P B (5.91)

Then,  
V̇ e, Θ̃ = −e Qe ≤ −λmin (Q) e 2 (5.92)

Therefore, e (t) and Θ (t) are bounded. Using the same argumentwith the
Barbalat’s lemma as in the previous sections, one can conclude that V̇ e, Θ̃ is
 
uniformly continuous so V̇ e, Θ̃ → 0 as t → ∞. Therefore, the tracking error is
asymptotically stable with e (t) → 0 as t → ∞.

5.5 Indirect MRAC for Second-Order SISO Systems

Indirect MRAC for second-order systems is similar to that for first-order systems.
Consider the second-order system in Sect. 5.4.1 with A and B unknown, but sign
of b is known. Assuming that there exist K x and kr that satisfy the model matching
conditions, and furthermore that Am and Bm have the same structures as those of A
and B, respectively, then A and B can be estimated. Let
       
0 1 0 0 1 0
A= , B= , Am = , B=
−ωn2 −2ζ ωn b −ωm2 −2ζm ωm bm
(5.93)
5.5 Indirect MRAC for Second-Order SISO Systems 107

The model matching conditions are

 (t) + B̂ (t) K x (t) = Am (5.94)

B̂ (t) kr (t) = Bm (5.95)

from which K x (t) and kr (t) are determined by

 −1    
  1   0 0
K x = B̂ B̂ B̂ Am − Â = 0 b̂
b̂2 −ωm2 + ω̂n2 −2ζm ωm + 2ζ̂ ω̂n
1 
= −ωm2 + ω̂n2 −2ζm ωm + 2ζ̂ ω̂n (5.96)
b̂
 −1  
  1   0 bm
kr = B̂ B̂ B̂ Bm = 0 b̂ = (5.97)
b̂ 2 bm b̂

where  (t), B̂ (t), ω̂n (t), and ζ̂ (t) are estimates of A, B, ωn , and ζ , respectively.
Let à (t) =  (t) − A and B̃ (t) = B̂ (t) − B be the estimation errors. Now, the
plant model is expressed as
   


ẋ = Â − Ã x + B̂ − B̃ u + Θ ∗ Φ (x) (5.98)

Then, substituting Eqs. (5.67), (5.96), and (5.97) into Eq. (5.98) yields
 
 

ẋ = Â − Ã x + B̂ K x x + kr r − Θ Φ (x) + Θ ∗ Φ (x)

 

− B̃ K x x + kr r − Θ Φ (x) + Θ ∗ Φ (x)
⎛ ⎞

= ⎝ Â + B̂ K x − Ã⎠ x + B̂kr r − B Θ̃ Φ (x) − B̃ (K x x + kr r )


(5.99)
   
Am Bm

Let
ū = K x (t) x + kr (t) r (5.100)

Then, the tracking error equation is established as


ė = ẋm − ẋ = Am e + Ãx + B̃ ū + B Θ̃ Φ (x) (5.101)

Proof Proceed as usual by choosing a Lyapunov candidate function

    B̃  B̃
V e, Ã, B̃, Θ̃ = e Pe + trace ÃΓ A−1 Ã + + |b| Θ̃  ΓΘ−1 Θ̃ (5.102)
γb
108 5 Model-Reference Adaptive Control

where Γ A = Γ A > 0 ∈ R2 × R2 is a positive-definite adaptation rate matrix for

 (t).
Note that the matrix trace operator is used in the Lyapunov function to map a
matrix product into a scalar quantity.
V̇ e, Ã, B̃, Θ̃ is evaluated as
 

V̇ e, Ã, B̃, Θ̃ = −e Qe + 2e P Ãx + B̃ ū + B Θ̃  Φ (x)
  2 B̃  B̃˙
+ trace 2 ÃΓ A−1 Ã˙  + + 2 |b| Θ̃  ΓΘ−1 Θ̃˙ (5.103)
γb
 
Now, consider the trace operator of a product of two vectors C = c1 c2 . . . cn ∈
 
Rn and D = d1 d2 . . . dn ∈ Rn . Note that C  D = D  C ∈ R and C D  ∈
R × R . Then, one of the identities of a trace operator is as follows:
n n

 
trace C D  = C  D = D  C (5.104)

This can be shown by evaluating both sides of the identity as


n
C  D = DC = ci di (5.105)
i=1

& '
C D  = ci d j , i, j = 1, 2, . . . , n (5.106)

The trace operator is the sum of all the diagonal elements. Therefore,

 
j

trace C D  = ci di = C  D = D  C (5.107)
i=1

Now, utilizing this trace identity, one can express

    
2 e P Ãx = trace 2 Ãxe P (5.108)

Also note that   

2 e P B̃ = 2 B̃  Pe (5.109)

and
 

2 e P B Θ̃  Φ (x) = 2Θ̃  Φ (x) e P B = 2Θ̃  Φ (x) e P̄ |b| sgn (b) (5.110)

since the terms e P B and e P B̃ are scalar quantities (verify!), where P̄ is defined
previously. Therefore,
5.5 Indirect MRAC for Second-Order SISO Systems 109
 
 
  B̃˙
V̇ e, Ã, B̃, Θ̃ = −e Qe + trace 2 Ã xe P + Γ A−1 Ã˙  +2 B̃  Peū +
γb


+ 2 |b| Θ̃  Φ (x) e P̄sgnb + ΓΘ−1 Θ̃˙ (5.111)

The following adaptive laws are then obtained:

Â˙  = −Γ A xe P (5.112)

B̂˙ = −γb Peū (5.113)

Θ̇ = −ΓΘ Φ (x) e P̄sgnb (5.114)

It follows that e (t), Ã (t), B̃ (t), and Θ̃ (t) are bounded since
 
V̇ e, Ã, B̃, Θ̃ = −e Qe ≤ −λmin (Q) e 2 (5.115)
 
V e, Ã, B̃, Θ̃ has a finite limit as t → ∞ since

( ∞
V (t → ∞) = V (t0 ) − λmin (Q) e 2 dt < ∞ (5.116)
t0

 
It can be shown that V̇ e, Ã, B̃, Θ̃ is uniformly continuous because
 
V̈ e, Ã, B̃, Θ̃ is bounded. Then, applying the Barbalat’s lemma, one can conclude
that the tracking error is asymptotically stable with e (t) → 0 as t → ∞.


Let
 
     
Āˆ = 0 1 Â = 0 1
0 1
= −ω̂n2 −2ζ̂ ω̂n (5.117)
−ω̂n2 −2ζ̂ ω̂n

and since  
    0
b̂ = 0 1 B̂ = 0 1 (5.118)
b̂

then the adaptive laws can be expressed in terms of the estimates of the unknown
quantities ωn and ζ as
 
˙
Āˆ  = −Γ A xe P
0
= −Γ A xe P̄ (5.119)
1
110 5 Model-Reference Adaptive Control

 
b̂˙ = −γb 0 1 Peū = −γb P̄  eū = −γb ūe P̄ (5.120)

Let  
γω 0
ΓA = >0 (5.121)
0 γζ

Then,
d  2
−ω̂n = −γω x1 e P̄ (5.122)
dt
or
γω x1 e P̄
ω̂˙ n = (5.123)
2ω̂n

and
d  
−2ζ̂ ω̂n = −2ω̂n ζ̂˙ − 2ζ̂ ω̂˙ n = −γζ x2 e P̄ (5.124)
dt
or  
γζ x2 ω̂n − γω x1 ζ̂ e P̄
ζ̂˙ = (5.125)
2ω̂n2

To prevent the possibility of ω̂n (t) = 0 or b̂ (t) = 0 that will cause the adaptive
laws to blow up, both the adaptive laws for estimating ω̂n (t) and b̂ (t) need to be
modified by the projection method according to
)
γω x1 e P̄
if ω̂n > ω0 > 0 or if ω̂n = ω0 and ω̂˙ n ≥ 0
ω̂˙ n = 2ω̂n (5.126)
0 otherwise
⎧ " " " " " "
" "
⎨  " " " " d "b̂"
−γb ūe P̄ if "b̂" > b0 or if "b̂" = b0 and ≥0
b̂˙ = dt (5.127)
⎩
0 otherwise

In the modified adaptive law for ω̂n (t), it is assumed that ω̂n (t) is always a positive
quantity for a physically realizable system.

5.6 Direct MRAC for MIMO Systems

Consider a MIMO system with a matched uncertainty

ẋ = Ax + BΛ [u + f (x)] (5.128)
5.6 Direct MRAC for MIMO Systems 111

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, A ∈ Rn × Rn

is a constant, known or unknown matrix, B ∈ Rn × Rm is a known matrix, Λ =
Λ = diag (λ1 , λ2 , . . . , λm ) ∈ Rm ×Rm is a control input uncertainty and a diagonal
matrix, and f (x) ∈ Rm is a matched uncertainty that can be linearly parametrized as

f (x) = Θ ∗ Φ (x) (5.129)

where Θ ∗ ∈ Rl × Rm is a constant, unknown matrix, and Φ (x) ∈ Rl is a vector of

known and bounded basis functions.
Furthermore, it is assumed that the pair (A, BΛ) is controllable. Recall that the
controllability condition ensures that the control input u (t) has a sufficient access to
the state space to stabilize all unstable modes of a plant. The controllability condition
can be checked by the rank condition of the controllability matrix C, where
 
C = BΛ | ABΛ | A2 BΛ | . . . | An−1 BΛ (5.130)

The pair (A, BΛ) is controllable if rank (C) = n.

The reference model is specified by

ẋm = Am xm + Bm r (5.131)

where xm (t) ∈ Rn is a reference state vector, Am ∈ Rn × Rn is known and Hurwitz,

Bm ∈ Rn × Rq is known, and r (t) ∈ Rq is a piecewise continuous and bounded
command vector.
The objective is to design a full-state adaptive controller to allow x (t) to follow
xm (t).

5.6.1 Case I: A and Λ Unknown, but B and Sign of Λ Known

Firstly, it must be assumed that there exist ideal control gains K x∗ and K r∗ such that
the following model matching conditions are satisfied:

A + BΛK x∗ = Am (5.132)

BΛK r∗ = Bm (5.133)

If Am and Bm have the same structures as those of A and BΛ, respectively, or

if BΛ is a square and invertible matrix, then there exist K x∗ and K r∗ that satisfy the
model matching conditions.

Example 5.12 A MIMO system and a reference model are specified as

       
1 1 11 0 1 20
A= , BΛ = , Am = , Bm =
−1 −1 01 −16 −2 01
112 5 Model-Reference Adaptive Control

Then,  
14 1
K x∗ = (BΛ)−1 (Am − A) =
−15 −1
 
2 −1
K r∗ = (BΛ) −1
Bm =
0 1


Define an adaptive controller as

u = K x (t) x + K r (t) r − Θ  Φ (x) (5.134)

where K x (t) ∈ Rm × Rn , K r (t) ∈ Rm × Rq , and Θ (t) ∈ Rl × Rm are estimates of

K x∗ , K r∗ , and Θ ∗ , respectively.
Let K̃ x (t) = K x (t) − K x∗ , K̃ r (t) = K r (t) − K r∗ , and Θ̃ (t) = Θ (t) − Θ ∗ be the
estimation errors. Then, the closed-loop plant model is expressed as
⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟
ẋ = ⎝ A + BΛK x∗ +BΛ K̃ x ⎠ x + ⎝ BΛK r∗ +BΛ K̃ r ⎠ r − BΛΘ̃  Φ (x) (5.135)
     
Am Bm

The closed-loop tracking error equation can now be formulated as

ė = ẋm − ẋ = Am e − BΛ K̃ x x − BΛ K̃ r r + BΛΘ̃  Φ (x) (5.136)

Proof To derive the adaptive laws, choose the following Lyapunov candidate func-
tion:
     
V e, K̃ x , K̃ r , Θ̃ = e Pe + trace |Λ| K̃ x Γx−1 K̃ x + trace |Λ| K̃ r Γr−1 K̃ r
 
+ trace |Λ| Θ̃  ΓΘ−1 Θ̃ (5.137)
 
V̇ e, K̃ x , K̃ r , Θ̃ is evaluated as
 

V̇ e, K̃ x , K̃ r , Θ̃ = −e Qe + 2e P −BΛ K̃ x x − BΛ K̃ r r + BΛΘ̃  Φ (x)
 
+ 2trace |Λ| K̃ x Γx−1 K̃˙ x
   
+ 2trace |Λ| K̃ r Γr−1 K̃˙ r + 2trace |Λ| Θ̃  ΓΘ−1 Θ̃˙
(5.138)
 
Utilizing the trace property trace C D  = D  C and Λ = sgnΛ |Λ| where
sgnΛ = diag(sgnλ1 , sgnλ2 , . . . , sgnλm ), then notice that
5.6 Direct MRAC for MIMO Systems 113
 
e P BΛ K̃ x x = e P BsgnΛ |Λ| K̃ x x = trace |Λ| K̃ x xe P BsgnΛ (5.139)
 
e P BΛ K̃ r r = trace |Λ| K̃ r r e P BsgnΛ (5.140)
 
e P BΛΘ̃  Φ (x) = trace |Λ| Θ̃  Φ (x) e P BsgnΛ (5.141)

Then,
  

V̇ e, K̃ x , K̃ r , Θ̃ = −e Qe + 2trace |Λ| K̃ x −xe P BsgnΛ + Γx−1 K̃˙ x


+ 2trace |Λ| K̃ r −r e P BsgnΛ + Γr−1 K̃˙ r


+ 2trace |Λ| Θ̃  Φ (x) e P BsgnΛ + ΓΘ−1 Θ̃˙ (5.142)

Thus, the adaptive laws are obtained as

K̇ x = Γx xe P BsgnΛ (5.143)

K̇ r = Γr r e P BsgnΛ (5.144)

Θ̇ = −ΓΘ Φ (x) e P BsgnΛ (5.145)

It follows that e (t), K̃ x (t), K̃ r (t), and Θ̃ (t) are bounded since
 
V̇ e, K̃ x , K̃ r , Θ̃ = −e Qe ≤ −λmin (Q) e 2 ≤ 0 (5.146)

Using the usual argument with the Barbalat’s lemma, the tracking error is asymp-
totically stable with e (t) → 0 as t → ∞.
   
Example 5.13 Let x (t) = x1 (t) x2 (t) , u (t) = u 1 (t) u 2 (t) , Φ (x) =
 2 2 
x1 x2 , A is unknown, Λ is unknown but Λ > 0 so sgnΛ = I , and B is known
and given by  
11
B=
01

A second-order reference model is specified by

ẍ1m + 2ζm ωm ẋ1m + ωm2 x1m = bm r

where ζm = 0.5, ωm = 2, bm = 4, and r (t) = sin 2t. For simulation purposes, the
true A, Λ, and Θ ∗ matrices are
114 5 Model-Reference Adaptive Control
  4   
1 1 0 0.2 0
A= , Λ = 5 4 Θ∗ =
−1 −1 0 5 0 −0.1

Since BΛ is non-singular and invertible, K x∗ and K r∗ exist and are equal to

$ %−1     $ 5 5 %
4 4
0 1 1 1
K x∗ = (BΛ)−1 (Am − A) = 5 5
− = 2 4
0 45 −4 −2 −1 −1 − 15 − 45
4

$ %−1    
4 4
0 −5
K r∗ = (BΛ)−1 Bm = 5 5
=
0 45 4 5

Let Q = I , then the solution of the Lyapunov equation is

$ %
3 1
P= 2 8
1 5
8 16

   
Let Γx = diag γx1 , γx2 , Γr = γr , and ΓΘ = diag γθ1 , γθ2 . Then, the adaptive
laws are
   $ %
  γx1 0 x1   23 13
K̇ x = Γx xe P B sgnΛ = e1 e2 8
   0 γx2 x2 1 7
8 16
I
$ 3   13 %
γx1 x 1 e1 + 18 e2 γx1 x1 8 e1 + 7
e
16 2
=  23   
γx2 x 2 e + 18 e2 γx2 x2 13
2 1
e +
8 1
7
e
16 2

$ %
  3 13
    
K̇ r 
= Γr r e P BsgnΛ = γr r e1 e2 2 8
1 7
= γr r 23 e1 + 18 e2 γr r 13
8 e1 +
7
16 e2
8 16

 2  $ %
x1   23 13
 γ θ1 0
Θ̇ = −ΓΘ Φ (x) e P BsgnΛ = − e1 e2 8
0 γ θ2 x22 1 7
8 16
$     %
γθ1 x12 23 e1 + 18 e2 γθ1 x12 13 e + 16
8 1
7
e2
=−    
γθ2 x2 2 e1 + 8 e2 γθ2 x2 8 e1 + 16 e2
2 3 1 2 13 7

For simulation, all adaptation rates are chosen to be 10 and all initial conditions
are set to zero. The results are shown in Figs. 5.6, 5.7, 5.8, and 5.9.
It can be seen that the tracking error tends to zero so that x (t) follows xm (t). Also
note that K r (t) and some elements of K x (t) do not converge to their ideal values.
5.6 Direct MRAC for MIMO Systems 115

0.5 2

2
e 0 0

e
−1

−0.5 −2
0 10 20 30 40 0 10 20 30 40
t t

2 x 4 x
x m x m
1 1 2 2
1 2
1

2
m

x , xm
x ,x

0 0
1

2
−1 −2

−2 −4
0 10 20 30 40 0 10 20 30 40
t t

Fig. 5.6 Adaptive control system response

0
x
K

−1

−2

−3
kx kx kx kx
11 12 21 22

−4
0 20 40 60 80 100
t

Fig. 5.7 Adaptive feedback gain K x

116 5 Model-Reference Adaptive Control

2
Kr

−2

−4
kr k
r
1 2

−6
0 20 40 60 80 100
t

Fig. 5.8 Adaptive command gain K r

1.5

0.5
Θ

−0.5

−1

−1.5
θ θ θ θ
11 12 21 22
−2
0 20 40 60 80 100
t

Fig. 5.9 Adaptive parameter Θ

5.6 Direct MRAC for MIMO Systems 117

5.6.2 Case II: A, B, Λ = I Known

The plant model is given by

 
ẋ = Ax + B u + Θ ∗ Φ (x) (5.147)

where both A and B are known.

Assuming that there exist K x and K r that satisfy the model matching conditions

A + B K x = Am (5.148)

B K r = Bm (5.149)

then an adaptive controller is designed as

u = K x x + K r r − Θ  (t) Φ (x) (5.150)

Let Θ̃ (t) = Θ (t) − Θ ∗ be the estimation error, then the closed-loop plant model
is ⎛ ⎞

ẋ = ⎝ A + B K x ⎠ x + B K r r − B Θ̃  Φ (x) (5.151)
   
Am Bm

The closed-loop tracking error equation is obtained as

ė = ẋm − ẋ = Am e + B Θ̃  Φ (x) (5.152)

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (5.153)

Then,
   
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) + 2trace Θ̃  Γ Θ̃˙


= −e Qe + 2trace Θ̃  Φ (x) e P B + Γ Θ̃˙ (5.154)

The adaptive law is

Θ̇ = −Γ Φ (x) e P B (5.155)

Using the Barbalat’s lemma, the tracking error can be shown to be asymptotically
stable with e (t) → 0 as t → ∞.
118 5 Model-Reference Adaptive Control

5.7 Indirect MRAC for MIMO Systems

For the system in Sect. 5.6.1 with A and Λ unknown but B and sign of Λ known,
assuming that BΛ ∈ Rn × Rm is invertible and n ≤ m, then there exist K x∗ and K r∗
that satisfy the model matching conditions such that

K x∗ = (BΛ)−1 (Am − A) (5.156)

K r∗ = (BΛ)−1 Bm (5.157)

If  (t) and Λ̂ (t) are estimates of A and Λ, then the estimates of K x∗ and K r∗ are
given by

−1

K x (t) = B Λ̂ (t) Am − Â (t) (5.158)

−1
K r (t) = B Λ̂ (t) Bm (5.159)

−1
Note that if n < m, then B Λ̂ (t) is defined by the right pseudo-inverse

−1
Λ̂ (t) B  B Λ̂ (t) Λ̂ (t) B  .
Let the adaptive controller be

u = K x (t) x + K r (t) r − Θ  (t) Φ (x) (5.160)

Let à (t) =  (t) − A and Λ̃ (t) = Λ̂ (t) − Λ be the estimation errors. Then, the
closed-loop plant model is expressed as
 

ẋ = Ax + B Λ̂ − Λ̃ K x x + K r r − Θ̃  Φ (x)
 
= A + Am − Â x + Bm r − B Λ̃ (K x x + K r r ) − BΛΘ̃  Φ (x) (5.161)

Let
ū = K x x + K r r (5.162)

Then, the tracking error equation is obtained as

ė = ẋm − ẋ = Am e + Ãx + B Λ̃ū + BΛΘ̃  Φ (x) (5.163)

Proof Choose a Lyapunov candidate function

       
V e, Ã, B̃, Θ̃ = e Pe +trace ÃΓ A−1 Ã +trace Λ̃ΓΛ−1 Λ̃ +trace |Λ| Θ̃  ΓΘ−1 Θ̃ (5.164)
5.7 Indirect MRAC for MIMO Systems 119

Then,
 

V̇ e, Ã, B̃, Θ̃ = −e Qe + 2e P Ãx + B Λ̃ū + BΛΘ̃  Φ (x)
     
+ 2trace ÃΓ A−1 Ã˙  + 2trace Λ̃ΓΛ−1 Λ̃˙  + 2trace |Λ| Θ̃  ΓΘ−1 Θ̃˙ (5.165)

Utilizing the following relationships:

 
e P Ãx = trace Ãxe P (5.166)
 
e P B Λ̃ū = trace Λ̃ūe P B (5.167)
 
e P BΛΘ̃  Φ (x) = e P Bsgn (Λ) |Λ| Θ̃  Φ (x) = trace |Λ| Θ̃  Φ (x) e P Bsgn (Λ) (5.168)

we obtain
  

V̇ e, Ã, B̃, Θ̃ = −e Qe + 2trace à xe P + Γ A−1 Ã˙ 


+ 2trace Λ̃ ūe P B + ΓΛ−1 Λ̃˙ 


+ 2trace |Λ| Θ̃  Φ (x) e P Bsgn (Λ) + ΓΘ−1 Θ̃˙ (5.169)

from which the adaptive laws are obtained as

Â˙  = −Γ A xe P (5.170)

Λ̂˙  = −ΓΛ ūe P B (5.171)

Θ̇ = −ΓΘ Φ (x) e P BsgnΛ (5.172)

   
Since V̇ e, Ã, B̃, Θ̃ ≤ −λmin (Q) e 2 ≤ 0 and V̈ e, Ã, B̃, Θ̃ ∈ L∞ , then
 
V̇ e, Ã, B̃, Θ̃ is uniformly continuous. In addition, V (t → ∞) ≤ V (t0 ). There-
fore, according to the Barbalat’s lemma, the tracking error is asymptotically stable
with e (t) → 0 as t → ∞.


Because Λ̂ (t) is involved in a matrix inversion operation, Λ̂ (t) cannot be singular

assuming B is non-singular. Therefore, the adaptive law for Λ̂ (t) needs to be modified
by the projection method if a priori knowledge of the bounds on the " elements
" of Λ
" "
is available. Suppose the diagonal elements are bounded by λi0 ≤ "λ̂ii " ≤ 1, and the
" "
" "
non-diagonal elements are bounded to be close to zero such that "λ̂i j " ≤ , i = j.
Then, the modified adaptive law for the diagonal elements can be expressed as
120 5 Model-Reference Adaptive Control
⎧ " " " "
" "
" " " "
" "
" "
⎨   " " " " d "λ̂ii " " " d "λ̂ii "
λ̂˙ ii = − ΓΛ ūe P B ii if 1 ≥ "λ̂ii " ≥ λi0 , or if "λ̂ii " = λi0 and dt ≥ 0, or if "λ̂ii " = 1 and dt ≤0
⎩
0 otherwise
(5.173)
and the modified adaptive law for the non-diagonal elements can be expressed as
⎧ " " " " " "
" "
⎨  
 " " " " d "λ̂i j "
˙λ̂ = − ΓΛ ūe P B ji if "λ̂i j " ≤ , or if "λ̂i j " =  and dt ≤ 0 (5.174)
ij
⎩
0 otherwise

5.8 Summary

In situations when the system uncertainty may become significant beyond a level of
desired tolerance that can adversely affect the performance of a controller, adaptive
control can play an important role in reducing the effects of the system uncertainty on
the controller performance. Situations that may warrant the use of adaptive control
could include unintended consequences of off-nominal modes of operation such
as system failures or highly uncertain operating conditions, and complex system
behaviors that can result in an increase in the complexity and hence cost of the
modeling efforts.
There are generally two classes of adaptive control: (1) direct adaptive control
and (2) indirect adaptive control. Adaptive control architectures that combine both
types of adaptive control are also frequently used and are referred to as compos-
ite, combined, or hybrid direct–indirect adaptive control. Adaptive control can deal
with either linear or nonlinear plants with various types of uncertainty which can be
structured uncertainty, unstructured uncertainty, or unmodeled dynamics. Matched
uncertainty is a type of structured uncertainty that can be matched by the control
input for a class of MIMO linear affine-in-control systems. Adaptive control sys-
tems can be designed to provide cancellation of a matched uncertainty. When an
uncertainty cannot be matched, it is called an unmatched uncertainty. Adaptive con-
trol systems can be designed to accommodate unmatched uncertainty, but cannot
cancel the unmatched uncertainty in general. Control input uncertainty is a type of
uncertainty that exists in the control input matrix for a class of MIMO linear affine-
in-control systems. Control input uncertainty can be in the amplitude or in the sign or
both. When the control input uncertainty is in the amplitude, a control saturation can
occur and may worsen the performance of a controller. When the input uncertainty is
in the sign, a control reversal can occur and potentially can cause instability. Control
input uncertainty in general presents more challenges to adaptive control designers.
A reference model is used to specify a desired response of an adaptive control
system to a command input. It is essentially a command shaping filter to achieve
a desired command following. Since adaptive control is formulated as a command
following or tracking control, the adaptation is operated on the tracking error between
the reference model and the system output. A reference model must be designed
properly for an adaptive control system to be able to follow.
5.8 Summary 121

Various direct and indirect model-reference adaptive control techniques for first-
order and second-order SISO systems and MIMO systems are presented. MRAC
can be shown to achieve asymptotic tracking, but it does not guarantee that adaptive
parameters converge to their true values. The Lyapunov stability theory shows that
adaptive parameter estimation errors are only bounded but not asymptotic.

5.9 Exercises

1. Consider a first-order nonlinear SISO system with a matched uncertainty

 
ẋ = ax + b u + θ ∗ φ (x)

where a is unknown, but b is known, θ ∗ is unknown, and φ (x) = x 2 .

A reference model is specified by

ẋm = am xm + bm r

where am < 0 and bm are known, and r (t) is a bounded command signal.
a. Design and implement in Simulink a direct adaptive controller that enables
the plant output x (t) to track the reference model signal xm (t), given b = 2,
am = −1, bm = 1, and r (t) = sin t. For adaptation rates, use γx = 1 and
γ = 1. For simulation purposes, assume a = 1 and θ ∗ = 0.2 for the
unknown parameters. Plot e (t), x (t), xm (t), u (t), and θ (t) for t ∈ [0, 50].
b. Show by the Lyapunov stability analysis that the tracking error is asymptot-
ically stable, i.e., e (t) → 0 as t → ∞.
c. Repeat part (a) for r (t) = 1 (t) where 1 (t) is the unit step function. Plot the
same sets of data as in part (a). Comment on the convergence of k x (t) and
θ (t) to the ideal values k x∗ and θ ∗ .
2. Consider the following first-order plant
 
ẋ = ax + b u + θ ∗ φ (x)

where a , b > 0, θ ∗ is unknown, and φ (x) = x 2 . Design an indirect adaptive

controller in Simulink by estimating a, b, and θ ∗ so that the plant follows a
reference model
ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t. For simulation purposes, use a = 1,

b = 1, θ ∗ = 0.1, x (0) = xm (0) = 1, â (0) = 0, b̂ (0) = 1.5, γa = γb = γθ = 1.
Also assume that a lower bound of b is b0 = 0.5. Plot the time histories of e (t),
x (t) vs. xm (t), â (t), b̂ (t), and θ̂ (t) for t ∈ [0, 50]
122 5 Model-Reference Adaptive Control

3. Derive direct MRAC laws for a second-order SISO system

 
ÿ + 2ζ ωn ẏ + ωn2 y = b u + Θ ∗ Φ (y)

where ζ and ωn are unknown, but b is known. Show by applying the Barbalat’s
lemma that the tracking error is asymptotically stable.
Design a direct adaptive controller for a second-order system using the following
information: b = 1, ζm = 0.5, ωm = 2, bm = 4, r (t) = sin 2t, and
 
1
Φ (y) =
y2

For simulation purposes, the  unknown parameters may be assumed to be ζ =

−0.5, ωn = 1, and Θ ∗ = 0.5 −0.1 , and all initial conditions are assumed to
be zero. Use Γx = ΓΘ = 100I . Plot the time histories of e (t), x (t) vs. xm (t),
K x (t), and Θ (t) for t ∈ [0, 100].
4. For Exercise 3, suppose b is unknown, but b > 0 is known. Design an indirect
adaptive controller in Simulink. For simulation purposes, all initial conditions are
assumed to be zero, except for ω̂n (0) = 0.8 and b̂ (0) = 0.6. For simplicity, use
the unmodified adaptive laws for ω̂n (t) and b̂ (t). Use γω = γζ = γb = 10 and
ΓΘ = 10I . Plot the time histories of e (t), x (t) vs. xm (t), ω̂n (t), ζ̂ (t), b̂ (t), and
Θ (t) for t ∈ [0, 100].
5. Thus far, we have considered adaptive control with a matched uncertainty as a
function of x. In physical systems, an external disturbance is generally a function
of t. Adaptive control can be used for disturbance rejection if the disturbance
structure is known. Suppose the matched uncertainty is a function of t, then all
the adaptive laws can still be used by just replacing Φ (x) by Φ (t), assuming
Φ (t) is known and bounded.
Consider the following first-order plant:
 
ẋ = ax + b u + θ ∗ φ (t)

where a , b, and θ ∗ are unknown, but b > 0 is known, and φ (t) = sin 2t − cos 4t.
Design an indirect adaptive controller in Simulink by estimating a, b, and θ ∗ so
that the plant follows a reference model

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t. For simulation purposes, use a = 1,

b = 1, θ ∗ = 0.1, x (0) = xm (0) = 1, â (0) = 0, b̂ (0) = 1.5, γa = γb = γθ = 1.
Also assume that a lower bound of b is b0 = 0.5. Plot the time histories of e (t),
x (t) vs. xm (t), â (t), b̂ (t), and θ̂ (t) for t ∈ [0, 50].
5.9 Exercises 123

6. Derive direct MRAC laws for a MIMO system

 
ẋ = Ax + B u + Θ ∗ Φ (x)

where A is unknown, but B is known. Show by applying the Barbalat’s lemma

that the tracking error is asymptotically stable.
     
Given x (t) = x1 (t) x2 (t) , u (t) = u 1 (t) u 2 (t) , Φ (x) = x12 x22 ,
and  
11
B=
01

design a direct adaptive controller in Simulink for the MIMO system to follow a
second-order SISO system specified by

ẋm = Am x + Bm r

where r (t) = sin 2t and

   
0 1 0
Am = , Bm =
−4 −2 4

For simulation purposes, the unknown parameters may be assumed to be

   
1 1 ∗ 0.2 0
A= ,Θ =
−1 −1 0 −0.1

and all initial conditions are assumed to be zero. Use Γx = ΓΘ = 10I . Plot the
time histories of e (t), x (t) vs. xm (t), K x (t), and Θ (t) for t ∈ [0, 100].

References

1. Åström, K. J., & Wittenmark, B. (2008). Adaptive control. New York: Dover Publications Inc.
2. Ioannu, P. A., & Sun, J. (1996). Robust adaptive control. Upper Saddle River: Prentice-Hall Inc.
3. Lavretsky, E. (2009). Combined/composite model reference adaptive control. IEEE Transactions
on Automatic Control, 54(11), 2692–2697.
4. Ishihara, A., Al-Ali, K., Kulkarni, N., & Nguyen, N. (2009). Modeling Error Driven Robot
Control, AIAA Infotech@Aerospace Conference, AIAA-2009-1974.
Chapter 6
Least-Squares Parameter Identification

Abstract This chapter presents the fundamental theory of least-squares parameter

identification. Least-squares methods are central to function approximation theory
and data regression analysis. Least-squares methods can also be used in adaptive
control as indirect adaptive control methods to estimate unknown system parameters
to provide the information for adjusting the control gains. The batch least-squares
method is often used for data regression analysis. Least-squares gradient and recur-
sive least-squares methods are well-suited for on-line time series analysis and adap-
tive control. The concept of persistent excitation is introduced as a fundamental
requirement for exponential parameter convergence of least-squares methods. Indi-
rect least-squares adaptive control theory is introduced. The adaptation signal is based
on the plant modeling error in contrast to the tracking error for model-reference adap-
tive control. An important notion to recognize is that the plant modeling error is the
source of the tracking error and not vice versa. The combined least-squares model-
reference adaptive control uses both the plant modeling error and tracking error for
adaptation. As a result, the adaptation mechanism is highly effective. Both the least-
squares gradient and recursive least-squares methods can also be used separately in
adaptive control without combining with model-reference adaptive control. A fun-
damental difference with the least-squares adaptive control methods from model-
reference adaptive control is that a parameter convergence to true system parameters
is guaranteed in the presence of a persistently exciting input signal.

Indirect model-reference adaptive control is one technique for identifying uncertain

parameters of a dynamical system. A more common method for parameter identi-
fication is the well-known least-squares method [1–3]. The least-squares approach
to function approximation and parameter identification is derived from the mini-
mization of the approximation error between a process and a model of the process.
Parameter convergence is usually obtained if the so-called persistent excitation con-
dition is met. The well-known recursive least-squares method is fundamental to many
system identification techniques.

© Springer International Publishing AG 2018 125

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_6
126 6 Least-Squares Parameter Identification

Least-squares methods can be used in adaptive control whereby the plant’s

unknown parameters are estimated by least-squares methods to provide the informa-
tion to adjust the control gains. Many least-squares adaptive control methods can be
found in the literature [4–8]. Artificial neural network-based adaptive control using
least-squares methods is also common in the literature [7, 9]. Chebyshev orthogo-
nal polynomials have been used as basis functions for least-squares adaptive control
[10]. Hybrid adaptive control is a method that combines a direct model-reference
adaptive control to reduce the tracking error with an indirect recursive least-squares
parameter estimation to reduce the plant modeling error simultaneously [11]. Con-
current learning adaptive control is a least-squares modification to adaptive control
problems whereby the uncertainty can be linearly parameterized, and the modified
weight training law uses an estimate of the ideal weights formed on-line by solving
a least-squares problem using recorded and current data concurrently [5].
In this chapter, the learning objectives are:
• To develop an understanding of least-squares parameter estimation and the notion
of persistent excitation for parameter convergence;
• To be able to apply least-squares methods for parameter estimation; and
• To recognize and be able to use least-squares methods as indirect adaptive control
techniques.

6.1 Least-Squares Regression

Suppose the input–output transfer function of a system is given by a set of mea-

surement data y (t) ∈ Rn as a function of some independent variable x (t) ∈ R p ,
expressed as data pairs (xi , yi ), i = 1, 2, . . . , N . Furthermore, suppose the transfer
function between x (t) and y (t) can be linearly parameterized as

y = Θ ∗ Φ (x) (6.1)

where Θ ∗ ∈ Rm × Rn is a matrix of constant but unknown coefficients and Φ (x) ∈

Rm is a bounded regressor (or basis function) vector and is assumed to be known.
Let ŷ be an estimate of y such that

ŷ = Θ  Φ (x) (6.2)

where Θ is the estimate of Θ ∗ .

Formulate an approximation error  as

 = ŷ − y = Θ  Φ (x) − y (6.3)
6.1 Least-Squares Regression 127

Consider the following cost function:

1 
N
J (Θ) =  i (6.4)
2 i=1 i

When J (Θ) is minimized, the approximation error is also minimized. Then, ŷ

approximates y in a least-squares sense. Thus, the parameter identification problem
is posed as a minimization problem.
The necessary condition is given by

∂ J  ∂i N  N


= ∇ JΘ (Θ) = 
i = Φ (xi ) Φ  (xi ) Θ − yi = 0 (6.5)
∂Θ i=1
∂Θ i=1

where ∇ JΘ (Θ) is called the gradient of J (Θ) with respect to Θ.

Thus, Θ can be found by solving the following least-squares regression equation:

Θ = A−1 B (6.6)

where

N
A= Φ (xi ) Φ  (xi ) (6.7)
i=1


N
B= Φ (xi ) yi (6.8)
i=1

and A is a non-singular matrix if there are sufficient unique data.

Example 6.1 Suppose y (t) ∈ R is a scalar variable which can be approximated as

a p-th degree polynomial in terms of x (t) ∈ R as


p
y = θ0 + θ1 x + · · · + θ p x p = θ j x j = Θ  Φ (x)
j=0

   
where Θ  = θ0 θ1 . . . θ p and Φ (x) = 1 x . . . x p .

The least-squares regression equation is expressed as

AΘ = B
128 6 Least-Squares Parameter Identification

where
⎡ ⎤
1
 N ⎢
 ⎥
p
N xi
⎢ ⎥
A= Φ (xi ) Φ  (xi ) = ⎢ .. ⎥ 1 xi · · · xi
⎣ . ⎦
i=1 i=1
p
xi
⎡ N N N p ⎤
i=1 1 i=1 x i ··· xi

⎢ N xi   Ni=1 p+1 ⎥  N
N 2
···
⎢ i=1 i=1 x i i=1 x i ⎥ j+k
=⎢ .. .. .. .. ⎥= xi
⎣ . . . . ⎦ i=1 jk
 N p  N p+1 N 2 p
i=1 x i i=1 x i ··· i=1 x i

⎡ ⎡ N⎤ ⎤
1 yi  N 
⎢  N xi yi ⎥
i=1
N N ⎢ ⎥xi  j
⎢ ⎥ ⎢ i=1 ⎥
B= Φ (xi ) yi = ⎢ ⎥ yi = ⎢
.. .. ⎥= xi yi
⎣ ⎦ ⎣
. . ⎦
i=1 i=1
p  N p
i=1 j
xi i=1 x i yi

This least-squares regression method is in fact a polynomial curve-fitting tech-

nique. For example, let p = 2, then the quadratic curve-fitting coefficients can be
found by
Θ = A−1 B

where ⎡ N N 2 ⎤ ⎡ N ⎤
N x xi i=1 yi
N  Ni=1 2i i=1 
A = ⎣ i=1 xi xi N
xi3 ⎦ , B = ⎣ i=1
N
x y ⎦
 N 2 i=1 N 3
i=1
N 4
 N 2i i
i=1 x i i=1 x i i=1 x i i=1 x i yi

When all available data are used in the least-squares regression method, it is some-
times called a batch least-squares method. This is usually when there are sufficient
data over a given time interval and the estimates of the unknown coefficients are not
needed immediately at each time step.

6.2 Convex Optimization and Least-Squares Gradient

Method

When the estimates of the unknown coefficients are needed at each time step, Θ can
be estimated recursively using each pair of data (xi , yi ) at each time step.
Consider the following cost function:

1 
J (Θ) =   (6.9)
2
6.2 Convex Optimization and Least-Squares Gradient Method 129

The gradient of the cost function with respect to Θ is given by

 
∂ J ∂
= ∇ JΘ (Θ) =   = Φ (x)   (6.10)
∂Θ  ∂Θ 

To determine a least-squares estimation based on a given data pair at each time

step, the concept of convex optimization is now introduced.
Definition 6.1 A subset S is said to be convex if there exist x, y in S and a constant
α ∈ [0, 1] such that αx +(1 − α) y is also in S . A function f (x) is said to be convex
in a convex set S if, for every x, y in S , then

f (αx + (1 − α) y) ≤ α f (x) + (1 − α) f (y) (6.11)


Note that J (Θ) is convex since

1
[α + (1 − α) 1 ] [α + (1 − α) 1 ]
2
1 1
= α 2    + α (1 − α)   1 + (1 − α)2 1 1
2 2
1 2   1
= α   − 2 1 + α 1 + (1 − α)2 1 1
 
(6.12)
2 2

But, α 2 ≤ α and (1 − α)2 ≤ 1 − α for all α ∈ [0, 1]. So,

1    1 1 1
α   − 2  1 + α  1 + (1 − α) 1 1 ≤ α    + (1 − α) 1 1 (6.13)
2 2 2 2

If f (x) ∈ C 1 , i.e., f (x) is differentiable at least once, then

f (y) ≥ f (x) + (∇ f (x)) (y − x) (6.14)

If f (x) ∈ C 2 , then f (x) is convex if ∇ 2 f (x) ≥ 0 where ∇ 2 f (x) is called the

Hessian of f (x).
Now, consider the minimization of J (Θ). Θ ∗ is said to be a global minimum
of J (Θ) if  
J Θ ∗ ≤ J (Θ) (6.15)

This implies that ∇ JΘ (Θ ∗ ) = 0 and ∇ 2 JΘ (Θ ∗ ) ≥ 0 since J (Θ) is twice-

differentiable with respect to Θ.
Utilizing the Taylor series expansion, one writes
       
∇ JΘ Θ ∗ = ∇ JΘ Θ ∗ + ΔΘ + ∇ 2 JΘ Θ ∗ + ΔΘ ΔΘ + O ΔΘ  ΔΘ (6.16)
  
≈0
130 6 Least-Squares Parameter Identification

Since ∇ JΘ (Θ ∗ ) = 0, ∇ JΘ (Θ ∗ + ΔΘ) = ∇ JΘ (Θ), and ∇ 2 JΘ (Θ ∗ + ΔΘ) =

∇ JΘ (Θ), then
2
 −1
ΔΘ = − ∇ 2 JΘ (Θ) ∇ JΘ (Θ) (6.17)

Equation (6.17) can be written in a discrete-time form as

 −1
Θi+1 = Θi − ∇ 2 JΘ (Θi ) ∇ JΘ (Θi ) (6.18)

This is known as a second-order gradient or Newton’s method for convex opti-

mization. It is noted that the inverse of the Hessian matrix is generally numeri-
cally intensive. So, a first-order approximation can be made by recognizing that
∇ 2 JΘ−1 (Θ) ≈ ∇ 2 JΘ−1 (Θ ∗ ) = ε ≥ 0, where ε is a small positive parameter, when Θ
is in the neighborhood of Θ ∗ . This approximation leads to the well-known steepest
descent or first-order gradient method for convex optimization given by

Θi+1 = Θi − ε∇ JΘ (Θi ) (6.19)

Now, dividing both sides by Δt and taking the limit as Δt → 0 yield

Θ̇ = −Γ ∇ JΘ (Θ) (6.20)

where Γ = Γ  > 0 ∈ Rm × Rm is a positive-definite adaptation rate matrix which

ε
effectively replaces Δt . This is the continuous-time version of the gradient method.
Returning to the minimization of J (Θ) for estimating Θ ∗ , the differential form of
the least-squares estimation of Θ (t) can be expressed using the gradient method as
Θ̇ = −Γ ∇ JΘ (Θ) = −Γ Φ (x)   (6.21)

Notice the resemblance of this least-squares gradient method to a model-reference

adaptive law, where the approximation error  (t) replaces the tracking error e (t).

Example 6.2 For Example 6.1, the least-squares gradient method is expressed as
⎡ ⎤ ⎛⎡ ⎤⎡ ⎤ ⎡ ⎤⎞
θ̇0 1 x ··· xp θ0 y
⎢ θ̇1 ⎥ ⎜⎢ x x 2 · · · x p+1 ⎥ ⎢ θ1 ⎥ ⎢ x y ⎥⎟
⎢ ⎥ ⎜⎢ ⎥⎢ ⎥ ⎢ ⎥⎟
Θ̇ = ⎢ . ⎥ = −Γ ⎜⎢ . .. . . .. ⎥ ⎢ .. ⎥ − ⎢ .. ⎥⎟
⎣ .. ⎦ ⎝⎣ .. . . . ⎦ ⎣ . ⎦ ⎣ . ⎦⎠
θ̇ p p p+1
x x ··· x 2p
θp p
x y

6.3 Persistent Excitation and Parameter Convergence

Let Θ̃ (t) = Θ (t) − Θ ∗ be the estimation error, then

 = Θ  Φ (x) Θ − y = Θ̃  Φ (x) (6.22)

6.3 Persistent Excitation and Parameter Convergence 131

The least-squares gradient method can be written as

Θ̃˙ = Θ̇ = −Γ Φ (x) Φ  (x) Θ̃ (6.23)

Proof Choose a Lyapunov candidate function

   
V Θ̃ = trace Θ̃  Γ −1 Θ̃ (6.24)

Then,
     
V̇ Θ̃ = 2trace Θ̃  Γ −1 Θ̃˙ = −2trace Θ̃  Φ (x) Φ  (x) Θ̃
= −2Φ  (x) Θ̃ Θ̃  Φ (x) = −2   = −2

2 ≤ 0 (6.25)
   
Note that V̇ Θ̃ can only be negative semi-definite because V̇ Θ̃ can be zero
 
when Φ (x) = 0 independent of Θ̃ (t). One can establish that V Θ̃ has a finite
limit as t → ∞ since
∞
V (t → ∞) = V (t0 ) − 2

2 dt < ∞ (6.26)
t0

which implies
∞
2

2 dt = V (t0 ) − V (t → ∞) < ∞ (6.27)
t0

Therefore,  (t) ∈ L2 ∩ L∞ . Moreover, since Φ (x) ∈ L∞ as a result of  (t) ∈

L2 ∩ L∞ , then Θ̃ (t) ∈ L∞ , but there is no assurance that Θ̃ (t) → 0 as t → ∞
which implies a parameter convergence.
 
One cannot conclude that V̇ Θ̃ is uniformly continuous since
  ! "
V̈ Θ̃ = −4  ˙ = −4  Θ̃˙  Φ (x) + Θ̃  Φ̇ (x) (6.28)

is not necessarily boundedbecause

 there is no other condition placed on Φ (x)
except Φ (x) ∈ L∞ . For V̇ Θ̃ to be uniformly continuous, an additional condition
Φ̇ 
(x) ∈ L∞ is required. Then, using the Barbalat’s lemma, one can conclude that
V̇ Θ̃ → 0 or  (t) → 0 which also implies that Θ̇ (t) → 0 as t → ∞. Note that,
from Eq. (6.23), Θ̇ (t) → 0 does not necessarily imply that Θ̃ (t) → 0 since Φ (x)
can also tend to zero instead of Θ̃ (t).
So, up to this point, one can only show that the approximation error  (t) can tend
to zero if Φ̇ (x) ∈ L∞ , but the estimation error Θ̃ (t) does not necessarily tend to
zero since Φ (x) can be a zero signal at some time interval. To examine the issue of
132 6 Least-Squares Parameter Identification

the parameter convergence, Eq. (6.23) can be solved as

# t $

Θ̃ (t) = exp −Γ Φ (x) Φ (x) dτ Θ̃0 (6.29)
t0

where Θ̃0 = Θ̃ (t0 ).

Note that x (t) is an independent variable as a function of t. Then, for Θ̃ (t) to
be exponentially stable which implies an exponential parameter convergence, the
following condition is required:
t+T
Φ (x) Φ  (x) dτ ≥ α0 I (6.30)
t

for all t ≥ t0 and some α0 > 0.

This condition is called the persistent excitation (PE) condition which essentially
requires an input signal to be persistently exciting (PE), that is, a signal that does
not go to zero after some finite time when a parameter convergence has not been
established [1]. Another interpretation of the persistent excitation condition is that,
for parameters to converge exponentially, an input signal must be sufficiently rich
to excite all system modes associated with the parameters to be identified. It should
be noted that while the persistent excitation is needed for a parameter convergence,
in practice, input signals that are persistently exciting can lead to unwanted conse-
quences such as exciting unknown or unmodeled dynamics that can have deleterious
effects on stability of a dynamical system.
Another observation to be made is that if x (t) is a state variable of a closed-
loop system, one cannot assume that the persistent excitation condition can easily be
satisfied. This can be explained as follows:
Suppose a parameter identification is used for adaptive control, then the closed-
loop stability usually implies a parameter convergence to the ideal values of the
unknown parameters. However, the parameter convergence requires the persistent
excitation condition which depends on x (t) which in turn depends on the parameter
convergence. This is a circular argument. Therefore, it is difficult to assert the PE
condition when Φ (x) depends on a feedback action. However, if x (t) is an indepen-
dent variable, then the persistent excitation condition can be verified. In practice, a
control system can have a persistent excitation if a command signal is PE. This is a
fundamental process of system identification.
Suppose the persistent excitation condition is satisfied, then the estimation error
is given by %% %% %% %% −γ alph
%%Θ̃ %% ≤ %%Θ̃0 %% e 0 , ∀t ∈ [t , t + T ], t > t
1 1 1 0 (6.31)

where γ = λmin (Γ ) is the smallest eigenvalue of Γ . Thus, Θ̃ (t) is exponentially

stable with Θ̃ (t) → 0 as t → ∞. Hence, the parameter convergence is established. It
follows that the approximation error is also asymptotically stable (but not necessarily
exponentially stable because of Φ (x)) with  (t) → 0 as t → ∞.
6.3 Persistent Excitation and Parameter Convergence 133
& & & &
Since &Φ  (x) Φ (x)& ≥ &Φ (x) Φ  (x)&, it implies

t+T t+T
α0 I ≤ Φ(x)Φ  (x) dτ ≤ Φ  (x) Φ(x)dτ I = β0 I (6.32)
t t
 
Then, if the persistent excitation condition is satisfied, V̇ Θ̃ is expressed as
 
  2α0 V Θ̃
V̇ Θ̃ ≤ −   (6.33)
λmax Γ −1

Thus, the least-squares gradient method is exponentially stable with the rate of
convergence of λ α0Γ −1 .
max ( )
Example 6.3 Consider a scalar estimation error equation

θ̃˙ = −γ φ 2 (t) θ̃

subject to θ̃ (0) = 1, where

1
φ (t) =
1+t

which is bounded. Then,

# $ #  $  
t
dτ 1 −γ t
θ̃ (t) = exp −γ = exp γ − 1 = exp
0 (1 + τ )2 1+t 1+t
 
γ −γ t
θ̃˙ (t) = − exp
(1 + t)2 1+t

As t → ∞, θ̃˙ (t) → 0 but θ̃ (t) → e−γ = 0. Thus, φ (t) is not PE and does not
guarantee a parameter convergence. The only way to ensure that θ̃ (t) → 0 is for
γ → ∞.
Consider another signal with zero value in some time interval such as

1 0≤t ≤1
φ (t) =
0 t >1

Then, 
e−γ t 0 ≤ t ≤ 1
θ̃ (t) =
e−γ t > 1

which leads to the same conclusion that φ (t) is not PE.

134 6 Least-Squares Parameter Identification
 
Example 6.4 Consider Φ (t) = 1 sin t . Then,
# $
1 sin t
Φ (t) Φ  (t) =
sin t sin2 t

Notice that Φ (t) Φ  (t) is singular at any instant of time t. The PE condition is
evaluated as

t+T t+T #1 sin τ

$ #
τ − cos τ
$t+T
Φ (τ ) Φ  (τ ) dτ = dτ = τ 1
t t sin τ sin2 τ − cos τ 2 − 4 sin 2τ t
# $
T cos (t) − cos (t + T )
= T 1 1
cos (t) − cos (t + T ) 2 − 4 sin 2 (t + T ) + 4 sin 2t

Now, T must be chosen such that the PE condition is satisfied. Let T = 2π which
is the period of the signal. Then, Φ (t) is PE since

# $
t+T 2π cos (t) − cos (t + 2π )
Φ (τ ) Φ  (τ ) dτ =
t cos (t) − cos (t + 2π ) 1 − 14 sin 2 (t + 2π ) + 41 sin 2t
# $ # $ # $
2π 0 2 0 1 0
= =π ≥π
0 π 0 1 0 1

Thus, α0 = π . The estimation error is exponentially stable and a parameter

convergence is guaranteed since
%% %% %% %% −πγ t
%%Θ̃ %% ≤ %%Θ̃0 %%e , ∀t ∈ [t1 , t1 + 2π ] , t1 > t0

6.4 Recursive Least-Squares

Consider the following cost function:

t
1
J (Θ) =   dτ (6.34)
2 t0

which is the continuous-time version of the cost function in Sect. 6.1.

The necessary condition is

∂ J t  
∇ JΘ (Θ) = = Φ (x) Φ  (x) Θ − y  dτ = 0 (6.35)
∂Θ  t0

from which Θ (t) is obtained as

6.4 Recursive Least-Squares 135

# t $−1 t

Θ= Φ (x) Φ (x) dτ Φ (x) y  dτ (6.36)
t0 t0

't
assuming the inverse of t0 Φ (x) Φ  (x) dτ exists. Note that the matrix Φ (x) Φ  (x)
is
' talways singular and is not invertible. However, if the PE condition is satisfied, then

t0 Φ (x) Φ (x) dτ is invertible.
Introducing a matrix R (t) = R  (t) > 0 ∈ Rm × Rm where
# t $−1
R= Φ (x) Φ  (x) dτ (6.37)
t0

then t
R −1 Θ = Φ (x) y  dτ (6.38)
t0

Upon differentiation, this yields

d R −1
R −1 Θ̇ + Θ = Φ (x) y  (6.39)
dt
From Eq. (6.37), we obtain

d R −1
= Φ (x) Φ  (x) (6.40)
dt
Therefore,
 
Θ̇ = −RΦ (x) Φ  (x) Θ − y  = −RΦ (x)   (6.41)

Now, since R (t) R −1 (t) = I , then

d R −1
Ṙ R −1 + R =0 (6.42)
dt
Thus,
Ṙ = −RΦ (x) Φ  (x) R (6.43)

Both Eqs. (6.41) and (6.43) constitute the well-known recursive least-squares
(RLS) parameter identification method. The time-varying matrix R (t) is called
the covariance matrix, and the RLS formula is similar to the Kalman filter where
Eq. (6.43) is a differential Riccati equation for a zero-order plant model. Compar-
ing Eqs. (6.21)–(6.41), R (t) plays the role of Γ as a time-varying adaptation rate
matrix and Eq. (6.43) is effectively an adaptive law for the time-varying adapta-
tion rate matrix.
136 6 Least-Squares Parameter Identification

Let Θ̃ (t) = Θ (t) − Θ ∗ be the estimation error. Since  (t) = Θ̃  (t) Φ (x), then

Θ̃˙ = −RΦ (x) Φ  (x) Θ̃ (6.44)

Proof Choose a Lyapunov candidate function

   
V Θ̃ = trace Θ̃  R −1 Θ̃ (6.45)

Then,
   
d R −1
V̇ Θ̃ = trace 2Θ̃  R −1 Θ̃˙ + Θ̃  Θ̃
dt
 
= trace −2Θ̃ Φ (x) Φ (x) Θ̃ + Θ̃  Φ (x) Φ  (x) Θ̃
 

 
= −trace Θ̃  Φ (x) Φ  (x) Θ̃ = −   = −

2 ≤ 0 (6.46)
 
One can establish that V Θ̃ has a finite limit as t → ∞ since

∞
V (t → ∞) = V (t0 ) −

2 dt < ∞ (6.47)
t0

Therefore,  (t) ∈ L2 ∩ L∞ . Since Φ (x) ∈ L∞ because  (t) ∈ L2 ∩ L∞ , then

Θ̃ (t) ∈ L∞ , but there is no guarantee that Θ̃ (t) → 0 as t → ∞ which implies a
parameter convergence,
  unless Φ (x) is PE.
Note that V̇ Θ̃ is not necessarily uniformly continuous since this would require
   
that V̈ Θ̃ is bounded. Evaluating V̈ Θ̃ as
  ! "
V̈ Θ̃ = −2  ˙ = −2  Θ̃˙  Φ (x) + Θ̃  Φ̇ (x)
! "
= −2  −Θ̃  Φ (x) Φ  (x) RΦ (x) + Θ̃  Φ̇ (x)
( # t $−1 )
    
= −2 −Θ̃ Φ (x) Φ (x) Φ (x) Φ (x) dτ Φ (x) + Θ̃ Φ̇ (x)
t0
(6.48)
 
Therefore, V̈ Θ̃ is bounded if the following conditions are met:

• Φ̇ (x) ∈ L∞ .
!' "−1
• t0 Φ (x) Φ  (x) dτ
t
is invertible which implies Φ (x) is PE.
6.4 Recursive Least-Squares 137

If these conditions are satisfied, then using the Barbalat’s lemma, it can be shown
that  (t) → 0 as t → ∞. In addition, Θ̃ (t) → 0 as t → ∞ and the parameter
convergence is achieved. 
Note that there are various versions of the RLS method. One popular version is
the RLS method with normalization where the adaptive law for R (t) is modified as
follows:
RΦ (x) Φ  (x) R
Ṙ = − (6.49)
1 + n2

where 1 + n 2 (x) = 1 + Φ  (x) RΦ (x) is called a normalization factor.

The time derivative of the Lyapunov candidate function for the RLS method with
normalization is
* +
  Θ̃  Φ (x) Φ  (x) Θ̃
 
V̇ Θ̃ = trace −2Θ̃ Φ (x) Φ (x) Θ̃ +
1 + n2
  
1 + 2n 2
= −trace Θ̃  Φ (x) Φ  (x) Θ̃
1 + n2
 2  2
 1 + 2n 2 1 + 2n
= −  = −

≤0 (6.50)
1 + n2 1 + n2
 
Note that V̇ Θ̃ is more negative with than without normalization. Therefore,
the effect of normalization is to make the adaptive law for R (t) more stable, but the
parameter convergence is slower.
Another popular version is the RLS method with a forgetting factor and normal-
ization, which is given without derivation by

RΦ (x) Φ  (x) R
Ṙ = β R − (6.51)
1 + n2

where 0 ≤ β ≤ 1 is called a forgetting factor.

6.5 Indirect Adaptive Control with Least-Squares

Parameter Identification

Consider the following MIMO system with a matched uncertainty in Sect. 5.7 with
A ∈ Rn × Rn unknown, but B ∈ Rn × Rm known with n ≤ m [4]:
 
ẋ = Ax + B u + Θ ∗ Φ (x) (6.52)

where Θ ∗ ∈ Rl × Rm and Φ (x) ∈ Rl is a bounded regressor (or basis function)

vector and is assumed to be known.
138 6 Least-Squares Parameter Identification

The objective is to design an indirect adaptive controller with least-squares para-

meter identification that follows a reference model

ẋm = Am xm + Bm r (6.53)

where Am ∈ Rn ×Rn is Hurwitz Bm ∈ Rn ×Rq , and r (t) ∈ Rq ∈ L∞ is a piecewise

continuous and bounded reference command signal.
 −1
Note that if n < m, then B −1 is defined by the right pseudo-inverse B  B B  .
Assuming B is invertible, then there exist K x∗ and K r∗ that satisfy the model matching
conditions. If  (t) is an estimate of A, then the estimate of K x∗ is given by
! "
K x (t) = B −1 Am − Â (t) (6.54)

Let the adaptive controller be

u = K x (t) x + K r r − Θ  (t) Φ (x) (6.55)

where K r = B −1 Bm is known.
Then, the closed-loop plant model is given by
   
ẋ = A + Am − Â x + Bm r + B Θ ∗ − Θ  Φ (x) (6.56)

where
ū = K x x + K r r (6.57)

If  (t) → A and Θ (t) → Θ ∗ , then ẋ (t) converges to

ẋd = Am x + Bm r = Âx + B ū (6.58)

which follows the reference model if x (t) → xm (t).

Now, define the plant modeling error as the difference between ẋd (t) and ẋ (t).
Then,
   
 = ẋd −ẋ = Âx+B ū−ẋ = Â − A x+B Θ  − Θ ∗ Φ (x) = Ãx+B Θ̃  Φ (x)
(6.59)
Then, the tracking error equation is expressed in terms of the plant modeling
error as
ė = ẋm − ẋ = ẋm − ẋd + ẋd − ẋ = Am e +  (6.60)
6.5 Indirect Adaptive Control with Least-Squares Parameter Identification 139

Thus, we see that the plant modeling error  (t) between the true plant and the
actual plant actually is what really affects the tracking error e (t). So, if we can
minimize the plant modeling error, then the tracking error will also be minimized.
   
Let Ω̃  (t) = Ã (t) B Θ̃  (t) ∈ Rn × Rn+l and Ψ (x) = x  Φ  (x) ∈
Rn+1 where à (t) =  (t) − A and Θ̃ (t) = Θ (t) − Θ ∗ . Then, the plant modeling
error is expressed as
 = Ω̃  Ψ (x) (6.61)

The tracking error equation is also expressed as

ė = Am e + Ω̃  Ψ (x) (6.62)

Consider the following cost function:

  1
J Ω̃ =    (6.63)
2
Then, the least-squares gradient adaptive law is given by

∂ J
Ω̇ = −Γ = −Γ Ψ (x)   (6.64)
∂ Ω̃ 
For the standard RLS adaptive law, the constant adaptation rate matrix Γ is
replaced by the covariance matrix R (t) as follows:

Ω̇ = −RΨ (x)   (6.65)

Ṙ = −RΨ (x) Ψ  (x) R (6.66)

We now combine the least-squares gradient adaptive law with the MRAC adap-
tive law to obtain the following adaptive law which we call combined least-squares
gradient MRAC adaptive law:
 
Ω̇ = −Γ Ψ (x)   + e P (6.67)

To compute  (t), and Θ (t) from Ω (t), it is noted that

# $ # $ # $ # $
Ω1 Â Â Ω1
Ω= = ⇒ = (6.68)
Ω2 Θ B Θ Ω2 B −

Proof To prove stability of the combined RLS MRAC adaptive laws, we choose a
Lyapunov candidate function
140 6 Least-Squares Parameter Identification

   
V e, Ω̃ = e Pe + trace Ω̃  R −1 Ω̃ (6.69)
 
Then, V̇ e, Ω̃ is evaluated as
* +
    dR −1
V̇ e, Ω̃ = −e Qe + 2e P + trace −2Ω̃  Ψ (x)   + e P + Ω̃  Ω̃
dt
 
= −e Qe + 2e P − 2   + e P  + Ψ  (x) Ω̃ Ω̃  Ψ (x)
     
 

= −e Qe −    ≤ −λmin (Q)

e
2 −

2 (6.70)

   
Note that V̇ e, Ω̃ ≤ 0 as opposed to V̇ e, Ω̃ < 0 because the condition of
Ψ (x) on which  (t) depends on has not been imposed. Therefore, e (t) ∈ L2 ∩ L∞ ,
 (t) ∈ L2 ∩ L∞, ė (t) ∈ L∞ , and Ω̃ ∈ L∞ . Applying the Barbalat’s lemma, we
evaluate V̈ e, Ω̃ as
  ! "
V̈ e, Ω̃ = −2e Q (Am e + ) −   Ω̃˙  Ψ (x) + Ω̃  Ψ̇ (x)
! "
= −2e Q (Am e + ) −   − (Pe + ) Ψ  (x) RΨ (x) + Ω̃  Ψ̇ (x)
(6.71)

But, by definition
# t $−1

R= Ψ (x) Ψ (x) dτ (6.72)
t0

 
If Ψ̇ (x) ∈ L∞ and if Ψ (x) satisfies the PE condition, then V̈ e, Ω̃ ∈ L∞ .
 
Therefore, V̇ e, Ω̃ is uniformly continuous. It follows that e (t) → 0,  (t) → 0,
and Ω̃ (t) → 0 as t → ∞. Since Ω̃ (t) tends to zero exponentially because of
the PE condition, e (t) also tends to zero exponentially. Thus, the system achieves
exponential stability.

Proof If Ψ (x) satisfies the PE condition, then we have

  & &2
& &
V̇ e, Ω̃ ≤ −λmin (Q)
e
2 − α0 &Ω̃ & (6.73)

α0 λmax (P)
Choose Q and β0 from Eq. (6.32) such that λmin (Q) = β0
. Then,
6.5 Indirect Adaptive Control with Least-Squares Parameter Identification 141
 
  α0 λmax (P) & &2 α0 V e, Ω̃
& &
V̇ e, Ω̃ ≤ −
e
2 − α0 &Ω̃ & ≤ − (6.74)
β0 β0

α0
Thus, the system is exponentially stable with the convergence rate of 2β0
.

Note that for the combined least-squares gradient MRAC adaptive laws, it can be
shown that
 
V̇ e, Ω̃ = −e Qe − 2   ≤ −λmin (Q)
e
2 − 2

2 (6.75)
 
V̈ e, Ω̃ is evaluated as

  ! "
V̈ e, Ω̃ = −2e Q (Am e + ) − 2  − (Pe + ) Ψ  (x) Γ Ψ (x) + Ω̃  Ψ̇ (x)
(6.76)
   
If Ψ̇ (x) ∈ L∞ , then V̈ e, Ω̃ is bounded. Therefore, V̇ e, Ω̃ is uniformly
continuous. It follows that e (t) → 0 and  (t) → 0 as t → ∞. The rate of con-
vergence of  (t) is twice as large as that of the combined RLS MRAC adaptive
laws. In addition, if Ψ (x) satisfies the PE condition, then the exponential parameter
convergence is achieved.
Example 6.5 Consider a first-order system with a matched uncertainty
 
ẋ = ax + b u + θ ∗ x 2

where a and θ ∗ are unknown, but b = 2. For simulation purposes, a = 1 and

θ ∗ = 0.2.
The reference model is given by

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t.

The combined RLS MRAC is computed as
# $ # $# $
â˙ r r x ( + e)
Ω̇ = = − 11 12
bθ̇ r12 r22 x 2 ( + e)
# $ # $# $# $
ṙ11 ṙ12 r11 r12 x2 x3 r11 r12
Ṙ = =−
ṙ12 ṙ22 r12 r22 x3 x4 r12 r22
142 6 Least-Squares Parameter Identification

with the initial conditions Ω (0) = 0 and R (0) = 10I , where

 = âx + bū − ẋ

am − â bm
ū = k x (t) x + kr r = x+ r
b b
Simulation results are shown in Figs. 6.1 and 6.2.
It can be seen that â (t) and θ (t) tends to their true values, and the tracking error
e (t) also tends to zero. The covariance matrix R (t) which acts as time-varying
adaptation rate matrix Γ converges to zero as t → ∞.
For comparison, the combined least-squares gradient MRAC method is used
# $ # $# $
â˙ γ11 0 x ( + e)
Ω̇ = =−
bθ̇ 0 γ22 x 2 ( + e)

where γa = γθ = 10.
The simulation results are shown in Fig. 6.3.
The tracking error tends to zero asymptotically much faster than the combined
RLS MRAC method. The combined least-squares gradient MRAC method gener-
ally provides a much better parameter convergence than the combined RLS MRAC
method. 

0.2 1
x xm
0.1 0.5
x, xm

0 0
e

−0.1 −0.5

−0.2 −1
0 10 20 30 40 0 10 20 30 40
t t

1.5 0.3

1 0.2

< 0.5 0.1

a

0 0

−0.5 −0.1
0 10 20 30 40 0 10 20 30 40
t t

Fig. 6.1 Combined RLS MRAC

6.5 Indirect Adaptive Control with Least-Squares Parameter Identification 143

10
r
11
r
12
r
21
r22

R 5

−5
0 5 10 15 20 25 30 35 40
t

Fig. 6.2 Covariance matrix R

0.05 1
x xm
0 0.5
x, xm

−0.05 0
e

−0.1 −0.5

−0.15 −1
0 10 20 30 40 0 10 20 30 40
t t

1.5 0.3

1 0.2

< 0.5 0.1

a

0 0

−0.5 −0.1
0 10 20 30 40 0 10 20 30 40
t t

Fig. 6.3 Combined least-squares gradient MRAC

While the combined least-squares MRAC method can provide an exponential

parameter convergence, we should note that the pure least-squares method can also be
used for adaptive control without combining with MRAC. Both Eqs. (6.64) and (6.65)
are pure least-squares adaptive laws without MRAC. Let us examine the behavior of
pure RLS adaptive control from Eqs. (6.65) and (6.66).
144 6 Least-Squares Parameter Identification

Proof Choose a Lyapunov candidate function

   
V e, Ω̃ = e Pe + μ trace Ω̃  R −1 Ω̃ (6.77)

 μ > 0.
with
V̇ e, Ω̃ is evaluated as

   
V̇ e, Ω̃ = −e Qe + 2e P + μtrace −2Ω̃  Ψ (x)   + Ω̃  Ψ (x) Ψ  (x) Ω̃
= −e Qe + 2e P − μ   ≤ −λmin (Q)
e
2
+ 2λmax (P)
e


− μ

2 (6.78)
 
V̇ e, Ω̃ is now not negative semi-definite unconditionally. We note that

# $2
λ2max (P) λmax (P)
−
e
+2λmax (P)
e


−δ

= −
2 2 2

e
− δ

≤ 0
δ2 δ
  (6.79)
Therefore, V̇ e, Ω̃ can be expressed as

  # $
λ2 (P)  
V̇ e, Ω̃ ≤ − λmin (Q) − max 2
e
2 − μ − δ 2

2 (6.80)
δ
  √
Then, V̇ e, Ω̃ ≤ 0 if √λλmax (P) < δ < μ. Furthermore, if Ψ (x) satisfies the
min (Q)
PE condition, then from Eq. (6.32) we have
# $
  λ2max (P)   & &2
2 & &
V̇ e, Ω̃ ≤ − λmin (Q) −
e
2
− α0 μ − δ &Ω̃ & (6.81)
δ2

λ2max (P) α0 (μ−δ 2 )λmax (P)

Choose Q and β0 from Eq. (6.32) such that λmin (Q) − δ2
= μβ0
.
Then,
     
  α0 μ − δ 2 λmax (P)   & &2 α0 μ − δ 2 V e, Ω̃
& &
V̇ e, Ω̃ ≤ −
e
2 − α0 μ − δ 2 &Ω̃ & ≤ −
μβ0 μβ0
(6.82)
α0 (μ−δ 2 )
Thus, the system is exponentially stable with the convergence rate of 2μβ0
.


From Sect. 6.3, for the least-squares gradient method, if Ψ̇ (x) ∈ L∞ , then the
plant modeling error is asymptotically stable with  (t) → 0 as t → ∞. This implies
that the tracking error is also asymptotically stable with e (t) → 0 as t → ∞.
6.5 Indirect Adaptive Control with Least-Squares Parameter Identification 145

However, for the RLS method, asymptotic tracking additionally requires the PE
condition, but it is important to recognize that the PE condition cannot be verified in
a closed-loop system. So, if the PE condition is not satisfied, then the plant modeling
error is only bounded for the RLS method. This implies that the tracking error is
only bounded but is not asymptotically stable.
Example 6.6 Consider Example 6.5
 
ẋ = ax + b u + θ ∗ x 2

where a and θ ∗ are unknown, but b = 2. For simulation purposes, a = 1 and

θ ∗ = 0.2.
π
Let r (t) = sin 10 t for t ∈ [0, 5] and r (t) = 1 for t ∈ [5, 40].
The RLS adaptive control method is implemented as
# $ # $# $
â˙ r r x
Ω̇ = = − 11 12
bθ̇ r12 r22 x 2

with R (0) = I .
The simulation results are shown in Fig. 6.4.
0 1.5

−0.2
1
m
x, x

−0.4
e

0.5
−0.6
x xm
−0.8 0
0 10 20 30 40 0 10 20 30 40
t t

1 0.4

0.5 0.2
<
a

0 0

−0.5 −0.2
0 10 20 30 40 0 10 20 30 40
t t

Fig. 6.4 RLS adaptive control

For comparison, the least-squares gradient adaptive control method is also imple-
mented as # $ # $# $
â˙ γ 0 x
Ω̇ = = − 11
bθ̇ 0 γ22 x 2

with Γ = I .
146 6 Least-Squares Parameter Identification

The simulation results are shown in Fig. 6.5. It can be seen that the least-squares
gradient adaptive control achieves asymptotic tracking with e (t) → 0 as t → ∞.
Both the RLS and least-squares gradient adaptive control do not achieve a parameter
convergence with this reference command signal.

0 1.5

−0.2
1

x, xm
−0.4
e

0.5
−0.6
x xm
−0.8 0
0 20 40 0 20 40
t t

1 0.4

0.5 0.2
<
a

0 0

−0.5 −0.2
0 20 40 0 20 40
t t

Fig. 6.5 Least-squares gradient adaptive control

6.6 Estimation of Time Derivative Signals

In least-squares adaptive control, the signal ẋ (t) is required in the computation of

the plant modeling error. In many applications, this signal may not be available.
Therefore, in these situations, the signal ẋ (t) needs to be estimated. One method of
estimating ẋ (t) is to use a backward finite-difference method as

x (t) − x (t − Δt)
x̂˙ (t) = (6.83)
Δt
to estimate ẋ (t) at a current time step using the current and past state information.
The disadvantage of numerical differentiation is the introduction of noise into the
signal.
Another approach is to sample data sufficiently and then use them to construct an
at least C 1 smooth time function using a cubic or B-spline method over a finite time
interval. This curve is then differentiated at the spline knots to find the estimated time
derivative values. This derivative estimate would tend to result in lower noise than the
backward finite-difference method [11]. In either case, the derivative computation
will introduce a noise source. Least-squares methods are generally robust to noise,
6.6 Estimation of Time Derivative Signals 147

especially if the noise source is a normal Gaussian distribution. Therefore, least-

squares adaptive control should be generally more effective than model-reference
adaptive control in the presence of noisy measurements. A possibility to suppress
noise source is to use a low-pass filter to smooth out the reconstructed signal from
numerical differentiation, but the use of a low-pass filter will also introduce a phase
delay into the signal which can have a stability implication.
An alternative approach of estimating the signal ẋ (t) without numerical differ-
entiation is via a model prediction method. Consider a predictor model of the plant
in Eq. (9.341) as follows [4]:
   
x̂˙ = Am x̂ + Â − Am x + B u + Θ  Φ (x) (6.84)

If  (t) → A and Θ (t) → Θ ∗ , then x̂˙ (t) → ẋ (t). Thus, x̂˙ (t) is in fact an
estimate of ẋ (t). The estimate x̂˙ (t) only depends on the information of the current
state, the control, and the adaptive parameters. The predictor model thus can be used
to provide the estimate of the signal ẋ (t) without differentiation.
As the adaptation proceeds, the predictor model should converge to the plant
model with x̂ (t) → x (t) + e p (t) as t → ∞, where e p (t) = x̂ (t) − x (t) is the
predictor error whose dynamics are described by

ė p = x̂˙ − ẋ = Am e p + Ãx + B Θ̃  Φ (x) (6.85)

Therefore, as  (t) → A and Θ (t) → Θ ∗ , e p (t) → 0 and x̂˙ (t) → ẋ (t).

The plant modeling error based on the predictor model is established by

 p = ẋd − x̂˙ = ẋd − ẋ − ė p =  − ė p (6.86)

If the predictor error converges to zero, i.e., e p (t) → 0, then  p (t) →  (t).
Therefore, the signal  p (t) could be used in least-squares adaptive control in lieu of
the signal  (t).

6.7 Summary

Least-squares parameter identification is central to function approximation theory

and data regression analysis. Least-squares methods for function approximation and
parameter identification are derived from the minimization of the approximation
error between a process and a model of the process. Least-squares methods can be
used in adaptive control whereby the plant’s unknown parameters are estimated by
least-squares methods to provide the information to adjust the control gains. Various
least-squares techniques are presented, including batch least-squares, least-squares
gradient for real-time update, and recursive least-squares.
148 6 Least-Squares Parameter Identification

Parameter convergence in least-squares real-time update laws depends on a con-

dition called persistent excitation (PE). This condition essentially requires an input
signal to be persistently exciting (PE), that is, a signal that does not go to zero after
some finite time when a parameter convergence has not been established. For para-
meter identification to converge exponentially, an input signal must be sufficiently
rich to excite all system modes associated with the parameters to be identified. While
the persistent excitation condition is needed for a parameter convergence, in practice,
input signals that are persistently exciting can lead to unwanted consequences such
as exciting unknown or unmodeled dynamics that can have deleterious effects on
stability of a dynamical system. In general, the persistent excitation condition for
closed-loop systems is not easily verified.
Least-squares indirect adaptive control methods are discussed to show how least-
squares methods can be used in adaptive control settings. Whereas in MRAC, the
tracking error is used to drive the adaptation, least-squares adaptive control pro-
vides an adaptation mechanism based on the plant modeling error. The combined
least-squares MRAC uses both the tracking error and plant modeling error for the
adaptation. Both the least-squares gradient and recursive least-squares adaptive con-
trol can achieve exponential tracking and parameter convergence if the persistent
excitation condition is satisfied. In general, the plant modeling error requires the
time derivative signals of the state information. In some applications, these signals
may not be available from measurements. The model prediction method could afford
an advantage in that the estimated time derivative signals do not require numerical
differentiation which could introduce noise.

6.8 Exercises

1. A process is represented by a set of data (t, x, y) given in the MATLAB file

“Process_Data.mat” where the output y (t) can be approximated by a f our th −
th degree polynomial in terms of x (t) with end point conditions y = 0 and
dy
dx
= 0 at x = 0. Determine numerically the matrix A and vector B and solve
for the coefficients θi , i = 2, 3, 4. Compare the result with the MATLAB function
“polyfit.”
2. Write MATLAB code to solve Exercise 6.1 using the least-squares gradient
method with Θ (0) = 0 and Γ = 10. Plot θi (t) versus t. Compare the result
with that in Exercise 6.1. Note that the Euler method for the least-squares gradient
method is expressed as
 
Θi+1 = Θi − ΔtΓ Φ (xi ) Φ  (xi ) Θi − yi

3. Determine if the following functions are persistently exciting (PE), and if so,
determine T and α.
6.8 Exercises 149

a. φ (t) = e#−t . (Hint:

$ find limit of θ̃ (t) as t → ∞)
cos π t
b. Φ (t) =
sin π t
4. Consider a first-order system with a matched uncertainty
 
ẋ = ax + b u + θ ∗ φ (t)

where a and θ ∗ are unknown, but b = 2, and φ (t) = sin t. For simulation
purposes, use a = 1 and θ ∗ = 0.2. The reference model is given by

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t.

Implement in Simulink an indirect adaptive control using the recursive least-
squares method with normalization. All initial conditions are zero. Use R (0) =
10. Plot e (t), x (t) versus xm (t), â (t), and θ (t), for t ∈ [0, 40].

References

1. Ioannu, P. A., & Sun, J. (1996). Robust adaptive control. Upper Saddle River: Prentice-Hall
Inc.
2. Slotine, J.-J., & Li, W. (1991). Applied nonlinear control. Upper Saddle River: Prentice-Hall
Inc.
3. Bobal, V. (2005). Digital self-tuning controllers: algorithms, implementation, and applications.
London: Springer.
4. Nguyen, N. (2013). Least-squares model reference adaptive control with chebyshev orthogonal
polynomial approximation. AIAA Journal of Aerospace Information Systems, 10(6), 268–286.
5. Chowdhary, G., & Johnson, E. (2011). Theory and flight-test validation of a concurrent-learning
adaptive controller. AIAA Journal of Guidance, Control, and Dynamics, 34(2), 592–607.
6. Chowdhary, G., Yucelen, T., & Nguyen, N. (2013). Concurrent learning adaptive control of the
aeroelastic generic transport model. In AIAA Infotech@Aerospace Conference, AIAA-2013-
5135.
7. Guo, L. (1996). Self-convergence of weighted least-squares with applications to stochastic
adaptive control. IEEE Transactions on Automatic Control, 41(1), 79–89.
8. Lai, T., & Wei, C.-Z. (1986). Extended least squares and their applications to adaptive control
and prediction in linear systems. IEEE Transactions on Automatic Control, 31(10), 898–906.
9. Suykens, J., Vandewalle, J., & deMoor, B. (1996). Artificial neural networks for modeling and
control of non-linear systems. Dordrecht: Kluwer Academic Publisher.
10. Zou, A., Kumar, K., & Hou, Z. (2010). Attitude control of spacecraft using chebyshev neural
networks. IEEE Transactions on Neural Networks, 21(9), 1457–1471.
11. Nguyen, N., Krishnakumar, K., Kaneshige, J., & Nespeca, P. (2008). Flight dynamics model-
ing and hybrid adaptive control of damaged asymmetric aircraft. AIAA Journal of Guidance,
Control, and Dynamics, 31(3), 751–764.
Chapter 7
Function Approximation and Adaptive
Control with Unstructured Uncertainty

Abstract This chapter presents the fundamental theories of least-squares func-

tion approximation and least-squares adaptive control of systems with unstructured
uncertainty. The function approximation theory based on polynomials, in particu-
lar the Chebyshev orthogonal polynomials, and neural networks is presented. The
Chebyshev orthogonal polynomials are generally considered to be optimal for func-
tion approximation of real-valued functions. The Chebyshev polynomial function
approximation is therefore more accurate than function approximation with regu-
lar polynomials. The neural network function approximation theory for a two-layer
neural network is presented for two types of activation functions: sigmoidal func-
tion and radial basis function. Model-reference adaptive control of systems with
unstructured uncertainty is developed in connection with the function approxima-
tion theory. Least-squares direct adaptive control methods with polynomial approx-
imation and neural network approximation are presented. Because the least-squares
methods can guarantee the parameter convergence, the least-squares adaptive con-
trol methods are shown to achieve uniform ultimate boundedness of control signals
in the presence of unstructured uncertainty. The standard model-reference adaptive
control can be used for systems with unstructured uncertainty using polynomial or
neural network approximation. Unlike the least-squares adaptive control methods,
boundedness of tracking error is guaranteed but boundedness of adaptive parame-
ters cannot be mathematically guaranteed. This can lead to robustness issues with
model-reference adaptive control, such as the well-known parameter drift problem.
In general, least-squares adaptive control achieves better performance and robustness
than model-reference adaptive control.

In many physical applications, there is no clear certainty about the structure

between the input and output of a process. In systems with unstructured uncer-
tainty, the mapping between the input and output is usually not known. Modeling
systems with unstructured uncertainty require function approximation. Polynomial
regression and neural networks are common methods for function approximation.
Adaptive control of systems with unstructured uncertainty can be addressed with
least-squares function approximation.

© Springer International Publishing AG 2018 151

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_7
152 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

In this chapter, the learning objectives are:

• To develop an understanding of function approximation using orthogonal polyno-
mials such as Chebyshev polynomials and neural networks using sigmoidal and
radial basis functions;
• To be able to apply least-squares methods for single-hidden layer neural network
function approximation;
• To develop a working knowledge of using various least-squares direct and indirect
adaptive control techniques for systems with unstructured uncertainty; and
• To recognize that the standard model-reference adaptive control is not robust for
systems with unstructured uncertainty due to the possibility of unbounded adaptive
parameters.

7.1 Polynomial Approximation by Least-Squares

Let y (t) ∈ Rn be the output of a process, expressed as

y = f (x) (7.1)

where x (t) ∈ R p is the input and f (x) ∈ Rn is an unknown function but assumed
to be bounded function.
Any sufficiently smooth function f (x) ∈ C q can be expanded as a Taylor series
about some x = x̄


p
∂ f (x̄) 1   ∂ f (x̄)
p p
f (x) = f (x̄)+ (xi − x̄i )+ (xi − x̄i )(x j − x̄ j )+· · · (7.2)
i=1
∂ xi 2 i=1 j=1 ∂ xi x j

f (x) can be represented as

f (x) = Θ ∗ Φ (x) − ε∗ (x) (7.3)

where Θ ∗ ∈ Rl × Rn is a matrix of constant but unknown coefficients, Φ (x) ∈ Rl

is a vector of regressors in terms of monomials of x
 
Φ (x) = 1 x1 x2 . . . x p x12 x1 x2 . . . x 2p . . . x1q x1 x2q−1 . . . x qp (7.4)

and ε∗ (x) is a function approximation error which depends on x.

f (x) is approximated by
ŷ = Θ  Φ (x) (7.5)

where Θ (t) ∈ Rl × Rn is the estimate of Θ ∗ .

7.1 Polynomial Approximation by Least-Squares 153

Then, ŷ (t) → f (x) as q → ∞. This means that any sufficiently smooth function
can be approximated by a polynomial of q-th degree. Then, the function approxima-
tion error could be made sufficiently small on a compact domain of x (t) such that
supx∈D ε∗ (x) ≤ ε0∗ for all x (t) ∈ D ⊂ R p .
Another type of polynomials that can be used for function approximation is a
class of orthogonal polynomials which form a true basis that spans a Hilbert inner
product space with the following inner product definition:
 b
w (x) pi (x) p j (x) d x = δi j (7.6)
a

where δi j = 1 if i = j and δi j = 0 if i
= j.
There are many orthogonal polynomial classes. The Chebyshev polynomials are
one such polynomial class that is frequently used in function approximation. In
particular, the first few Chebyshev polynomials of the first kind are shown as follows:

T0 (x) = 1
T1 (x) = x
T2 (x) = 2x 2 − 1
T3 (x) = 4x 3 − 3x (7.7)
T4 (x) = 8x 4 − 8x 2 + 1
..
.
Tn+1 (x) = 2x Tn (x) − Tn−1 (x)

One advantage of orthogonal polynomials over regular polynomials is that lower-

degree orthogonal polynomials can provide a good function approximation with
minimal loss of accuracy as compared to higher-degree regular polynomials [1, 2].
Thus, it is economical in function approximation to use the Chebyshev polynomials.
The coefficient vector Θ can be computed using various least-squares methods
such as the batch least-squares, least-squares gradient method, or RLS method. Note
that since Θ  Φ (x) is an approximation of an unknown function f (x), the approx-
imation error will not be asymptotic regardless whether or not Φ (x) is PE.

Example 7.1 Approximate

y = sin x + cos 2x + e−x

2

over a compact domain x (t) ∈ [−1, 1], where x (t) = sin 10t, with a 4th -degree
Chebyshev polynomial using the least-squares gradient method with ΓΘ = 2I and
Θ (0) = 0.
The approximation results are shown in Fig. 7.1 for 1000 s which shows an excel-
lent parameter convergence after only less than 100 s. The root mean square error is
3.8 × 10−4 .
154 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

Θ
0

−1
θ1 θ2 θ3 θ4 θ5
−2
0 200 400 600 800 1000
t

2
y, ΘTΦ(x)

0 y
ΘTΦ(x)
−1
−1 −0.5 0 0.5 1
x

Fig. 7.1 Chebyshev polynomial approximation

7.2 Neural Network Approximation

Neural network is a concept pioneered in the 1960s in the artificial intelligence fields
as an attempt to represent a simple model of how a human brain neuron connectivity
works. Neural networks have been used in many applications such as classifiers,
pattern recognition, function approximation, and adaptive control [1, 3–12]. The
terminology “intelligent control” has been used loosely to describe adaptive control
with neural networks. A neural network is a network of connectivity of neurons
that form various layers of neurons to describe a complex input–output mapping.
Figure 7.2 is a two-layer neural network that performs input–output mapping. A
neural network with connectivity progressing from inputs to an output is called
a feedforward neural network. A neural network with a closed-loop connectivity
between some elements of the neural network is called a recurrent neural network.

Fig. 7.2 Feedforward neural network

Feedforward neural networks have been shown to be capable of approximating

a large class of nonlinear functions on a compact domain to within any specified
accuracy. That is,
7.2 Neural Network Approximation 155

f (x) = fˆ (x) − ε∗ (x) (7.8)

such that supx∈D ε∗ (x) ≤ ε0 for all x ∈ D ⊂ Rn .

Each neuron in a layer has two main components: a weighted summing junc-
tion and an activation function, which can be linear or, most commonly, nonlinear.
Figure 7.3 illustrates a neuron with a set of inputs passing through an activation
function to form a set of outputs.

Fig. 7.3 Neuron and activation function

If x (t) ∈ R p is an input to a neuron and y (t) ∈ Rn is an output of a neuron, then

a neuron performs the following computation:
 
y = f W x + b (7.9)

where W is called a weight matrix and b is a constant vector called a bias vector.

Example 7.2 A second-degree polynomial regression can be represented by a neural

network as follows:
The inputs are x 2 , x, and 1, and the output is y as shown in Fig. 7.4. The activation
function is linear, that is,

y = W  Φ (x) = w2 x 2 + w1 x + w0

where Φ (x) is the input function. 

Fig. 7.4 Polynomial neural network representation

The most common activation functions are sigmoidal and radial basis functions
(RBF) since these functions possess a universal approximation property, which
156 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

σ(x)
0.5

0
0
x

Fig. 7.5 Sigmoidal function

implies that any nonlinear function can be approximated by these functions. Note
that a polynomial in some sense can be viewed as a universal approximator since
any sufficiently smooth function can be represented by a Taylor series.
A sigmoidal function or also called a logistic function as shown in Fig. 7.5 is
described by
1
σ (x) = (7.10)
1 + e−ax

An unknown function can be approximated by a sigmoidal neural network as

   
ŷ = fˆ (x) = V  σ Wx x + W0 + V0 = Θ  Φ W  x̄ (7.11)

where V ∈ Rm × Rn andWx ∈ R n × Rm are weight matrices,  W0 ∈ R and

m
V0 ∈ Rn
   
are bias vectors, Θ = V0 V ∈ R × R , W = W0 Wx ∈ R × Rn+1 ,
n m+1 m
         
x̄ = 1 x  ∈ Rn+1 , Φ W  x̄ = 1 σ  W  x̄ ∈ Rm+1 , and σ W  x̄ =
          
σ W1 x̄ σ W2 x̄ . . . σ Wm x̄ ∈ Rm , where W j ∈ Rn+1 , j = 1, . . . , m are
column vectors of W .
A radial basis function (RBF), more commonly known as a Gaussian normal
distribution or “bell-shaped” curve as shown in Fig. 7.6, is given by

ψ (x) = e−ax
2
(7.12)

An unknown function can be approximated by an RBF neural network as

   
ŷ = fˆ (x) = V  ψ Wx x + W0 + V0 = Θ  Φ W  x̄ (7.13)

As with a linear regression, the coefficients of a neural network can be estimated

from the input–output data. The process of estimation is called “training” or “teach-
ing” a neural network. This is essentially a least-squares estimation. A “pre-trained”
7.2 Neural Network Approximation 157

ψ(x)
0.5

0
0
x

Fig. 7.6 Radial basis function

or “off-line” neural network is one whose coefficients are estimated a priori from data
or model information. An “on-line learning” neural network is one whose coefficients
are recursively estimated usually from a gradient method such as “back propagation”
in real time.
Consider a sigmoidal neural network approximation. The gradient method can be
applied to derive least-squares learning laws for a neural network to minimize the
following cost function:

1  1          
J= ε ε= Θ Φ W x̄ − y Θ Φ W x̄ − y (7.14)
2 2
Taking the partial derivatives of J with respect to the weight matrices and bias
vectors yields
∂ J ∂ε   
∇ JΘ = 
= 
ε = Φ W  x̄ ε (7.15)
∂Θ ∂Θ
 
∂J ∂ W  x̄  ∂ε   
∇ JW = = 
ε    = x̄ε V  σ W  x̄ (7.16)
∂W ∂W ∂ W x̄
          
where σ W  x̄ = diag σ W1 x̄ , σ W2 x̄ , . . . , σ Wm x̄ is a diagonal
matrix with
 ae−ax
σ (x) = (7.17)
(1 + e−ax )2

Then, the least-squares gradient learning laws for the neural network are given by
 
Θ̇ = −ΓΘ Φ W  x̄ ε (7.18)
  
Ẇ = −ΓW x̄ε V  σ W  x̄ (7.19)
158 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

Example 7.3 Consider Example 7.1. Approximate

y = sin x + cos 2x + e−x

2

over a compact domain x (t) ∈ [−1, 1], where x (t) = sin 10t.
y (t) is approximated by the following sigmoidal neural network:
     
ŷ = v0 + v1 σ w01 + w1 x + v2 σ w02 + w2 x + v3 σ w03 + w3 x

⎤ ⎡
v0 ⎡ ⎤
v1  
⎢ v1 ⎥
Θ=⎢ ⎥ , V = ⎣ v2 ⎦ , W = w01 w02 w03
⎣ v2 ⎦ w1 w2 w3
v3
v3
⎡ ⎤
   1 
1   ⎢ σ w01 + w1 x ⎥
x̄ = , Φ W x = ⎢  ⎥
⎣ σ w02 + w2 x ⎦ ,
x  
σ w03 + w3 x
⎡   ⎤
σ w01 + w1 x 0 0
    
σ W  x̄ = ⎣ ⎦

0 σ w02 + w2 x 0
  
0 0 σ w03 + w3 x

Note that the initial conditions of Θ (t) and W (t) are usually set by a random
number generator. Otherwise, if they are zero, then V (t) and W (t) will have identical
elements, that is, v1 = v2 = v3 , w01 = w02 = w03 , and w1 = w2 = w3 . For this
example, let ΓΘ = ΓW = 10I . The neural network approximation is shown in
Fig. 7.7.

5
Θ, W

−5
0 200 400 600 800 1000
t

2
y, Θ Φ(W x)

|
T

1
T

0 y
Θ Φ(W −
T T
x)
−1
−1 −0.5 0 0.5 1
t

Fig. 7.7 Sigmoidal neural net approximation

7.2 Neural Network Approximation 159

Note that the neural network approximation converges much more slowly than the
Chebyshev polynomial approximation in Example 7.1. Some of the neural network
weights are still converging even after 1000 s. The root mean square error is 6.9×10−4
which is about twice that for the Chebyshev polynomial approximation. The neural
network has ten weighting coefficients while the Chebyshev polynomial has only
five coefficients.

7.3 Adaptive Control with Unstructured Uncertainty

Consider the following system with a matched unstructured uncertainty:

ẋ = Ax + B [u + f (x)] (7.20)

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, A ∈ Rn × Rn is

a constant, but unknown matrix, B ∈ Rn × Rm with n ≤ m is a known matrix such
that (A, B) is controllable, and f (x) ∈ Rm is a matched unstructured uncertainty.
The unknown function f (x) can be approximated by a polynomial or any function
that can be expressed as

f (x) = Θ ∗ Φ (x) − ε∗ (x) (7.21)

where Θ ∗ ∈ Rl × Rm is an unknown constant ideal weight matrix, Φ (x) ∈ Rl is a

vector of orthogonal basis functions, and ε∗ (x) ∈ Rm is the function approximation
error.
Alternatively, the unknown function f (x) can also be approximated by a neural
network  
f (x) = Θ ∗ Φ W ∗ x̄ − ε∗ (x) (7.22)

where Θ ∗ ∈ Rl+1 × Rm and W ∗ ∈ Rn+1 × Rl are unknown constant ideal weight

       
matrices, x̄ = 1 x  ∈ Rn+1 , Φ W ∗ x̄ = 1 σ  W  x̄ ∈ Rl+1 .
∗
ε (x) can be made sufficiently small in a compact domain of x (t) such that
supx∈D ε∗ (x) ≤ ε0∗ for all x (t) ∈ D ⊂ Rn by a suitable selection of Φ (x) or
Φ W ∗ x̄ .
The objective is to design a full-state feedback in order for x (t) to follow a
reference model described by

ẋm = Am xm + Bm r (7.23)

where xm (t) ∈ Rn is a reference state vector, Am ∈ Rn × Rn is known and Hurwitz,

Bm ∈ Rn × Rq is known, and r (t) ∈ Rq ∈ L∞ is a piecewise continuous and
bounded reference command vector.
160 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

7.3.1 Recursive Least-Squares Direct Adaptive Control

with Matrix Inversion

We assume that there exist ideal control gains K x∗ and K r∗ that satisfy the model
matching conditions
A + B K x∗ = Am (7.24)

B K r∗ = Bm (7.25)

Let the adaptive controller be

u = K x (t) x + K r r − Θ  Φ (x) (7.26)

where K r = K r∗ is known.
Let  (t) be an estimate of A, then

 (t) = Am − B K x (t) (7.27)

The closed-loop plant is expressed as

  
ẋ = Am − B K x∗ x + B K x x + K r r − Θ  Φ (x) + Θ ∗ Φ (x) − ε∗ (x) (7.28)

Let K̃ x (t) = K x (t) − K x∗ and Θ̃ (t) = Θ (t) − Θ ∗ be the estimation errors, then
 
ẋ = Am + B K̃ x x + Bm r − B Θ̃  Φ (x) − Bε∗ (7.29)

Define a desired plant model as

ẋd = Am x + Bm r (7.30)

Then, formulate the plant modeling error as

ε = ẋd − ẋ = Âx + B ū − ẋ = −B K̃ x x + B Θ̃  Φ (x) + Bε∗ (7.31)

  
Let Ω̃  (t) = −B K̃ x (t) B Θ̃  (t) ∈ Rn ×Rn+l , and Ψ (x) = x  Φ  (x) ∈
Rn+l . Then, the plant modeling error can be expressed as

ε = Ω̃  Ψ (x) + Bε∗ (7.32)

The tracking error equation can be expressed in terms of the plant modeling error
as
ė = ẋm − ẋ = ẋm − ẋd + ẋd − ẋ = Am e + ε (7.33)
7.3 Adaptive Control with Unstructured Uncertainty 161

The pure RLS adaptive laws are

Ω̇ = −RΨ (x) ε (7.34)

Ṙ = −RΨ (x) Ψ  (x) R (7.35)

K x (t) and Θ (t) can be computed from Ω (t) as

       
Ω1 −K x B  K x −Ω1 B −
Ω= = ⇒ = (7.36)
Ω2 Θ B Θ Ω2 B −

Note that K x (t) and Θ (t) require the matrix inverse B −1 . This sometimes can
cause some issues which will be discussed later.
The estimation error equation for the RLS adaptive law is obtained as
 
Ω̃˙ = −RΨ (x) Ψ  (x) Ω̃ + ε∗ B  (7.37)

The RLS adaptive laws can be shown to result in a parameter convergence.

Proof Choose a Lyapunov candidate function

   
V Ω̃ = trace Ω̃  R −1 Ω̃ (7.38)
 
Then, V̇ Ω̃ is evaluated as
⎛ ⎞
 
⎜ ⎟
V̇ Ω̃ = trace ⎝−2 Ω̃  Ψ (x) ε + Ω̃  Ψ (x) Ψ  (x) Ω̃ ⎠
        
ε−Bε∗ ε−Bε∗ ε −ε∗ B 
    
= −2ε ε − Bε∗ + ε − ε∗ B  ε − Bε∗


= −ε ε + ε∗ B  Bε∗ ≤ − ε2 + B2 ε0∗2

 2  
   
≤ − Ψ (x)2 Ω̃  + 2 Ψ (x) Ω̃  B ε0∗ (7.39)
   
  2Bε0∗
We see that V̇ Ω̃ ≤ 0 if ε ≥ B ε0∗ and Ω̃  ≥ where Ψ0 = Ψ (x).
Ψ0
 
Therefore, ε (t) ∈ L∞ , Ω̃ (t) ∈ L∞ , and Ψ (x) ∈ L∞ . We evaluate V̈ Ω̃ as
162 7 Function Approximation and Adaptive Control with Unstructured Uncertainty
   
V̈ Ω̃ = −2ε ε̇ = −2ε Ω̃˙  Ψ (x) + Ω̃  Ψ̇ (x)
 
= −2ε −εΨ  (x) RΨ (x) + Ω̃  Ψ̇ (x)
⎡ ⎤
⎢  t −1 ⎥
⎢ ⎥
= −2ε ⎢−εΨ  (x) Ψ (x) Ψ  (x) dτ Ψ (x) + Ω̃  Ψ̇ (x)⎥
⎣ t0 ⎦
  
R
(7.40)
 
V̈ Ω̃ is bounded if Ψ (x) satisfies the PE condition and Ψ̇ (x) ∈ L∞ . Applying
 
  2Bε∗
the Barbalat’s lemma, we conclude that ε → B ε0∗ and Ω̃  → Ψ0 0 as
t → ∞. 

The tracking error can now be shown to be uniformly ultimate bounded.

Proof Choose a Lyapunov candidate function

   
V e, Ω̃ = e Pe + μtrace Ω̃  R −1 Ω̃ (7.41)

 μ > 0.
with
V̇ e, Ω̃ is evaluated as
 
V̇ e, Ω̃ = −e Qe + 2e Pε − με ε + με∗ B  Bε∗
≤ −λmin (Q) e2 + 2λmax (P) e ε − μ [[ε]] 2 + μ B2 ε0∗2
(7.42)

Using the inequality

λ2max (P)
2λmax (P) e ε ≤ e2 + δ 2 ε2 (7.43)
δ2
 
then V̇ e, Ω̃ can be expressed as

   
λ2max (P)
V̇ e, Ω̃ ≤ − λmin (Q) − e2 −(μ−δ 2 )Ψ (x)2 Ω̃2 +μ B2 ε0∗2
δ2
  (7.44)
λmax (P) √
Thus, V̇ e, Ω̃ ≤ 0 if λ (Q) < δ < μ and is further bounded by
√
min
 
λ2max (P)
V̇ (e, Ω̃) ≤ − λmin (Q) − e2 + μB2 ε0∗2 (7.45)
δ2
7.3 Adaptive Control with Unstructured Uncertainty 163

Then, the largest lower bound of e is obtained by letting V̇ (e, Ω̃) ≤ 0 as
√
μ B ε0∗
e ≥  =p (7.46)
λ2max (P)
λmin (Q) − δ2

 
Similarly, V̇ e, Ω̃ is also bounded by
 
V̇ e, Ω̃ ≤ −(μ − δ 2 )Ψ (x)2 Ω̃2 + μB2 ε0∗2 (7.47)
 
 
Then, the largest lower bound of Ω̃  is obtained as

  √
  μBε0∗
Ω̃  ≥ =α (7.48)
Ψ0 μ − δ 2

Let D be a subset that contains all trajectories of e (t) and Ω̃ (t). Let Rα ⊂ D
and Rβ ⊂ D be two subsets in D defined by the lower bounds and upper bounds of
 
e and Ω̃ , respectively, where

!   "
 
Rα = e (t) ∈ Rn , Ω̃ (t) ∈ Rn+l × Rn : e ≤ p, Ω̃  ≤ α (7.49)
!   "
 
Rβ = e (t) ∈ Rn , Ω̃ (t) ∈ Rn+l × Rn : e ≤ ρ, Ω̃  ≤ β (7.50)
 
 
where ρ and β are the upper bounds of e and Ω̃ , respectively.
 
Then, we see that V̇ e, Ω̃ ≤ 0 for all e (t) ∈ D − Rα , that is, a region outside
Rα . Let Sα be the smallest subset that encloses Rα such that Rα ⊆ Sα . That is, Sα
is a hypersphere that circumscribes Rα which is given by
!     "
Sα = e (t) ∈ Rn , Ω̃ (t) ∈ Rn+l × Rn : V e, Ω̃ ≤ λmax (P) p 2 + μλmax R −1 α 2
(7.51)
where λmax (R −1 ) is the largest eigenvalue of R −1 (t) for all t.
Let Sβ be the largest subset inside Rβ that encloses Sα such that Sβ ⊆ Rβ .
That is, Sβ is an inscribed hypersphere in Rβ which is given by
!     "
Sβ = e (t) ∈ Rn , Ω̃ (t) ∈ Rn+l × Rn : V e, Ω̃ ≤ λmin (P) ρ 2 + μλmin R −1 β 2
(7.52)
where λmin (R −1 ) is the smallest eigenvalue of R −1 (t) for all t.
We see that Sα ⊆ Sβ . Then, the ultimate bound of e can be found by setting
Sα = Sβ [13]. Thus,
164 7 Function Approximation and Adaptive Control with Unstructured Uncertainty
   
λmin (P) ρ 2 + μλmin R −1 β 2 = λmax (P) p 2 + μλmax R −1 α 2 (7.53)

Fig. 7.8 Bounding sets for uniform ultimate boundedness

 
 
Since e ≤ ρ and Ω̃  ≤ β in Sβ , it follows that [13]

 
λmin (P) e2 ≤ λmax (P) p 2 + μλmax R −1 α 2 (7.54)

 
 2
  
μλmin R −1 Ω̃  ≤ λmax (P) p 2 + μλmax R −1 α 2 (7.55)
 
 
Then, the ultimate bounds of e and Ω̃  are obtained as

#  
λmax (P) p 2 + μλmax R −1 α 2
e ≤ =ρ (7.56)
λmin (P)
#  
  λmax (P) p 2 + μλmax R −1 α 2
 
Ω̃  ≤   =β (7.57)
μλmin R −1

The bounding sets for uniform ultimate boundedness are illustrated in Fig. 7.8.

7.3.2 Modified Recursive Least-Squares Direct Adaptive

Control without Matrix Inversion

In the RLS adaptive laws, the matrix inverse B −1 is used to compute the adaptive
parameters K (t) and Θ (t). This sometimes can cause issues when B −1 is small
7.3 Adaptive Control with Unstructured Uncertainty 165

or when B ∈ Rn × Rm with n > m. The RLS adaptive control can be modified to

eliminate the need for thes matrix inversion of the B matrix in the case of B ∈ Rn ×Rm
with n > m.   
Let Ω̃  (t) = − K̃ x (t) Θ̃  (t) ∈ Rm × Rn+1 and Ψ (x) = x  Φ  (x) ∈
Rn+l . Then, the plant modeling error can be expressed as

ε = B Ω̃  Ψ (x) + Bε∗ (7.58)

The modified RLS adaptive laws are

Ω̇ = −RΨ (x) ε B (7.59)

Ṙ = −η RΨ (x) Ψ  (x) R (7.60)

where η > 0 is an adjustment factor.

K x (t) and Θ (t) can be computed from Ω (t) as
       
Ω1 −K x K x −Ω1
Ω= = ⇒ = (7.61)
Ω2 Θ Θ Ω2

Note that K x (t) and Θ (t) no longer require the matrix inverse B −1 .
The estimation error equation for the modified RLS adaptive law is obtained as
 
Ω̃˙ = −RΨ (x) Ψ  (x) Ω̃ + ε∗ B  B (7.62)

The tracking error equation can be expressed in terms of the plant modeling
error as
ė = ẋm − ẋ = ẋm − ẋd + ẋd − ẋ = Am e + ε = Am e + B Ω̃  Ψ (x) + Bε∗ (7.63)

The RLS adaptive

 laws can be shown to result in a parameter convergence if
η < 2λmin B  B as follows:

Proof Choose a Lyapunov candidate function

   
V Ω̃ = trace Ω̃  R −1 Ω̃ (7.64)
 
Then, V̇ Ω̃ is evaluated as
166 7 Function Approximation and Adaptive Control with Unstructured Uncertainty
   
V̇ Ω̃ = trace −2Ω̃  Ψ (x) ε B + ηΩ̃  Ψ (x) Ψ  (x) Ω̃
= −2ε B Ω̃  Ψ (x) + ηΨ  (x) Ω̃ Ω̃  Ψ (x)
= −2ε ε + 2ε Bε∗ + ηΨ  (x) Ω̃ Ω̃  Ψ (x)
 2
 
≤ −2 ε2 + 2 B ε ε0∗ + η Ψ (x)2 Ω̃ 
    2    
   
≤ − 2λmin B  B − η Ψ (x)2 Ω̃  + 2λmax B  B Ψ (x) Ω̃  ε0∗
(7.65)
     
  c ε∗
Thus, V̇ Ω̃ ≤ 0 if η < 2λmin B  B and Ω̃  ≥ Ψ1 00 where c1 = 2λ maxB( B −η
2λ B B)
.
min ( )
We note that
   2  
 
λmin B  B Ψ (x)2 Ω̃  ≤ ε2 + 2 B ε ε0∗ + λmax B  B ε0∗2 (7.66)
 
Therefore, V̇ Ω̃ is also expressed in terms of ε as
$ % $ %
  η η
V̇ Ω̃ ≤ − 2 −   ε2 + 2 1 +   B ε ε0∗
λmin B  B λmin B  B
ηλmax (B  B) ∗2
+ ε ≤0 (7.67)
λmin (B  B) 0
 & 
λmin ( B  B )+η
1 + 1 + [ min ( ) 2] .
η 2λ B  B −η
which yields ε ≥ c2 B ε0∗ where c2 = 2λmin ( B  B )−η [λmin ( B B )+η
]
Therefore, ε (t) ∈ L∞ , Ω̃ (t) ∈ L∞ , and Ψ (x) ∈ L∞ . We evaluate V̈ Ω̃ as
     
V̈ Ω̃ = −2 Ψ  (x) Ω̃ 2B  B − ηI + ε∗ B  B
⎡ ⎤
⎢  t −1 ⎥
⎢    ⎥
× ⎢−B εΨ (x) Ψ (x) Ψ (x) dτ Ψ (x) + Ω̃  Ψ̇ (x)⎥
⎣ t0 ⎦
  
R
(7.68)
 
V̈ Ω̃ is bounded if Ψ (x) satisfies the PE condition and Ψ̇ (x) ∈ L∞ . Applying
 
  c Bε∗
the Barbalat’s lemma, we conclude that ε → c2 B ε0∗ and Ω̃  → 1 Ψ0 0 as
t → ∞. 
The tracking error can now be shown to be uniformly ultimately bounded.
Proof Choose a Lyapunov candidate function
   
V e, Ω̃ = e Pe + μtrace Ω̃  R −1 Ω̃ (7.69)
7.3 Adaptive Control with Unstructured Uncertainty 167

with μ
 > 0.
V̇ e, Ω̃ is evaluated as
 
V̇ e, Ω̃ = −e Qe + 2e P B Ω̃  Ψ (x) + 2e P Bε∗
− 2με ε + 2με Bε∗ + μηΨ  (x) Ω̃ Ω̃  Ψ (x)
 
 
≤ −λmin (Q) e2 + 2 P B e Ψ (x) Ω̃  + 2 P B e ε0∗
   2
   
− μ 2λmin B  B − η Ψ (x)2 Ω̃ 
   
 
+ 2μλmax B  B Ψ (x) Ω̃  ε0∗ (7.70)

Using the inequality

  P B2
 
2 P B e Ψ (x) Ω̃  ≤ e2 + δ 2 Ψ (x)2 Ω̃2 (7.71)
δ2
 
V̇ e, Ω̃ is expressed as

  ' (
P B2
V̇ e, Ω̃ ≤ − λmin (Q) − e2 + 2P Beε0 ∗
δ2
)    *  2
 
− μ 2λmin B  B − η − δ 2 Ψ (x)2 Ω̃ 
   
 
+ 2μλmax B  B Ψ (x) Ω̃  ε0∗ (7.72)
 
Thus, V̇ e, Ω̃ ≤ 0 if √P B < δ < μ[2λmin (B  B) − η]. Upon completing
λmin (Q)
the square, we obtain
     2
 
V̇ e, Ω̃ ≤ −c1 (e − c2 )2 + c1 c22 − c3 Ψ (x)2 Ω̃  − c4 + c3 c42 (7.73)

P B2 P Bε0∗ 
where c1 = λmin (Q) − δ2
, c2 = c1
, c3 = μ 2λmin (B  B) − η − δ 2 , and
μλ (B  B)ε∗
c4 = max c3 0
.
Then, the largest lower bound of e is obtained from

V̇ (e, Ω̃) ≤ −c1 (e − c2 )2 + c1 c22 + c3 c42 ≤ 0 (7.74)

which yeilds #
c3 c42
e ≥ c2 + c22 + =p (7.75)
c1
168 7 Function Approximation and Adaptive Control with Unstructured Uncertainty
 
 
Similarly, the largest lower bound of Ω̃  is obtained from

     2
 
V̇ e, Ω̃ ≤ −c3 Ψ (x)2 Ω̃  − c4 + c1 c22 + c3 c42 ≤ 0 (7.76)

which yields  2
  c1 c2
  c4 + + c42
Ω̃  ≥ c3
=α (7.77)
  Ψ0
  !
Thus, V̇ e, Ω̃ ≤ 0 outside a compact set defined by Rα = e (t) ∈ Rn , Ω̃ (t) ∈
  "
 
Rn+l × Rn : e ≤ p, Ω̃  ≤ α . Therefore, the tracking error is uniformly
ultimately bounded with the ultimate bound
#
λmax (P) p 2 + μλmax (R −1 )α 2
e ≤ =ρ (7.78)
λmin (P)

Note that since n > m, then η can be made independent of the B matrix by scaling
the modified RLS adaptive laws by the left pseudo-inverse as
 −1
Ω̇ = −RΨ (x) ε B B  B (7.79)

Then, the estimation error equation for the scaled RLS adaptive law becomes
 
Ω̃˙ = −RΨ (x) Ψ  (x) Ω̃ + ε (7.80)

A similar Lyapunov stability analysis can be performed to show that 0 ≤ η < 2,

which is independent of B for the scaled RLS adaptive laws.

7.3.3 Least-Squares Gradient Direct Adaptive Control

Least-squares gradient adaptive laws can be designed to not require the use of the
matrix inverse B −1 . The modified least-squares gradient adaptive laws are given by

K̇ x = Γx xε B (7.81)

Θ̇ = −ΓΘ Φ (x) ε B (7.82)

7.3 Adaptive Control with Unstructured Uncertainty 169

Note the resemblance of least-squares gradient adaptive laws to MRAC, where

the plant modeling error ε (t) replaces the tracking error e (t). The tracking error
equation can be expressed in terms of the plant modeling error as

ė = Am e + ε = Am e − B K̃ x x + B Θ̃  Φ (x) + Bε∗ (7.83)

The parameter convergence and uniform ultimate boundedness can be established

by the following proof:

Proof Choose a Lyapunov candidate function

     
V K̃ x , Θ̃ = trace K̃ x Γx−1 K̃  + trace Θ̃  ΓΘ−1 Θ̃ (7.84)
 
Then, V̇ K̃ x , Θ̃ is evaluated as
     
V̇ K̃ x , Θ̃ = trace 2 K̃ x xε B + trace −2Θ̃  Φ (x) ε B
 
= −2ε −B K̃ x x + B Θ̃  Φ (x)
= −2ε ε + 2ε Bε∗ ≤ −2 ε2 + 2 B ε ε0∗ (7.85)
 
We see that V̇ K̃ x , Θ̃ ≤ 0 if ε ≥ B ε0∗ . If Ψ̇ (x) ∈ L∞ , then ε → B ε0∗
as t → ∞.
For the tracking error, we use the Lyapunov candidate function
     
V e, K̃ x , Θ̃ = e Pe + μtrace K̃ x Γx−1 K̃  + μtrace Θ̃  ΓΘ−1 Θ̃ (7.86)

with μ >0. 
Then, V̇ e, K̃ x , Θ̃ is evaluated as

     
V̇ e, K̃ x , Θ̃ = −e Qe + 2e Pε + μtrace 2 K̃ x xε B + μtrace −2Θ̃  Φ (x) ε B

= −e Qe + 2e Pε − 2με ε + 2με Bε∗

≤ −λmin (Q) e2 + 2λmax (P) e ε − 2μ ε2 + 2μ B ε ε0∗
(7.87)

Using the inequality

λ2max (P)
2λmax (P)eε ≤ e2 + δ 2 ε2 (7.88)
δ2
and completing the square, we obtain
170 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

   
λ2max (P)
V̇ e, K̃ x , Θ̃ ≤ − λmin (Q) − e2
δ2
 
μBε0∗ 2 μ2 B2 ε0∗2
− (2μ − δ ) ε −
2
+ (7.89)
2μ − δ 2 (2μ − δ 2 )
  √
Then, V̇ e, K̃ x , Θ̃ ≤ 0 if √λλmax (P) < δ < 2μ. The largest lower bounds of
min (Q)
e and ε and then obtained as

μ B ε0∗
e ≥ &  =p (7.90)
λ2max (P)
(2μ − δ 2 ) λmin (Q) − δ2

2μBε0∗
ε ≥ =α (7.91)
2μ − δ 2
 
Since V̇ e, K̃ x , Θ̃ ≤ 0 outside a compact set, the closed-loop adaptive system
is uniformly ultimately bounded.

Example 7.4 Consider a first-order system with unstructured uncertainty

ẋ = ax + b [u + f (x)]

where a and f (x) are unknown, but b = 1. For simulation purposes, a = 1 and
f (x) = 0.2 sin 2x + cos 4x + e−x .
2

The reference model is given by

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t.

Since f (x) is unknown, a polynomial of q-degree is used to approximate f (x)
as
f (x) = a0 + a1 x + · · · + aq x q − ε∗ (x) = Θ ∗ Φ (x) − ε∗ (x)

where ai , i = 0, 1, . . . , q are constant unknown coefficients.

The adaptive controller is designed as

u = k x (t) x + kr r − Θ  (t) Φ (x)

where k x (t) and Θ (t) are computed by the least-squares gradient adaptive law as
   
k̇ −xεb
Ω̇ = x = −Γ
Θ̇ Φ (x) εb
7.3 Adaptive Control with Unstructured Uncertainty 171

with Ω (0) = 0 and Γ = I , where

ε = âx + bū − ẋ

â = am − bk x

ū = k x x + kr r

The tracking error for q = 1, 2, 3, 4 is shown in Figs. 7.9 and 7.10. Note that the
tracking error improves for q ≥ 2. Even the tracking error improves, the function
f (x) does not seem to be well approximated as shown in Fig. 7.11. This is also due
to k x (t) and Θ (t) not quite converging.
Now, suppose the Chebyshev orthogonal polynomials are used instead. Then,

f (x) = a0 + a1 T1 (x) + · · · + aq Tq (x) − ε∗ (x) = Θ ∗ Φ (x) − ε∗ (x)

The simulation results are as shown in Fig. 7.12. For q = 1, the result is the
same as that for the regular polynomial. However, it can be seen that the tracking
error significantly reduces for q = 2 with the Chebyshev polynomial and is even
smaller than that for q = 4 with the regular polynomial. For q = 4, the Chebyshev
polynomial approximation results in a very small tracking error. The parameter con-
vergence of the Chebyshev polynomial coefficients is excellent for q = 4 as shown
in Fig. 7.13. The unknown function is very well approximated by a fourth-degree
Chebyshev polynomial as shown in Fig. 7.14.

1 1

0.5 0.5

0 0
e

q=1 q=2
−0.5 −0.5

−1 −1
0 50 100 0 50 100
t t

1 1

0.5 0.5

0 0
e

q=3 q=4
−0.5 −0.5

−1 −1
0 50 100 0 50 100
t t

Fig. 7.9 Tracking error with regular q-th degree polynomial approximation
172 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

2
1
1

kx, Θ

k ,Θ
0
0 q=2

x
q=1 −1
−1

−2 −2
0 50 100 0 50 100
t t

2 2

1 1
k ,Θ

k ,Θ
0 0 q=4
q=3
x

x
−1 −1

−2 −2
0 50 100 0 50 100
t t

Fig. 7.10 k x and Θ with regular q-th degree polynomial approximation

1.5

1
ax+bf(x), ax+bΘTΦ(x)

0.5

0
<

−0.5

−1
ax+bf(x)
^
ax+bΘ
T
Φ(x) q=1
−1.5
^ T
ax+bΘ Φ(x) q=2
^ T
−2 ax+bΘ Φ(x) q=3
T
^
ax+bΘ Φ(x) q=4
−2.5
−1 −0.5 0 0.5 1
x

Fig. 7.11 Function approximation at t = 100 with regular q-th degree polynomial

1 1

0.5 0.5

0 0
e

q=1 q=2
−0.5 −0.5

−1 −1
0 50 100 0 50 100
t t

1 1

0.5 0.5

0 0
e

q=3 q=4
−0.5 −0.5

−1 −1
0 50 100 0 50 100
t t

Fig. 7.12 Tracking error with Chebyshev q-th degree polynomial approximation
7.3 Adaptive Control with Unstructured Uncertainty 173

2
1
1 q=2

k ,Θ

k ,Θ
0
0

x
q=1 −1
−1

−2 −2
0 50 100 0 50 100
t t

2 2

1 1
q=3 q=4
kx, Θ

k ,Θ
0 0

x
−1 −1

−2 −2
0 50 100 0 50 100
t t

Fig. 7.13 k x and Θ with Chebyshev q-th degree polynomial approximation

1.5

1
ax+bf(x), ax+bΘTΦ(x)

0.5

< 0

−0.5
ax+bf(x)
−1 ^ax+bΘTΦ(x) q=1
^ T
ax+bΘ Φ(x) q=2
−1.5 ^ T
ax+bΘ Φ(x) q=3
T
^
ax+bΘ Φ(x) q=4
−2
−1 −0.5 0 0.5 1
x

Fig. 7.14 Function approximation at t = 100 with Chebyshev q-th degree polynomial

7.3.4 Least-Squares Gradient Direct Adaptive Control with

Neural Network Approximation

If f (x) is approximated by a neural network, then the adaptive controller is expressed

as  
u = K x (t) x + K r r − Θ  Φ W  x̄ (7.92)

Then, the plant modeling error is expressed as

   
ε = ẋd − ẋ = Âx + B ū − ẋ = −B K̃ x x + BΘ  Φ W  x̄ − BΘ ∗ Φ W ∗ x̄ + Bε∗
(7.93)
Utilizing the Taylor series expansion, this can be written as
174 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

   
ε = −B K̃ x x + BΘ  Φ W  x̄ − B Θ  − Θ̃ 
       
× Φ W  x̄ + Φ W  x̄ W ∗ x̄ − W  x̄ + · · · + Bε∗ (7.94)

Combining the higher-order terms with the neural network approximation error
ε∗ (x), then
    
ε = −B K̃ x x + B Θ̃  Φ W  x̄ + BΘ  Φ W  x̄ W̃  x̄ + Bδ (7.95)

where δ (x) is the combined approximation error due to the Taylor series truncation
and the neural network approximation.
This becomes
    
ε = −B K̃ x x + B Θ̃  Φ W  x̄ + BV  σ W  x̄ W̃  x̄ + Bδ (7.96)

The least-squares gradient adaptive laws for K x (t), Θ (t), and W (t) are then
given by
K̇ x = Γx xε B (7.97)
 
Θ̇ = −ΓΘ Φ W  x̄ ε B (7.98)
  
Ẇ = −ΓW x̄ε BV  σ W  x̄ (7.99)

We will examine the parameter convergence and uniform ultimate boundedness

with the general least-squares gradient adaptive laws by performing the following
proof:
Proof We choose a Lyapunov candidate function
       
V K̃ x , Θ̃, W̃ = trace K̃ Γx−1 K̃  + trace Θ̃  ΓΘ−1 Θ̃ + trace W̃  ΓW−1 W̃
(7.100)
Then,
       
V̇ K̃ x , Θ̃, W̃ = 2trace K̃ x xε B + 2trace −Θ̃  Φ W  x̄ ε B
   
+ 2trace −W̃  x̄ε BV  σ W  x̄
    
= 2ε B K̃ x x − 2ε B Θ̃  Φ W  x̄ − 2ε BV  σ W  x̄ W̃  x̄
⎡ ⎤
⎢      ⎥
= −2ε ⎣−B K̃ x x + B Θ̃  Φ W  x̄ + BV  σ W  x̄ W̃  x̄ ⎦
  
ε−Bδ
 
= −2ε ε + 2ε Bδ ≤ −2 ε + 2 B ε δ0
2
(7.101)

where supx∈D δ (x) ≤ δ0 .

7.3 Adaptive Control with Unstructured Uncertainty 175
 
We see that V̇ K̃ x , Θ̃, W̃ ≤ 0 if ε ≥ B δ0 . Therefore, ε (t) ∈ L∞ , K̃ x (t) ∈
 
L∞ , Θ̃ (t) ∈ L∞ , W̃ (t) ∈ L∞ , x (t) ∈ L∞ , and Φ W  x̄ ∈ L∞ . If ẋ (t) ∈ L∞
  
and Φ̇ W x̄ ∈ L∞ , then using the Barbalat’s lemma, we can show that ε →
B δ0 as t → ∞. Note that we do not need to show for least-squares  gradient

adaptive laws that the PE condition must be satisfied in order for V̇ K̃ x , Θ̃, W̃ → 0
as t → ∞, but for exponential parameter convergence the PE condition is still
required.
The tracking error behavior is now examined next. Choose a Lyapunov candidate
function
     
V e, K̃ x , Θ̃, W̃ = e Pe + μtrace K̃ x Γx−1 K̃ x + μtrace Θ̃  ΓΘ−1 Θ̃
 
−1
+ μtrace W̃  ΓW W̃ (7.102)

Then,
 
V̇ e, K̃ x , Θ̃, W̃ = −e Qe + 2e Pε − 2με ε + 2με Bδ
≤ −λmin (Q) e2 + 2λmax (P) e ε − 2μ ε2
+ 2μ B ε δ0 (7.103)
 
Using the results in Sect. 7.3.3, V̇ e, K̃ x , Θ̃, W̃ ≤ 0 outside a compact set
defined by the following lower bounds of e and ε:

μ B δ0
e ≥ &  =p (7.104)
λ2max (P)
(2μ − δ 2 ) λmin (Q) − δ2

2μBδ0
ε ≥ =α (7.105)
2μ − δ 2

Therefore, the closed-loop adaptive system is uniformly ultimately bounded.

7.3.5 MRAC with Neural Network Approximation

By now, it can be seen that there are at least two approaches to adaptive control. One
approach is based on the Lyapunov method which is the basis for MRAC and another
approach is based on the least-squares adaptive control. The underlying principle of
both the MRAC and the least-squares adaptive control is to reduce system uncertainty
176 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

by either parameter identification of parametric structured uncertainty or function

approximation of unstructured uncertainty.
Consider the previous system with a matched unstructured uncertainty

ẋ = Ax + B [u + f (x)] (7.106)

where the unknown function f (x) is approximated by a linearly parametrized func-

tion
f (x) = Θ ∗ Φ (x) − ε∗ (x) (7.107)

Then, it can be shown that all the adaptive laws developed in Sects. 5.6 and 5.7
could be used to estimate Θ ∗ with some caveats as explained later. For example, if
A and B are known, then Eq. (5.155) from Sect. 5.6.2 can be used to update Θ (t) as

Θ̇ = −Γ Φ (x) e P B (7.108)

Since A and B are known, assuming there exist constant control gain matrices K x
and K r that satisfy the model matching conditions

A + B K x = Am (7.109)

B K r = Bm (7.110)

then the adaptive controller is given by

u = K x x + K r r − Θ  (t) Φ (x) (7.111)

where Θ (t) ∈ Rl × Rm is the estimate of Θ ∗ .

The closed-loop plant model becomes
⎛ ⎞

ẋ = ⎝ A + B K x ⎠ x + B K r r + B −Θ  Φ (x) + Θ ∗ Φ (x) − ε∗ (x) (7.112)
   
Am Bm

Let Θ̃ (t) = Θ (t) − Θ ∗ . The tracking error equation is obtained as

ė = ẋm − ẋ = Am e + B Θ̃  Φ (x) + Bε∗ (7.113)

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (7.114)

Then,
7.3 Adaptive Control with Unstructured Uncertainty 177
     
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) + Bε∗ − 2trace Θ̃  Φ (x) e P B
 
= −e Qe + 2e P B Θ̃  Φ (x) + Bε∗ − 2e P B Θ̃  Φ (x)
= −e Qe + 2e P Bε∗ ≤ −λmin (Q) e2 + 2 P B e ε0∗ (7.115)
 
Thus, V̇ e, Θ̃ ≤ 0 if

2 P B ε0∗
− λmin (Q) e2 + 2 P B e ε0∗ ≤ 0 ⇒ e ≥ =p (7.116)
λmin (Q)

At a first glance, it may be tempted to conclude that the tracking error is uniformly
ultimately bounded, but this conclusion would be incorrect since the tracking error is
only bounded. The reason for the least-squares adaptive laws to result in a uniformly
ultimately bounded tracking error is that any least-squares method guarantees that the
approximation error is at least bounded. This implies that the parameter estimation
error Θ̃ (t) is also bounded. In addition, if the uncertainty is structured and the basis
function is PE, then the parameter estimation error tends to zero, which implies a
parameter convergence. If both the tracking error and parameter estimation error can
be shown to be bounded, then the system
 is uniformly ultimately bounded.
For MRAC, the condition for V̇ e, Θ̃ ≤ 0 in Eq. (7.116) does not say anything
about the parameter estimation error Θ̃ (t). Let
) *
Br = e (t) ∈ Rn : e ≤ p (7.117)

Since Br does not contain any information on Θ̃ (t), Br is an open

 set withΘ̃ (t)
free to grow inside Br as illustrated in Fig. 7.15. A trajectory of e (t) , Θ̃ (t) may
 
go outside of Br , but V̇ e, Θ̃ ≤ 0 is outside of Br , so the trajectory will go back

Fig. 7.15 Trajectory of MRAC with unstructured uncertainty

178 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

toward Br . However, if Θ̃ (t) grows too large inside Br , the small approximation
error assumption may be violated. Consequently, the boundary of Br may also grow
and therefore e (t) can also grow. The possibility of a parameter drift that can lead to
a point where the system becomes unstable may exist when an adaptive law does not
provide a mathematically guaranteed bound on adaptive parameters. As a result, the
standard MRAC as developed in Chap. 5 for systems with unstructured uncertainty
is non-robust. Instability due to a parameter drift when adaptive parameters are not
bounded is a well-known issue with adaptive control. In this regard, least-squares
adaptive control offers an advantage over the standard MRAC since the adaptive
parameters are bounded. Robust modification of MRAC is another way to address
this issue. In summary, it is generally not a good practice to use the standard MRAC
for systems with unstructured uncertainty.

Example 7.5 Consider Example 7.4. The MRAC laws are given by

k̇ x = γx xeb

Θ̇ = −ΓΘ Φ (x) eb

where e (t) = xm (t) − x (t), γx = 1, and ΓΘ = I .

Let Φ (x) be a vector of the basis Chebyshev polynomials up to T4 (x). The
simulation results are shown in Figs. 7.16, 7.17, and 7.18.
Comparing to Figs. 7.12, 7.13, and 7.14, the tracking error is not as good with the
MRAC laws as with the least-squares gradient adaptive laws. The parameters k x (t)
and Θ (t) are more oscillatory and do not converge. For some reason, the second-
degree Chebyshev polynomial produces the best tracking for the MRAC laws. The
function approximation by the MRAC laws is poorer than that by the least-squares
gradient method. This is not surprising since MRAC is not intended to be a parameter
identification method. Its sole purpose is to reduce the tracking error. 

1 1

0.5 0.5

0 0
e

q=1 q=2
−0.5 −0.5

−1 −1
0 50 100 0 50 100
t t

1 1

0.5 0.5

0 0
e

q=3 q=4
−0.5 −0.5

−1 −1
0 50 100 0 50 100
t t

Fig. 7.16 Direct MRAC with Chebyshev fourth-degree polynomial

7.3 Adaptive Control with Unstructured Uncertainty 179

2
1
1

kx, Θ

k ,Θ
0
0

x
q=1
−1 −1
q=2

−2 −2
0 50 100 0 50 100
t t

2 2

1 1
k ,Θ

k ,Θ
0 0
x

x
−1 −1
q=3 q=4
−2 −2
0 50 100 0 50 100
t t

Fig. 7.17 k x and Θ with Chebyshev q-th degree polynomial approximation

2
ax+bf(x), ax+bΘ Φ(x)

1
T

< 0

−1
ax+bf(x)
^ax+bΘTΦ(x) q=1
^ax+bΘTΦ(x) q=2
−2
^ T
ax+bΘ Φ(x) q=3
T
^
ax+bΘ Φ(x) q=4
−3
−1 −0.5 0 0.5 1
x

Fig. 7.18 Function approximation at t = 100 with Chebyshev q-th degree polynomial

Now, consider neural network MRAC when f (x) is approximated by a neural

network. The adaptive controller is then modified as
 
u = K x x + K r r − Θ  Φ W  x̄ (7.118)

The tracking error equation is obtained as

    
ė = Am e + B Θ̃  Φ W  x̄ + BV  σ W  x̄ W̃  x̄ + Bδ (7.119)

Proof Choose a Lyapunov candidate function

     
V e, Θ̃, W̃ = e Pe + trace Θ̃  ΓΘ−1 Θ̃ + trace W̃  ΓW−1 W̃ (7.120)
180 7 Function Approximation and Adaptive Control with Unstructured Uncertainty
 
Differentiating V e, Θ̃, W̃ yields
        
V̇ e, Θ̃, W̃ = −e Qe + 2e P B Θ̃  Φ W  x̄ + BV  σ W  x̄ W̃  x̄ + Bδ
   
+ 2trace Θ̃  ΓΘ−1 Θ̃˙ + 2trace W̃  ΓW−1 W̃˙ (7.121)

Note that
     
e P B Θ̃  Φ W  x̄ = trace Θ̃  Φ W  x̄ e P B (7.122)

      
e P BV  σ W  x̄ W̃  x̄ = trace W̃  x̄e P BV  σ W  x̄ (7.123)

Hence,
      
V̇ e, Θ̃, W̃ = −e Qe + 2e P Bδ + 2trace Θ̃  Φ W  x̄ e P B + ΓΘ−1 Θ̃˙
     
+ 2trace W̃  x̄e P BV  σ W  x̄ + ΓW−1 W̃˙ (7.124)

Setting the trace terms to zero yields the neural network adaptive laws
 
Θ̇ = −ΓΘ Φ W  x̄ e P B (7.125)
  
Ẇ = −ΓW x̄e P BV  σ W  x̄ (7.126)

Then,
 
V̇ e, Θ̃, W̃ = −e Qe + 2e P B ε̄ ≤ −λmin (Q) e + 2 P B e δ0 (7.127)
 
Therefore, V̇ e, Θ̃, W̃ ≤ 0 if

2 P B δ0
e ≥ =p (7.128)
λmin (Q)

As discussed previously, stability of the tracking error cannot be guaranteed due

to the potential for a parameter drift. However, since the neural network activation
function is a bounded function; therefore, in comparison with other types of function
approximation, neural network MRAC tends to be more robust than the standard
MRAC.

Example 7.6 Consider Example 7.4. f (x) is to be approximated by a neural net-

work. The MRAC neural network adaptive laws are given by

k̇ x = γx xeb
7.3 Adaptive Control with Unstructured Uncertainty 181
 
Θ̇ = −ΓΘ Φ W  x̄ eb
  
Ẇ = −ΓW x̄ebV  σ W  x̄

where Θ (t) ∈ R5 , W (t) ∈ R2 × R4 and ΓΘ = ΓW = 10I . The initial conditions

are given by
⎡
⎤
−0.2
⎢ −0.1 ⎥  
⎢ ⎥ 0 0.1 0.2 0.3
Θ (0) = ⎢
⎢ 0 ⎥ , W (0) =
⎥
⎣ 0.1 ⎦ −0.3 −0.2 −0.1 0
−0.2

The neural network topology is shown in Fig. 7.19. The simulation results are
shown in Fig. 7.20.

Fig. 7.19 Neural network topology

2 1

1 0
x, xm
e

0 −1
x xm
−1 −2
0 50 100 0 50 100
t t

1 2

0
0
kx, Θ
W

−1
−2
−2

−3 −4
0 50 100 0 50 100
t t

Fig. 7.20 Neural net adaptive control system response

182 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

The neural network adaptive control appears to reduce the tracking error a bit
better than the direct MRAC with Chebyshev polynomial approximation. However,
there is a large transient at the beginning. This is probably due to the initial conditions
of Θ (t) and W (t) being chosen arbitrarily, and the poor initial approximation that
causes the tracking error to grow before the learning becomes effective in reducing
the tracking error. The neural network adaptive control does not seem to perform as
well as the least-squares gradient method with the Chebyshev polynomials. 

In the context of least-squares adaptive control, the plant modeling error ε (t)
represents a modeling error between the true plant and the desired plant. Modeling
errors in dynamical systems usually arise from the presence of system uncertainty.
In model-reference adaptive control, the tracking error e (t) is exclusively used as a
control signal to drive the adaptation so as to cancel out undesired effects of system
uncertainty on closed-loop systems. Philosophically, the tracking error is a mani-
festation of the plant modeling error which in turn is a direct consequence of the
system uncertainty. Thus, the tracking error is a result of the plant modeling error,
but not vice versa. One can separate control and system identification as two related
dynamical actions. A control action is generally designed to reduce the tracking
error to ensure that a closed-loop system follows a reference model as closely as
possible. The objective of adaptive control is not so much to estimate the uncertainty
itself but rather to achieve a desired reference model, regardless whether or not a
parameter convergence is achieved. Example 7.5 clearly demonstrates that model-
reference adaptive control does a rather poor job on parameter estimation even with
the Chebyshev polynomials.
On the other hand, a system identification action uses the plant modeling error
to estimate the uncertainty directly by least-squares estimation methods. The goal
of a system identification is first and foremost to achieve a parameter convergence.
The system identification action can achieve a better parameter convergence than
MRAC. System identification is usually performed in an open-loop process without
directly influencing a control action. Both the system identification and control can
also be combined together to provide a very effective adaptation strategy. Using
parameter estimates, least-squares adaptive control can be formulated to cancel out
effects of system uncertainty by feeding back the parameter estimates into an adaptive
control law. In so doing, least-squares adaptive control achieves bounded tracking and
bounded adaptive parameters, whereas MRAC can only achieve bounded tracking.

7.4 Summary

In many physical applications, there is no clear certainty about the structure between
the input and output of a process. In systems with unstructured uncertainty, the map-
ping between the input and output is usually not known. Polynomial approximation
by least-squares methods is a well-known regression method. Regular polynomials
are frequently used in least-squares data regression. Other types of polynomials can
7.4 Summary 183

also be used. One class of polynomials that can offer better function approximation
over the regular polynomials is orthogonal polynomials. Orthogonal polynomials
form true basis functions that span the Euclidean space. In particular, the Chebyshev
orthogonal polynomials are generally considered to be optimal that can accurately
approximate real-valued functions with the smallest degree for the same accuracy as
a regular polynomial approximation.
Neural networks have been used in many applications such as classifiers, pat-
tern recognition, function approximation, and adaptive control. A neural network is
a nonlinear function approximation that models a complex input–output nonlinear
mapping. Two basic types of neural networks are discussed: sigmoidal or logistic
functions and radial basis functions. Various adaptive control methods for unstruc-
tured uncertainty are presented. Both least-squares and model-reference adaptive
control with polynomial and neural network function approximation are shown to
be able to handle adaptive control systems with unstructured uncertainty. Least-
squares adaptive control, in general, exhibits better performance and robustness than
MRAC. Robustness of least-squares adaptive control is achieved through the nat-
ural parameter identification process which ensures parameter boundedness, whereas
MRAC cannot guarantee boundedness of adaptive parameters. This leads to robust-
ness issues, such as a parameter drift.

7.5 Exercises

1. Approximate
y = 0.1 sin 0.4x + cos2 2x

where x (t) = sin t for t ∈ [0, 60], by a fourth-degree Chebyshev polynomial

using the least-squares gradient method with Γ = 100I and Δt = 0.001. Ini-
tialize Θ (t) with zero. Plot Θ (t) versus t. Plot y (t) and ŷ (t) versus x (t) on the
same plot. Compute the root mean square error between y (t) and ŷ (t).
2. Implement a sigmoidal neural network
   
ŷ = fˆ (x) = V  σ Wx x + W0 + V0 = Θ  Φ W  x̄

where
1
σ (x) =
1 + e−x

to approximate y (t) in Exercise 7.1 with Θ (t) ∈ R5 , W (t) ∈ R2 × R4 and

ΓΘ = ΓW = 100I and Δt = 0.001. The initial conditions Θ (0) and W (0) are
to be generated by a random number generator. Plot Θ (t) and W (t) versus t.
Plot y (t) and ŷ (t) versus x (t) on the same plot. Compute the root mean square
error between y (t) and ŷ (t).
3. Consider a first-order system with a matched unstructured uncertainty
184 7 Function Approximation and Adaptive Control with Unstructured Uncertainty

ẋ = ax + b [u + f (x)]

where a and f (x) are unknown, but b = 2. For simulation purposes, a = 1 and
f (x) = 0.1 sin 0.4x + cos2 2x.
The reference model is given by

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t.

Implement in Simulink a direct adaptive control using a least-squares gradient
method to approximate f (x) by a 4th -degree Chebyshev polynomial. All initial
conditions are zero. Use Γ = 0.2I . Plot e (t), x (t) versus xm (t), k x (t), and Θ (t)
for t ∈ [0, 60].

References

1. Lee, T., & Jeng, J. (1998). The chebyshev-polynomials-based unified model neural networks
for function approximation. IEEE Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics, 28(6), 925–935.
2. Mason, J., & Handscomb, D. (2002). Chebyshev polynomials. Boca Raton: Chapman and
Hall/CRC.
3. Cybenko, G. (1989). Approximation by superpositions of a sigmoidal function. Mathematics
of Control Signals Systems, 2, 303–314.
4. Suykens, J., Vandewalle, J., & deMoor, B. (1996). Artificial neural networks for modeling and
control of non-linear systems. Dordrecht: Kluwer Academic Publisher.
5. Wang, X., Huang, Y., & Nguyen, N. (2010). Robustness quantification of recurrent neural
network using unscented transform. Elsevier Journal of Neural Computing, 74(1–3).
6. Calise, A. J., & Rysdyk, R. T. (1998). Nonlinear adaptive flight control using neural networks.
IEEE Control System Magazine, 18(6), 1425.
7. Ishihara, A., Ben-Menahem, S., & Nguyen, N. (2009). Protection ellipsoids for stability analysis
of feedforward neural-net controllers. In International Joint Conference on Neural Networks.
8. Johnson, E. N., Calise, A. J., El-Shirbiny, H. A., & Rysdyk, R. T. (2000). Feedback linearization
with neural network augmentation applied to X-33 attitude control. AIAA Guidance, Naviga-
tion, and Control Conference, AIAA-2000-4157.
9. Kim, B. S., & Calise, A. J. (1997). Nonlinear flight control using neural networks. Journal of
Guidance, Control, and Dynamics, 20(1), 26–33.
10. Lam, Q., Nguyen, N., & Oppenheimer, M. (2012). Intelligent adaptive flight control using
optimal control modification and neural network as control augmentation layer and robustness
enhancer. In AIAA Infotech@Aerospace Conference, AIAA-2012-2519.
11. Lewis, F. W., Jagannathan, S., & Yesildirak, A. (1998). Neural network control of robot manip-
ulators and non-linear systems. Boca Raton: CRC.
12. Zou, A., Kumar, K., & Hou, Z. (2010). Attitude control of spacecraft using chebyshev neural
networks. IEEE Transactions on Neural Networks, 21(9), 1457–1471.
13. Khalil, H. K. (2001). Nonlinear systems. Upper Saddle River: Prentice-Hall.
Chapter 8
Robustness Issues with Adaptive Control

Abstract This chapter discusses limitations and weaknesses of model-reference

adaptive control. Parameter drift is the result of the lack of a mathematical guarantee
of boundedness of adaptive parameters. Systems with bounded external disturbances
under feedback control actions using model-reference adaptive control can experi-
ence a signal growth of a control gain or an adaptive parameter even though both
the state and control signals remain bounded. This signal growth associated with the
parameter drift can cause instability of adaptive systems. Model-reference adaptive
control for non-minimum phase systems presents a major challenge. Non-minimum
phase systems have unstable zeros in the right half plane. Such systems cannot toler-
ate large control gain signals. Model-reference adaptive control attempts to seek the
ideal property of asymptotic tracking. In so doing, an unstable pole-zero cancelation
occurs that leads to instability. For non-minimum phase systems, adaptive control
designers generally have to be aware of the limiting values of adaptive parameters in
order to prevent instability. Time-delay systems are another source of challenge for
model-reference adaptive control. Many real systems have latency which results in a
time delay at the control input. Time delay is caused by a variety of sources such as
communication bus latency, computational latency, transport delay, etc. Time-delay
systems are a special class of non-minimum phase systems. Model-reference adaptive
control of time-delay systems is sensitive to the amplitude of the time delay. As the
time delay increases, robustness of model-reference adaptive control decreases. As
a consequence, instability can occur. Model-reference adaptive control is generally
sensitive to unmodeled dynamics. In a control system design, high-order dynamics
of internal states of the system sometimes are neglected in the control design. The
neglected internal dynamics, or unmodeled dynamics, can result in loss of robust-
ness of adaptive control systems. The mechanism of instability for a first-order SISO
system with a second-order unmodeled actuator dynamics is presented. The insta-
bility mechanism can be due to the frequency of a reference command signal or an
initial condition of an adaptive parameter that coincides with the zero phase mar-
gin condition. Fast adaptation is referred to the use of a large adaptation rate to
achieve the improved tracking performance. An analogy of an integral control action
of a linear time-invariant system is presented. As the integral control gain increases,
the cross-over frequency of the closed-loop system increases. As a consequence,
the phase margin or time-delay margin of the system decreases. Fast adaptation

© Springer International Publishing AG 2018 185

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_8
186 8 Robustness Issues with Adaptive Control

of model-reference adaptive control is analogous to the integral control of a linear

control system whereby the adaptation rate plays the equivalent role as the integral
control gain. As the adaptation rate increases, the time-delay margin of an adaptive
control system decreases. In the limit, the time-delay margin tends to zero as the
adaptation rate tends to infinity. Thus, the adaptation rate has a strong influence on
the closed-loop stability of an adaptive control system.

Model-reference adaptive control (MRAC) can be used in an adaptive control system

to track a reference command signal by estimating directly the control gains or
indirectly the adaptive parameters so as to cancel out the unwanted effect of the system
uncertainty. Asymptotic tracking can be achieved if the uncertainty is structured.
However, with unstructured uncertainty, MRAC is generally non-robust since the
bounds on the adaptive parameters cannot be established by the Lyapunov stability
analysis. Consequently, a parameter drift can result that can cause an adaptive control
algorithm to blow up [1]. As the complexity of a plant increases, robustness of MRAC
becomes increasingly difficult to ascertain. In practical applications, the knowledge
of a real plant can never be established precisely. Thus, a model of a physical plant
often cannot fully capture all the effects due to unmodeled dynamics, unstructured
uncertainty, and exogenous disturbances that may exist in a real plant. All these
effects can produce closed-loop dynamics that can lead to instability when MRAC
is used in an adaptive controller [2]. Some of the causes of robustness issues with
adaptive control are discussed in this chapter.
The learning objectives of this chapter are:
• To understand robustness issues with the parameter drift due to unbounded
adaptive parameters in the presence of exogenous disturbances and unstruc-
tured uncertainty;
• To develop an understanding of and to know how to deal with the non-minimum
phase behaviors due to unstable zeros that cause instability of closed-loop adaptive
systems resulting from an unstable pole-zero cancellation;
• To be able to recognize and to deal with time-delay systems which can cause
instability of closed-loop adaptive systems when adaptive parameters are permitted
to be unconstrained;
• To understand robustness issues with unmodeled dynamics due to complex behav-
iors of systems which impose design constraints that adaptive control fails to
account for; and
• To be familiar with the concept of fast adaptation which results in loss of robustness
as the time-delay margin of a closed-loop adaptive system tends to zero.

8.1 Parameter Drift

Parameter drift is a consequence when the bound on an adaptive parameter cannot

be established by the Lyapunov stability analysis. Consider a MIMO system with an
unknown bounded disturbance
8.1 Parameter Drift 187

ẋ = Ax + Bu + w (8.1)

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, A ∈ Rn × Rn is a

constant but unknown matrix, B ∈ Rn × Rm is a known matrix such that (A, B) is
controllable, and w (t) ∈ Rn is an unknown bounded exogenous disturbance.
An adaptive controller is designed to achieve tracking of a reference model

ẋm = Am + Bm r (8.2)

where xm (t) ∈ Rn is a reference state vector, Am ∈ Rn × Rn is known and Hurwitz,

Bm ∈ Rn × Rq is known, and r (t) ∈ Rq is a piecewise continuous and bounded
reference command vector.
The ideal feedback gain K x∗ and feedforward gain K r can be determined from the
model matching conditions as

A + B K x∗ = Am (8.3)

B K r = Bm (8.4)

Suppose the disturbance is not accounted for, then one can design an adaptive
controller as
u = K x (t) x + K r r (8.5)

The feedback gain K x (t) is computed by the following MRAC law:

K̇ x = Γx xe P B (8.6)

where e (t) = xm (t) − x (t) is the tracking error, Γx = Γx > 0 ∈ Rn × Rn is an

adaptation rate matrix, and P = P  > 0 ∈ Rn × Rn solves the Lyapunov equation

P Am + A
m P = −Q (8.7)

where Q = Q  > 0 ∈ Rn × Rn is a positive-definite matrix.

The closed-loop plant is expressed as

ẋ = (A + B K x ) x + B K r r + w (8.8)

Let K̃ x (t) = K x (t) − K x∗ . Then, the tracking error equation is obtained as

ė = Am e − B K̃ x x − w (8.9)

Stability of the adaptive system is analyzed by the Lyapunov’s direct method.

188 8 Robustness Issues with Adaptive Control

Proof Choose a Lyapunov candidate function

   
V e, K̃ x = e Pe + trace K̃ x Γx−1 K̃ x (8.10)

Then,
 
V̇ e, K̃ x = −e Qe − 2e Pw ≤ −λmin (Q) e2 + 2λmax (P) e w0 (8.11)

where w0 = maxw∞.
Therefore, V̇ e, K̃ x ≤ 0 if

2λmax (P) w0
e ≥ =p (8.12)
λmin (Q)

However, the Lyapunov stability analysis provides no information on the bound

of K̃ x (t). Therefore, K̃ x (t) can potentially become unbounded in certain situations.


To further illustrate this point, consider a first-order SISO system

ẋ = ax + bu + w (8.13)

where a is unknown and b is known.

An adaptive controller is designed to regulate x (t) to follow a zero reference
model whereby xm (t) = 0 for all t as

u = k x (t) x (8.14)

Then, the adaptive law for k x (t) is given by

k̇ x = −γx x 2 b (8.15)

It can be shown that it is possible for a bounded response of x (t) to exist that
would result in an unbounded signal k x (t).

Example 8.1 Consider the following solution:

x = (1 + t)n

which is bounded for n ≤ 0.

Then,
 t
(1 + t)2n+1 − 1
k x − k x (0) = −γx b (1 + τ )2n dτ = −γx b
0 2n + 1
8.1 Parameter Drift 189

For k x (t) to be bounded, we require 2n + 1 < 0 or n < − 21 .

The control signal is computed as
 
(1 + t)2n+1 − 1
u = − γx b − k x (0) (1 + t)n
2n + 1

The disturbance w (t) that results in this particular solution of x (t) can be found by

 
(1 + t)2n+1 − 1
w = ẋ − ax − bu = n (1 + t)n−1 − a (1 + t)n + b γx b − k x (0) (1 + t)n
2n + 1

Both the control signal and the disturbance are bounded if 3n + 1 ≤ 0 or n ≤ − 13

and n = − 21 . Thus, the system is completely bounded if n < − 21 . However, if − 21 <
n ≤ − 13 , then x (t), u (t), and w (t) are bounded, but k x (t) becomes unbounded as
t → ∞. The boundedness of the adaptive system is illustrated in Table 8.1.
The response of the closed-loop system with a = 1, b = 1, n = − 12 5
, γx = 10,
x (0) = 1, and k x (0) = 0 is shown in Fig. 8.1.
This example demonstrates that it is possible to have bounded disturbances or
control signals but yet unbounded adaptive parameters.

One explanation of the parameter drift can be given as follows:
For a sufficiently large disturbance, the adaptive controller attempts to generate
a high-gain control in order to reduce the effect of the disturbance. In the limit, the
steady-state solution of x (t) is obtained as
w
x∗ = − (8.16)
a + k x∗

As w (t) is sufficiently large, k x (t) tends to a large negative value in order to keep
x (t) close to zero. This high-gain control can cause k x (t) to become unbounded. In
the limit as x (t) → 0 to achieve the ideal asymptotic tracking property of MRAC,
k x (t) → ∞. While Example 8.1 shows that x (t) can remain bounded even though
k x (t) is unbounded, in practice, a high-gain control with a large value of k x (t) can be

Table 8.1 Boundedness of example adaptive system

x (t) u (t) w (t) k x (t)
n>0 ∈
/ L∞ ∈
/ L∞ ∈
/ L∞ ∈
/ L∞
− 13 < n ≤ 0 ∈ L∞ ∈
/ L∞ ∈
/ L∞ ∈
/ L∞
− 21 < n ≤ − 13 ∈ L∞ ∈ L∞ ∈ L∞ ∈
/ L∞
n = − 21 ∈ L∞ ∈
/ L∞ ∈
/ L∞ ∈
/ L∞
n < − 21 ∈ L∞ ∈ L∞ ∈ L∞ ∈ L∞
190 8 Robustness Issues with Adaptive Control

1 0

−5
0.5

u
x
−10

0 −15
0 2500 5000 0 2500 5000
t t

15 0

10 −50

kx
w

5 −100

0 −150

−5 −200
0 2500 5000 0 2500 5000
t t

Fig. 8.1 Effect of parameter drift for n = − 12

5
and γx = 10

problematic since real plants can include other closed-loop behaviors that can lead to
instability when k x (t) becomes large. Thus, a parameter drift can lead to instability
in practical applications.

8.2 Non-minimum Phase Systems

A transfer function of a SISO system is called minimum phase if all of its poles and
zeros are in the left half plane. If any of the zeros lie in the right half plane, then
the system is called non-minimum phase. A non-minimum phase system is more
difficult to control since the existence of the unstable zeros can limit the values of
stable feedback gains. Increasing feedback gains beyond certain limits will cause the
closed-loop poles to become unstable.

Example 8.2 The system

y (s) s+1
=
u (s) (s + 2) (s + 3)
y=x

is minimum phase. For a feedback with u (s) = k x x (s) + kr r (s), the closed-loop
transfer function
y (s) kr
 G (s) = 2
r (s) s + (5 − k x ) s + 6 − k x

has stable closed-loop poles for all values of k x ∈ (−∞, 5).

On the other hand, the system
8.2 Non-minimum Phase Systems 191

x (s) s−1
=
u (s) (s + 2) (s + 3)

is non-minimum phase. The closed-loop transfer function

kr
G (s) =
s 2 + (5 − k x ) s + 6 + k x

has an unstable closed-loop pole at s = 1 when k x → −∞. The zero cross-over

value of k x when the pole becomes unstable occurs at k x = −6.


If the system exhibits a non-minimum phase behavior, then the adaptation can
result in instability due to MRAC attempting to perform an unstable pole-zero can-
cellation in order to achieve asymptotic tracking.

Example 8.3 Consider the system

ẋ = ax + bu − 2z + w

ż = −z + u

y=x

where a < 0 is unknown and w (t) is a disturbance.

The system is non-minimum phase with a transfer function

x (s) s−1
=
u (s) (s − a) (s + 1)

If the same adaptive controller is used for the same disturbance in Example 8.1
with n = −1, then the disturbance w (t) is bounded and the response of the closed-
loop system in Example 8.1 is completely bounded. However, this is no longer the
case for the non-minimum phase system in this example. Let a = −1, b = 1, γx = 1,
x (0) = 1, and k x (0). The closed-loop system is unstable as shown in Fig. 8.2.

In fact, the closed-loop system is unstable even when the disturbance is zero and
γx = 10 as shown in Fig. 8.3.
Now, suppose the plan is minimum phase with

ẋ = ax + u + 2z + w

ż = −z + u

y=x
192 8 Robustness Issues with Adaptive Control

4 0

3
−20
2

u
x
−40
1

0 −60
0 5 10 0 5 10
t t

0.8 0

0.6 −5

kx
w

0.4 −10

0.2 −15

0 −20
0 5 10 0 5 10
t t

Fig. 8.2 Response of non-minimum phase system with disturbance, γx = 1

6 0

4 −500
u
x

2 −1000

0 −1500
0 5 10 0 5 10
t t

1 0

0.5 −100
x
w

0 −200
k

−0.5 −300

−1 −400
0 5 10 0 5 10
t t

Fig. 8.3 Response of non-minimum phase system with zero disturbance, γx = 10

The closed-loop system now is stable with the same adaptive controller as shown
in Fig. 8.4.

1 0.5

0
0.5
−0.5
u
x

0
−1

−0.5 −1.5
0 5 10 0 5 10
t t

1 0

0.5
−1
kx
w

0
−2
−0.5

−1 −3
0 5 10 0 5 10
t t

Fig. 8.4 Response of minimum phase system with zero disturbance, γx = 10

8.3 Time-Delay Systems 193

8.3 Time-Delay Systems

Many real systems have latency which results in a time delay at the control input.
Time delay is caused by a variety of sources such as communication bus latency,
computational latency, transport delay, etc. Time-delay systems are a special class
of non-minimum phase systems which exhibit some similar behaviors with respect
to adaptive control. Model-reference adaptive control of time-delay systems is chal-
lenging as no well-established theory is available. Instability can result when an
adaptive parameter exceeds a certain limit.
Consider a time-delay MIMO system with a matched uncertainty

ẋ = Ax + B u (t − td ) + Θ ∗ Φ (x) (8.17)

where td is a known time delay.

If A and B are known, then an adaptive controller can be designed for the delay-
free system as
u = K x x + K r r − Θ  Φ (x) (8.18)

Θ̇ = −Γ Φ (x) e P B (8.19)

When a time delay is present, there exists an upper limit on the value of the adap-
tation rate matrix Γ that ensures stable adaptation. Thus, the adaptation rate can no
longer be chosen arbitrary but rather by a careful consideration of the effect of the
time delay present in the system. Unfortunately, no well-established theory currently
exists that would enable a systematic design of adaptive control for time-delay sys-
tems. As a result, the design is relied upon extensive Monte-Carlo simulations to
ensure that a proper adaptation rate is selected.
Consider a time-delay SISO system

ẋ = ax + bu (t − td ) (8.20)

where x (t) ∈ R, u (t) ∈ R, and b > 0.

The open-loop transfer function

x (s) be−td s
= (8.21)
u (s) s−a

is non-minimum phase since there exists a multiplicity of zeros in the right half plane
due to the term e−td s .
The system has a feedback control

u = kx x (8.22)

with a closed-loop pole s = a + bk x < 0.

194 8 Robustness Issues with Adaptive Control

The feedback gain k x is designed to be robust in the presence of the time delay to
ensure stability of the closed-loop system.
The closed-loop transfer function

kr
G (s) = (8.23)
s − a − bk x e−td s

is not absolutely stable but conditionally stable for k xmin < k x < − ab .
The system pole can be computed by letting s = σ + jω. Then,

σ + jω − a − bk x e−td σ e− jωtd = 0 (8.24)

The zero cross-over value of k x is established by setting σ = 0. Using the Euler’s

formula
e− jωtd = cos ωtd − j sin ωtd , (8.25)

the real part and imaginary parts are equated to zero so that the following equations
are obtained:
− a − bk x cos ωtd = 0 (8.26)

ω + bk x sin ωtd = 0 (8.27)

The solutions of these equations yield the jω-axis cross-over frequency from
which k xmin can be determined as

ω= b2 k x2min − a 2 (8.28)

 
a + bk xmin cos b2 k x2min − a 2 td = 0 (8.29)

Thus, the closed-loop system is stable for all values of k xmin < k x < − ab .
Suppose the constant a is unknown and the adaptive controller in Eqs. (8.14) and
(8.15) is used to regulate x (t) to zero. Without the time delay, the adaptation rate can
be arbitrarily chosen. However, in the presence of the time delay, there exists an upper
limit on γx that ensures stability of the closed-loop system. Thus, the closed-loop
adaptive system is stable if γx ≤ γxmax such that
 t
k x = k x (0) − γxmax x 2 bdτ > k xmin (8.30)
0

Example 8.4 Suppose a = −1, b = 1, and td = 0.5 s. The minimum feedback

gain is computed to be k xmin = −3.8069. The response of the closed-loop adaptive
system is shown in Fig. 8.5 for a time step Δt = 0.001. For γx = 7 and k x (0) = 0,
k x (t) → −3.5081 > k xmin as t → ∞. Thus, the closed-loop system is stable.
For γx = 7.6, k x (t) becomes unbounded and the closed-loop system is unstable.
Therefore, the maximum adaptation rate γxmax is some value between 7 and 7.6.
8.4 Unmodeled Dynamics 195

1 2

0
0.5
γ =7 γ =7.6
x −2 x

x
0
−4

−0.5 −6
0 5 10 0 5 10
t t

0 0

−1 −20

γ =7 γx=7.6
kx

kx
−2 x −40

−3 −60

−4 −80
0 5 10 0 5 10
t t

Fig. 8.5 Response of time-delay system

8.4 Unmodeled Dynamics

Non-minimum phase systems are a special class of systems with unmodeled dynam-
ics. Unmodeled dynamics are dynamics that actually exist in a plant but are not
accounted for in a control design. This is a very common situation since most con-
trol designs are based on mathematical models of some real plants that most likely
do not fully capture all the complex physics. In some cases, the neglected dynamics
do not present problems in a design, but in other cases, controllers that are designed
based on slightly inaccurate mathematical models can lead to catastrophic conse-
quences. Therefore, it is a good practice to include stability margins in any control
design to account for effects of unmodeled dynamics. Traditional gain and phase
margins are well-established design criteria for linear control design for SISO sys-
tems. For adaptive control, there is no such equivalent stability margin nor is there
any well-established theory for analysis of stability margins of adaptive systems.
Consider the system
ẋ = Ax + Bu + Δ (x, z, u) (8.31)

ż = f (x, z, u) (8.32)

y=x (8.33)

where z (t) is an internal state vector which may be unmeasurable or unobservable,

Δ (x, z, u) is the plant model error that is unknown and not accounted for, ż (t) is
the unmodeled dynamics, and y (t) is the plant output vector which is equal to the
measurable state vector x (t).
If model-reference adaptive control is used in a control design by assuming
Δ (x, z, u) = 0, then it is clear that such a control design is non-robust as demon-
strated by instability phenomena of MRAC for non-minimum phase systems.
196 8 Robustness Issues with Adaptive Control

Research in robust adaptive control was motivated by instability phenomena of

adaptive control. In fact, instability of adaptive control in the early 1960s which con-
tributed to the crash of one of the NASA X-15 hypersonic vehicles caused a great deal
of concern about the viability of adaptive control. Rohrs et al. investigated various
instability mechanisms of adaptive control in the 1980s [2]. The Rohrs counterex-
ample exposes the weakness of MRAC in its inability to provide robustness in the
presence of unmodeled dynamics.
To illustrate the lack of robustness to unmodeled dynamics associated with
MRAC, consider the following first-order SISO system with a second-order SISO
actuator plant:
ẋ = ax + bu (8.34)

ü + 2ζ ωn u̇ + ωn2 u = ωn2 u c (8.35)

where a < 0, b > 0, ζ > 0, ω > 0, and u c (t) is the actuator command.
Suppose we want to design an adaptive controller to track a first-order reference
model
ẋm = am xm + bm r (8.36)

where am < 0.
Without knowing or ignoring the actuator dynamics, the adaptive controller is
designed to be of the form

u c = k y (t) y + kr (t) r (8.37)

where k x (t) and kr (t) are adjusted by the following MRAC adaptive laws:

k̇ y = γx ye (8.38)

k̇r = γr r e (8.39)

where e (t) = ym (t) − y (t).

Let us now consider a non-adaptive control design with k y and kr constant. The
open-loop transfer function of the plant with actuator dynamics is expressed as

y (s) bωn2
=  (8.40)
u c (s) (s − a) s 2 + 2ζ ωn s + ωn2

We introduce a time delay at the input u c (t). Then, the closed-loop transfer
function is obtained as

bωn2 kr e−ωtd
G (s) =  (8.41)
(s − a) s 2 + 2ζ ωn s + ωn2 − bωn2 k y e−ωtd
8.4 Unmodeled Dynamics 197

Let s = jω. Then, the characteristic equation of the closed-loop transfer function
is computed as

− jω3 −(2ζ ωn − a) ω2 + ωn2 − 2aζ ωn jω−aωn2 −bωn2 k y (cos ωtd − j sin ωtd ) = 0
(8.42)
Separating the real part and imaginary part leads to the following equations:

− (2ζ ωn − a) ω2 − aωn2 − bωn2 k y cos ωtd = 0 (8.43)


− ω3 + ωn2 − 2aζ ωn ω + bωn2 k y sin ωtd = 0 (8.44)

The phase margin of the closed-loop plant is then obtained as

  
−1 ω3 − ωn2 − 2aζ ωn ω
φ = ωtd = sin (8.45)
bωn2 k y

or alternatively as
  
−1 ω3 − ωn2 − 2aζ ωn ω
φ = ωtd = tan (8.46)
− (2ζ ωn − a) ω2 − aωn2

The system has zero phase margin if


ω  ω0 = ωn2 − 2aζ ωn (8.47)

The feedback gain k y affects the cross-over frequency of the closed-loop system
according to the following equation:
  
ω6 + a 2 + 4ζ 2 ωn2 − 2ωn2 ω4 + ωn2 + 4a 2 ζ 2 − 2a 2 ωn2 ω2 + a 2 − b2 k 2y ωn4 = 0
(8.48)
We now return to the adaptive control design. Suppose the reference command
signal is specified as
r = r0 sin ω0 t (8.49)

Since the reference command signal provides frequency input that matches the
cross-over frequency at zero phase margin, the closed-loop adaptive system will be
unstable.

Example 8.5 Let a = −1, b = 1, am = −2, bm = 2, ωn = 5 rad/s, and ζ =√ 0.5.

Then the cross-over frequency at zero phase margin is computed to be ω0 = 30
rad/s. The reference command signal is then specified as
√
r = sin 30t
198 8 Robustness Issues with Adaptive Control

1 5

0.5
0
0

u
x
−5
−0.5

−1 −10
0 5 10 0 5 10
t t

−0.5 4

−1
3
ky

r
−1.5

k
2
−2

−2.5 1
0 5 10 0 5 10
t t

Fig. 8.6 Instability of MRAC due to √

unmodeled dynamics with reference command at zero phase
margin cross-over frequency r = sin 30t, k y (0) = −1

We choose γx = 1, γr = 1, and the initial conditions k y (0) = −1 and

kr (0) = 1. The closed-loop adaptive system is indeed unstable as shown in Fig. 8.6.
This illustrates the lack of robustness of MRAC to unmodeled dynamics.
Suppose we change the frequency of the reference command signal to 3 rad/s.
The adaptive system is now stable with k y (t) → −3.5763 and kr (t) → 1.2982 as
shown in Fig. 8.7.
The closed-loop stability of an adaptive system also depends on the initialization
of the adaptive feedback gain k y (t). To illustrate this, we compute the feedback
√ gain
corresponding to the cross-over frequency at zero phase margin ω0 = 30. This
results in k y0 = − 315
= −6.2. For the reference command signal r (t) = sin 3t
which results in a stable closed-loop adaptive system as shown in Fig. 8.7, suppose
we initialize the adaptive law with k y (0) = −6.2. The closed-loop system is now
on the verge of instability as shown in Fig. 8.8.

1 4

0.5 2

0 0
u
x

−0.5 −2

−1 −4
0 20 40 0 20 40
t t

0 3

−1 2.5
ky

−2 2
k

−3 1.5

−4 1
0 20 40 0 20 40
t

Fig. 8.7 Stable response with reference command r = sin 3t, k y (0) = −1
8.4 Unmodeled Dynamics 199

0.5 2

0 0

u
x
−1

−0.5 −2
0 20 40 0 20 40
t t

−5 2

−5.5
ky

r
1.5

k
−6

−6.5 1
0 20 40 0 20 40
t t

Fig. 8.8 Incipient instability of MRAC due to unmodeled dynamics with reference command
r = sin 3t, k y (0) = −6.2 at zero phase margin

Example 8.6 The Rorhs counterexample

√is defined with the following parameters:
a = −1, b = 2, am = −3, bm = 3, ω = 229, and ζ = 2√30229 [2].

5 100

50

0 0
u
x

−50

−5 −100
0 10 20 0 10 20
t t

0 20

15
−5
y

kr

10
k

−10
5

−15 0
0 10 20 0 10 20
t t

Fig. 8.9 Instability of MRAC by Rohrs counterexample

The adaptive laws are initialized with k y (0) = −0.65 and kr (0) = 1.14. The
adaptation rates are chosen to be γ√ x = 1 and γr = 1. The zero phase-margin
frequency is computed to be ω0 = 259 = 16.1 rad/s. The reference command
is given by
r = 0.3 + 1.85 sin 16.1t

As a result, instability occurs as shown in Fig. 8.9.


Typically, actuator dynamics must be sufficiently faster than the plant dynamics
in order to maintain enough stability margins for a controller. This usually means
that the actuator frequency bandwidth ωn must be greater than the control frequency
200 8 Robustness Issues with Adaptive Control

bandwidth ω, that is, ωn  ω. Then, the closed-loop transfer function in Eq. (8.41)
may be approximated by

bkr e−ωtd
G (s) = (8.50)
s − a − bk y e−ωtd

This results in the following cross-over frequency and phase margin:


ω= b2 k 2y − a 2 (8.51)
 
ω
φ = ωtd = sin−1 − (8.52)
bk y
 
for bk y  > |a|. As k y → −∞, φ → π
2
. The closed-loop system is therefore robustly
stable in the presence of high gain.

8.5 Fast Adaptation

Model-reference adaptive control is designed to achieve asymptotic tracking for sys-

tems with matched structured uncertainty. In practice, asymptotic tracking is a highly
demanding requirement that can be difficult to meet. For systems with unknown dis-
turbances and unstructured uncertainty which are more frequently encountered in
practice, only bounded tracking could be achieved with MRAC, but there is no guar-
antee of boundedness of adaptive parameters. As the adaptation rate increases, the
tracking error reduces and the tracking performance is improved. This is usually
referred to as fast adaptation. However, as seen previously, for systems with non-
minimum phase behaviors, time delay, and unmodeled dynamics, increasing adap-
tation rate beyond a certain limit can result in instability. Robustness of MRAC, in
general, requires a small adaptation rate. Therefore, tracking performance and robust-
ness are often viewed as two directly competing requirements in adaptive control.
This viewpoint is also valid for any linear control design.
Consider a first-order SISO system

ẋ = ax + bu (8.53)

A linear proportional-integral (PI) controller is designed to enable the plant output

x (t) to track a reference signal xm (t) as
 t
u = k x x + ki (x − xm ) dτ (8.54)
0

The closed-loop plant is expressed as

8.5 Fast Adaptation 201

 t
ẋ = a + bk p x + bki (x − xm ) dτ (8.55)
0

Differentiating yields

ẍ − a + bk p ẋ − bki x = −bki xm (8.56)

The closed-loop plant is stable with a + bk p < 0 and ki < 0. As ki → ∞,

x (t) → xm (t). Thus, increasing the integral gain improves the tracking performance.
The frequency of the closed-loop system is ωn2 = −bki . Thus, as ki → ∞ and
ωn → ∞, the closed-loop response becomes highly oscillatory. It is of interest to
find out what happens to the stability margins of the closed-loop system as ki → ∞.
Suppose the input is delayed by an amount td so that the closed-loop plant is given
by
bki xm e−td s
x =− 2 (8.57)
s − as − bk p se−td s − bki e−td s

The stability of the closed-loop plant can be analyzed by setting s = jω which

yields 
− ω2 − a jω − bk p jω + bki (cos ωtd − j sin ωtd ) = 0 (8.58)

Equating the real and imaginary parts to zero results in the following equations:

− ω2 − bk p ω sin ωtd − bki cos ωtd = 0 (8.59)

− aω − bk p ω cos ωtd + bki sin ωtd = 0 (8.60)

These equations can be recast as

b2 k 2p ω2 sin2 ωtd + b2 k p ki ω sin 2ωtd + b2 ki2 cos2 ωtd = ω4 (8.61)

b2 k 2p ω2 cos2 ωtd − b2 k p ki ω sin 2ωtd + b2 ki2 sin2 ωtd = a 2 ω2 (8.62)

Adding the two equations together yields


ω4 + a 2 − b2 k 2p ω2 − b2 ki2 = 0 (8.63)

The cross-over frequency is then computed as

 ⎡  ⎤
 
 b2 k 2 − a 2 
 ⎣1 + 1 + 
2
4b ki 2
2 ⎦
p
ω= (8.64)
2 b k 2p − a 2
2

Multiplying Eq. (8.59) by bki and Eq. (8.60) by bk p ω and then adding the resulting
equations together yield
202 8 Robustness Issues with Adaptive Control
 
abk p + bki ω2 + b2 k 2p ω2 + b2 ki2 cos ωtd = 0 (8.65)

The closed-loop plant is then stable for all t ≤ td where td is called a time-delay
margin which is computed as
  
1 −1 abk p + bki ω2
td = cos − 2 2 2 (8.66)
ω b k p ω + b2 ki2

Now, consider a limiting case when ωn2 = −bki → ∞. Then,

ω → ωn → ∞ (8.67)

1 ω4
td → cos−1 n4 → 0 (8.68)
ω ωn

The phase margin of the closed-loop plant is computed to be

φ = ωtd → 0 (8.69)

Thus, as the integral gain increases, the closed-loop response becomes more oscil-
latory and the time-delay margin and phase margin of the closed-loop plant decrease.
In the limit, these margins tend to zero as ki → ∞. On the other hand, the tracking
performance is improved with increasing the integral gain. Clearly, the two require-
ments cannot be simultaneously met. Thus, a control design typically strives for a
balance between these two competing requirements by a judicious choice of ki .
Suppose the plant becomes uncertain with a being an unknown parameter. A
reference model is specified by

ẋm = am xm + bm r (8.70)

where r (t) is a constant signal.

Then, an adaptive controller is designed as

u = k x (t) x + kr r (8.71)

k̇ x = γx x (xm − x) b (8.72)

where kr = bbm .
Differentiating the closed-loop plant yields

ẍ − [a + bk x (t)] ẋ + b2 γx x 3 = b2 γx x 2 xm (8.73)

Contrasting Eq. (8.73) with Eq. (8.56), one can conclude that the effect of adaptive
control is very similar to that of the linear integral control. In fact, let
8.5 Fast Adaptation 203

ki (x) = −γx x 2 b (8.74)

Then, the closed-loop plant can be expressed as

ẍ − [a + bk x (t)] ẋ − bki (x) x = −bki (x) xm (8.75)

Differentiating the adaptive controller yields

u̇ = ki (x) (x − xm ) + k x (t) ẋ + kr ṙ (8.76)

Then, integrating yields

 t  t
u= k x (t) ẋdτ + ki (x) (x − xm ) dτ + kr r (8.77)
0 0

Therefore, adaptive control in some way can be viewed as a nonlinear inte-

gral control. This leads to an observation that as γx increases, the closed-loop
response becomes more oscillatory and the time-delay margin of the closed-loop
plant decreases. In the limit, as γx → ∞, the time-delay margin tends to zero.
Robustness of MRAC therefore decreases as the adaptation rate increases.

Example 8.7 Let a = 1, b = 1, and xm (t) = 1. To track a reference command,

a linear PI controller is designed with k p = −2 and ki = −500. The cross-over
frequency and time-delay margin are computed to be ω = 22.3942 rad/s and td =
0.0020 s. Thus, the closed-loop system is stable for any time delay up to 0.0020 s.
The cross-over frequency ω and time-delay margin td as a function of the integral
gain ki are shown in Fig. 8.10. As can be seen, the cross-over frequency increases and
the time-delay margin decreases as the magnitude of the integral gain ki increases.
Therefore, high frequency integral gain should be avoided in any design.

30

20
ω

10

0 0 1 2 3
10 10 10 10
−ki

0.8

0.6
td

0.4

0.2

0 0 1 2 3
10 10 10 10
−k
i

Fig. 8.10 Cross-over frequency and time-delay margin as function of integral gain
204 8 Robustness Issues with Adaptive Control

Suppose an adaptive controller is used instead. Let am = −1, bm = 1, and r (t) =

1. Choose a large value of γx , say, γx = 500 to achieve fast adaptation. With the
initial conditions x (0) = 0 and k x (0) = 0, the responses of the closed-loop adaptive
system and linear system are shown in Fig. 8.11. Note that the responses exhibit high
frequency oscillations as indicated by the analysis. As x (t) → xm (t) = 1, then
ki (x) →≈ −γx xm2 (t) b ≈ −500. Thus, the adaptation rate γx = 500 is effectively
equivalent to the integral gain ki = −500. Therefore, the closed-loop adaptive system
may tend to exhibit similar behaviors as the closed-loop linear system. In fact, if a
time delay td = 0.0020 s is introduced at the input, both systems become unstable
and have similar responses as shown in Fig. 8.12.

1.5

1
x

0.5 Adaptive Controller

0
0 2 4 6 8 10
t

1.5

1
x

0.5 Linear PI Controller

0
0 2 4 6 8 10
t

Fig. 8.11 Responses of adaptive systems and linear systems with fast adaptation

1.5

1
x

0.5
Adaptive Controller
0
0 2 4 6 8 10
t

1.5

1
x

0.5
Linear PI Controller
0
0 2 4 6 8 10
t

Fig. 8.12 Responses of adaptive systems and linear systems with fast adaptation and time delay
8.6 Summary 205

8.6 Summary

Model-reference adaptive control is designed to achieve asymptotic tracking. This

ideal property of MRAC is a benefit but also a problem. MRAC is generally non-
robust since the bound on an adaptive parameter cannot be established by the
Lyapunov stability analysis. A parameter drift can result that can cause an adap-
tive control algorithm to blow up. As the complexity of a plant increases, robustness
of MRAC becomes increasingly difficult to ascertain. In practical applications, the
knowledge of a real plant can never be established precisely. Thus, a model of a
physical plant often cannot fully capture all the effects due to unmodeled dynamics,
unstructured uncertainty, and exogenous disturbances that may exist in a real plant.
All these effects can produce closed-loop dynamics that can lead to instability when
MRAC is used in an adaptive controller.
Examples of the lack of robustness of MRAC to the parameter drift, time delay,
non-minimum phase plants, unmodeled dynamics, and fast adaptation are illustrated.
For a sufficiently large disturbance, MRAC attempts to seek the ideal asymptotic
tracking property. In theory, to reject the disturbance, MRAC generates a high-gain
control in order to reduce the effect of the disturbance. This high-gain control can
cause adaptive parameters to become unbounded. A parameter drift can lead to
instability in practical situations.
Non-minimum phase systems are more difficult to control since the existence of
the unstable zeros imposes limiting values on stable feedback gains. MRAC always
seeks the ideal asymptotic tracking property by attempting to perform an unstable
pole-zero cancellation, which leads to instability in non-minimum phase systems.
For non-minimum phase systems, adaptive control designers generally have to be
aware of the limiting values of adaptive parameters in order to prevent potential
instability.
Many real systems have latency which results in a time delay at the control input.
Time delay is caused by a variety of sources such as communication bus latency,
computational latency, transport delay, etc. Time-delay systems are a special class
of non-minimum phase systems. Instability can result when an adaptive parameter
exceeds a certain limit. Without the time delay, the adaptation rate in MRAC can be
arbitrarily chosen. However, in the presence of the time delay, there exists an upper
limit on the adaptation rate that can ensure the closed-loop stability of an adaptive
system. Thus, the closed-loop adaptive system is stable if the adaptation rate is kept
below this limit.
Unmodeled dynamics are dynamics that actually exist in a plant but are not
accounted for in a control design. This is a very common situation since most control
designs are based on mathematical models of some real plants that most likely do not
fully capture all the complex physics. In some cases, the neglected dynamics do not
present problems in a design, but in other cases, controllers that are designed based
on slightly inaccurate mathematical models can lead to catastrophic consequences.
MRAC is known to be non-robust to unmodeled dynamics. When a reference com-
mand signal contains a frequency close to a zero phase margin due to unmodeled
206 8 Robustness Issues with Adaptive Control

dynamics, instability can occur. Initialization of adaptive parameters can also affect
the closed-loop stability if the initial condition coincides with a near zero phase
margin.
Fast adaptation refers to the notion that a large adaptation rate can be used to
achieve fast reference model following. In the limit, as the adaptation rate increases
to a theoretically infinite value, robustness of MRAC as measured by the time-delay
margin tends to zero. As with the case with non-minimum phase, time delay, and
unmodeled dynamics, MRAC can cause instability when a large adaptation rate is
used for fast adaptation.

8.7 Exercises

1. Consider a first-order SISO system

ẋ = ax + bu + w

where w (t) is a bounded disturbance and u (t) is an adaptive controller defined

as
u = kx x

k̇ x = −γx x 2 b

Suppose the solution of x (t) is given by

x = t (1 + t) p

a. Analyze parameter drift behaviors of the closed-loop system by finding all

values of p that result in unbounded feedback gain k x (t) and all values of
p that result in a completely bounded system.
b. Implement the adaptive controller in Simulink using the following informa-
tion: a = 1, b = 1, γx = 1, x (0) = 0, and k x (0) = 0 with a time step
Δt = 0.001 s for two different values of p: one for unbounded k x (t) and
the other for all bounded signals. Plot the time histories of x (t), u (t), w (t),
and k x (t) for each of the values of p for t ∈ [0, 20] s.
2. Consider a time delay second-order SISO system

ÿ − ẏ + y = u (t − td )

where td is an unknown time delay.

The unstable open-loop plant is stabilized with a linear derivative controller

u = kd∗ ẏ
8.7 Exercises 207

where kd∗ = −7.

a. Calculate analytically the cross-over frequency ω and the time-delay margin
td that corresponds to neutral stability of the closed-loop system.
b. Now, suppose an adaptive controller is designed to follow the delay-free
closed-loop system with the linear derivative controller as the reference
model
ÿm + 6 ẏm + ym = 0


Let x (t) = y (t) ẏ (t) ∈ R2 , then the open-loop plant is designed with
an adaptive derivative controller

u = Kx x

K̇ x = −Γx x x  P B

where K x (t) = 0 kd (t) and Γx = diag(0, γx ) and γx is an adaptation

rate.
Implement the adaptive controller in Simulink using the following infor-
mation: Q = I , y (0) = 1, ẏ (0) = 0, and K x (0) = 0 with a time step
Δt = 0.001 s. Determine γxmax that causes the closed-loop system to be on
the verge of instability by trial-and-error to within 0.1 accuracy. Calculate
kdmin that corresponds to γxmax . Plot the time histories of x (t), u (t), and kd (t)
for t ∈ [0, 10] s.
3. For the Rohrs’ counterexample, stability of the closed-loop system is affected
by the frequency of the reference command signal r (t). Write the closed-loop
transfer function from r (t) to y (t). Then, compute the cross-over frequency ω
for the reference command signal

r = 0.3 + 1.85 sin ωt

to give a 60◦ phase margin. Also compute the ideal feedback gain k ∗y corresponding
to this phase margin. Implement in Simulink the Rohrs’ counterexample using
the same initial conditions k y (0) and kr (0) with γ y = γr = 1 and Δt = 0.001 s.
Plot the time histories of y (t), u (t), k y (t), and kr (t) for t ∈ [0, 60] s.

References

1. Ioannu, P. A., & Sun, J. (1996). Robust adaptive control. Upper Saddle River: Prentice-Hall,
Inc.
2. Rohrs, C. E., Valavani, L., Athans, M., & Stein, G. (1985). Robustness of continuous-time
adaptive control algorithms in the presence of unmodeled dynamics. IEEE Transactions on
Automatic Control, AC-30(9) 881–889.
Chapter 9
Robust Adaptive Control

Abstract This chapter presents several techniques for improved robustness of

model-reference adaptive control. These techniques, called robust modification,
achieve increased robustness through two general principles: (1) limiting adaptive
parameters and (2) adding damping mechanisms in the adaptive laws to bound adap-
tive parameters. The dead-zone method and the projection method are two common
robust modification schemes based on the principle of limiting adaptive parameters.
The dead-zone method prevents the adaptation when the tracking error norm falls
below a certain threshold. This method prevents adaptive systems from adapting to
noise which can lead to a parameter drift. The projection method is widely used
in practical adaptive control applications. The method requires the knowledge of a
priority bounds on system parameters. Once the bounds are given, a convex set is
established. The projection method then permits the normal adaptation mechanism
of model-reference adaptive control as long as the adaptive parameters remain in-
side the convex set. If the adaptive parameters reach the boundary of the convex
set, the projection method changes the adaptation mechanism to bring the adaptive
parameters back into the set. The σ modification and e modification are two well-
known robust modification techniques based on the principle of adding damping
mechanisms to bound adaptive parameters. These two methods are discussed, and
the Lyapunov stability proofs are provided. The optimal control modification and
the adaptive loop recovery modification are two recent robust modification meth-
ods that also add damping mechanisms to model-reference adaptive control. The
optimal control modification is developed from the optimal control theory. The prin-
ciple of the optimal control modification is to explicitly seek a bounded tracking
as opposed to the asymptotic tracking with model-reference adaptive control. The
bounded tracking is formulated as a minimization of the tracking error norm bounded
from an unknown lower bound. A trade-off between bounded tracking and stabil-
ity robustness can therefore be achieved. The damping term in the optimal control
modification is related to the persistent excitation condition. The optimal control
modification exhibits a linear asymptotic property under fast adaptation. For linear
uncertain systems, the optimal control modification causes the closed-loop systems
to tend to linear systems in the limit. This property can be leveraged for the design
and analysis of adaptive control systems using many existing well-known linear con-
trol techniques. The adaptive loop recovery modification is designed to minimize the

© Springer International Publishing AG 2018 209

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_9
210 9 Robust Adaptive Control

nonlinearity in a closed-loop plant so that the stability margin of a linear reference

model could be preserved. This results in a damping term proportional to the square
of the derivative of the input function. As the damping term increases, in theory
the nonlinearity of a closed-loop system decreases so that the closed-loop plant can
follow a linear reference model which possesses all the required stability margin
properties. The L1 adaptive control has gained a lot of attention in the recent years
due to its ability to achieve robustness with fast adaptation for a given a priori bound
on the uncertainty. The underlying principle of the L1 adaptive control is the use
of fast adaptation for improved transient or tracking performance coupled with a
low-pass filter to suppress high-frequency responses for improved robustness. As a
result, the L1 adaptive control can be designed to achieve stability margins under fast
adaptation for a given a priori bound on the uncertainty. The basic working concept of
the L1 adaptive control is presented. The bi-objective optimal control modification
is an extension of the optimal control modification designed to achieved improved
performance and robustness of systems with input uncertainty. The adaptation mech-
anism relies on two sources of errors: the normal tracking error and the predictor
error. A predictor model of a plant is constructed to estimate the open-loop response
of the plant. The predictor error is formed as the difference between the plant and the
predictor model. This error signal is then added to the optimal control modification
adaptive law to enable the input uncertainty to be estimated. Model-reference adap-
tive control of singularly perturbed systems is presented to address slow actuator
dynamics. The singular perturbation method is used to decouple the slow and fast
dynamics of a plant and its actuator. The asymptotic outer solution of the singularly
perturbed system is then used in the design of model-reference adaptive control. This
modification effectively modifies an adaptive control signal to account for slow ac-
tuator dynamics by scaling the adaptive law to achieve tracking. Adaptive control of
linear uncertain systems using the linear asymptotic property of the optimal control
modification method is presented for non-strictly positive real (SPR) systems and
non-minimum phase systems. The non-SPR plant is modeled as a first-order SISO
system with a second-order unmodeled actuator dynamics. The plant has relative de-
gree 3 while the first-order reference model is SPR with relative degree 1. By invoking
the linear asymptotic property, the optimal control modification can be designed to
guarantee a specified phase margin of the asymptotic linear closed-loop system.
For non-minimum phase systems, the standard model-reference adaptive control is
known to be unstable due to the unstable pole-zero cancellation as a result of the
ideal property of asymptotic tracking. The optimal control modification is applied
as an output feedback adaptive control design that prevents the unstable pole-zero
cancellation by achieving bounded tracking. The resulting output feedback adaptive
control design, while preventing instability, can produce poor tracking performance.
A Luenberger observer state feedback adaptive control method is developed to im-
prove the tracking performance. The standard model-reference adaptive control still
suffers the same issue with the lack of robustness if the non-minimum phase plant
is required to track a minimum phase reference model. On the other hand, the opti-
9 Robust Adaptive Control 211

mal control modification can produce good tracking performance for both minimum
phase and non-minimum phase reference models.

Research in robust adaptive control was motivated by instability phenomena of adap-

tive control. In fact, instability of adaptive control in the early 1960s which con-
tributed to the crash of one of the NASA X-15 hypersonic vehicles caused a great
deal of concern about the viability of adaptive control. Rohrs et al. investigated in-
stability mechanisms of adaptive control due to unmodeled dynamics in the 1980s
[1]. As a result, various robust modification schemes had since been developed to
ensure boundedness of adaptive parameters. The σ modification and the e modifica-
tion are two well-known robust modification methods [2, 4]. Other techniques such
as the dead-zone method and the projection method are also used to improve robust-
ness of adaptive control algorithms [5]. In recent years, there have been numerous
new advanced robust adaptive control methods developed, such as the adaptive loop
recovery [6], composite model-reference adaptive control [7], L1 adaptive control
[8], optimal control modification [9], derivative-free adaptive control [10], and many
others. The reader is referred to Chap. 1 for a more extensive list of these recent
adaptive control methods. The principle of robust modification is based on two cen-
tral themes: (1) limiting adaptive parameters and (2) adding damping mechanisms
to model-reference adaptive control. The robustness issues with the parameter drift,
non-minimum phase behaviors, time delay, unmodeled dynamics, and fast adaptation
are largely ameliorated with these robust modification schemes, but are not entirely
eliminated if the nature of the uncertainty is not completely known.
In this chapter, a variety of robust adaptive control methods will be covered. The
learning objectives of this chapter are:
• To develop an understanding and to know how to apply common techniques for
enhancing robustness in model-reference adaptive control, namely the dead-zone
method, the projection method, σ modification, and e modification;
• To be able to understand and apply new modern robust adaptive control tech-
niques, namely the optimal control modification, the bi-objective optimal control
modification, the adaptive loop recovery, and the L1 adaptive control;
• To develop an understanding of the linear asymptotic property of the optimal
control modification that can be used to estimate the time-delay margin of a closed-
loop adaptive system;
• To understand and to know how to apply time-varying adaptation rate techniques
of normalization and covariance adjustment to improve robustness;
• To recognize robustness issues with slow actuators and to be able to apply the
singular perturbation method for systems with slow actuator dynamics; and
• To be able to address systems with unmodeled dynamics and non-minimum phase
plants with relative degree 1 by applying the linear asymptotic property of the
optimal control modification.
212 9 Robust Adaptive Control

9.1 Dead-Zone Method

Parameter drift can occur when noise is present in a signal. If the signal-to-noise
ratio is sufficiently small, adaptive control would attempt to reduce noise instead of
the tracking error. As a result, adaptive parameters would continue to be integrated,
thereby leading to a parameter drift. Thus, the dead-zone method can be used to stop
the adaptation process when the tracking error is reduced below a certain level [5].
The dead-zone method for the adaptive law in Eq. (8.19) is stated as

−Γ Φ (x) e P B e > e0
Θ̇ = (9.1)
0 e ≤ e0

where e0 is an adaptation threshold to be selected.

Assuming that e0 is properly chosen, then the dead-zone method will improve
robustness of MRAC since the adaptive parameters will be “frozen” when the adap-
tation process ceases. The adaptive parameters can be shown to be bounded by the
Lyapunov stability analysis as follows:

Proof When e > e0 , the adaptation process is active. The usual result of the
Lyapunov stability analysis was given in Sect. 8.1 with a bounded tracking error.
When e ≤ e0 , the adaptation is shut off to prevent the potential parameter drift as
e (t) → 0. Then, the following Lyapunov candidate function is used:

V (e) = e Pe (9.2)

The tracking error equation is given by

ė = Am e + B Θ̃  Φ (x) − w (9.3)

where w (t) is the noise disturbance.

Then,
V̇ (e) = −e Qe + 2e PBΘ̃  Φ (x) − 2e Pw (9.4)

V̇ (e) ≤ 0 requires

2e P B Θ̃  Φ (x) ≤ e Qe + 2e Pw (9.5)

or
  λ
  max (Q) e + 2λmax (P) w0 λmax (Q) e0 + 2λmax (P) w0
Θ̃  ≤ ≤ (9.6)
2 P B Φ (x) 2 P B Φ0
9.1 Dead-Zone Method 213
 
 
Since e ≤ e0 , it implies that Φ (x) ≤ Φ0 is bounded. Thus, Θ̃  is also
bounded from above. The adaptive law with the dead-zone method therefore is robust
and can prevent a parameter drift.

The difficulty with the dead-zone method usually lies with the selection of e0 which
is generally by trial-and-error. If the threshold is set too large, then the adaptation
process may be prematurely terminated by the dead-zone method. On the other hand,
if it is set too low, then the integral action of the adaptive law can lead to a parameter
drift.
Example 9.1 For Example 8.1 with k x (t) unbounded for n = − 12 5
, using the dead-
zone method with |x| > 0.2, the value of k x (t) is limited to −54.2195. Since
w (t) → 0 as t → ∞, x (t) will also tend to zero with the dead-zone method.
However, it tends to zero at a slower rate than that with MRAC. The response of
the closed-loop system with the dead-zone method is compared to that with MRAC
as shown in Fig. 9.1. A time delay is injected at the input. The closed-loop system
can tolerate up to td = 0.0235 s with MRAC and td = 0.0285 s with the dead-zone
method. This indicates that the dead-zone method is more robust than MRAC.

1
Deadband w/ |e|>0.2
MRAC

0.5
x

0
0 20 40 60 80 100
t

0
Deadband w/ |e|>0.2
−20 MRAC
kx

−40

−60

−80
0 20 40 60 80 100
t

Fig. 9.1 MRAC with dead-zone method

9.2 Projection Method

The projection method is used to limit adaptive parameters based on a priori knowl-
edge of the uncertainty. This may be a contradiction since uncertainty is usually
not known a priori. However, an adaptive controller can be designed under a cer-
tain assumption of the uncertainty. As long as the uncertainty does not violate the
assumption, then the projection method is guaranteed to provide robustness.
214 9 Robust Adaptive Control

To motivate the concept, consider a convex optimization problem

min J (Θ) = f (Θ) (9.7)

where Θ (t) ∈ Rn and f (Θ) ∈ R is a convex function, i.e.,

f (αx + (1 − α) y) ≤ α f (x) + (1 − α) f (y) (9.8)

for some α ∈ [0, 1], subject to an equality constraint

g (Θ) = 0 (9.9)

where g (Θ) ∈ R is a convex function.

This is a constrained optimization problem. We can solve this problem by using
the Lagrange multiplier method. Introducing a Lagrange multiplier λ (t), also called
an adjoint variable, then the augmented cost function is expressed as

J (Θ) = f (Θ) + λg (Θ) (9.10)

The second term is the augmented term which does not change the cost function
since g (Θ) = 0.
The necessary condition of optimality is obtained as

∂J
∇ JΘ (Θ) = =0 (9.11)
∂Θ
This yields
∇ f Θ (Θ) + λ∇gΘ (Θ) = 0 (9.12)

where ∇ f Θ (Θ) ∈ Rn and ∇gΘ (Θ) ∈ Rn .

Pre-multiplying Eq. (9.12) by ∇  gΘ (Θ) yields

∇  gΘ (Θ) ∇ f Θ (Θ) + λ∇  gΘ (Θ) ∇gΘ (Θ) = 0 (9.13)

Using the pseudo-inverse, λ (t) is obtained as

∇  gΘ (Θ) ∇ f Θ (Θ)
λ=− (9.14)
∇  gΘ (Θ) ∇gΘ (Θ)

Thus, the adjoint variable λ (t) is negative of the projection of the vector ∇ JΘ (Θ)
onto the vector ∇gΘ (Θ).
Now, the gradient update law for Θ (t) is given by

Θ̇ = −Γ ∇ JΘ (Θ) = −Γ [∇ f Θ (Θ) + λ∇gΘ (Θ)] (9.15)

9.2 Projection Method 215

Substituting λ (t) into the gradient update law results in

 
∇gΘ (Θ) ∇  gΘ (Θ) ∇ f Θ (Θ)
Θ̇ = −Γ ∇ f Θ (Θ) − (9.16)
∇  gΘ (Θ) ∇gΘ (Θ)

The update law can also be expressed by a projection operator as

 
∇gΘ (Θ) ∇  gΘ (Θ)
Θ̇ = Pro (Θ, −Γ ∇ JΘ (Θ)) = −Γ I −  ∇ JΘ (Θ) (9.17)
∇ gΘ (Θ) ∇gΘ (Θ)

The projection thus restricts of the optimal solution to only permissible values of
Θ (t) that lie on the constraint set g (Θ) = 0.
For example, consider the following constraint:
   
g (Θ) = Θ − Θ ∗ Θ − Θ ∗ − R2 = 0 (9.18)

which is a circle with the center at Θ(t) = Θ ∗ and the radius R.

g (Θ) can be verified to be a convex function. Then,
 
∇gΘ (Θ) = 2 Θ − Θ ∗ (9.19)

The least-squares gradient adaptive law with this constraint is given by


(Θ − Θ ∗ ) (Θ − Θ ∗ )
Θ̇ = −Γ I− Φ (x) ε (9.20)
R2

A more common situation in adaptive control typically involves an inequality

constraint defined by a convex set


S = Θ (t) ∈ Rn : g (Θ) ≤ 0 (9.21)

The optimization solution is then restricted to all permissible values of Θ (t)

inside the compact set S . If the initial values of Θ (t) is inside S , then there are
two situations to consider:
1. If Θ (t) remains inside S where g (Θ) < 0, then the optimization is uncon-
strained and the standard gradient update law is used.
2. If Θ (t) is on the boundary of S where g (Θ) = 0, then two situations can occur. If
Θ (t) proceeds in a direction toward S , then the standard gradient update law for
the unconstrained optimization is used. The direction toward S is such that ∇gΘ
points toward S and generally is in the opposite direction to that of −∇ JΘ . This
implies the dot product between ∇gΘ and −∇ JΘ is negative, or −∇  JΘ ∇gΘ ≤ 0.
If Θ (t) moves away from S in a direction such that −∇  JΘ ∇gΘ > 0, then the
projection method is used to bring Θ (t) back onto a tangent plane of S . Θ (t)
would then proceed in the direction on the boundary of S .
216 9 Robust Adaptive Control

Fig. 9.2 Projection method

From Fig. 9.2, when Θ (t) moves away from S , it follows that
 
Θ − Θ ∗ ∇gΘ (Θ) > 0 (9.22)

The projection method is then stated as [5]

Θ̇ = Pro (Θ, −Γ ∇ JΘ (Θ))

⎧
⎪
⎪ −Γ ∇ JΘ (Θ) if g (Θ) < 0 or if g (Θ) = 0
⎪
⎪
⎪
⎨ and − ∇  JΘ ∇gΘ (Θ) ≤ 0
  
= ∇gΘ (Θ) ∇ gΘ (Θ)
⎪
⎪ −Γ I −  ∇ JΘ (Θ) if g (Θ) ≥ 0
⎪
⎪ ∇ gΘ (Θ) ∇gΘ (Θ)
⎪
⎩
and − ∇  JΘ (Θ) ∇gΘ (Θ) > 0
(9.23)

Using the projection method for MRAC, the adaptive law is expressed as
 
Θ̇ = Pro Θ, −Γ Φ (x) e P B
⎧
⎪
⎪ −Γ Φ (x) e PB if g (Θ) < 0 or if g (Θ) = 0
⎪
⎪  
⎪
⎨ and − Φ (x) e P B ∇gΘ (Θ) ≤ 0
 
= ∇gΘ (Θ) ∇  gΘ (Θ)
⎪
⎪ −Γ I −  Φ (x) e PB if g (Θ) ≥ 0
⎪
⎪ ∇ gΘ (Θ) ∇gΘ (Θ)
⎪
⎩  
and − Φ (x) e P B ∇gΘ (Θ) > 0
(9.24)

The projection method can be shown to achieve uniform ultimate boundedness

by the Lyapunov stability analysis as follows:
Proof When g (Θ) < 0 or g (Θ) = 0 and Θ (t) moves toward S , then the usual
result of the Lyapunov analysis in Sect. 8.1 is obtained. When g (Θ) ≥ 0 and Θ (t)
moves away from S , then the projection method is used.
Choose a Lyapunov candidate function
 
V e, Θ̃ = e Pe + Θ̃  Γ −1 Θ̃ (9.25)
9.2 Projection Method 217

Then,
  ∇gΘ (Θ) ∇  gΘ (Θ)
V̇ e, Θ̃ = −e Qe − 2e Pw + 2Θ̃   Φ (x) e P B (9.26)
∇ gΘ (Θ) ∇gΘ (Θ)
 
The direction of Θ (t) away from S is defined by − Φ (x) e P B ∇gΘ (Θ) >
0. Then, let ∇  gΘ (Θ) Φ (x) e P B = −c0 < 0 where c0 > 0. It follows from
Eq. (9.22) that

  2c0 Θ̃  ∇gΘ (Θ)

V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2λmax (P) e w0 − (9.27)
∇  gΘ (Θ) ∇gΘ (Θ)

Suppose g (Θ) is defined by Eq. (9.18). Then,

2Θ̃  ∇gΘ (Θ) = 4Θ̃  Θ̃ = 4R 2 = ∇  gΘ (Θ) ∇gΘ (Θ) (9.28)

So,
 
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2λmax (P) e w0 − 2c0
≤ −λmin (Q) e2 + 2λmax (P) e w0 (9.29)
 
Thus, V̇ e, Θ̃ ≤ 0 if

2λmax (P) w0
e ≥ =p (9.30)
λmin (Q)
   
   
Since g (Θ) ≥ 0, then Θ̃  is bounded from below by Θ̃  ≥ R. We want the
 
trajectory of Θ̃ (t) to return to the set S . Thus, V̇ e, Θ̃ ≥ 0 inside a compact set
     
   
S = e , Θ̃  : e ≤ p and Θ̃  ≤ R (9.31)
 
but V̇ e, Θ̃ ≤ 0 outside S to ensure that both e (t) and Θ̃ (t) are uniformly
ultimately bounded.
The ultimate bounds can be found from the following inequalities:

 
 2

λmin (P) e2 ≤ λmin (P) e2 + λmin Γ −1 Θ̃ 
   
≤ V e, Θ̃ ≤ λmax (P) p 2 + λmax Γ −1 R 2 (9.32)
218 9 Robust Adaptive Control

 
 2
   
 2
 
λmin Γ −1 Θ̃  ≤ λmin (P) e2 + λmin Γ −1 Θ̃  ≤ V e, Θ̃
 
≤ λmax (P) p 2 + λmax Γ −1 R 2 (9.33)

Thus,   
λmax (P) p 2 + λmax Γ −1 R 2
p ≤ e ≤ =ρ (9.34)
λmin (P)
  
  λmax (P) p 2 + λmax Γ −1 R 2
 
R ≤ Θ̃  ≤   =β (9.35)
λmin Γ −1

Therefore, the projection method improves robustness of MRAC by constraining

the adaptive parameters to remain within a convex set. The boundedness of the
adaptive parameters causes an increase in the ultimate bound of the tracking error.
Since g (Θ) is a scalar function that constrains the adaptive parameters, the challenge
is to obtain such a function. For example, the convex function g (Θ) in Eq. (9.18)
can also be expressed as


n
 2
g (Θ) = θi − θi∗ − R 2 ≤ 0 (9.36)
i=1

Then,
θimin = θi∗ − R ≤ θi ≤ θi∗ + R = θimax (9.37)

The problem is that the range of θi which is equal to θimax − θimin is determined
by a single parameter R when in practice each θi could have a different range. For
example, if a priori bounds on the adaptive parameters can be determined such that

θimin = θi∗ − ri ≤ θi ≤ θi∗ + ri = θimax (9.38)

Then, R is determined as
n
R = min (ri ) (9.39)
i=1

This would then result in an over-constrained convex set wherein one or more
adaptive parameters will be over-constrained. Thus, a convex scalar function may
not be sufficiently flexible for handling more than one adaptive parameter. In such
situations, a convex vector function could be used.
9.2 Projection Method 219

Suppose a convex vector function is defined as

⎡ 2 ⎤
θ1 − θ1∗ − r12
⎢ .. ⎥
g (Θ) = [gi (θi )] = ⎢
⎣ .
⎥
⎦ (9.40)
 2
θm − θm∗ − rm2

where m ≤ n.
The augmented cost function is expressed as


m
J (Θ) = f (Θ) + g  (Θ) λ = f (Θ) + λi gi (θi ) (9.41)
i=1

where λ (t) ∈ Rm .
Then,
∇ f Θ (Θ) + ∇  g (Θ) λ = 0 (9.42)

λ (t) can be solved by the pseudo-inverse as

 −1
λ = − ∇gΘ (Θ) ∇  gΘ (Θ) ∇g (Θ) ∇ f Θ (Θ) (9.43)

The projection operator then becomes

Θ̇ = Pro (Θ, −Γ ∇ JΘ (Θ))

  −1 
= −Γ I − ∇  g (Θ) ∇gΘ (Θ) ∇  gΘ (Θ) ∇g (Θ) ∇ JΘ (Θ) (9.44)
 
if gi (θi ) ≥ 0 and − ∇  JΘ (Θ) ∇  gΘ (Θ) i > 0, i = 1, . . . , m.
For the special case when m = n, then

λ = −∇ − g (Θ) ∇ f Θ (Θ) (9.45)

The projection operator turns out to be equal to zero since


Θ̇ = Pro (Θ, −Γ ∇ JΘ (Θ)) = −Γ I − ∇  g (Θ) ∇ − g (Θ) ∇ JΘ (Θ) = 0
  (9.46)
if gi (θi ) ≥ 0 and − ∇  JΘ (Θ) ∇  gΘ (Θ) i > 0, i = 1, . . . , m.
This leads to a more intuitive projection method for MRAC which can be stated
in a column-wise format as
⎧      
⎪  ∗ 2 ∗ 2
⎨− Γ Φ (x) e P B i if θi − θi < ri orif  θi − θi  = ri
2 2

θ̇i = and − Φ (x) e P B i θi − θi∗ ≤ 0 (9.47)
⎪
⎩
0 otherwise
220 9 Robust Adaptive Control

The same projection method can be used when Θ (t) is a matrix for each element
θi j (t).
The challenge with the projection method is to be able to determine θi∗ and ri∗ a
priori. Since technically θi∗ is an unknown ideal parameter, the projection method
has the same shortcoming as the dead-zone method in that a priori knowledge of the
uncertainty must be established. This illustrates an important point that robustness
can never be attained for any adaptive control design if the uncertainty is completely
unknown. Thus, with some knowledge on the bounds of the adaptive parameters, the
projection method can be used as an effective robust modification to MRAC.

Example 9.2 For Example 8.4, the largest negative feedback gain that the system
can tolerate is k xmin = −3.8069. If a priori knowledge of this limiting feedback gain
is known, then the projection method can be used to enforce the constraint on k x (t)
during the adaptation process. Suppose this constraint is specified as
 
k x − k ∗  ≤ r
x

where k x∗ = 0 and r = 3.
Using an adaptation rate γx = 7.6, which, in Example 8.4, would result in insta-
bility, the closed-loop system is now stable with k x (t) → −3 when the adaptation
is shut off as shown in Fig. 9.3.

0.5
x

−0.5
0 2 4 6 8 10
t

−1
x

−2
k

−3

−4
0 2 4 6 8 10
t

Fig. 9.3 MRAC with projection method

9.3 σ Modification

Rohrs’ investigation of robustness of adaptive control in the presence of unmodeled

dynamics exhibits the various instability mechanisms of model-reference adaptive
9.3 σ Modification 221

control [1]. Since then, researchers began to examine ways to modify the standard
MRAC to improve robustness. A number of modifications were developed in the
1980s, among which the σ modification is perhaps the simplest modification method
that demonstrates the potential for improved robustness. The σ modification method
was developed by Ioannu and Kokotovic [2]. It is widely used in adaptive control
practice due to its simplicity.
Consider a MIMO system with a parametric matched uncertainty and an unknown
bounded disturbance
 
ẋ = Ax + B u + Θ ∗ Φ (x) + w (9.48)

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, A ∈ Rn × Rn is

known, B ∈ Rn × Rm is also known such that (A, B) is controllable, Θ ∗ ∈ R p × Rm
is the unknown parameter, Φ (x) ∈ R p is a known, bounded function, and w (t) ∈ Rn
is an unknown bounded disturbance.
The reference model is given by

ẋm = Am xm + Bm r (9.49)

where Am ∈ Rn × Rn is known and Hurwitz, Bm ∈ Rn × Rr is known, and r (t) ∈ Rr

is a bounded reference command signal.
We assume that K x and K r can be determined from the model matching conditions

A + B K x = Am (9.50)

B K r = Bm (9.51)

Then, the adaptive controller is designed as

u = K x x + K r r − Θ  Φ (x) (9.52)

The σ modification to the adaptive law that estimates Θ (t) is described by

 
Θ̇ = −Γ Φ (x) e P B + σ Θ (9.53)

where σ > 0 is the modification parameter.

The modification term effectively introduces a constant damping into the adap-
tive law, thereby providing a mechanism to bound the adaptive parameters. The σ
modification is quite effective and yet simple to implement. It is frequently used in
adaptive control to ensure robustness. By bounding the adaptive parameters, the ideal
asymptotic tracking property of MRAC is no longer preserved with the σ modifica-
tion.
222 9 Robust Adaptive Control

The σ modification can be shown to be stable by the Lyapunov’s direct method

as follows:

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.54)

The tracking error equation is obtained as

ė = Am e + B Θ̃  Φ (x) − w (9.55)
 
Evaluating V̇ e, Θ̃ yields
    
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) − 2e Pw + trace −2Θ̃  Φ (x) e P B + σ Θ
(9.56)

This can be simplified as

   
V̇ e, Θ̃ = −e Qe − 2e Pw − 2σ trace Θ̃  Θ
 
= −e Qe − 2e Pw − 2σ trace Θ̃  Θ̃ + Θ̃  Θ ∗ (9.57)
 
Therefore, V̇ e, Θ̃ is bounded by

   2  
   
V̇ e, Θ̃ ≤ −λmin (Q) e2 +2 e λmax (P) w0 −2σ Θ̃  +2σ Θ̃  Θ0 (9.58)

where w0 = max w∞ and Θ0 = Θ ∗ .  

To show that the solution is bounded, we need to show that V̇ e, Θ̃ > 0 inside
 
a compact set but V̇ e, Θ̃ ≤ 0 outside it. To that end, completing the square yields

   
λmax (P) w0 2
V̇ e, Θ̃ ≤ −λmin (Q) e −
λmin (Q)
!  "2
λ 2
(P) w02
  Θ0 σ Θ02
+ max − 2σ Θ̃  − + (9.59)
λmin (Q) 2 2
9.3 σ Modification 223
 
Thus, V̇ e, Θ̃ ≤ 0 if


λmax (P) w0 λ2max (P) w02 σ Θ02
e ≥ + + =p (9.60)
λmin (Q) λ2min (Q) 2λmin (Q)

or 
  Θ λ2max (P) w02 Θ2
  0
Θ̃  ≥ + + 0 =α (9.61)
2 2σ λmin (Q) 4
 
V̇ e, Θ̃ ≤ 0 outside a compact set S defined by
   !  "2
   λmax (P) w0 2
   2 Θ0
S = e , Θ̃  : λmin (Q) e − + 2σ Θ̃  −
λmin (Q) 2
#
λ2 (P) w02 σ Θ02
≤ max + (9.62)
λmin (Q) 2

Therefore, the solution is uniformly ultimately bounded with the ultimate bounds
determined by
  
λmax (P) p 2 + λmax Γ −1 α 2
p ≤ e ≤ =ρ (9.63)
λmin (P)
  
  λmax (P) p 2 + λmax Γ −1 α 2
 
α ≤ Θ̃  ≤   =β (9.64)
λmin Γ −1

We see that the ultimate bound for the tracking error increases due to the σ modifi-
cation term. This is a consequence of the trade-off between the tracking performance
and robustness. Another interesting
 observation is that, if σ = 0, then the standard
 
MRAC is recovered and Θ̃  is unbounded if w0 is not zero. This is consistent with
the parameter drift behavior in the standard MRAC.
If the adaptive controller is a regulator type and the disturbance is zero, then the
closed-loop system is autonomous and is expressed as
 
ẋ = Am x − B Θ − Θ ∗ Φ (x) (9.65)
 
Θ̇ = −Γ −Φ (x) x  P B + σ Θ (9.66)

Let Γ → ∞, then the equilibrium of this system is determined as

224 9 Robust Adaptive Control
 
1
x̄ = A−1
m B Φ (x̄) x̄  P B − Θ ∗ Φ (x̄) (9.67)
σ

1
Θ̄ = Φ (x̄) x̄  P B (9.68)
σ

The equilibrium adaptive parameter Θ̄ thus is inversely proportional to the mod-

ification parameter σ . The equilibrium state x̄ is also inversely proportional to σ .
As the value of the modification parameter σ increases, the tracking error becomes
larger. The reduced tracking performance of the σ modification offers better robust-
ness to unmodeled dynamics in return. Because the adaptive parameter is proven to
be bounded, the parameter drift issue is eliminated.

Example 9.3 Consider the parameter drift example in Example 8.1. Let

x = (1 + t)n

The σ modification adaptive law is given by

 
k̇ x = −γx x 2 b + σ k x

which could also be expressed as

d  γx σ t 
e k x = −eγx σ t γx x 2 b
dt
The solution of this equation is

eγx σ t (1 + t)2n − 1 2neγx σ t (1 + t)2n−1 − 2n
eγx σ t k x − k x (0) = −γx b −
γx σ γx2 σ 2

2n (2n − 1) eγx σ t (1 + t)2n−2 − 2n (2n − 1)
+ − ···
γx3 σ 3

which yields
 ! "
−γx σ t 1 2n 2n (2n − 1)
kx = e k x (0) − γx b − + 2 2− + ···
γx σ γx σ γx3 σ 3
 
(1 + t)2n 2n (1 + t)2n−1 2n (2n − 1) (1 + t)2n−2
− γx b − + − ···
γx σ γx2 σ 2 γx3 σ 3
9.3 σ Modification 225

1.2 0

1.1 −5

u
x
1 −10

0.9 −15
0 10 20 0 10 20
t t

15 0

10
−5

x
w

k
−10
0

−5 −15
0 10 20 0 10 20
t t

Fig. 9.4 σ modification

Note that not only x (t) is bounded for all n < 0, but k x (t) is also bounded. Thus,
the σ modification has eliminated the parameter drift issue.
The response of the closed-loop plant for the same disturbance in this example
with n = − 125
, x (0) = 1, k x (0) = 0, γx = 10, and σ = 0.1 is shown in Fig. 9.4.
Note that with MRAC, k x (t) is unbounded but now becomes bounded with the σ
modification.

Consider a MIMO system

ẋ = Ax + Bu (9.69)

where A is unknown.
The plant is designed with the following adaptive controller:

u = K x (t) x + K r r (9.70)

to follow the reference model in Eq. (9.49).

The σ modification adaptive law for K x (t) is given by
 
K̇ x = Γx xe P B − σ K x (9.71)

Differentiating the closed-loop plant gives

ẍ = A ẋ + B K̇ x x + B K x ẋ + B K r ṙ (9.72)

Then,  
ẍ − (A + B K x ) ẋ − B B  Pex  + σ K x Γx x = B K r ṙ (9.73)

The second term on the left-hand side acts as a damping term due to the propor-
tional control action. The third term is effectively a nonlinear integral control. Without
226 9 Robust Adaptive Control

the σ modification term, as Γx increases, robustness of the closed-loop plant is re-

duced. In the limit as Γx → ∞, the time-delay margin of the closed-loop plant goes
to zero. In the presence of the σ modification term, the integral control action tends
to zero as Γx → ∞, thereby improving robustness of the adaptive controller.
Consider the SISO plant in Sect. 8.5, the σ modification adaptive law for k x (t) is

k̇ x = γx (xeb − σ k x ) (9.74)

The closed-loop plant is then expressed as

ẍ − (a + bk x ) ẋ − bγx (xeb − σ k x ) x = bkr ṙ (9.75)

The third term in the left-hand side is the nonlinear integral control

ki (x) = γx (xeb − σ k x ) = k̇ x (9.76)

Thus, the σ modification changes the ideal integral control action of the adaptive
controller.
Suppose γx → ∞ and r (t) is a constant reference command signal, then k x (t)
tends to an equilibrium
x̄ ēb
k̄ x → (9.77)
σ
This effectively reduces the integral gain ki to zero. Then, the closed-loop system
tends to ! "
x̄ ēb2
ẋ = a + x + bkr r (9.78)
σ

The cross-over frequency and time-delay margin are obtained as

$
x̄ 2 ē2 b4
ω= − a2 (9.79)
σ2
1  σa 
td = cos−1 − (9.80)
ω x̄ ēb2
It can be seen that as γx → ∞, the time-delay margin of the closed-loop system
remains finite. In contrast, for MRAC, the time-delay margin goes to zero as γx → ∞.
In fact, by setting σ = 0 for MRAC, we see that ω → ∞ and td → 0. Since the
time-delay margin is a measure of robustness, the σ modification clearly is more
robust than the standard MRAC.
The equilibrium state is determined by letting γx → ∞ in the adaptive law for
k x (t) so that
σ k̄ x
x̄ 2 − x̄m x̄ + =0 (9.81)
b
9.3 σ Modification 227

which yields ⎛ ⎞

x̄m ⎝ 4σ k̄ x ⎠
x̄ = 1+ 1− (9.82)
2 b x̄m2

The σ modification therefore causes x (t) to not follow xm (t). So, the ideal prop-
erty of asymptotic tracking of MRAC is not preserved. This is a trade-off with
improved robustness.
The equilibrium value of k x (t) can be determined from the state equation

ẋ = (a + bk x ) x + bkr r (9.83)

where r (t) is a constant reference command signal.

This results in
kr r a
k̄ x = − − (9.84)
x̄ b
Since the reference model is given by

ẋm = am xm + bm r (9.85)

therefore a + bk x∗ = am and bkr = bm by the model matching conditions. Then,

the equilibrium reference state is related to the constant reference command signal
according to
am x̄m
r =− (9.86)
bm

Substituting into the equilibrium value of k x (t) gives

am x̄m a
k̄ x = − (9.87)
b x̄ b

If x̄ → x̄m , then ideally k̄ x → k x∗ . With the σ modification, the adaptive parameter

does not tend to its ideal value because the tracking is not asymptotic.
Using the equilibrium value of k x (t), the equilibrium state can be determined
explicitly as a function of the equilibrium reference state x̄m from the following
equation:
b2 x̄ 3 − b2 x̄m x̄ 2 − σ a x̄ + σ am x̄m = 0 (9.88)

Because the tracking is not asymptotic and the adaptive parameter does not tend
to its ideal value, a switching modification to the σ modification has been proposed
by Ioannu [3]. The switching σ modification turns on the σ modification when the
adaptive parameter exceeds a certain threshold and switches off when the adaptive
parameter falls below the threshold. When the σ modification is switched off, the
adaptive law will attempt to restore its ideal property of asymptotic tracking. In
practice, the switching can cause transients and high-frequency oscillations that could
present a problem in itself.
228 9 Robust Adaptive Control

The switching σ modification is stated as


σ0 Θ ≥ Θ0
σ = (9.89)
0 Θ < Θ0

As with the dead-zone method, the selection of the adaptive parameter threshold
is usually not guided by systematic design approaches, but rather by trial-and-error.
Example 9.4 Consider Example 8.7. Keeping everything the same but using a value
of σ = 0.1, the responses of the closed-loop system with and without a time delay of
0.0020 s are shown in Fig. 9.5. Note that the two responses are very similar, whereas
the response with the standard MRAC with a time delay of 0.0020 s shows that the
closed-loop system is unstable. All high-frequency oscillations are no longer present.

1.5 1.5
x xm x xm

1 1
x

0.5 0.5
td = 0 t =0.002 sec
d

0 0
0 5 10 0 5 10
t t

0 1

−0.5
0
x

−1
k

−1
−1.5 td = 0 t =0.002 sec
d

−2 −2
0 5 10 0 5 10
t t

Fig. 9.5 Responses due to σ modification with and without time delay

The ideal feedback gain is computed to be k x∗ = −2. The equilibrium state is then
computed from
x̄ 3 − x̄ 2 − 0.1x̄ − 0.1 = 0

which yields x̄ = 1.1604.

The equilibrium value of k x (t) is computed to be –1.8617. The simulation gives
the same results as the analytical solutions. The time delay that the system can tolerate
is 0.145 s estimated from the simulation.

9.4 e Modification

The e modification is another popular robust modification adaptive law. Developed

by Narendra and Annaswammy [4], it is supposed to have overcome a limitation with
the σ modification in that it can achieve asymptotic tracking under some conditions
while improving robustness.
9.4 e Modification 229

The e modification adaptive law is stated as

   
Θ̇ = −Γ Φ (x) e P B + μ e P B  Θ (9.90)

where μ > 0 is the modification parameter.

The goal of the e modification is to reduce the damping term proportionally to
the norm of the tracking error. As the tracking error tends to zero in the ideal case
with MRAC, the damping term is also reduced to zero, thereby restoring the ideal
asymptotic tracking property of MRAC. However, the stability analysis shows that
the e modification only achieves bounded tracking. The ideal property of MRAC
can no longer be preserved in all cases. So, asymptotic tracking cannot be achieved
in general with increased robustness. This is a trade-off between performance and
robustness that typically exists in all control designs.
The stability proof is given as follows:

Proof Choose the usual Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.91)
 
Evaluating V̇ e, Θ̃ yields
 
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) − 2e Pw
    
+ trace −2Θ̃  Φ (x) e P B + μ e P B  Θ (9.92)

This can be simplified as

     
V̇ e, Θ̃ = −e Qe − 2e Pw − 2μ e P B  trace Θ̃  Θ̃ + Θ̃  Θ ∗ (9.93)
 
V̇ e, Θ̃ is bounded by

   2
 
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2 e λmax (P) w0 − 2μ e P B Θ̃ 
 
 
+ 2μ e P B Θ̃  Θ0 (9.94)

where w0 = max w∞ and Θ0 = Θ ∗ .

 the largest lower bound of e, we want
To find  to maximize  the
 bound on
 
V̇ e, Θ̃ . Taking the partial derivative of V̇ e, Θ̃ with respect to Θ̃  and setting
it equal to zero yield
 
 
− 4μ e P B Θ̃  + 2μ e P B Θ0 = 0 (9.95)
230 9 Robust Adaptive Control
   
 
Solving for Θ̃  that maximizes V̇ e, Θ̃ gives

  Θ
  0
Θ̃  = (9.96)
2
Then,
  μ e P B Θ02
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2 e λmax (P) w0 + (9.97)
2
 
V̇ e, Θ̃ ≤ 0 if

4λmax (P) w0 + μ P B Θ02

e ≥ =p (9.98)
2λmin (Q)
 
Also, note that V̇ e, Θ̃ is bounded by

   2  
   
V̇ e, Θ̃ ≤ 2 e λmax (P) w0 − 2μ e P B Θ̃  + 2μ e P B Θ̃  Θ0
  (9.99)
Then, V̇ e, Θ̃ ≤ 0 if


  Θ Θ02 λmax (P) w0
  0
Θ̃  ≥ + + =α (9.100)
2 4 μ P B
 
Therefore, V̇ e, Θ̃ ≤ 0 outside a compact set S defined by
     
   
S = e , Θ̃  : e ≤ p and Θ̃  ≤ α (9.101)
 
 
Therefore, both e and Θ̃  are uniformly ultimate bounded with the ultimate
bounds determined from Eqs. (9.63) and (9.64). Thus, the e modification results in
bounded tracking. As can be seen, the duality of tracking performance and robustness
cannot usually be simultaneously met.

Example 9.5 Consider the parameter drift example in Example 8.1. The response
of the closed-loop plant for the same disturbance in this example with p = − 12 5
,
x (0) = 1, k x (0) = 0, γx = 10, and μ = 0.1 is shown in Fig. 9.6. The e modification
produces all bounded signals of the closed-loop system.


9.4 e Modification 231

1.2 0

1.1 −5

u
x
1 −10

0.9 −15
0 10 20 0 10 20
t t

15 0

10
−5

x
w

k
−10
0

−5 −15
0 10 20 0 10 20
t t

Fig. 9.6 e modification

Consider the SISO plant in Sect. 8.5, the e modification adaptive law for k x (t) is

k̇ x = γx (xeb − μ |eb| k x ) (9.102)

The closed-loop plant is expressed as

ẍ − (a + bk x ) ẋ − bγx (xeb − μ |eb| k x ) x = bkr ṙ (9.103)

The nonlinear integral control with a nonlinear integral gain is obtained as

ki (x) = γx (xeb − μ |eb| k x ) = k̇ x (9.104)

As γx → ∞, the equilibrium state can be determined by

b x̄ 2 − b x̄ x̄m + μ |ēb| k̄ x = 0 (9.105)

where x̄m is the equilibrium reference state for a constant reference command signal
r (t).
Then,  
x̄ 2 − x̄m ± μk̄ x x̄ ± μk̄ x x̄m = 0 (9.106)

which yields ⎡ ) ⎤
*
x̄m ± μk̄ x ⎣ * 4μ k̄ x̄
1 + +1 ∓  2 ⎦
x m
x̄ = (9.107)
2 x̄m ± μk̄ x

where the upper sign is for eb > 0 and the lower sign is for eb < 0.
Using the equilibrium value of k x (t), the equilibrium state can be determined
explicitly as a function of the equilibrium reference state x̄m from the following
232 9 Robust Adaptive Control

equation:

b x̄ 3 + (−b x̄m ± μa) x̄ 2 ∓ μ (a + am ) x̄m x̄ ± μam x̄m2 = 0 (9.108)

This equation leads to an interesting observation. Since there are multiple roots,
only one of these roots is a valid solution. One possible solution is x̄ = x̄m , which
corresponds to asymptotic tracking. The other solution could be one of the other roots,
depending on the nature of the equilibrium. The closed-loop system is linearized as
 
x̃˙ = a + bk̄ x x̃ + b x̄ k̃ (9.109)
 
k̃˙ x = γx b x̄m − 2 x̄ ± μk̄ x x̃ ∓ γx bμ (x̄m − x̄) k̃ x (9.110)

Then, the Jacobian is obtained as

 
 
J x̄, k̄ x =  a + bk̄ x  b x̄
(9.111)
γx b x̄m − 2 x̄ ± μk̄ x ∓γx bμ (x̄m − x̄)

The solution converges to a stable equilibrium which can be evaluated by exam-

ining of the eigenvalues of the Jacobian matrix.

Example 9.6 Consider Example 8.7 with a = 1, b = 1, and am = −1, the responses
of the closed-loop system with a constant reference command signal r (t) = 1 for
μ = 0.2 and μ = 0.8 are shown in Fig. 9.7.
For μ = 0.2, the solution of x (t) tends to the equilibrium reference state x̄m = 1,
and k x (t) tends to its the equilibrium value at k x∗ = −2. The roots of the polynomial
of x̄ for μ = 0.2 assuming eb < 0

x̄ 3 − 1.2 x̄ 2 + 0.2 = 0

are 1, 0.5583, and −0.3583. The eigenvalues of the Jacobian are (−0.5000 ±
17.3133i), (−4.3332, 46.7162), and (4.6460, 133.9710) for the respective roots.
Therefore, the equilibrium state is x̄ = 1 since it is the only stable equilibrium. The
tracking is therefore asymptotic as shown in Fig. 9.7.
For μ = 0.8, x (t) → x̄ = 1.3798, and k x (t) → k¯x = −1.7247. The roots of the
polynomial of x̄ for μ = 0.8 assuming eb < 0

x̄ 3 − 1.8x̄ 2 + 0.8 = 0

are 1.3798, 1, and −0.5798. The eigenvalues of the Jacobian are (−2.4781,
−150.1650), (−0.5000 ± 17.3133i), and (2.4523, 631.1908) for the respective
roots. The first equilibrium is a stable node and the second equilibrium is a stable
focus. The stable node is more attractive because all trajectories converge exponen-
tially faster to the first equilibrium than to the second equilibrium. Therefore, the
equilibrium state is x̄ = 1.3798. The tracking is no longer asymptotic as shown in
9.4 e Modification 233

Fig. 9.7. The equilibrium value of k x (t) is computed to be –1.7247 which agrees
with the simulation result.
Now, consider a time-varying reference command signal r (t) = sin 2t − cos 4t.
The responses of the closed-loop system μ = 0.2 and μ = 0.8 are shown in Fig. 9.8.
The tracking error is no longer asymptotic as can be seen. The closed-loop system is
non-autonomous, so it does not have a true equilibrium. As μ increases, the tracking
performance is degraded, while robustness increases. For small values of μ, it appears
that asymptotic tracking is almost achieved.

1.5 1.5

1 1
x

x
0.5 μ=0.2 0.5 μ=0.8
x x x x
m m
0 0
0 5 10 0 5 10
t t

0 0

−0.5
−1 μ=0.2
μ=0.8
kx

−1
k

−2
−1.5

−3 −2
0 5 10 0 5 10
t t

Fig. 9.7 e modification for constant reference command signal

2 2
x x x x
m m

1 1
x, xm

x, xm

0 0
μ=0.2 μ=0.8
−1 −1
0 5 10 0 5 10
t t

2 2

0 0
kx

x
k

−2 −2
μ=0.2 μ=0.8
−4 −4
0 5 10 0 5 10
t t

Fig. 9.8 e modification for time-varying reference command signal

234 9 Robust Adaptive Control

9.5 Optimal Control Modification

Robust adaptive control achieves robustness by compromising the ideal asymptotic

tracking property of MRAC. Almost all robust modifications in adaptive control re-
sult in bounded tracking. Recognizing this important requirement of robust adaptive
control, the optimal control modification is a robust modification in adaptive control
that addresses adaptive control in the context of optimal control. More specifically,
the optimal control modification is designed as an optimal control solution to min-
imize the tracking error norm bounded away from the origin by an unknown lower
bound. By not allowing the tracking error to tend to the origin asymptotically, in-
creased robustness can be achieved. This method was developed by Nguyen in 2008
[9, 11]. The method has been subjected to various rigorous validation tests includ-
ing pilot-in-loop flight simulation experiments [12, 13] and flight tests on NASA
F/A-18A aircraft [14–17].
The optimal control modification is formulated from the optimal control theory
as an adaptive optimal control method. The optimal control modification adaptive
law is given by
 
Θ̇ = −Γ Φ (x) e P − νΦ  (x) Θ B  P A−1
m B (9.112)

where ν > 0 is the modification parameter.

To motivate the development of the method, we will review some basic principles
in optimal control theory.

9.5.1 Optimal Control

In optimal control, we are interested in finding an extremal function that maximizes

or minimizes an integral [19]
, tf
J= L (x, u) dt (9.113)
t0

where x (t) ∈ Rn is a state vector, and u (t) ∈ R p is a control vector.

This is a functional extremization or optimization. The optimization finds a func-
tion that maximizes or minimizes the integral J , which is most commonly referred to
as a cost functional or simply cost function. The integrand L () is called an objective
function. The state vector x (t) and the control vector u (t) belong to a dynamical
process described by
ẋ = f (x, u) (9.114)

The solution of the optimization must satisfy the state equation (9.114). There-
fore, the state equation imposes a dynamical constraint on the optimization. Thus,
the optimization is referred to as a constrained optimization. A standard problem
definition in optimal control can usually be stated as follows:
9.5 Optimal Control Modification 235

Find an optimal control u ∗ (t) that minimizes the cost function

, tf
min J = L (x, u) dt (9.115)
t0

subject to Eq. (9.114).

From calculus, the Lagrange multiplier method is used to minimize a function
with constraints. The cost function is augmented by a Lagrange multiplier as
⎧ ⎫
, tf ⎨ ⎬
J= L (x, u) + λ [ f (x, u) − ẋ] dt (9.116)
t0 ⎩ - ./ 0⎭
0

where λ (t) ∈ Rn is the Lagrange multiplier and is often called an adjoint or co-state
vector.
Note that the augmented term is zero, so in effect the cost function does not
change. We define a Hamiltonian function as

H (x, u) = L (x, u) + λ f (x, u) (9.117)

The standard tool in optimal control is calculus of variations. The cost function is
perturbed by a small variation δx in the state vector and variation δu in the control
vector. These variations result in the cost function being perturbed by a small variation
such that
, tf
 
J + δJ = H (x + δx, u + δu) − λ (ẋ + δ ẋ) dt (9.118)
t0

Equation (9.118) can be evaluated as

,  
tf
∂H ∂H  
J + δJ = H (x, u) + δx + δu − λ ẋ − λ δ ẋ dt (9.119)
t0 ∂x ∂u

Consider the term λ δ ẋ. Integrating by parts yields

, ,
tf     tf
λ δ ẋdt = λ t f δx t f − λ (t0 ) δx (t0 ) − λ̇ δxdt (9.120)
t0 t0

Then,
, tf  
J + δJ = H (x, u) − λ ẋ dt
t
-0 ./ 0
J
236 9 Robust Adaptive Control
,  
tf
∂H ∂H    
+ δx + δu + λ̇ δx dt − λ t f δx t f + λ (t0 ) δx (t0 )
t0 ∂x ∂u
(9.121)

The variation in the cost function then becomes

, t f ! " 
∂H ∂H    
δJ = δx + λ̇ δx + δu dt − λ t f δx t f + λ (t0 ) δx (t0 )
t0 ∂x ∂u
(9.122)
J is rendered minimum when its variation is zero, i.e., δ J = 0, which results in
the adjoint equation
∂H 
λ̇ = − = −∇ Hx (9.123)
∂x
and the necessary condition of optimality

∂H 
= ∇ Hu = 0 (9.124)
∂u
Since the initial condition x (t0 ) = x0 is specified for the initial value problem,
 there is no variation in x (t) at t = t0 , i.e., δx (t0 ) = 0. At the final time t = t f ,
then
x t f is generally unknown, and so δx t f = 0. Therefore, the optimality enforces
a transversality condition  
λ tf = 0 (9.125)

which is the final time condition for the adjoint equation.

The optimization is described by a two-point boundary value problem

ẋ = f (x, u ∗ ) x (t0 ) = x0
  (9.126)
λ̇ = −∇ Hx (x, u ∗ , λ) λ t f = 0

with the optimal control u ∗ (t) obtained from the optimality condition (9.124) when
there is no constraint on u (t), or more generally from the celebrated Pontryagin’s
minimum principle
u ∗ (t) = arg min H (x, u) (9.127)
u∈U

which states that the optimal control u ∗ (t) which belongs to an admissible set of
control U is one that minimizes H (x, u).
The Pontryagin’s maximum principle can deal with constraints on the control
such as u min ≤ u (t) ≤ u max . So, in determining an optimal control, we must also
evaluate the Hamiltonian with the bounding values of u (t) if u (t) is constrained.

Example 9.7 Consider an ordinary function minimization with H (u) = −u 2 + u

and −1 ≤ u ≤ 1. Then,
9.5 Optimal Control Modification 237

1
∇ Hu = −2u + 1 = 0 ⇒ u =
2

If u is unconstrained, then the answer is u ∗ = 21 . However, u is constrained to

be between −1 and 1. So, we need to evaluate H (u) at u = 21 as well as at the two
 
bounding values u = ±1. Then, H 21 = 14 , H (−1) = −2, and H (1) = 0. So, the
optimal value of u that minimizes H (u) is

u ∗ = arg min H (u) = −1

u∈(−1, 13 ,1)


The two-point boundary value problem is so-called because half of the boundary
conditions are at the initial time t = t0 and the other half are at the final time t = t f .
The solution method is generally quite complicated and usually must be solved
by numerical methods such as the shooting method. For a general optimization
problem, the gradient method has a certain appeal in that the system equations are
solved exactly at each iteration, with the control variable u ∗ (t) being perturbed
from one iteration to the next to “home in” on the optimal solution. The solution
method usually starts with an initial guess of the control. This then allows the state
vector to be computed by integrating the state equation forward in time using the
specified initial conditions. Once the state vector is computed, the adjoint equation is
integrated backward in time using the final time transversality conditions. The control
is then updated for the next iteration. The whole solution process is iterated until a
convergence on the control is obtained. A schematic of the gradient optimization is
shown in Fig. 9.9. The gradient method is given by

u i+1 = u i − ε∇ Hu (9.128)

where ε = ε > 0 is a positive-definite control-gradient weighting matrix and i is

the iteration index.

Fig. 9.9 Schematic diagram of gradient method

238 9 Robust Adaptive Control

This update is also called the steepest descent method. Visualizing a bowl-shaped
Hamiltonian function H , the gradient ∇ Hu defines the magnitude and direction of
the local slope of the Hamiltonian function as illustrated in Fig. 9.10. Perturbing the
control by some function of ∇ Hu moves the control toward the bottom of the bowl. A
proper selection of the step size is critical for a rapid convergence to the minimizing
control. If ε is too small, the convergence may require a large number of iterations.
On the other hand, if ε is too large, the convergence may not occur at all. Therefore,
the effectiveness of a gradient method is predicated upon a judicious choice of the
weighting matrix ε. If the weighting matrix ε is the inverse of the Hessian matrix
∇ 2 Hu of the Hamiltonian function, then the gradient method is known as the second-
order gradient or Newton–Raphson method.
 −1
u i+1 = u i − ∇ 2 Hu ∇ Hu (9.129)

The continuous time version of the steepest descend method is the gradient search
method
u̇ = −Γ ∇ Hu (9.130)

where Γ = Γ  > 0 is the adaptation rate matrix.

Fig. 9.10 Steepest descent approach to a minimum Hamiltonian function

Now, consider a Linear Quadratic Regulator (LOR) optimal control design for a
MIMO system
ẋ = A p x + B p u (9.131)

y = Cx (9.132)

where x (t) ∈ Rn , u (t) ∈ R p , and y (t) ∈ Rm , with m ≤ p ≤ n

The objective is to design an optimal control that enables the state vector y (t) to
track a command r (t). To design a tracking controller for a step command where
r (t) is a constant signal, we introduce an integral error state
9.5 Optimal Control Modification 239
, t
e= (y − r ) dτ (9.133)
0

Then,
ė = y − r = C x − r (9.134)

If the controller is a stabilizing controller, then as t → ∞, e (t) → 0, and so

y (t) → r (t).
The integral error state is then augmented to the plant model as
        
ė 0 C e 0 I
= + u− r (9.135)
ẋ 0 Ap x Bp 0
-./0 - ./ 0 -./0 - ./ 0 -./0
ż A z B D

 
Let z (t) = e (t) x  (t) , then the augmented plant is expressed as

ż = Az + Bu − Dr (9.136)

We design an optimal control that minimizes the following linear quadratic cost
function: ,
1 tf   
J = lim z Qz + u  Ru dt (9.137)
t→∞ 2 0

Applying the optimal control method, the Hamiltonian function is defined as

1  1
H (z, u) = z Qz + u  Ru + λ (Az + Bu − Dr ) (9.138)
2 2
The adjoint equation and the necessary condition of optimality are obtained as

λ̇ = −∇ Hz = −Qz − A λ (9.139)

 
subject to the transversality condition λ t f = 0, and

∇ Hu = Ru + B  λ = 0 ⇒ u ∗ = −R −1 B  λ (9.140)

The optimal control problem can be solved by the “sweep” method [18] which
assumes an adjoint solution of λ (t) of the form

λ = Wz + Vr (9.141)

Then, substituting the assumed solution of λ (t) into the adjoint equation with the
optimal control u ∗ (t), we get
240 9 Robust Adaptive Control
⎡ ⎤

Ẇ z + W ⎣ Az − B R −1 B  (W z + V r ) − Dr ⎦ + V̇ r = −Qz − A (W z + V r )
- ./ 0
ż
(9.142)
Separating terms by z (t) and r (t) yields the following equations:

Ẇ + W A + A W − WBR−1 B  W + Q = 0 (9.143)

V̇ + A V − W B R −1 B  V − W D = 0 (9.144)
   
subject to the transversality conditions W t f = V t f = 0.
Equation (9.143) is the celebrated differential Riccati equation in optimal control.
Since t f → ∞, this is called an infinite-time horizon control problem. Equations
(9.143) and (9.144) have the following algebraic solutions:

W A + A W − W B R −1 B  W + Q = 0 (9.145)
 −1
V = A − W B R −1 B  WD (9.146)

The optimal control is then obtained as

u = K z z + Kr r (9.147)

where the optimal control gains K x and K r are given by

K z = −R −1 B  W (9.148)
 −1
K r = −R −1 B  A − W B R −1 B  WD (9.149)

The closed-loop plant becomes

ż = Am z + Bm r (9.150)

where
Am = A − B R −1 B  W (9.151)
 −1
Bm = −B R −1 B  A − W B R −1 B  WD − D (9.152)

9.5.2 Derivation of Optimal Control Modification

Given the MIMO system in Sect. 9.3, the optimal control modification seeks to min-
imize the L2 norm of the tracking error bounded away from the origin with a cost
9.5 Optimal Control Modification 241

function [9] , tf
1
J = lim (e − Δ) Q (e − Δ) dt (9.153)
t f →∞ 2 0

where Δ (t) represents an unknown lower bound of the tracking error, subject to
tracking error dynamics described by

ė = Am e + B Θ̃  Φ (x) − w (9.154)

The cost function J is convex and represents the weighted norm square measured
from a point on the trajectory of e (t) to the normal surface of a hypersphere BΔ =
{e (t) ∈ Rn : e ≤ Δ} ⊂ D ⊂ Rn as illustrated in Fig. 9.11. The cost function
is designed to provide robustness by not seeking the ideal property of asymptotic
tracking of MRAC whereby e (t) → 0 as t → ∞, but rather a bounded tracking
whereby the tracking error tends to some lower bound Δ (t) away from the origin.
By not requiring asymptotic tracking, the adaptation therefore can be made more
robust. Thus, the tracking performance can be traded with robustness by a suitable
design of the optimal control modification adaptive law.

Fig. 9.11 Tracking error bound

Using the optimal control framework, we define the Hamiltonian function as

  1 
H e, Θ̃ = (e − Δ) Q (e − Δ) + λ Am e + B Θ̃  Φ (x) − w (9.155)
2
where λ (t) ∈ Rn is an adjoint vector.
Then, the adjoint equation is found by

λ̇ = −∇ He = −Q (e − Δ) − A
mλ (9.156)
 
with the transversality condition λ t f → ∞ = 0 since e (0) is known.
Treating Θ̃ (t) as a control variable, the necessary condition is obtained as

∇ HΘ̃ = Φ (x) λ B (9.157)

242 9 Robust Adaptive Control

The optimal control modification adaptive law is then expressed by the gradient
update law as
Θ̃˙ = −Γ ∇ HΘ̃ = −Γ Φ (x) λ B (9.158)

Note that the adaptive law is dependent upon the adjoint vector which is governed
by the adjoint equation. To eliminate λ (t), we use the “sweep” method by assuming
the following adjoint solution:

λ = Pe + SΘ  Φ (x) (9.159)

Substituting λ (t) in the adjoint equation yields

  
  d Θ  Φ (x)
Ṗe + P Am e + B Θ − Θ ∗ Φ (x) − w + ṠΘ  Φ (x) + S
  dt
= −Q (e − Δ) − Am Pe + SΘ 
Φ (x) (9.160)

Separating the equation that contains e (t), Θ  (t) Φ (x), and the remaining terms
result in
Ṗ + P Am + A mP + Q = 0 (9.161)

Ṡ + A
mS + PB = 0 (9.162)
 
∗ d Θ  Φ (x)
QΔ + P BΘ Φ (x) + Pw − S =0 (9.163)
dt
   
subject to the transversality conditions P t f → ∞ = 0 and S t f → ∞ = 0.
Equation (9.161) is the differential Lyapunov equation. Equation (9.163) shows
that the unknown lower bound Δ (t) of the tracking error is a function of the para-
metric uncertainty Θ ∗ and the unknown disturbance w (t). Thus, as long as the
uncertainty and disturbance exist, the lower bound will be finite, thereby causing the
tracking error to be bounded away from the origin as opposed to tending to the origin
asymptotically as required by MRAC. Robustness thereby is achieved.
For the infinite-time horizon optimal control problem as t → ∞, the solutions
to both Eqs. (9.161) and (9.162) tend to their constant solutions at t = 0 which are
given by
P Am + A mP + Q = 0 (9.164)

A
mS + PB = 0 (9.165)

Equation (9.164) is the familiar Lyapunov equation in adaptive control (see

Sect. 4.2.7), where now the weighting matrix Q is clearly shown to be connected
to the cost function of minimizing the tracking error. The solution to Eq. (9.165) is
also given by
S = −A− m PB (9.166)
9.5 Optimal Control Modification 243

The adjoint solution is then obtained as

λ = Pe − A− 
m P BΘ Φ (x) (9.167)

Substituting the adjoint solution into the gradient update law in Eq. (9.158) yields
the “ideal” optimal control modification adaptive law
 
Θ̇ = −Γ Φ (x) e P − Φ  (x) Θ B  P A−1
m B (9.168)

The first term in the adaptive law is the familiar MRAC. The second term is the
damping term of the optimal control modification. So, this analysis is able to show
the connection between adaptive control and optimal control.
In any design, the ability to adjust a controller to obtain a desired tracking per-
formance and robustness is important. Therefore, suboptimal solutions may provide
a more flexible, yet practical approach to a control design by allowing a trade-off
between the optimality and other design considerations. Thus, to enable the optimal
control modification adaptive law to be sufficiently flexible for a control design, a
modification parameter ν > 0 is introduced as a gain to allow for the adjustment
of the optimal control modification term. The role of the modification parameter ν
is important. If the tracking performance is more desired in a control design than
robustness, then ν could be selected to be a small value. In the limit when ν = 0,
the standard MRAC is recovered and asymptotic tracking performance is achieved
but at the expense of robustness. On the other hand, if robustness is a priority in a
design, then a larger value of ν should be chosen.
Thus, the solution of S is modified as

S = −ν A−
m PB (9.169)

This results in the final form of the optimal control modification

 
Θ̇ = −Γ Φ (x) e P − νΦ  (x) Θ B  P A−1
m B (9.170)

The bound on Δ (t) as t f → ∞ can be estimated by

  4
     d Θ  Φ (x) 
1  ∗   −  
Δ ≤ P B Θ Φ (x) + λmax (P) w0 + ν Am P B   
λmin (Q)  dt 
(9.171)

Since the optimal control modification provides a damping mechanism to MRAC,

the modification term must be negative definite with respect to Θ. Examining the
modification term Γ νΦ (x) Φ  (x) Θ B  P A−1 
m B, the product Φ (x) Φ (x) is a pos-
 −1
itive semi-definite null matrix, and Γ ν > 0. Therefore, B P Am B is required to be
negative definite.
Any square real matrix C can be decomposed into a symmetric part M and anti-
symmetric part N
244 9 Robust Adaptive Control

C=M+N (9.172)

where
1 
M = M = C + C (9.173)
2
1 
N = −N  = C − C (9.174)
2

Consider the quadratic scalar function x  C x for any arbitrary x(t) ∈ Rn .

x C x = x  M x + x  N x (9.175)

Recall that for any anti-symmetric matrix N , x  N x = 0. Therefore,

x C x = x  M x (9.176)

This expression is useful for determining the sign definiteness of a matrix. It

follows that, if the symmetric part M is positive (negative) definite, then C is also
positive (negative) definite. Utilizing this property, the symmetric part of P A−1
m is
negative definite since

1  1 −
M= P A−1 − −1
m + Am P = − Am Q Am < 0 (9.177)
2 2

Therefore, x  B  P A−1 1   − −1
m Bx = − 2 x B Am Q Am Bx < 0. Thus, the optimal
control modification term is negative definite with respect to Θ (t).
Consider a SISO system in Sect. 8.1. The optimal control modification adaptive
law is given by  
k̇ x = −γx x 2 b − νx 2 am−1 b2 k x (9.178)

Since am < 0, the modification term is negative with respect to k x (t). The feed-
back gain k x (t) can be computed as

k̇ x
= −γ x 2 b (9.179)
1 − νam−1 bk x

The standard MRAC is

k̇ x∗ = −γ x 2 b (9.180)

where k x∗ denotes the ideal feedback gain obtained from MRAC.

Therefore,
k̇ x
= k̇ x∗ (9.181)
1 − νam−1 bk x
9.5 Optimal Control Modification 245

This equation can easily be solved. The solution is

1 1 − νam−1 bk x
− ln = k x∗ − k x∗ (0) (9.182)
νam−1 b 1 − νam−1 bk x (0)

This yields

1 1 − νam−1 bk x (0)
 
kx = − exp −νam−1 b k x∗ − k x (0) (9.183)
νam−1 b νam−1 b

In the case of the parameter drift, k x∗ (t) → −∞ for b > 0 with the standard
MRAC, therefore −νam−1 bk x∗ (t) → −∞. This implies that k x (t) is bounded with
the optimal control modification.
The optimal control modification exhibits some nice properties that make it easy
to analyze. Since k x (t) is bounded, the equilibrium value of k x (t) can be analytically
obtained by letting γx → ∞ which yields

1
k̄ x = (9.184)
νam−1 b

The equilibrium value of k x (t) can also be obtained from the solution of k x (t) as
k x∗ (t) → −∞ which yields the same result.
The equilibrium value of k x (t) thus is inversely proportional to the optimal control
modification parameter ν.
k x (t) can then be expressed as
   ∗ 
k x (0) k x − k x (0)
k x = k̄ x − k̄ x 1 − exp − (9.185)
k̄ x k̄ x

The closed-loop plant with k x (t) → k̄ x is

 am 
ẋ = a + x +w (9.186)
ν
The closed-loop plant is stable if
am am
a+ <0⇒ν<− (9.187)
ν a
This yields two cases 
< − aam a > 0
ν (9.188)
>0 a<0

The first case is when the open-loop plant is stable, i.e., a < 0. Then, the optimal
control modification parameter ν can take on any positive value. Thus, if the plant
246 9 Robust Adaptive Control

is open-loop stable, then the optimal control modification is guaranteed to be stable

for all values of ν > 0.
The second case is when the open-loop plant is unstable, i.e., a > 0, then there
exists an upper limit of the modification parameter ν as indicated in Eq. (9.188). This
implies that the modification parameter ν has to be chosen based on a priori knowl-
edge of a. This is perhaps counterintuitive since the problem statement indicates that
a is unknown which implies that the upper limit of ν is also unknown. However, this
is consistent with the robust control framework. Robust stability generally requires
a priori knowledge of the upper bound of the uncertainty. With a complete lack of
knowledge of the uncertainty, it is difficult to ensure robustness of any control sys-
tems. Therefore, the optimal control modification can be designed to guarantee the
closed-loop stability for a given specification of the open-loop plant uncertainty.

Example 9.8 Consider the parameter drift example in Example 8.1, k x∗ (t) is given
by
(1 + t)2n+1 − 1
k x∗ − k x∗ (0) = −γx b
2n + 1

k x∗ (t) → −∞ if n > − 21 .
With the optimal control modification, k x (t) is obtained as
 
1 1 − νam−1 bk x (0) (1 + t)2n+1 − 1
kx = − exp γx νam−1 b2
νam−1 b νam−1 b 2n + 1

Therefore, k x (t) is bounded for all n if the modification parameter ν is selected

such that a + aνm < 0. Then, the optimal control modification has eliminated the
parameter drift issue.
Let a = 1, b = 1, and am = −1. Then, ν = 0.1 is chosen to satisfy the limiting
condition which dictates that ν < 1 for the closed-loop stability. The response of the
closed-loop plant for the same disturbance in this example with n = − 12 5
, x (0) = 1,
k x (0) = 0, and γx = 10 is shown in Fig. 9.12. The equilibrium value is calculated
to be k̄ x = −10 which agrees exactly with the simulation result.
1.3 0

1.2
−5
1.1
u
x

−10
1

0.9 −15
0 10 20 0 10 20
t t

15 0

10
x
w

5 −5
k

−5 −10
0 10 20 0 10 20
t t

Fig. 9.12 Optimal control modification

9.5 Optimal Control Modification 247

9.5.3 Lyapunov Stability Analysis

The optimal control modification can be shown to achieve stable and bounded track-
ing by the Lyapunov’s direct method as follows:

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.189)
 
Differentiating V e, Θ̃ yields
 
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) − 2e Pw
   
− 2trace Θ̃  Φ (x) e P − νΦ  (x) Θ B  P A−1
m B
= −e Qe − 2e Pw + 2νΦ  (x) Θ B  P A−1 
m B Θ̃ Φ (x) (9.190)

Then,
 
V̇ e, Θ̃ = −e Qe − 2e Pw + 2νΦ  (x) Θ̃ B  P A−1 
m B Θ̃ Φ (x)

+ 2νΦ  (x) Θ ∗ B  P A−1 

m B Θ̃ Φ (x) (9.191)

But, B  P A−1
m B < 0. So,

2νΦ  (x) Θ̃ B  P A−1    − −1 

m B Θ̃ Φ (x) = −νΦ (x) Θ̃ B Am Q Am B Θ̃ Φ (x)
  (9.192)
Then, V̇ e, Θ̃ is bounded by
 
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2λmax (P) e w0
   2
2 
− νλmin B  A− −1
m Q Am B Φ (x) Θ̃ 
   
+ 2ν  B  P A−1 B  Φ (x)2  
Θ̃  Θ0 (9.193)
m

 
Let c1 = λmin (Q), c2 = λmaxc(P)w 1
0
, c3 = λmin B  A− −1
m Q Am B , and c4 =
m B Θ0
 B  P A−1
c3
. Then, upon completing the square, we get
    2
 
V̇ e, Θ̃ ≤ −c1 (e − c2 )2 +c1 c22 −νc3 Φ (x)2 Θ̃  − c4 +νc3 c42 Φ (x)2
  (9.194)
V̇ e, Θ̃ ≤ 0 implies
248 9 Robust Adaptive Control

νc3 c42 Φ (x)2
e ≥ c2 + c22 + =p (9.195)
c1

  c1 c22
 
Θ̃  ≥ c4 + c42 + =α (9.196)
νc3 Φ (x)2

Note that the lower bounds p and α are dependent on Φ (x). Therefore, to prove
boundedness, we also need to show that Φ (x) is bounded. If Φ (x) is a bounded
function such as sin x, sigmoidal or radial basis functions, then Φ (x) ≤ Φ0 and
the solution is uniformly ultimately bounded. The ultimate bounds can be established
as   
λmax (P) p 2 + λmax Γ −1 α 2
e ≤ ρ = (9.197)
λmin (P)
  
  λmax (P) p 2 + λmax Γ −1 α 2
 
Θ̃  ≤ β =   (9.198)
λmin Γ −1

Otherwise, we consider the following cases:

1. The closed-loop plant with the nominal controller and no disturbance is stable.
That is, Θ (t) = 0 and w (t) = 0. This is the same as stating that the uncertainty
is non-destabilizing. The tracking error equation becomes

ė = Am e − BΘ ∗ Φ (x) (9.199)

Choose a Lyapunov candidate function

V (e) = e Pe (9.200)

Then,

V̇ (e) = −e Qe − 2e P BΘ ∗ Φ (x) ≤ −λmin (Q) e2 + 2 e P B Θ0 Φ (x)

(9.201)
Since the plant is stable, V̇ (e) ≤ 0 which implies Φ (x) is bounded by

λmin (Q) e

Φ (x) ≤ (9.202)
2 P B Θ0

Substituting Eq. (9.202) into Eq. (9.194) yields

9.5 Optimal Control Modification 249

  λ2 (Q) e2   2
 
V̇ e, Θ̃ ≤ −c1 (e − c2 )2 + c1 c22 − νc3 min  Θ̃  − c4
4 P B2 Θ02
λ2min (Q) e2
+ νc3 c42 (9.203)
4 P B2 Θ02

V̇ (e) ≤ 0 requires the coefficient of e2 to be negative. Therefore, there exists

a maximum value νmax for which ν < νmax where
 
4λmin B  A− −1
m Q Am B P B
2
νmax =   (9.204)
λmin (Q)  B  P A−1
m B
2

 that νmax is not dependent on the bound of the parametric uncertainty Θ0 .

Note
V̇ e, Θ̃ then becomes

  ! " 
ν λ2 (Q) e2  
2
V̇ e, Θ̃ ≤ −c1 1 − e2 + 2c1 c2 e − νc3 min  Θ̃  − c4
νmax 4 P B2 Θ02
(9.205)

Upon completing the square, we get

  ! "5 62
ν c2
V̇ e, Θ̃ ≤ −c1 1 − e − ν
νmax 1 − νmax
c1 c22 νc3 λ2min (Q) e2   2
 
+ ν −  Θ̃  − c4 (9.206)
1 − νmax 4 P B2 Θ02
 
The lower bound of e for which V̇ e, Θ̃ ≤ 0 is determined from
! "
ν
− c1 1 − e2 + 2c1 c2 e ≤ 0 (9.207)
νmax

which yields
2c2
e ≥ ν = p (9.208)
1 − νmax
 
 
The lower bound of Θ̃  is determined by setting e = c2
ν that renders
1−
νmax
  c1 c22 νc3 λ2min (Q) c22   2
 
V̇ Θ̃ ≤ ν −  2 Θ̃  − c4 (9.209)
1− ν
4 1 − νmax P B2 Θ02
νmax
250 9 Robust Adaptive Control
 
Therefore, V̇ Θ̃ ≤ 0 implies

)  
*
  * 4c1 1 − ν P B2 Θ 2
  + νmax 0
Θ̃  ≥ c4 + =α (9.210)
νc3 λmin (Q)
2

Then, the closed-loop system is uniformly ultimately bounded with the ultimate
bounds given by Eqs. (9.197) and (9.198).
2. The closed-loop plant with the nominal controller has no stability guarantee. The
 If ν = 0, then from Eqs. (9.195) and (9.196),
uncertainty can be destabilizing.
 
we see that e is bounded but Θ̃  is unbounded. This illustrates the parameter
 
 
drift behavior of MRAC. On the other hand, if ν → ∞, then Θ̃  is bounded,
but e can be unbounded. Thus, there exists a maximum value νmax such that
ν < νmax which corresponds to the largest value of Φ (x) ≤ Φ0 that renders
the largest ultimate bound of e. From Eq. (9.194), the largest ultimate bound
of e is determined from
 
V̇ e, Θ̃ ≤ −c1 (e − c2 )2 + c1 c22 + νc3 c42 Φ (x)2 (9.211)

But,

− e Qe = − (xm − x) Q (xm − x) = −x  Qx + 2xm Qx − xm Qxm (9.212)

Therefore,

− c1 e2 ≤ −c1 x2 + 2c5 x xm  − c1 xm 2 (9.213)

where c5 = λmax (Q).

Let

ϕ (x , xm  , Q, ν, w0 , Θ0 ) = −c1 x2 + 2 (c1 c2 + c5 xm ) x

+ 2c1 c2 xm  − c1 xm 2 + νc3 c42 Φ (x)2
(9.214)
   
where ϕ () is the largest upper bound of V̇ e, Θ̃ such that V̇ e, Θ̃ ≤
ϕ (x , xm  , Q, ν, w0 , Θ0 ).
Then, for any ν < νmax , x can be determined by

x = ϕ −1 (xm  , Q, ν, w0 , Θ0 ) (9.215)

where ϕ −1 () is an inverse function of ϕ ().

It follows that Φ (x) is bounded such that
9.5 Optimal Control Modification 251
  
Φ (x) = Φ ϕ −1 (xm  , Q, ν, w0 , Θ0 )  = Φ0 (9.216)

Then, the closed-loop system is uniformly ultimately bounded with

   
νc3 c42 Φ02 λmax (P) r 2 + λmax Γ −1 α 2
c2 + c22 + = p ≤ e ≤ ρ =
c1 λmin (P)
(9.217)
   
c1 c22   λmax (P) r 2 + λmax Γ −1 α 2
 
c4 + c4 +
2
= α ≤ Θ̃  ≤ β =  
νc3 Φ02 λmin Γ −1
(9.218)
for any 0 < ν < νmax .
Consider a special case when Φ (x) belongs to a class of regressor functions such
that Φ (x) ≤ x. This class
  regressor functions also includes Φ (x) = x.
of
Then, from Eq. (9.214), V̇ e, Θ̃ is bounded by

 
ϕ (x , xm  , Q, ν, w0 , Θ0 ) = − c1 − νc3 c42 x2 + 2 (c1 c2 + c5 xm ) x
+ 2c1 c2 xm  − c1 xm 2 (9.219)

x is obtained from

x = ϕ −1 (xm  , Q, ν, w0 , Θ0 )
7   
c2 + c5 xm  + (c1 c2 + c5 xm )2 + 4 c1 + νc3 c42 2c1 c2 xm  − c1 xm 2
=
c1 − νc3 c42
(9.220)

for any 0 < ν < νmax such that c1 − νc3 c42 > 0 which yields
 
c1 λmin (Q) λmin B  A− −1
m Q Am B
νmax = =  2 2 (9.221)
c3 c42  B  P A−1
m B
 Θ
0

Note that νmax is now dependent on the upper bound of the parametric uncertainty
Θ0 and is a conservative estimate since the negative definite term with ||Θ̃||2 in
Eq. (9.194) is neglected. As the bound of the uncertainty increases, νmax must
be reduced to ensure the stability of the closed-loop system. Thus, the stability
of the optimal control modification depends on the bound of the uncertainty. The
adaptive law is guaranteed to be stable if a priori knowledge of the bound of the
uncertainty exists.

In contrast to the σ modification and e modification which do not impose a limit

on the modification parameters σ and μ, respectively, it may seem that the optimal
252 9 Robust Adaptive Control

control modification adaptive law is more restrictive. However, the lack of a well-
defined limit on the modification parameter can sometimes be a source of stability
issues. Numerical simulations using the Rohrs counterexample have shown that both
the σ modification and e modification as well as the optimal control modification all
exhibit limiting values on the modification parameters σ , μ, and ν, respectively.
Is should be noted that the limiting value νmax derived from the Lyapunov analysis,
in general, is a conservative estimate of the true limiting value νmax . Consider the
following first-order SISO system with an adaptive regulator controller using the
optimal control modification:
 
ẋ = ax + b u + θ ∗ x (9.222)

u = k x x − θ (t) x (9.223)
 
θ̇ = −γ −x 2 b − νx 2 am−1 b2 θ (9.224)

with a and b known, and am < 0.

The closed-loop system is
 
ẋ = am − bθ + bθ ∗ x (9.225)

Note that the adaptive law implicitly uses p = 1 where p is the solution of the
scalar Lyapunov equation
2 pam = −q (9.226)

By letting γ → ∞, the equilibrium value of θ (t) is found to be

1
θ̄ = − (9.227)
νam−1 b

Using the equilibrium value θ̄ , the closed-loop system becomes

 am 
ẋ = am + + bθ ∗ x (9.228)
ν

If the uncertainty is stabilizing, then bθ ∗ < 0. Then, the closed-loop system is

stable for any ν > 0. Yet, using Eq. (9.204) from the Lyapunov analysis, ν should
be limited by

4b2 am−2 b2
νmax = =4 (9.229)
b4 am−2

If the uncertainty is destabilizing but the closed-loop plant with the nominal
(non-adaptive) controller u (t) = k x x (t) is stable, then 0 < bθ ∗ < −am . Then, the
closed-loop system is still stable for any ν > 0. Yet, using Eq. (9.221), ν should be
limited by
9.5 Optimal Control Modification 253
! "2
4am2 b2 am−2 2am
νmax = = (9.230)
b4 am−2 θ ∗2 bθ ∗

Finally, if the uncertainty is destabilizing and the closed-loop plant with the nom-
inal controller is unstable, then bθ ∗ > −am . The closed-loop plant is stable if
am
am + + bθ ∗ < 0 (9.231)
ν
Then, ν is limited by
am
νmax = − (9.232)
am + bθ ∗

Combining with the Lyapunov expression, νmax is expressed as

! "2
2am am
νmax = min ,− (9.233)
bθ ∗ am + bθ ∗

 m 2  2
Let bθ ∗ = −αam where α > 1, then 2a bθ ∗
→ α2 and − am +bθ am
∗ → α−1 .
1
 2 2
Since α < α−1 1
for all α > 1, therefore the most conservative value of νmax is
established by the Lyapunov analysis which is equal to

4
νmax = (9.234)
α2
On the other hand, the least conservative value of νmax is the true value of νmax
which is equal to
1
νmax = (9.235)
α−1

While the optimal control modification is defined according to Eq. (9.170), there
are other variances of the optimal control modification. These variances are:
 
Θ̇ = −Γ Φ (x) e P + νΦ  (x) Θ B  A− −1
m Q Am B (9.236)
 
Θ̇ = −Γ Φ (x) e P + νΦ  (x) Θ B  R B (9.237)
 
Θ̇ = −Γ Φ (x) e P B + νΦ  (x) Θ R (9.238)
 
Θ̇ = −Γ Φ (x) e P B + νΦ  (x) Θ (9.239)

where R = R  > 0 is a positive-definite matrix.

254 9 Robust Adaptive Control

9.5.4 Linear Asymptotic Property

Consider a linear uncertain MIMO system

 

ẋ = Ax + B u + Θ ∗ x (9.240)

The plant is designed with the following adaptive controller:

u = K x x + K r r − Θ  (t) x (9.241)

to follow a reference model

ẋm = Am xm + Bm r (9.242)

where Am = A + B K x and Bm = B K r .
The optimal control modification adaptive law for Θ (t) is given by
 
Θ̇ = −Γ xe P B − νx x  Θ B  P A−1
m B (9.243)

It has been shown that MRAC is non-robust with fast adaptation. The time-delay
margin of MRAC tends to zero as the adaptation rate tends to infinity. The optimal
control modification exhibits a linear asymptotic property as Γ → ∞ since the
equilibrium value Θ (t) is independent of x (t) if xm (t) = 0, as has been shown for
simple SISO systems [20, 21]. The equilibrium value of Θ  (t) x (t) is given by

1   − −1 
Θ̄  x = B Am P B B Pe (9.244)
ν

Then, the closed-loop system tends to the following asymptotic linear system as
Γ → ∞ in the limit:
 
1  −1  ∗
ẋ = Am + B B  A− P B B P + BΘ x
ν m

1  −1 
− B B  A− m PB B P xm + Bm r (9.245)
ν

This system is stable for all ν > 0 if Am + BΘ ∗ is Hurwitz and stable for
ν < νmax if Am + BΘ ∗ is an unstable matrix.
9.5 Optimal Control Modification 255

Consider a special case when ν = 1 corresponding to the optimal control solution

of the adaptive law and B is a square invertible matrix, then
 −1 
Am + B B  A−
m PB B P = P −1 P Am + BB−1 P −1 A
mB
− 
B P
−1
 
 −1
=P P Am + Am P = −P Q (9.246)

The tracking error equation tends to in the limit

 
 
ė = − P −1 Q − BΘ ∗ e − BΘ ∗ xm (9.247)

Since P > 0 and Q > 0, −P −1 Q < 0. The closed-loop poles of the ideal system
with Θ ∗ = 0 are all negative real. Therefore, the closed-loop system has the best
stability margin. The ideal system is exponentially stable with no high-frequency
oscillations. By choosing Q appropriately, the closed-loop system is guaranteed to

be stable if −P −1 Q + BΘ ∗ is Hurwitz.
The linear asymptotic property of the optimal control modification is quite useful
since it can be used for stability analysis using many available linear analysis tools.
Another advantage that comes with the linear asymptotic property is the scaled input–
output behavior of the closed-loop system. That is, if r (t) is scaled by a multiplier
c, then x (t) is scaled by the same amount. To see this, we express the asymptotic
closed-loop system as a transfer function
 
1   − −1  ∗
sx = Am + B B Am P B B P + BΘ x
ν
1  −1 
− B B  A−
m PB B P xm + Bm r (9.248)
ν
From the reference model, we get

sxm = Am xm + Bm r ⇒ xm = (s I − Am )−1 Bm r (9.249)

Therefore,
 −1
1   − −1  ∗
x = s I − Am − B B Am P B B P − BΘ
ν
 
1   − −1  −1
× − B B Am P B B P (s I − Am ) + I Bm r (9.250)
ν

If x (t) = x0 (t) is the response due to r (t) = r0 (t) and if r (t) is scaled by a
constant multiplier c, i.e., r (t) = cr0 (t), then it follows that x (t) is also scaled by
the same multiplier c, i.e., x (t) = cx0 (t). The scale input–output behavior makes
the optimal control modification more predictable than the standard MRAC which
does not possess this linear asymptotic property.
256 9 Robust Adaptive Control

If r (t) is a constant signal, then the equilibrium value of x (t) can be found by
setting s = 0 as t → ∞. Then,
 −1
1  −1  ∗
x̄ = − Am + B B  A− P B B P + BΘ
ν m
 
1   − −1 
× B B Am P B B P A−1 + I Bm r (9.251)
ν m

If ν = 0, then the ideal property of asymptotic tracking of MRAC is recovered

since
 −1
1   − −1  ∗
x̄ = − lim Am + B B Am P B B P + BΘ
ν→0 ν
 
1   − −1  −1
× B B Am P B B P Am + I Bm r
ν
 
1   − −1  −1 1   − −1 
= − lim B B Am P B B P B B Am P B B P A−1
m Bm r
ν→0 ν ν
= −A−1
m Bm r = x̄ m (9.252)

The equilibrium value of the tracking error is given by

  −1
1   − −1  ∗
ē = x̄m − x̄ = −A−1 + Am + B B Am P B B P + BΘ
m
ν
 #
1   − −1 
× B B Am P B B P A−1 + I Bm r (9.253)
ν m

The largest norm of ē (t) can be interpreted as a steady-state error when Γ → ∞

which is given by
  −1
 1   − −1 
 −1 ∗
ē = −Am Bm + Am + B B Am P B B P + BΘ
 ν
  
1   − −1  
× B B Am P B B P Am + I Bm 
−1
 r  (9.254)
ν

The linear asymptotic property also affords another advantage in that the stability
margins of the system in the limit can be computed.
Consider a first-order time-delay SISO system with an optimal control modifica-
tion adaptive controller
 
ẋ = ax + b u (t − td ) + θ ∗ x (9.255)
9.5 Optimal Control Modification 257

u = k x x + kr r − θ (t) x (9.256)
 
θ̇ = −γ xeb − νx 2 am−1 b2 θ (9.257)

with a and b known and am = a + bk x < 0.

For fast adaptation with γ → ∞ and a constant command signal r (t), the equi-
librium value of θ (t) x (t) is
xm − x
θ̄ x = (9.258)
νam−1 b

Then, the closed-loop plant tends to in the limit

   am  am
ẋ = a + bθ ∗ x + bk x + x (t − td )− xm (t − td )+bkr r (t − td ) (9.259)
ν ν
To simplify the analysis, let r (t) = 1 and xm (t) = 1. Then, the characteristic
equation with s = jω is
   am 
jω − a + bθ ∗ − bk x + (cos ωtd − j sin ωtd ) = 0 (9.260)
ν
which results in the following equations:
   am 
− a + bθ ∗ − bk x + cos ωtd = 0 (9.261)
ν
 am 
ω + bk x + sin ωtd = 0 (9.262)
ν
Then, the cross-over frequency and time-delay margin are computed as
$
am 2
ω= bk x + − (a + bθ ∗ )2 (9.263)
ν
⎛ ⎞
∗
1 ⎜ a + bθ ⎟
td = cos−1 ⎝− a ⎠ (9.264)
ω bk x +
m
ν

If ν = 0, then the optimal control modification reverts to the standard MRAC.

Then, the time-delay margin tends to zero precisely as follows:
$
am 2
ω = lim bk x + − (a + bθ ∗ )2 → ∞ (9.265)
ν→0 ν
! "
1 a + bθ ∗
td = lim cos−1 − =0 (9.266)
ν→0 ω bk x + aνm
258 9 Robust Adaptive Control

This is the expected result for MRAC. For any 0 < ν < νmax , the optimal
control modification yields a non-zero time-delay margin with fast adaptation. This
is a robustness property of any robust adaptive control with the ability to provide
sufficient stability margins under fast adaptation. For a given time delay td and a priori
knowledge of θ ∗ , the modification parameter ν thus can be computed to guarantee
stability of the closed-loop system.

Example 9.9 Let a = 1, am = −1, b = bm = 1, θ ∗ = 0.2, and r (t) = 1. This

is the same example as Example 8.7 which has a time-delay margin of 0.0020 s for
γx = 500. The open-loop system is unstable. So, the limiting value of ν is computed
as ! "
2am 2
νmax = =1
bθ ∗

Choose ν = 0.1 < 1. The time-delay margin for the closed-loop system with the
optimal control modification is calculated as
$
1
ω= − 1 = 9.9499 rad/s
ν2

1
td = 7 cos−1 ν = 0.1478 sec
1
ν2
−1

Since the time-delay margin td decreases with increasing the adaptation rate γ ,
therefore, for finite γ , the time-delay margin estimated with γ → ∞ is the lowest
estimate of the time-delay margin for any finite γ < ∞. In other words, the estimated
time-delay margin td = 0.1478 s using the linear asymptotic property provides a
lower bound on the time-delay margin. The actual time-delay margin for any arbitrary
adaptation rate γ should be greater than this value in theory. We see that, for even
a small value of ν, a significant increase in the time-delay margin can be achieved
as compared to the time-delay margin of 0.0020 s for the standard MRAC with
γx = 500. In the limit as ν → 0, td → 0 corresponding to the standard MRAC as
γx → ∞.
The steady-state error is estimated to be
  −1 ! 1 "
−1 am ∗
ē = −am + am + + bθ + 1 bm r = −0.0185
ν ν

The equilibrium value of θ (t) is computed to be

ē ē
θ̄ = = = 0.1818
νam−1 b x̄ νam−1 b (x̄m − ē)
9.5 Optimal Control Modification 259

We choose γ = 100 and θ (0) = 0. The results of the simulation with the optimal
control modification are shown in Fig. 9.13. The closed-loop system is completely
stable. The tracking error e (t)and the adaptive parameter θ (t) converge to –0.0185
and 0.1818, respectively, at 10 s. Thus, the simulation results agree exactly with the
equilibrium values of ē and θ̄ .

1.5 0

−0.005
1
x, xm

−0.01

e
0.5
x −0.015
xm
0 −0.02
0 5 10 0 5 10
t t

1 0.4

0.3
0
0.2
u

−1
0.1

−2 0
0 5 10 0 5 10
t t

Fig. 9.13 Optimal control modification with γ = 100

When a time delay equal to 0.1477 s is injected at the input, the closed-loop system
is on the verge of instability as shown in Fig. 9.14. In fact, for γ = 1000 using a
time step Δt = 0.0001, the closed-loop system becomes unstable at 0.1478 s as
predicted. Thus, the numerical evidence of the time-delay margin is in agreement
with the analytical prediction.
To illustrate the linear asymptotic property, we consider the asymptotic linear
system in the limit as follows:

xm − x
ẋ = ax + bθ ∗ x − b
νam−1 b

Figure 9.15 shows the closed-loop response with the optimal control modification
for r (t) = 1 + sin 2t + cos 4t. The closed-loop response follows the asymptotic
linear system almost perfectly as both responses are on top of each other.
To illustrate the scaled input–output linear behavior of the optimal control modi-
fication, the reference command signal r (t) = 1 + sin 2t + cos 4t is doubled so that
the new command is r1 (t) = 2 + 2 sin 2t + 2 cos 4t. Let x1 (t) be the closed-loop
response to r1 (t). Then, the response of x1 (t) is twice that of x (t) as shown in
Fig. 9.16.
260 9 Robust Adaptive Control

3 2
x x
m
2 1

x, xm
1 0

e
0 −1

−1 −2
0 5 10 0 5 10
t t

20 40

10 20

0 0
u

θ
−10 −20

−20 −40
0 5 10 0 5 10
t t

Fig. 9.14 Optimal control modification with γ = 1000 for td = 0.1477 s

1.6

1.4

1.2

|
x, x

0.8

0.6

0.4

0.2
x
−
x
0
0 2 4 6 8 10
t

Fig. 9.15 Scaled input–output property of adaptive control

4
x x
1
3
1
x, x

0
0 2 4 6 8 10
t

2.005

2
x1/x

1.995

1.99
0 2 4 6 8 10
t

Fig. 9.16 Scaled input–output property of adaptive control

9.6 Adaptive Loop Recovery Modification 261

9.6 Adaptive Loop Recovery Modification

Adaptive loop recovery modification is another recent robust modification adaptive

law developed by Calise and Yucelen in 2009 [6]. The adaptive loop recovery mod-
ification is intended to achieve asymptotic tracking while preserving the stability
margin of the reference model to the degree possible even under uncertainty. The
modification is given by
 
Θ̇ = −Γ Φ (x) e P B + ηΦx (x) Φx (x) Θ (9.267)

where η > 0 is a modification parameter and Φx (x) = dΦ(x) dx

.
The adaptive loop recovery modification is based on the idea of seeking an adaptive
law that minimizes the nonlinearity in a closed-loop plant so that the stability margin
of a linear reference model could be preserved. Consider the tracking error equation
with no disturbance
ė = Am e + B Θ̃  Φ (x) (9.268)

The tracking error equation can be linearized as

Δė = Am Δe + BΔΘ̃  Φ (x̄) + B Θ̃  Φx (x̄) (x − x̄) (9.269)

where x̄ is the equilibrium state.

If ΔΘ̃ (t) = 0 and Θ̃  (t) Φx (x̄) = 0, then

Δė = Am Δe (9.270)

Then, the closed-loop plant follows the reference model exactly and the stability
of the reference model is preserved.
This leads to a hypothesis that if the quantity Θ  (t) Φx (x) is minimized, as-
ymptotic tracking is achieved, and the stability margin of the reference model is
preserved. This can be established by minimizing the following cost function:

1 
J= Φ (x) ΘΘ  Φx (x) (9.271)
2 x
as an unconstrained optimization.
The gradient of the cost function is computed as

∇ JΘ = Φx (x) Φx (x) Θ (9.272)

The gradient update law then is expressed as

Θ̇ = −Γ ∇ JΘ = −Γ Φx (x) Φx (x) Θ (9.273)

262 9 Robust Adaptive Control

Adding the standard MRAC to the update law results in the adaptive loop recovery
modification.
The original stability proof is based on a singular perturbation approach [6, 22].
The reader is referred to the original work [6, 22] for the formal proof. It can be shown
that the stability proof of the optimal control modification can also be applied to the
adaptive loop recovery modification since both adaptive laws have their damping
terms proportional to some positive-definite function of x (t). The alternate proof is
as follows:

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.274)
 
V̇ e, Θ̃ is evaluated as
 
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) − 2e Pw
  
− 2trace Θ̃  Φ (x) e P B + ηΦx (x) Φx (x) Θ
= −e Qe − 2e Pw − 2ηΦx (x) Θ Θ̃  Φx (x) (9.275)
 
Then, V̇ e, Θ̃ is bounded by

   2
 
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2λmax (P) e w0 − 2η Φx (x)2 Θ̃ 
 
 
+ 2η Φx (x)2 Θ̃  Θ0 (9.276)

λmax (P) w0
Let c1 = λmin (Q) and c2 = . Upon completing the square, we get
c1
  !  "2
  Θ0 η
V̇ e, Θ̃ ≤ −c1 (e − c2 )2 + c1 c22 − 2η Φx (x)2 Θ̃  − + Φx (x)2 Θ02
2 2
(9.277)
 
Then, V̇ e, Θ̃ ≤ 0 implies


η Φx (x)2 Θ02
e ≥ c2 + c22 + =p (9.278)
2c1

  Θ Θ02 c1 c22
  0
Θ̃  ≥ + + =α (9.279)
2 4 2η Φx (x)2
9.6 Adaptive Loop Recovery Modification 263

Note that the lower bounds p and α are dependent on Φx (x). Therefore, to
prove boundedness, we need to show that Φx (x) is bounded.  
There exists a maximum value ηmax such that η < ηmax for which V̇ e, Θ̃ ≤ 0.
Let

ϕ (x , xm  , Q, η, w0 , Θ0 ) = −c1 x2 + 2 (c1 c2 + c3 xm ) x

η
+ 2c1 c2 xm  − c1 xm 2 + Φx (x)2 Θ02
2
(9.280)

where c3 = λmax (Q) > 0.

Then, Φx (x) is bounded by
  
Φx (x) = Φx ϕ −1 (xm ∞ , Q, ν, w0 , Θ0 )  = Φx0 (9.281)

It follows that the closed-loop system is also uniformly ultimately bounded for
any 0 < η < ηmax .
Consider a special case when Φ (x) belongs to a class of regressor functions such
that Φ (x) ≤ x2 or Φx (x) ≤ 2 x. Then, it can be shown that the adaptive
loop recovery modification is stable and bounded for any 0 ≤ η < ηmax with a
conservative estimate of ηmax given by

λmin (Q)
ηmax = (9.282)
2Θ02

Consider another special case when Φ (x) belong to a class of regressor func-
tions such that Φ (x) ≤ x or Φx (x) ≤ 1. Then, the adaptive loop recovery
modification is unconditionally stable.

Since the possibility of Φx (x) being unbounded exists, the adaptive loop recovery
modification only works for any Φ (x) such that Φx (x) is bounded. For example, if
Φ (x) = x p , then the adaptive law is valid only for p ≥ 1. If p = 1, then Φx (x) = 1,
and the adaptive loop recovery modification becomes the σ modification.
Example 9.10 Consider a first-order SISO plant
 
ẋ = ax + b u + θ ∗ φ (x)

with an adaptive controller

u = k x x + kr r − θ (t) φ (x)

Suppose φ (x) = x 2 . Then, the adaptive loop recovery modification is expressed

as
264 9 Robust Adaptive Control
 
θ̇ = −γ x 2 eb + 4ηx 2 θ

Consider the fast adaptation condition when γ → ∞, then the asymptotic behav-
ior of the adaptive loop recovery modification is given by

(xm − x) bx 2
θ̄ x 2 = −
4η

(xm − x) bx 2
ū = k x x + kr r +
4η

The closed-loop plant then becomes as γ → ∞

 
(xm − x) bx 2
ẋ = am x + bm r + b + θ∗x2
4η

Let r (t) be a constant command signal. Then, the equilibrium value of x (t) can
be determined from
 
−am b2 x̄ 3 + 4ηbam θ ∗ − b2 bm r x̄ 2 + 4ηam2 x̄ + 4ηam bm r = 0

For example, let r (t) = 1, a = 1, b = 1, θ ∗ = 0.5, am = −2, bm = 2, and

η = 0.1. Then, x̄ = 1.1223 and θ̄ = 0.3058. The closed-loop response of the
adaptive loop recovery modification with k x = −3, kr = 2, and γ = 1 is shown in
Fig. 9.17. The simulation results of x̄ and θ̄ agree exactly with the analytical values.

1.5 0

−0.1
1
x, xm

−0.2
e

0.5
−0.3
x x
m
0 −0.4
0 5 10 0 5 10
t t

2 0.4

1 0.3

0 0.2
u

−1 0.1

−2 0
0 5 10 0 5 10
t t

Fig. 9.17 Closed-loop response with adaptive loop recovery modification


9.6 Adaptive Loop Recovery Modification 265

Consider a special case when the asymptotic behavior of the adaptive loop re-
covery modification results in a linear controller for a certain function φ (x). The
adaptive law is expressed in general as
 
θ̇ = −γ φ (x) eb + ηφx2 (x) θ (9.283)

For the linear asymptotic property to exist,

φx2 (x) = c2 φ 2 (x) (9.284)

where c > 0, as γ → ∞.
This results in the differential equation

dφ
= cd x (9.285)
φ

which has the following bounded solution:

φ (x) = e−cx (9.286)

Then, the adaptive loop recovery modification and the optimal control modifica-
tion have the same linear asymptotic property if ηc2 = −νb2 am−1 . Note that even
though the adaptive controller tends to a linear controller asymptotically, the closed-
loop plant is still nonlinear since φ (x) is nonlinear.
Example 9.11 Consider a time-delay first-order SISO system
 
ẋ = ax + b u (t − td ) + θ ∗ e−x

The adaptive controller with the adaptive loop recovery modification is given by

u = k x x + kr r − θ (t) e−x
 
θ̇ = −γ e−x eb + ηe−2x θ

where a + bk x = am and bkr = bm .

Since φ (x) = e−x , the adaptive loop recovery modification is the same as the
optimal control modification. The linear asymptotic adaptive controller is obtained
eb
by noticing that θ (t) φ (x) → − in the limit as γ → ∞ so that
η

eb
u → k x x + kr r +
η

Even though the adaptive controller is linear in the limit, the closed-loop plant is
nonlinear due to the nonlinear uncertainty since
266 9 Robust Adaptive Control
! " 
b b
ẋ = ax + b kx − x (t − td ) + kr r (t − td ) + xm (t − td ) + θ ∗ e−x
η η

Let a = 1, b = 1, θ ∗ = 2, td = 0.1, am = −1, bm = 1, r (t) = 1, η = 0.1, and

γ = 10. The response is shown in Fig. 9.18.

1.5 0.2

0
1
m
x, x

−0.2

e
0.5
−0.4
x xm
0 −0.6
0 5 10 0 5 10
t t

1 3

0 2
u

−1 1

−2 0
0 5 10 0 5 10
t t

Fig. 9.18 Closed-loop response with adaptive loop recovery modification with td = 0.1 s

It is instructive at this point to summarize the four robust modification schemes

for model-reference adaptive control that add damping mechanisms to the adap-
tation process, namely the σ modification, the e modification, the optimal control
modification, and the adaptive loop recovery modification. Table 9.1 shows the four
modification adaptive laws. Each of these modification adaptive laws has its own
relative merit to one another, but all in general will improve robustness of model-
reference adaptive control in their own way by providing a damping mechanism to
bound adaptive parameters from the possibility of a parameter drift for systems with
exogenous disturbances, non-minimum phase behaviors, time delay, and unmodeled
dynamics.

Table 9.1 Four robust modification adaptive laws

 
σ modification Θ̇ = −Γ Φ (x) e P B + σ Θ
   
e modification Θ̇ = −Γ Φ (x) e P B + μ e P B  Θ
 
Optimal control modification Θ̇ = −Γ Φ (x) e P B − νΦ (x) Φ  (x) Θ B  P A−1
m B
 
Adaptive loop recovery modifica- Θ̇ = −Γ Φ (x) e P B + ηΦx (x) Φx (x) Θ
tion
9.7 L1 Adaptive Control 267

9.7 L1 Adaptive Control

The L1 adaptive control has gained a considerable attention in recent years due to
its ability to achieve robustness with fast adaptation for a given a priori bound on
the uncertainty. The L1 adaptive control was developed by Hovakimyan and Cao in
2006 [8, 23, 24]. Since then, it has been widely used in adaptive control practice.
The underlying principle of the L1 adaptive control is the use of fast adaptation for
improved transient or tracking performance coupled with a low-pass filter to suppress
high-frequency responses for improved robustness. As a result, the L1 adaptive
control can be designed to achieve stability margins under fast adaptation for a given
a priori bound on the uncertainty. Theoretical proofs of guaranteed robustness bounds,
stability margins, and tracking performance in transient and steady-state responses
for the L1 adaptive control are well-established [8].
Consider the following linear SISO uncertain system with a matched uncertainty:
 
ẋ = Am x + B u + Θ ∗ (t) (t) x + d (t) (9.287)

y = Hx (9.288)

subject to x (0) = x0 , where x (t) ∈ Rn is a state vector, u (t) ∈ R is a control input,

y (t) ∈ R is a plant output, Am ∈ Rn × Rn is a known Hurwitz matrix, B ∈ Rn is
a known column vector, H ∈ Rn is a known row vector, Θ ∗ (t) ∈ Rn is unknown
time-varying state uncertainty, and d (t) ∈ R is a matched bounded time-varying
disturbance.
We assume that Θ ∗ (t) ∈ SΘ ∗ and d (t) ∈ Sd where SΘ ∗ and Sd are known 
compact sets defined a priori. Since Θ ∗ (t) is time-varying, we also require Θ̇ ∗  ≤
Δ. Further, we assume
 that the disturbance and its derivatives are bounded with
|d| ≤ d0 and ḋ  ≤ δ. Without loss of generality, the following known constraint
sets can be used to bound the adaptive parameters:
⎧ ⎡ θ ∗ +θ ∗
2  θ ∗ −θ ∗ 2 ⎤ ⎫
⎪
⎪ θ1∗ − 1max 2 1min − 1max 2 1min ⎪
⎪
⎪
⎨ ⎢ ⎥ ⎪
⎬
∗
 ∗
 ⎢ . ⎥
SΘ ∗ = Θ (t) ∈ R : g Θ = ⎢
n .
. ⎥≤0
⎪
⎪ ⎣   θ ∗ −θ ∗ 2 ⎦ ⎪
⎪
⎪
⎩ ∗ θn∗max +θn∗min 2 ⎪
⎭
θn − 2
− n max
2
n min

(9.289)
 ! " ! " 4
dmax + dmin dmax − dmin 2
Sd = d (t) ∈ R : gd (d) = d − − ≤0
2 2
(9.290)
A state-predictor model is defined as

x̂˙ = Am x̂ + B u + Θ  (t) x + d̂ (t) (9.291)
268 9 Robust Adaptive Control

with the initial condition x̂ (0) = x0 , where x̂ (t), Θ (t) and d̂ (t) are estimates of
x (t), Θ ∗ (t), and d (t), respectively.
We define the state-predictor error as e p (t) = x̂ (t) − x (t). Then, the estimates of
Θ ∗ (t), and d (t) are computed from the following adaptive laws with the projection
method:
⎧
⎪ 
  ⎨−Γ xe p P B if g (Θ) < 0 or if 
g (Θ) = 0
 
Θ̇ = Pro Θ, −Γ xe p P B = and − xe p P B ∇gΘ (Θ) ≤ 0
⎪
⎩
0 otherwise
⎧     (9.292)
⎪ 
  ⎪ ⎨−γd e p P B if gd d̂ < 0 or ifgd d̂ = 0
˙ 
d̂ = Pro d̂, −γd e p P B =
⎪ and − ep P B∇gd d̂ ≤ 0
⎪
⎩
0 otherwise
(9.293)
with the initial conditions Θ (0) = Θ0 and d̂ (0) = d̂0 .
The adaptive laws can be shown to be stable as follows:

Proof The predictor error equation is established as

 
ė p = x̂˙ − ẋ = Am e p + B Θ̃  x + d̃ (9.294)

where Θ̃ (t) = Θ (t) − Θ ∗ (t) and d̃ (t) = d̂ (t) − d (t).

Choose a Lyapunov candidate function

  d̃ 2
V e p , Θ̃, d̃ = ep Pe p + Θ̃  Γ −1 Θ̃ + (9.295)
γd

The projection
 method  guarantees that Θ (t) and d̂ (t) stay inside their constraint
sets. Then, V̇ e p , Θ̃, d̃ is evaluated as

  2d̃ ḋ
V̇ e p , Θ̃, d̃ = −ep Qe p − 2Θ̃  Γ −1 Θ̇ ∗ −
γd
 
 
 2  −1  



2 d̃  δ
≤ −λmin (Q) e p  + 2λmax Γ Θ̃  Δ + (9.296)
γd

Θ̃ (t) and d̃ (t) are constrained by

θimin − θi∗ ≤ θ̃i ≤ θimax − θi∗ (9.297)

dmin − d ≤ d̃ ≤ dmax − d (9.298)

9.7 L1 Adaptive Control 269

This implies
           
 
θ̃i  ≤ max θimin − θi∗  , θimax − θi∗  ≤ max θimin  , θimax  + max θi∗  (9.299)
 
 
d̃  ≤ max (|dmin − d| , |dmax − d|) ≤ max (|dmin | , |dmax |) + max |d| = d̃0
       (9.300)
Let Θ̃0 = max max θimin  , θimax  + max θi∗  . Then,
   2  
V̇ e p , Θ̃, d̃ ≤ −λmin (Q) e p  + 2λmax Γ −1 Θ̃0 Δ + 2γd−1 d̃0 δ (9.301)
 
V̇ e p , Θ̃, d̃ ≤ 0 implies

 
  2λmax Γ −1 Θ̃0 Δ + 2γd−1 d̃0 δ
e p  ≥ =p (9.302)
λmin (Q)
   
 
There exists an upper bound of e p  along with the upper bounds of Θ̃  and
   
 
d̃  that establish a compact set outside which V̇ e p , Θ̃, d̃ ≤ 0. Therefore, all
signals are bounded.

The L1 adaptive controller is defined as

u (s) = −k D (s) w (s) (9.303)

where k > 0 is a feedback gain, D (s) is a proper transfer function and w (s) is the
Laplace transform of

w (t) = u (t) + Θ  (t) x + d̂ (t) − kr r (t) (9.304)

where the command feedforward gain kr is given by

1
kr = − (9.305)
H A−1
m B

The ideal controller is given by

u ∗ (t) = −Θ ∗ (t) x (t) − d (t) + kr r (t) (9.306)

The ideal closed-loop plant then becomes

ẋ = Am x + Bkr r (9.307)
270 9 Robust Adaptive Control

The L1 reference controller is then given by the filtered version of the ideal
controller
u m (s) = C (s) u ∗ (s) (9.308)

where C (s) is a low-pass filter which can be obtained as

k D (s)
C (s) = (9.309)
1 + k D (s)

We require C (0) = 1. Thus, D (s) can be any suitable transfer function that
1
satisfies lims→0 D(s) = 0.
The design of the L1 adaptive control then essentially comes down to the design
of the feedback gain k and the transfer function D (s) which defines the low-pass
filter C (s).
The closed-loop reference system formed by the L1 reference controller is then
obtained as
 
sx (s) = Am x (s) + B C (s) u ∗ (s) + kr r (s) − u ∗ (s) (9.310)

Let the largest bound of Θ ∗ be defined using the L1 norm definition as

  
n
 ∗
L = Θ ∗ 1 = max θ 
i (9.311)
i=1

Then, for the closed-loop reference system to be stable, the L1 norm of the system
transfer function G (s) must satisfy the following condition:
 
G (s) = (s I − Am )−1 B [1 − C (s)]1 L < 1 (9.312)

This condition is the stability condition on the unit circle for discrete-time which
is equivalent to the Hurwitz stability condition of the matrix Am + B [1 − C (s)] Θ ∗
for continuous time.
The choices of k and D (s) can greatly influence the performance and stability of
the L1 adaptive control. For example, consider the following transfer function:

1
D (s) = (9.313)
s
which satisfies the ideal closed-loop transfer function

k 
C (0) = =1 (9.314)
s + k s=0

Then, the L1 adaptive controller is given by

9.7 L1 Adaptive Control 271

k
u = − w (s) (9.315)
s
or, in the time domain,

u̇ = −ku − k Θ  (t) x + d̂ (t) − kr r (9.316)

Consider the case when Θ ∗ and d are constant. Then, the closed-loop reference
system becomes
      
ẋ Am + BΘ ∗ B x Bd
= + (9.317)
u̇ −kΘ ∗ −k u −k (d − kr r )

The stability condition requires

 
Am + BΘ ∗ B
Ac = (9.318)
−kΘ ∗ −k

to be Hurwitz by a suitable selection of the feedback gain k > 0.

Consider a first-order SISO plant as
 
ẋ = am x + b u + θ ∗ x + d (9.319)

where θ ∗ and d are unknown constants with given known bounds.

The predictor model is
 
x̂˙ = am x̂ + b u + θ x + d̂ (9.320)

1
Choose D (s) = . Then, the L1 adaptive controller is defined as
s
k
u = − w (s) (9.321)
s
 
θ̇ = Pro θ, −γ xe p b (9.322)
 
d̂˙ = Pro d̂, −γd e p b (9.323)

Then, the closed-loop plant matrix is

 
a + bθ ∗ b
Ac = m (9.324)
−kθ ∗ −k
272 9 Robust Adaptive Control

The closed-loop characteristic equation is then obtained as

 
 s − am − bθ ∗ −b   

det (s I − Ac ) =   = s 2 + k − am − bθ ∗ s − kam = 0
kθ ∗ s +k
(9.325)
Comparing the characteristic equation to that of a second-order system

s 2 + 2ζ ωn s + ωn2 = 0 (9.326)

then
ωn2 = −kam (9.327)

2ζ ωn = k − am − bθ ∗ (9.328)

For a specified damping ratio ζ , k can be computed from

   2
k 2 − 2k am + bθ ∗ − 2ζ 2 am + am + bθ ∗ = 0 (9.329)

which yields
7  
k = am + bθ ∗ − 2ζ 2 am + 2ζ am − (am + bθ ∗ ) + ζ 2 am (9.330)

Consider two cases:

1. The open-loop plant with uncertainty is stable with am + bθ ∗ < 0. Then, k ≥
− (am + bθ ∗ ) > 0 if 
am + bθ ∗
ζ ≥ (9.331)
am

2. The open-loop plant with uncertainty is unstable with am + bθ ∗ > 0. Then, the
closed-loop poles are computed as
 
k − am − bθ ∗ 4kam
s=− 1± 1+ (9.332)
2 (k − am − bθ ∗ )2

which lie in the left half plane if k > am + bθ ∗ .

Example 9.12 Consider the parameter drift example in Example 8.1. Let am = −1,
w (t) 5
b = 1, θ ∗ = 2, and d (t) = with p = − .
b 12
1
Choose D (s) = . Since am + bθ ∗ = 1 > 0, we choose k = 2 > 1. Then, the
s
L1 adaptive regulator controller is specified by
9.7 L1 Adaptive Control 273
 
u̇ = −2u − 2 θ x + d̂

⎧
⎪−γ xe p b if θmin < θ < θmax or if θ = θmin
  ⎨
θ̇ = Pro θ, −γ xe p b = and θ̇ ≥ 0 or if θ = θmax and θ̇ ≤ 0
⎪
⎩
0 otherwise
⎧
 ⎨−γd e p b if dmin ≤ d̂ ≤ dmax or if d̂ = dmin
⎪ 
˙
d̂ = Pro d̂, −γd e p b = and d̂˙ ≥ 0 or if d̂ = dmax and d̂˙ ≤ 0
⎪
⎩
0 otherwise

3 5
x
2 x^ 0
<
x, x

1 −5
u

0 −10

−1 −15
0 50 100 0 50 100
t t

15 4

10 3
<
d, d

5 2
θ

0 d 1
d^
−5 0
0 50 100 0 50 100
t t

Fig. 9.19 Closed-loop response with L1 adaptive control, θ → θmax

3 5
x
2 x^ 0
<
x, x

1 −5
u

0 −10

−1 −15
0 50 100 0 50 100
t t

15 6

10
4
<
d, d

5
θ

2
0 d
d^
−5 0
0 50 100 0 50 100
t t

Fig. 9.20 Closed-loop response with L1 adaptive control, θ → θ̄ < θmax

274 9 Robust Adaptive Control

We choose θmin = 0, θmax = 3, dmin = −2, and dmax = 11. To ensure fast
adaptation, γ and γd must be chosen to be large values. The time step must be chosen
sufficiently small if γ and γd are large. Choose Δt = 0.001 and γ = γd = 100. The
initial conditions are x (0) = 0, x̂ (0) = 0, θ (0) = 0, and d̂ (0) = 0. The response of
the closed-loop system with the L1 adaptive controller is shown in Fig. 9.19. Notice
that θ (t) → θmax because the upper bound of θ (t) is set too small. Nonetheless,
x (t) → ε as t → ∞ and d̂ (t) → d (t) + δ.
Suppose the upper bound is increased to θmax = 10. Then, the response with the
new upper bound of θ (t) is shown in Fig. 9.20. The response shows that θ (t) →
θ̄ = 5.1011 < θmax . Note that θ (t) does not tend to θ ∗ .

9.8 Normalization

Fast adaptation can lead to loss of robustness of MRAC. Fast adaptation is usually
associated with the use of a large adaptation rate. However, this is not the only effect
of fast adaptation. When the amplitude of the regressor function Φ (x) is large, the
effect is equivalent. Normalization is a technique that can be used to achieve more
robust adaptation for large amplitude inputs [5]. The objective of the normalization
is to reduce the effect of fast adaptation based on the amplitude of the regressor
function. Normalized adaptation can achieve a significant increase in the time-delay
margin of a closed-loop adaptive system. The normalized MRAC is given by

Γ Φ (x) e P B
Θ̇ = − (9.333)
1 + Φ  (x) RΦ (x)

where R = R  > 0 is a positive-definite weighting matrix.

To prevent the parameter drift, the projection method can be used to bound adaptive
parameters. Thus, the normalized MRAC with the projection method is described by
! "
Γ Φ (x) e P B
Θ̇ = Pro Θ, − (9.334)
1 + Φ  (x) RΦ (x)
 
 
where Θ (t) is constrained to stay within a constraint set such that Θ̃  ≤ Θ̃0 .
The quantity 1+Φ  (x) RΦ (x) is called the normalization factor. If the amplitude
of the regressor function Φ (x) is large when the closed-loop system exhibits poor
robustness, the normalization effectively reduces the adaptation process to improve
robustness. The stability proof of the normalized optimal control modification is
provided as follows:

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.335)
9.8 Normalization 275

There are two cases to consider.

 

1. Consider the case when g (Θ) < 0 or g (Θ) = 0 and − Φ (x) e P B ∇gΘ
(Θ) ≤ 0. Then, evaluating V̇ e, Θ̃ yields

  2e P B Θ̃  Φ (x) Φ  (x) RΦ (x)

V̇ e, Θ̃ = −e Qe − 2e Pw +
1 + Φ  (x) RΦ (x)
≤ −λmin (Q) e + 2 e λmax (P) w0
2

2e P B Θ̃  Φ (x) λmax (R) Φ (x)2

+ (9.336)
m 2 (x)

where m 2 (x) = 1 + λmin (R) Φ (x)2 .

Suppose g (Θ) = (Θ − Θ ∗ ) (Θ − Θ ∗ ) − β 2 ≤ 0. Then, ∇gΘ (Θ) = 2Θ̃. So,
 
− Φ (x) e P B ∇gΘ (Θ) = −2e P B Θ̃  Φ (x) = −c0 ≤ 0 (9.337)

where c0 > 0.
Therefore,
  c0 λmax (R) Φ (x)2
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2 e λmax (P) w0 +
m 2 (x)
  (9.338)
Note that Φ(x)
2

m 2 (x)
≤ 1
λmin (R)
. Then, it follows that V̇ e, Θ̃ ≤ 0 if

c0 λmax (R)
− c1 e2 + 2c1 c2 e + ≤0 (9.339)
λmin (R)
or 
c0 λmax (R)
e ≥ c2 + c2 + =p (9.340)
c1 λmin (R)

where c1 = λmin (Q) and c2 = λmax (P)w0

.
  λmin (Q)  
   
Since g (Θ) ≤ 0, then Θ̃  has an upper bound where Θ̃  ≤ β. Then, the
closed-loop system is uniformly ultimately bounded with the following ultimate
bound:   
λmax (P) p 2 + λmax Γ −1 α 2
p ≤ e ≤ =ρ (9.341)
λmin (P)
  
  λmax (P) p 2 + λmax Γ −1 α 2
 
α ≤ Θ̃  ≤ =β (9.342)
λmin (P)
276 9 Robust Adaptive Control
 
2. Consider the case when g (Θ) ≥ 0 and − Φ (x) e P B ∇gΘ (Θ) > 0. Then,
 
V̇ e, Θ̃ = −e Qe − 2e Pw
∇gΘ (Θ) ∇  gΘ (Θ) Φ (x) e P BΦ  (x) RΦ (x)
+ 2Θ̃ 
∇  gΘ (Θ) ∇gΘ (Θ) 1 + Φ  (x) RΦ (x)
(9.343)

For the same g (Θ), 2e P B Θ̃  Φ (x) = −c0 < 0 where c0 > 0. Then,

∇gΘ (Θ) ∇  gΘ Θ̃ Θ̃ 
= (9.344)
∇  gΘ (Θ) ∇gΘ (Θ) Θ̃  Θ̃

Therefore,
  c0 λmin (R) Φ (x)2
V̇ e, Θ̃ ≤ −λmin (Q) e2 + 2 e λmax (P) w0 −
n 2 (x)
≤ −λmin (Q) e2 + 2 e λmax (P) w0 (9.345)

where n 2(x)= 1 + λmax (R) Φ (x)2 .

Thus, V̇ e, Θ̃ ≤ 0 if
2λmax (P) w0
e ≥ =p (9.346)
λmin (Q)
   
   
Since g (Θ) ≥ 0, then Θ̃  is bounded from below such that Θ̃  ≥ α. Therefore,
the closed-loop system is uniformly ultimately bounded.

Note that the normalization technique will result in poorer tracking as λmax (R)
increases, but at the same time improve robustness in the presence of large amplitude
inputs or fast adaptation.
For the normalized MRAC without the projection method, if a disturbance does
not exist, the tracking error is asymptotic, but in the presence of a disturbance, the
Lyapunov stability proof does not guarantee boundedness of the adaptive parameters.
Therefore, the normalized MRAC without the projection method is not robust to the
parameter drift.
In lieu of the projection method, the normalization technique can be used in
conjunction with any one of the robust modification schemes such as the σ modifica-
tion, e modification, and optimal control modification. For example, the normalized
σ modification is given by
 
Γ Φ (x) e P B + σ Θ
Θ̇ = − (9.347)
1 + Φ  (x) RΦ (x)
9.8 Normalization 277

Since there is no projection, R and σ have to be chosen such that the closed-loop
system is stable. This can be determined from the Lyapunov stability analysis as
follows:

Proof Choose the usual Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.348)

  Φ (x)2
Then, evaluating V̇ e, Θ̃ and recognizing that 2 ≤ 1
λmin (R)
yield
m (x)
 
V̇ e, Θ̃ = −e Qe + 2e P B Θ̃  Φ (x) − 2e Pw
5  6
2Θ̃  Φ (x) e P B + σ Θ
− trace
1 + Φ  (x) RΦ (x)
≤ −λmin (Q) e2 + 2 e λmax (P) w0
   2  
     
2λmax (R) e P B Θ̃  Φ (x) 2σ Θ̃  2σ Θ̃  Θ0
+ − 2 +
λmin (R) n (x) m 2 (x)
(9.349)
 
Utilizing the inequality 2 a b ≤ a2 + b2 , V̇ e, Θ̃ is bounded by
 2
 
  λmax (R) P B Θ̃  Φ (x)2
V̇ e, Θ̃ ≤ −c1 (e − c2 )2 + c1 c22 +
λmin (R)
  2
 
− c3 Θ̃  − c4 + c3 c42 (9.350)

λmax (R) P B λmax (P) w0 2σ

where c1 = λmin (Q) − , c2 = , c3 = 2 , and
λmin (R) c1 n (x)
n 2 (x) Θ0
c4 = .
2m (x)
2
  
  n 2 (x)
V̇ e, Θ̃ can be made larger by setting Θ̃  = c4 . Recognizing that m 2 (x)
≤
λmax (R) n (x)
2

λmin (R)
and m 4 (x)
≤ 1, then

  λ3 (R) P B Θ02 Φ (x)2 σ Θ02

V̇ e, Θ̃ ≤ −c1 (e − c2 )2 + c1 c22 + max +
4λmin (R)
3 2
(9.351)
Therefore, R and σ are chosen such that the following inequality is satisfied:

ϕ (x , xm  , Q, R, w0 , Θ0 ) = −c1 x2 + 2 (c1 c2 + c5 xm ) x

+ 2c1 c2 xm  − c1 xm 2 + c6 Φ (x)2 + c7 ≤ 0
(9.352)
278 9 Robust Adaptive Control

λmax (R) P B λ3 (R) P B Θ02 σ Θ02

where c5 = λmax (Q) − , c6 = max 3 , and c7 = .
λmin (R) 4λmin (R) 2

The use of the normalization in conjunction with a robust modification scheme can
significantly improve robustness to the parameter drift, time delay, and unmodeled
dynamics [16, 25].
Consider a first-order SISO system with an adaptive controller using the normal-
ized MRAC without the projection method

ẋ = ax + bu (9.353)

u = k x (t) x + kr r (9.354)

γx x (xm − x) b
k̇ x = (9.355)
1 + Rx 2

where r (t) is a constant reference command signal.

Differentiating the closed-loop plant gives

γx x 2 b γx x 2 x m b
ẍ − [a + bk x (t)] ẋ + b x =b (9.356)
1 + Rx 2 1 + Rx 2

The effect of the normalized MRAC is to reduce the nonlinear integral control
action of MRAC. The effective nonlinear integral gain of the normalized MRAC is
given by
γx x 2 b
ki (x) = (9.357)
1 + Rx 2

Then, the closed-loop plant is expressed as

ẍ − [a + bk x (t)] ẋ + bki (x) x = bki (x) xm (9.358)

Note that ki (x) is a bounded function where

γx b
0 < ki (x) ≤ (9.359)
R
Thus, the maximum closed-loop frequency is independent of x (t). So, any large
amplitude input will not affect the nonlinear integral control action. Nonetheless,
ki (x) can still tend to a large value with γx .
Example 9.13 Consider the following first-order SISO system:
 
ẋ = ax + b u + θ ∗ x 2

where a = 1 and b = 1 are known, and θ ∗ = 2 is unknown.

9.8 Normalization 279

An adaptive controller is designed as

u = k x x + kr r − θ (t) x 2

using the normalized MRAC without the projection method

γ x 2 eb
θ̇ = −
1 + Rx 4

The reference model is given by

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = 1.

Choose γ = 1000, R = 100, and x (0) = 1. The closed-loop responses with
the standard MRAC and normalized MRAC are shown in Fig. 9.21. The response
with the standard MRAC exhibits high-frequency oscillations due to the use a large
adaptation rate. On the other hand, the normalized MRAC effectively eliminates
the effect of fast adaptation. Both the tracking error and parameter estimation error
converge to zero since there is no disturbance.

1.5 1.5
x x x x
m m

1 1
x, xm

x, xm

0.5 0.5
γ=1000 R=0 γ=1000 R=100
0 0
0 5 10 15 20 0 5 10 15 20
t t

40 6

20 4
θ

0 2
γ=1000 R=0 γ=1000 R=100
−20 0
0 5 10 15 20 0 5 10 15 20
t t

Fig. 9.21 Closed-loop responses with standard MRAC and normalized MRAC

Example 9.14 Consider Example 8.1

ẋ = ax + bu + w

where
 
(1 + t)2 p+1 − 1
w = p (1 + t) p−1
− a (1 + t) + b γx bp
− k x (0) (1 + t) p
2p + 1
280 9 Robust Adaptive Control

and an adaptive regulator controller

u = k x (t) x

where k x (t) is computed by the normalized MRAC without the projection method

γx x 2 b
k̇ x = −
1 + Rx 2

Let a = 1, b = 1, n = − 12 5
, γx = 10, and R = 1. The closed-loop response
with the normalized MRAC is shown in Fig. 9.22. Notice that the response exhibits a
parameter drift as k x (t) → −∞. Thus, the normalized MRAC without the projection
method is not robust to the parameter drift.

1.5

1
x

0.5

0
0 10 20 30 40 50
t

−20
x
k

−40

−60
0 10 20 30 40 50
t

Fig. 9.22 Parameter drift with normalized MRAC

9.9 Covariance Adjustment of Adaptation Rate

Fast adaptation is needed to reduce the tracking error rapidly when a large uncertainty
is present in a system. However, in most situations, when the adaptation process
has achieved sufficiently a desired level of tracking performance, fast adaptation is
usually no longer needed. Maintaining fast adaptation even when after the adaptation
has achieved its objective can result in persistent learning. At best, persistent learning
would do nothing to further improve the tracking performance once the adaptation has
achieved its objective. At worst, persistent learning reduces robustness of an adaptive
controller which is not highly desired. Therefore, the adaptation rate can be reduced
by a covariance adjustment method. The adjustment allows for an arbitrarily large
initial adaptation rate to be used to enable fast adaptation. The covariance adjustment
then adjusts the adaptation rate toward a lower value to achieve improved robustness
9.9 Covariance Adjustment of Adaptation Rate 281

as the adaptation proceeds. By reducing the adaptation rate, improved robustness can
be achieved with the covariance adjustment, while the tracking performance during
the initial adaptation is retained [5, 16, 25].
The standard MRAC with the covariance adjustment of the adaptation rate is
described by
Θ̇ = −Γ (t) Φ (x) e P B (9.360)

Γ˙ = −ηΓ Φ (x) Φ  (x) Γ (9.361)

where η > 0 is an adjustment parameter.

The adjustment method is the same as the update law for the covariance matrix
R (t) in the recursive least-squares method.
Note that there are other variances of the covariance adjustment method. The
covariance adjustment with a forgetting factor is given by

Γ˙ = βΓ − ηΓ Φ (x) Φ  (x) Γ (9.362)

where β > 0 is a forgetting factor.

The covariance adjustment with the normalization is given by

βΓ − ηΓ Φ (x) Φ  (x) Γ
Γ˙ = (9.363)
1 + Φ  (x) RΦ (x)

The projection method can be used to bound the adaptive parameter. Thus, the
MRAC with the projection method and covariance adjustment of the adaptation rate
is described by  
Θ̇ = Pro Θ, −Γ (t) Φ (x) e P B (9.364)

The stability proof of the covariance adjustment of the adaptation rate is as follows:

Proof Choose a Lyapunov candidate function

   
V e, Θ̃ = e Pe + trace Θ̃  Γ −1 Θ̃ (9.365)

 

 when g (Θ) < 0 or g (Θ) = 0 and − Φ (x) e P B ∇gΘ
Consider the case
(Θ) ≤ 0. V̇ e, Θ̃ is evaluated as

  ! −1 "
   dΓ
V̇ e, Θ̃ = −e Qe − 2e Pw + trace Θ̃ Θ̃ (9.366)
dt

But,
Γ Γ −1 = I (9.367)
282 9 Robust Adaptive Control

Then,
dΓ −1
Γ˙ Γ −1 + Γ =0 (9.368)
dt
So,
dΓ −1
= −Γ −1 Γ˙ Γ −1 = ηΦ (x) Φ  (x) (9.369)
dt
Therefore,
 
V̇ e, Θ̃ = −e Qe − 2e Pw + ηΦ  (x) Θ̃ Θ̃  Φ (x) ≤ −λmin (Q) e2
 2
 
+ 2 e λmax (P) w0 + η Θ̃  Φ (x)2 (9.370)
 
 
Since Θ̃  ≤ β where β is the a priori bound from the projection method, then
 
the parameter η is chosen to ensure that V̇ e, Θ̃ ≤ 0 or equivalently the following
inequality is satisfied:
  
 
ϕ x , xm  , Q, η, w0 , Θ̃  = −c1 x2 + 2 (c1 c2 + c5 xm ) x
+ 2c1 c2 xm  − c1 xm 2 + ηβ 2 Φ (x)2 ≤ 0
(9.371)


Note that if a disturbance does not exist, then the tracking error converges as-
ymptotically to zero, but the adaptive parameter is bounded. In the presence of a
disturbance, a parameter drift can still exist even though the adaptation rate is con-
tinuously adjusted toward zero. In theory, the adaptive parameter would stop drift-
ing when Γ (t) → 0 as t → ∞. In practice, the adaptive parameter can reach a
sufficiently large value that can cause robustness issues. Therefore, the covariance
adjustment method needs to be used in conjunction with the projection method or
any robust modification schemes to prevent a parameter drift.
When used in conjunction with a robust modification scheme, the covariance
adjustment method can further improve robustness. For example, the optimal control
modification with the covariance adjustment is described by [25]
 
Θ̇ = −Γ (t) Φ (x) e P − νΦ  (x) Θ B  P A−1
m B (9.372)

Γ˙ = −ηΓ Φ (x) Φ  (x) Γ (9.373)

 
Then, V̇ e, Θ̃ is given by

    2
 
V̇ e, Θ̃ ≤ −c1 (e − c2 )2 +c1 c22 −νc3 Φ (x)2 Θ̃  − c4 +νc3 c42 Φ (x)2
(9.374)
9.9 Covariance Adjustment of Adaptation Rate 283

  η

where c1 , c2 , c4 are defined in Sect. 9.5, c3 = λmin B  A− −1
m Q Am B 1 − ηmax ,
  − −1

and 0 ≤ η < ηmaxwith ηmax = νλmin B Am Q Am B .
Thus, V̇ e, Θ̃ ≤ 0 outside a compact set. Therefore, the closed-loop adaptive
system is completely bounded.
Consider a first-order SISO system with an adaptive controller with the covariance
adjustment of the adaptation rate

ẋ = ax + bu (9.375)

u = k x (t) x + kr r (9.376)

k̇ x = γx xeb (9.377)

γ̇x = −ηγx2 x 2 (9.378)

The adaptation rate can be integrated as

, t
dγx 1 1
− = ηx 2 dt ⇒ − =η x 2 (τ ) dτ (9.379)
γx2 γx γx (0) 0

This yields
γx (0)
γx = :t (9.380)
1 + γx (0) η 0 x 2 (τ ) dτ

Then, the adaptive law becomes

γx (0) xeb
k̇ x = :t (9.381)
1 + γx (0) η 0 x 2 (τ ) dτ

which is effectively a normalized adaptive law with an integral square normalization

factor.
Let r (t) be a constant reference command signal. Differentiating the closed-loop
plant gives
ẍ − [a + bk x (t)] ẋ + bki (x) x = bki (x) xm (9.382)

where ki (x) is the nonlinear integral gain

γx (0) x 2 b
ki (x) = :t (9.383)
1 + γx (0) η 0 x 2 (τ ) dτ
:t
As t → ∞, ki (x) → 0 because 0 x 2 (τ ) dτ → ∞ since r (t) is a constant
command signal for which x (t) ∈
/ L2 . Therefore, the effect of fast adaptation is
entirely eliminated.
284 9 Robust Adaptive Control

The effect of the covariance adjustment of the adaptation rate is to gradually shut
down the adaptation process. A problem associated with the covariance adjustment is
that, once the adaptation rate is reduced to a value near zero, the adaptation process has
to be restarted manually if a new uncertainty becomes present, whereas the adaptation
rate with any other adaptive control methods will be finite and the adaptation process
can be made to always remain active. It is possible to design a resetting adaptation
process using the covariance adjustment. The adaptation process can be reset with
a new initial value of the adaptation rate whenever a threshold criterion is satisfied.
For example, a covariance adjustment resetting algorithm could be expressed as

Γ˙ with Γ (te ) = Γe t ≥ te when e (t) > e0 at t = te
Γ˙ = (9.384)
Γ˙ with Γ (0) = Γ0 otherwise

The threshold should be chosen judiciously so that the trigger would occur appro-
priately to prevent false triggering. Also, when a switching action occurs with a large
change in the value of the initial condition of Γ (t), to prevent transient behaviors,
a filter can be used to phase in the new initial condition of Γ (t). For example, if a
first-order filter is used, then
Γ˙ = −λ (Γ − Γe ) (9.385)

for t ∈ [te , te + Δt], where λ > 0, Γ (te ) = Γ (t) at t = te is computed from the
previous covariance adjustment, and Γe is the new reset initial condition of Γ (t).

Example 9.15 Consider the same first-order SISO system in Example 9.13
 
ẋ = ax + b u + θ ∗ x 2

The MRAC adaptive law with the covariance adjustment without the projection
method is given by
θ̇ = −γ x 2 eb

γ̇ = −ηγ 2 x 4

Let γ (0) = 1000 and η = 0.1. The response of the closed-loop system is shown
in Fig. 9.23. Comparing to Fig. 9.21, the response with the covariance adjustment is
free of high-frequency oscillations. Both the tracking error and adaptive parameter
estimation error tend to zero asymptotically. The value of the adaptation rate at
t = 20 s is γ (t) = 0.6146 which is substantially less than the initial value of
γ (0) = 1000.
9.9 Covariance Adjustment of Adaptation Rate 285

2 1

1.5
0
1

e
x
−1
0.5

0 −2
0 10 20 0 10 20
t t
4
20 10

2
10 10
θ

γ
0
0 10

−2
−10 10
0 10 20 0 10 20
t t

Fig. 9.23 Closed-loop response with covariance adjustment of adaptation rate

Example 9.16 Consider the same parameter drift example in Example 9.14 with
the covariance adjustment without the projection method

k̇ x = −γx x 2 b

γ̇x = −ηγx2 x 2

2.5

1.5
x

0.5
0 100 200 300 400 500
t

0
x

−5
k

−10
0 100 200 300 400 500
t

Fig. 9.24 Parameter drift with MRAC with covariance adjustment

Let γx (0) = 10 and η = 1. The response is shown in Fig. 9.24. Note that even
though the value of k x (t) appears to reach a steady-state value, in fact, k x (t) is just
drifting very slowly because the adaptation rate is very small and never reaches zero
in a finite time.
286 9 Robust Adaptive Control

9.10 Optimal Control Modification for Systems with

Control Input Uncertainty

In certain situations, the control effectiveness of a control system may be impaired

due to failures. When an uncertainty exists in the control input, the system can
undergo significant changes in its closed-loop characteristics that can compromise
stability and performance of the control system. The control signal must be modified
accordingly to produce achievable dynamics in the presence of the reduced control
effectiveness.
Consider a linear plant with a control input uncertainty and a matched uncertainty
 

ẋ = Ax + BΛ u + Ω ∗ x (9.386)

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, A ∈ Rn × Rn

and B ∈ Rn × Rm are constant and known matrices such that the pair (A, B) is
controllable, Λ ∈ Rm × Rm is an unknown control effectiveness diagonal matrix,
and Ω ∗ ∈ Rn×m is an unknown constant matrix.
Let ΔA = BΛΩ ∗ and ΔB = B (Λ − I ). Then, the plant can also be expressed
as
ẋ = (A + ΔA) x + (B + ΔB) u (9.387)

A nominal fixed gain controller is designed to stabilize the plant with no uncer-
tainty and enable it to track a reference command r (t)

ū = K x x + K r r (9.388)

where r (t) ∈ Rq ∈ L∞ is a piecewise continuous and bounded reference command

vector.
The closed-loop nominal plant without uncertainty is obtained as

ẋ = Am x + Bm r (9.389)

where Am = A + B K x ∈ Rn × Rn is Hurwitz and Bm = B K r ∈ Rn × Rq is a matrix

with q ≤ n.
This closed-loop nominal plant is then used to specify a reference model

ẋm = Am xm + Bm r (9.390)

where xm (t) ∈ Rn is a reference state vector.

The objective is to design a full-state feedback adaptive augmentation controller
to enable x (t) to follow xm (t) in the presence of uncertainty due to ΔA and ΔB
with the following controller:

u = ū + ΔK x x + ΔK r r − Ω  x (9.391)
9.10 Optimal Control Modification for Systems with Control Input Uncertainty 287

For convenience. u (t) can be expressed as

u = ū − Θ  Φ (x, r ) (9.392)
 
where Θ  (t) = −ΔK x (t) + Ω  (t) −ΔK r (t) ∈ Rm × Rn+q and Φ (x, r ) =
   
x r ∈ Rn+q .
It is assumed that there exist constant and unknown matrices ΔK x∗ and ΔK r∗ such
that the following model matching conditions are satisfied:
 
A + BΛ K x + ΔK x∗ = Am (9.393)
 
BΛ K r + ΔK r∗ = Bm (9.394)

Let Λ̃ (t) = Λ̂ (t) − Λ, Δ K̃ x (t) = ΔK x (t) − ΔK x∗ , Δ K̃ r (t) = ΔK r (t) − ΔK r∗ ,

Ω̃ (t) = Ω (t) − Ω ∗ , and Θ̃ (t) = Θ (t) − Θ ∗ be the estimation errors. Then, the
closed-loop plant becomes

ẋ = Am x + Bm r − B Λ̂Θ̃  Φ (x, r ) + B Λ̃Θ̃  Φ (x, r ) (9.395)

Defining the tracking error as e (t) = xm (t) − x (t), then the tracking error
equation becomes
ė = Am e + B Λ̂Θ̃  Φ (x, r ) + Bε (9.396)

where
ε = −Λ̃Θ̃  Φ (x, r ) (9.397)

If the sign of Λ is known, then the standard MRAC law for adjusting Θ (t) is
given by
Θ̇ = −ΓΘ Φ (x, r ) e P BsgnΛ (9.398)

This MRAC law, while providing asymptotic tracking, does not provide robustness
to potential unmodeled dynamics. The optimal control modification adaptive law can
be used to provide robust adaptation as [21]

Θ̇ = −ΓΘ Φ (x, r ) e P − νΦ  (x, r ) Θ Λ̂ B  P A−1
m B Λ̂ (9.399)

It is noted that the adaptive law depends on the estimate of Λ which needs to be
computed. Toward this end, a predictor model of the plant is defined as
 
x̂˙ = Am x̂ + (A − Am ) x + B Λ̂ u + Ω  x (9.400)

Define the predictor error as e p (t) = x̂ (t) − x (t), then

288 9 Robust Adaptive Control
 
ė p = Am e p + B Λ̂Ω̃  x + B Λ̃ u + Ω  x + Bε p (9.401)

where
ε p = −Λ̃Ω̃  x (9.402)

Then, Λ can be estimated by the following adaptive laws:

 
Ω̇ = −ΓΩ x ep P − νx  Ω Λ̂ B  P A−1
m B Λ̂ (9.403)

   
Λ̂˙  = −ΓΛ u + Ω  x ep P − ν u  + x  Ω Λ̂ B  P A−1
m B (9.404)

The Lyapunov stability proof is provided as follows:

Proof Choose a Lyapunov candidate function

   
V e, e p Θ̃, Ω̃, Λ̃ = e Pe + ep Pe p + trace Θ̃  ΓΘ−1 Θ̃
   
+ trace Ω̃  ΓΩ−1 Ω̃ + trace Λ̃ΓΛ−1 Λ̃ (9.405)
 
Evaluating V̇ e, e p , Θ̃, Ω̃, Λ̃ yields
 
V̇ e, e p Θ̃, Ω̃, Λ̃ = −e Qe − ep Qe p + 2e P Bε + 2ep P Bε p
 
+ 2νtrace Θ̃  Φ (x, r ) Φ  (x, r ) Θ Λ̂ B  P A−1
m B Λ̂
 
+ 2νtrace Ω̃  x x  Ω Λ̂ B  P A−1
m B Λ̂
    
+ 2νtrace Λ̃ u + Ω  x u  + x  Ω Λ̂ B  P A−1 m B
(9.406)
⎡ ⎤
  Θ (t) Λ̂ (t) 0 0
Let B̄ = B B B ∈ Rn ×R3m , Π (t) = ⎣ 0 Ω (t) Λ̂ (t) 0 ⎦ ∈
0 0 Λ̂ (t)
 
R2n+m+q × R3m , Ψ (x, r ) = Φ  (x, r ) x  u  + x  Ω ∈ R2n+m+q , then
 
trace Π̃  Ψ (x, r ) Ψ  (x, r ) Π B̄  P A1m B̄
 
= trace Θ̃  Φ (x, r ) Φ  (x, r ) Θ Λ̂ B  P A−1
m B Λ̂
 
+ trace Ω̃  x x  Ω Λ̂ B  P A−1 m B Λ̂
     
+ trace Λ̃ u + Ω x u + x  Ω Λ̂ B  P A−1

m B (9.407)
9.10 Optimal Control Modification for Systems with Control Input Uncertainty 289
 
V̇ e, e p , Θ̃, Ω̃, Λ̃ can be expressed as
 
V̇ e, e p Π̃ = −e Qe − ep Qe p + 2e P Bε + 2ep P Bε p
− νΨ  (x, r ) Π̃ B̄  A− −1 
m Q Am B̄ Π̃ Ψ (x, r )

+ 2νΨ  (x, r ) Π ∗ B̄  P A−1 

m B̄ Π̃ Ψ (x, r ) (9.408)
 
Then, V̇ e, e p Π̃ is bounded by
    2     
V̇ e, e p Π̃ ≤ −λmin (Q) e2 + e p  + 2 P B e ε0 + e p  ε p0
   2
2 
− νλmin B̄  A−
m Q A −1
m B̄ Ψ (x, r ) Π̃ 
    
2 
+ 2ν  B̄ P A−1 
m B̄ Ψ (x, r ) Π̃  Π0 (9.409)
 
where sup∀x∈D ε (x) ≤ ε0 , sup∀x∈D ε p (x) ≤ ε p0 , and Π0 = Π ∗ .
P B ε0 P B ε p0  
Let c1 = λmin (Q), c2 = , c3 = , c4 = λmin B̄  A− −1
m Q Am B̄
 λmin (Q)  λmin (Q)
 B̄ P A−1 B̄  Π0
m
Ψ (x, r )2 , and c5 =   . Then,
λmin B̄  A− −1
m Q Am B̄

    2
V̇ e, e p Π̃ ≤ −c1 (e − c2 )2 + c1 c22 − c1 e p  − c3
  2
 
+ c1 c32 − νc4 Π̃  − c5 + νc4 c52 (9.410)
 
Thus, V̇ e, e p , Π̃ ≤ 0 outside a compact set S defined by
    2
S = e (t) , e p (t) , Π̃ (t) : c1 (e − c2 )2 + c1 e p  − c3
  2 #
 
+νc4 Π̃  − c5 ≤ c1 c22 + c1 c32 + νc4 c52 (9.411)

This implies

νc4 c52
e ≥ c2 + c22 + c32 + =p (9.412)
c1

  νc4 c52
e p  ≥ c3 + c22 + c32 + =q (9.413)
c1
290 9 Robust Adaptive Control

  c1 c22 + c1 c32
 
Π̃  ≥ c5 + c52 + =α (9.414)
νc4

There exists Ψ0 such that Ψ (x, r ) ≤ Ψ0 for any 0 < ν < νmax that satisfies the
following inequalities:

ϕ (x , xm  , Q, ν, ε0 , Π0 ) = −c1 x2 + 2 (c1 c2 + λmax (Q) xm ) x

+ 2c1 c2 xm  − c1 xm 2 + c1 c32
+ νc4 (Ψ (x, r )) c52 ≤ 0 (9.415)

    2  
φ x p  , xm  , Q, ν, ε p0 , Π0 = −c1 x p  + 2 (c1 c3 + λmax (Q) x) x p 
+ 2c1 c3 x − c1 x2 + c1 c22
+ νc4 (Ψ (x, r )) c52 ≤ 0 (9.416)

Then, the lower bounds which are dependent on Ψ (x,  r ) also exist. Since
V̇ e, e p , Π̃ ≤ 0 outside the compact set S , V e, e p , Π̃ ≤ V0 , where V0 is the
 
smallest upper bound of V e, e p , Π̃ which is given by

        
V0 = λmax (P) p 2 + q 2 + λmax ΓΘ−1 + λmax ΓΩ−1 + λmax ΓΛ−1 α 2 (9.417)

Then,  
λmin (P) e2 ≤ V e, e p , Π̃ ≤ V0 (9.418)

 2  
λmin (P) e p  ≤ V e, e p , Π̃ ≤ V0 (9.419)

Therefore, the closed-loop system is uniformly ultimately bounded with the

following ultimate bounds: 
V0
e ≤ (9.420)
λmin (P)

  V0
e p  ≤ (9.421)
λmin (P)
9.11 Bi-Objective Optimal Control Modification for Systems … 291

9.11 Bi-Objective Optimal Control Modification for

Systems with Control Input Uncertainty

Consider the following MIMO system with a control input uncertainty and a matched
uncertainty:  
ẋ = Ax + BΛ u + Θ ∗ Φ (x) + w (9.422)

where x (t) ∈ Rn is a state vector, u (t) ∈ Rm is a control vector, A ∈ Rn × Rn is

known, B ∈ Rn × Rm is also known such that (A, B) is controllable, Λ = Λ >
0 ∈ Rm × Rm is a constant unknown diagonal matrix which represents a control
input uncertainty, Θ ∗ ∈ R p × Rm is the unknown parameter, Φ (x) ∈ R p is a
known regressor function, and w (t) ∈ Rn is a bounded exogenous disturbance with
a bounded time derivative, i.e., sup w ≤ w0 and sup ẇ ≤ δ0 .
The closed-loop plant is designed to follow a reference model specified as

ẋm = Am xm + Bm r (9.423)

where xm (t) ∈ Rn is a reference state vector, Am ∈ Rn × Rn is Hurwitz, and

Bm ∈ Rn × Rq is a matrix associated with a piecewise continuous and bounded
reference command vector r (t) ∈ Rq .
In the presence of both the control input uncertainty and matched uncertainty due
to Λ and Θ ∗ , an adaptive controller is designed as

u = K x (t) x + K r (t) r − Θ  (t) Φ (x) (9.424)

where K x (t) ∈ Rm × Rn is an adaptive feedback gain, K r (t) ∈ Rm × Rq is an

adaptive command feedforward gain, and Θ (t) ∈ R p × Rm is the estimate of Θ ∗ .
We assume that there exist constant and unknown matrices K x∗ and K r∗ such that
the following model matching conditions are satisfied:

A + BΛK x∗ = Am (9.425)

BΛK r∗ = Bm (9.426)

If Λ is unknown but sign of Λ is known, then the standard MRAC adaptive laws
are given by

K̇ x = Γx xe P BsgnΛ (9.427)

K̇ r = Γr r e P BsgnΛ (9.428)

Θ̇ = −ΓΘ Φ (x) e P BsgnΛ (9.429)

292 9 Robust Adaptive Control

However, the standard MRAC is non-robust. To improve robustness, the adaptive

laws should include any one of the robust modification schemes or the projection
method. For example, the σ modification adaptive laws are given by
 
K̇ x = Γx xe P BsgnΛ − σ K x (9.430)
 
K̇ r = Γr r e P BsgnΛ − σ K r (9.431)
 
Θ̇ = −ΓΘ Φ (x) e P BsgnΛ + σ Θ (9.432)

If Λ is completely unknown, then we need to consider other approaches such

as the method in Sect. 9.10. We now introduce a new optimal control modification
method that uses two types of errors for the adaptation: tracking error and predictor
error. We call this bi-objective optimal control modification adaptive law [26–29].
Model-reference adaptive control that uses both the tracking error and the predictor
error or the plant modeling error is also called composite adaptive control [30]. Other
works that have also used this type of adaptation include hybrid adaptive control [31]
and composite model-reference adaptive control [7].
Let Λ̃ (t) = Λ̂ (t) − Λ, K̃ x (t) = K x (t) − K x∗ , K̃ r (t) = K r (t) − K r∗ , and
Θ̃ (t) = Θ (t) − Θ ∗ be the estimation errors. Then, the closed-loop plant becomes
 
ẋ = Am x + Bm r + B Λ̂ − Λ̃ K̃ x x + K̃ r r − Θ̃  Φ (x) + w (9.433)

The tracking error equation is obtained as


ė = Am e + B Λ̂ − K̃ x x − K̃ r r + Θ̃  Φ (x) − w + Bε (9.434)

where ε (t) ∈ Rm is the residual estimation error of the plant model


ε = Λ̃ K̃ x x + K̃ r r − Θ̃  Φ (x) (9.435)

such that sup ε ≤ ε0 .

Consider a predictor model of the plant as
 
x̂˙ = Am x̂ + (A − Am ) x + B Λ̂ u + Θ  Φ (x) + ŵ (9.436)

where ŵ (t) is the estimate of the disturbance w (t).

We define the predictor error as e p (t) = x̂ (t) − x (t), then
 
ė p = Am e p + B Λ̃ u + Θ  Φ (x) + B Λ̂Θ̃  Φ (x) + w̃ + Bε p (9.437)

where w̃ (t) = ŵ (t) − w (t) is the disturbance estimation error, and ε p (t) ∈ Rm is
the residual estimation error of the predictor model
9.11 Bi-Objective Optimal Control Modification for Systems … 293

ε p = −Λ̃Θ̃  Φ (x) (9.438)

 
such that sup ε p  ≤ ε p0 .

Proposition The following bi-objective optimal control modification adaptive laws

can be used to compute K x (t), K r (t), and Θ (t):
 
K̇ x = Γx x e P + νu  Λ̂ B  P A−1
m B Λ̂ (9.439)
 
K̇ r = Γr r e P + νu  Λ̂ B  P A−1
m B Λ̂ (9.440)


Θ̇ = −ΓΘ Φ (x) e P + νu  Λ̂ B  P A−1 
m + ep W
   
− η u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1
m B Λ̂ (9.441)

 ! ; # "
Λ̂˙  = −ΓΛ u + Θ  Φ (x) e W − η u + 2Θ  Φ (x)  Λ̂ B  + ŵ W A−1 B
p m
(9.442)
    
ŵ˙  = −γw ep W − η u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1
m (9.443)

where Γx = Γx > 0 ∈ Rn × Rn , Γr = Γr > 0 ∈ Rq × Rq , ΓΘ = ΓΘ > 0 ∈

R p × R p , ΓΛ = ΓΛ > 0 ∈ Rm × Rm , and γw > 0 are adaptation rate matrices;
ν > 0 ∈ R and η > 0 ∈ R are the optimal control modification parameters; and
P = P  > 0 ∈ Rn × Rn and W = W  > 0 ∈ Rn × Rn are the solutions of the
following Lyapunov equations:

P Am + A
m P = −Q (9.444)

W Am + A
m W = −R (9.445)

where Q = Q  > 0 ∈ Rn × Rn and R = R  > 0 ∈ Rn × Rn are positive-definite

weighting matrices.
The bi-objective optimal control modification adaptive laws are considered to be
the generalization of the optimal control modification adaptive laws developed in
Sect. 9.10.
The optimal control modification adaptive laws are called bi-objective because
they use both the tracking error and the predictor error for adaptation and are derived
from the following infinite-time horizon cost functions:
, tf
1
J1 = lim (e − Δ1 ) Q (e − Δ1 ) dt (9.446)
t f →∞ 2 0
294 9 Robust Adaptive Control
,
1 tf    
J2 = lim e p − Δ2 R e p − Δ2 dt (9.447)
t f →∞ 2 0

subject to Eqs. (9.434) and (9.437), where Δ1 (t) and Δ2 (t) represent the unknown
lower bounds of the tracking error and the predictor error, respectively.
The cost functions J1 and J2 are combined into the following bi-objective cost
function:
J = J1 + J2 (9.448)

The bi-objective cost function J combines both the objectives of minimiza-

tion of the tracking error and the predictor error bounded away from the origin.
Geometrically, it represents the sum of the weighted norm squares measured from
 the
trajectories of e (t) and e p (t) to the normal surface of a hypersphere BΔ = e (t) ∈
    
Rn , e p (t) ∈ Rn : (e − Δ1 ) Q (e − Δ1 ) + e p − Δ2 R e p − Δ2 ≤ Δ2 ⊂
D ⊂ Rn . The bi-objective cost function is designed to provide stability robustness
by not seeking asymptotic tracking which results in e (t) → 0 and e p (t) → 0 as
t f → ∞, but rather bounded tracking which results in the errors tending to some
lower bounds. By not requiring asymptotic tracking, the adaptation can be made
more robust. Therefore, the tracking performance can be traded with stability ro-
bustness by a suitable selection of the modification parameters ν and η. Increasing
the tracking performance by reducing ν and or η will decrease stability robustness
of the adaptive laws to unmodeled dynamics, and vice versa.
The adaptive laws are derived from the Pontryagin’s minimum principle as
follows:

Proof Using the optimal control framework, the Hamiltonian of the cost function is
defined as
1 1   
H= (e − Δ1 ) Q (e − Δ1 ) + e p − Δ2 R e p − Δ2
2   2 
+ λ Am e + B Λ̂ − K̃ x x − K̃ r r + Θ̃  Φ (x) − w + Bε


   
+ μ Am e p + B Λ̃ u + Θ  Φ (x) + B Λ̂Θ̃  Φ (x) + ŵ − w + Bε p
(9.449)

where λ (t) ∈ Rn and μ (t) ∈ Rn are adjoint variables.

The adjoint equations can be obtained from the necessary conditions of optimality
as

λ̇ = −∇ He = −Q (e − Δ1 ) − A
mλ (9.450)
 
μ̇ = −∇ Hep = −R e p − Δ2 − A
mμ (9.451)
9.11 Bi-Objective Optimal Control Modification for Systems … 295
   
subject to the transversality conditions λ t f → ∞ = 0 and μ t f → ∞ = 0 since
both e (0) and e p (0) are given.
Treating K̃ x (t), K̃ r (t), Θ̃ (t), Λ̃ (t), and ŵ (t) as control variables, then the opti-
mal control solutions are obtained by the following gradient-based adaptive laws:

K̇ x = K̃˙ x = −Γx ∇ HK̃ x = Γx xλ B Λ̂ (9.452)

K̇ r = K̃˙ r = −Γr ∇ HK̃r = Γr r λ B Λ̂ (9.453)

 
Θ̇ = Θ̃˙ = −ΓΘ ∇ HΘ̃ = −ΓΘ Φ (x) λ + μ B Λ̂ (9.454)

 
Λ̂˙  = Λ̃˙  = −ΓΛ ∇ HΛ̃ = −ΓΛ u + Θ  Φ (x) μ B (9.455)

ŵ˙  = −γw ∇ Hŵ = −γw μ (9.456)

The closed-form solutions can be obtained by eliminating the adjoint variables

λ (t) and μ (t) using the “sweep” method with the following assumed solutions of
the adjoint equations:
 
λ = Pe + S Λ̂ −K x x − K r r + Θ  Φ (x) (9.457)
 
μ = W e p + T Λ̂ u + 2Θ  Φ (x) + V ŵ (9.458)

Substituting the adjoint solutions back into the adjoint equations yields
 
Ṗe + P Am e + P B Λ̂ −K x x − K r r + Θ  Φ (x)
 
− P B Λ̂ −K x∗ x − K r∗r + Θ ∗ Φ (x) − Pw + P Bε
 
+ Ṡ Λ̂ −K x x − K r r + Θ  Φ (x)

 
d Λ̂ −K x x − K r r + Θ  Φ (x)
+S
dt
 
= −Q (e − Δ1 ) − A  
m Pe − Am S Λ̂ −K x x − K r r + Θ Φ (x)
(9.459)

 
Ẇ e p + W Am e p + W B Λ̂ u + Θ  Φ (x) − W BΛ u + Θ  Φ (x) + W B Λ̂Θ  Φ (x)

− W B Λ̂Θ ∗ Φ (x) + W ŵ − W w + W Bε p
  
 d Λ̂ u + 2Θ  Φ (x)
+ Ṫ Λ̂ u + 2Θ  Φ (x) + T + V̇ ŵ
dt
  
+ V ŵ˙ = −R e p − Δ2 − A   
m W e p − Am T Λ̂ u + 2Θ Φ (x) − Am V ŵ
(9.460)
296 9 Robust Adaptive Control

Equating terms yields the following equations:

Ṗ + P Am + A
mP + Q = 0 (9.461)

Ṡ + A
mS + PB = 0 (9.462)

 
QΔ1 + P B Λ̂ −K x∗ x − K r∗r + Θ ∗ Φ (x) + Pw

 
d Λ̂ −K x x − K r r + Θ  Φ (x)
− P Bε − S =0 (9.463)
dt

Ẇ + W Am + A
mW + R = 0 (9.464)

Ṫ + A
mT + W B = 0 (9.465)

V̇ + A
mV + W = 0 (9.466)

 
RΔ2 + W BΛ u + Θ  Φ (x) + W B Λ̂Θ ∗ Φ (x) + W w − W Bε p

 
d Λ̂ u + 2Θ  Φ (x)
−T − V ŵ˙ = 0 (9.467)
dt
   
subject
 to the
 transversality
 conditions
 P t f → ∞  = 0, S t f → ∞ = 0,
W t f → ∞ = 0, T t f → ∞ = 0, and V t f → ∞ = 0.
The infinite-time horizon solutions of P (t) and W (t) as t f → ∞ tend to their
equilibrium solutions at t = 0 and are given by the algebraic Lyapunov equations

P Am + A
mP + Q = 0 (9.468)

W Am + A
mW + R = 0 (9.469)

The solutions of S (t), T (t), and V (t) also tend to their equilibrium solutions

A
mS + PB = 0 (9.470)

A
mT + W B = 0 (9.471)

A
mV + W = 0 (9.472)

As with any control design, performance and robustness are often considered as
two competing design requirements. Increasing robustness tends to require a com-
promise in performance and vice versa. Thus, to enable the bi-objective optimal
control modification adaptive laws to be sufficiently flexible for a control design, the
9.11 Bi-Objective Optimal Control Modification for Systems … 297

modification parameters ν > 0 and η > 0 are introduced as design free parameters
to allow for the adjustment of the bi-objective optimal control modification terms in
the adaptive laws.
Thus, the solutions of S (t), T (t), and V (t) are given by

S = −ν A−
m PB (9.473)

T = −η A−
m WB (9.474)

V = −η A−
m W (9.475)

Using the expression of u (t), the adjoint solutions are then obtained as

λ = Pe + ν A−
m P B Λ̂u (9.476)
 
μ = W e p − η A−  −
m W B Λ̂ u + 2Θ Φ (x) − η Am W ŵ (9.477)

Substituting the adjoint solutions into the gradient-based adaptive laws yields the
bi-objective optimal control modification adaptive laws in Eqs. (9.452)–(9.456).

We note that K x (t) and K r (t) are adapted based on the tracking error, Λ̂ (t) and
ŵ (t) are adapted based on the predictor error, and Θ (t) is adapted based on both
the tracking error and the predictor error.
The bounds on Δ1 (t) and Δ2 (t) as t f → ∞ can be estimated by

 
1  
Δ1  ≤ P B Λ̂ −K x∗ x − K r∗ r + Θ ∗ Φ (x) + λmax (P) w0 + P B ε0
λmin (Q)
    ⎤
  d Λ̂ −K x x − K r r + Θ  Φ (x)  
 
+ ν A−
m P B 
⎦
 (9.478)
 dt 

    
1    
Δ2  ≤ W BΛ u + Θ  Φ (x) + W B Λ̂ Θ ∗ Φ (x) + λmax (W ) w0
λmin (R)
   
  d Λ̂ u + 2Θ  Φ (x) 
   
  +η − W  ŵ
˙
+ W B ε p0 + η A−
m W B   A m 
 dt 
(9.479)

which are dependent upon the modification parameters, control effectiveness uncer-
tainty, matched uncertainty, and residual tracking error and predictor error.
298 9 Robust Adaptive Control

Note that if R = Q and η = ν, then the bi-objective optimal control modification

adaptive laws for Θ (t), Λ̂ (t), and ŵ (t) become
   
Θ̇ = −ΓΘ Φ (x) e + ep P − ν 2Φ  (x) Θ Λ̂ B  + ŵ P A−1
m B Λ̂
(9.480)
 ! ; # "

Λ̂˙  = −ΓΛ u + Θ  Φ (x) e
p P −ν u + 2Θ  Φ (x) Λ̂ B  + ŵ P A−1
m B
     (9.481)
ŵ˙  = −γw ep P − ν u + 2Θ  Φ (x) Λ̂ B  + ŵ P A−1
m (9.482)

The bi-objective optimal control modification adaptive laws can be shown to be

stable as follows:
Proof Choose a Lyapunov candidate function
   
V e, e p , K̃ x , K̃ r , Θ̃, Λ̃, w̃ = e Pe + ep W e p + trace K̃ x Γx−1 K̃ x
   
+ trace K̃ r Γr−1 K̃ r + trace Θ̃  ΓΘ−1 Θ̃
 
+ trace Λ̃ΓΛ−1 Λ̃ + w̃ γw−1 w̃ (9.483)
 
Evaluating V̇ e, e p , K̃ x , K̃ r , Θ̃, Λ̃, w̃ yields
  
V̇ e, e p , K̃ x , K̃ r , Θ̃, Λ̃, w̃ = −e Qe + 2e P B Λ̂ − K̃ x x − K̃ r r + Θ̃  Φ (x)

− 2e Pw + 2e P Bε
  
− e   
p Re p + 2e p W B Λ̃ u + Θ Φ (x) + Λ̂Θ̃ Φ (x)

+ 2e 
p W w̃ + 2e p W Bε p
   
+ 2trace K̃ x x e P + νu  Λ̂ B  P A−1
m B Λ̂
   
+ 2trace K̃ r r e P + νu  Λ̂ B  P A−1
m B Λ̂
 
− 2trace Θ̃  Φ (x) e P + νu  Λ̂ B  P A−1m + ep W


;   "

− η u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1 m B Λ̂
 
− 2trace Λ̃ u + Θ  Φ (x)
! ; # " "

e
pW −η u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1 m B
! ; # "

− 2ep W w̃ + 2 η u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1 m w̃

− 2ẇ γw−1 w̃ (9.484)

9.11 Bi-Objective Optimal Control Modification for Systems … 299
 
V̇ e, e p , K̃ x , K̃ r , Θ̃, Λ̃, w̃ can be further simplified as
 
V̇ e, e p , K̃ x , K̃ r , Θ̃, Λ̃, w̃ = −e Qe − 2e Pw + 2e P Bε − e   −1
p Re p + 2e p W Bε p − 2ẇ γw w̃

+ 2νu  Λ̂ B  P A−1

m B Λ̂ũ
! ; # "

+ 2ηtrace Θ̃  Φ (x) u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1
m B Λ̂
 
+ 2ηtrace Λ̃ u + Θ  Φ (x)
; # "

u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1
m B
! ; # "

+ 2ηtrace w̃ u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1
m (9.485)

where ũ (t) = K̃ x (t) x (t) + K̃ r (t) r (t) − Θ̃  (t)⎡Φ (x). ⎤

  Θ (t) Λ̂ (t) 0 0
Let B̄ = B B I ∈ Rn × R2m+n , Ω (t) = ⎣ 0 Λ̂ (t) 0 ⎦ ∈
0 0 ŵ (t)
⎡ ⎤
Φ (x)
R p+m+1 × R2m+n , Ψ (x, r ) = ⎣ u + Θ  Φ (x) ⎦ ∈ R p+m+1 . Then,
1
 
trace Ω̃  Ψ (x, r ) Ψ  (x, r ) Ω B̄  W A−1
m B̄
    
= trace Θ̃  Φ (x) u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1 m B Λ̂
      
× trace Λ̃ u + Θ  Φ (x) u + 2Θ  Φ (x) Λ̂ B  + ŵ W A−1m B
      
  −1
+ trace w̃ u + 2Θ Φ (x) Λ̂ B + ŵ W Am (9.486)
⎡⎤
Θ∗ 0 0
where Ω̃ (t) = Ω (t) − Ω ∗ and Ω ∗ (t) = ⎣ 0 Λ 0 ⎦ ∈ R p+m+1 × R2m+n .
0 0 w (t)
Thus,
 
V̇ e, e p , ũ, Ω̃ = −e Qe − 2e Pw + 2e P Bε − e   −1
p Re p + 2e p W Bε p − 2ẇ γw w̃

+ 2νu  Λ̂ B  P A−1   −1 

m B Λ̂ũ + 2ηΨ (x, r ) Ω B̄ W Am B̄ Ω̃ Ψ (x, r )
(9.487)

Note that B  P A−1   −1

m B and B̄ W Am B̄ are both negative definite matrices, there-
fore
300 9 Robust Adaptive Control
 
V̇ e, e p , Ω̃, ũ = −e Qe − 2e Pw + 2e P Bε − e   −1
p Re p + 2e p W Bε p − 2ẇ γw w̃

− ν ũ  Λ̂ B  A− −1 ∗   −1
m Q Am B Λ̂ũ + 2νu Λ̂ B P Am B Λ̂ũ
− ηΨ  (x, r ) Ω̃ B̄  A− −1 
m R Am B̄ Ω̃ Ψ (x, r )
+ 2ηΨ  (x, r ) Ω ∗ B̄  W A−1 
m B̄ Ω̃ Ψ (x, r ) (9.488)

⎤ ⎡
  x
Let K (t) = K x (t) K r (t) −Θ  (t) ∈ Rm × Rn+q+ p and z (x, r ) = ⎣ r ⎦
Φ (x)
 
∈Rn+q+ p
. Then, V̇ e, e p , Ω̃, ũ is bounded by
 
V̇ e, e p , K̃ , Ω̃ ≤ −λmin (Q) e2 + 2 e λmax (P) w0 + 2 e P B ε0
 2  
− λmin (R) e p  + 2 e p  W B ε p0
       
  2  2  2
+ 2γw−1 Ω̃  δ0 − νλmin B  A− −1
m Q Am B z (x, r ) Λ̂  K̃ 
   2  
    
+ 2ν z (x, r )2 B  P A−1
m B  Λ̂  K̃  K 0
   
− ηλmin B̄  A− −1 2  2
m R Am B̄ Ψ (x, r ) Ω̃ 
   
  2 
+ 2η  B̄  W A−1
m B̄  Ψ (x, r ) Ω̃  Ω0 (9.489)

where K 0 = K ∗  and Ω0 = sup Ω ∗ .

W Bε
Let c1 = λmin (Q), c2 = λmax (P)w 0 +P Bε0
λmin (Q)
, c3 = λmin (R), c4 = λmin (R)p0 ,
  m BK0
 B  P A−1
c5 = λmin B  A− −1
m Q Am B z (x, r ) , c6 = λmin ( B  A−
2
, c7 = λmin
m Q Am B )
−1

  −  m B̄ Ω0
 B̄  W A−1 γ −1 δ
B̄ Am R A−1m B̄ Ψ (x, r ) , and c8 = λmin ( B̄  A−
2
+ w 0 2 . Then,
m R Am B̄ )
−1
ηc7  B̄ 

    2
V̇ e, e p , K̃ , Ω̃ ≤ −c1 (e − c2 )2 + c1 c22 − c3 e p  − c4 + c3 c42
 2   2  2
     
− νc5 Λ̂  K̃  − c6 + νc5 c62 Λ̂
  2
 
− ηc7 Ω̃  − c8 + ηc7 c82 (9.490)
   
   
Note that Λ̃ ≤ Ω̃ . Then,

 2  2    2
     
Λ̂ = Λ + Λ̃ ≤ Λ + Ω̃  (9.491)
   
 
Let Ω̃  = c8 . Then, V̇ e, e p , K̃ , Ω̃ is bounded by
9.11 Bi-Objective Optimal Control Modification for Systems … 301
    2
V̇ e, e p , K̃ , Ω̃ ≤ −c1 (e − c2 )2 + c1 c22 − c3 e p  − c4
  2
 
+ c3 c42 − νc5 Λ20  K̃  − c6 + νc5 c62 Λ20 + ηc7 c82 (9.492)

where Λ0 =(Λ + c8 )2.

Thus, V̇ e, e p , K̃ , Ω̃ ≤ 0 outside a compact set S defined as

; 
S = e (t) , e p (t) , K̃ (t) , Ω̃ (t) : c1 (e − c2 )2
  2   2
 
+ c3 e p  − c4 + νc5 Λ20  K̃  − c6
  2 #
 
+ ηc7 Ω̃  − c8 ≤ c1 c22 + c3 c42 + νc5 c62 Λ20 + ηc7 c82 (9.493)

This implies

c3 c42 + νc5 c62 Λ20 + ηc7 c82
e ≥ c2 + c22 + =p (9.494)
c1

  c1 c22 + νc5 c62 Λ20 + ηc7 c82
e p  ≥ c4 + c42 + =q (9.495)
c3

  c1 c22 + c3 c42 + ηc7 c82
 
 K̃  ≥ c6 + c62 + =α (9.496)
νc5 Λ20

  c1 c22 + c3 c42 + νc5 c62 Λ20
 
Ω̃  ≥ c8 + c82 + =β (9.497)
ηc7

There exist z 0 and Ψ0 such that z (x, r ) ≤ z 0 and Ψ (x, r ) ≤ Ψ0 for any
0 < ν < νmax and 0 < η < ηmax that satisfy the following inequalities:

ϕ (x , xm  , Q, ν, w0 , ε0 , Λ0 , K 0 ) = −c1 x2 + 2 (c1 c2 + λmax (Q) xm ) x

+ 2c1 c2 xm  − c1 xm 2
+ c3 c42 + νc5 (z (x, r )) c62 Λ20
+ ηc7 (Ψ (x, r )) c82 ≤ 0 (9.498)
302 9 Robust Adaptive Control

    2  
φ x p  , xm  , R, η, ẇ0 , ε p0 , B0 , Ω0 = −c3 x p  + 2 (c3 c4 + λmax (R) x) x p 
+ 2c3 c4 x − c3 x2
+ c1 c22 + νc5 (z (x, r )) c62 Λ20
+ ηc7 (Ψ (x, r )) c82 ≤ 0 (9.499)

Then, the lower bounds which  are dependent on z (x, r ) and Ψ (x, r )
also exist. Since V̇ e, e p , K̃ , Ω̃ ≤ 0 outside the compact set S , limt→∞ V
   
e, e p , K̃ , Ω̃ ≤ V0 , where V0 is the smallest upper bound of V e, e p , K̃ , Ω̃
which is given by
 
V0 = λmax (P) p 2 + λmax (W ) q 2 + λmax Γx−1 α 2
    
+ λmax Γr−1 α 2 + λmax ΓΘ−1 α 2 + β 2
 
+ λmax ΓΛ−1 β 2 + γw−1 β 2 (9.500)

Then,  
λmin (P) e2 ≤ V e, e p , K̃ , Ω̃ ≤ V0 (9.501)

 2  
λmin (W ) e p  ≤ V e, e p , K̃ , Ω̃ ≤ V0 (9.502)

Therefore, the closed-loop system is uniformly ultimately bounded with the

following ultimate bounds: 
V0
e ≤ (9.503)
λmin (P)

  V0
e p  ≤ (9.504)
λmin (W )

Example 9.17 Consider a first-order SISO plant

 
ẋ = ax + bλ u (t − td ) + θ ∗ x 2 + w

where a = 1 and b = 1 are known, λ = −1 and θ ∗ = 0.2 are unknown, td = 0.1 s

is a known time delay, and w (t) = 0.01 (sin t + cos 2t).
9.11 Bi-Objective Optimal Control Modification for Systems … 303

The reference model is given by

ẋm = am xm + bm r

where am = −2, bm = 2, and r (t) = sin t.

The nominal control input effectiveness is equal to unity. So, λ = −1 represents
a full control reversal which is generally challenging to any control systems.
The baseline nominal control gains are k̄ x = amb−a = −3 and kr = bbm = 2. With
the full control reversal, the ideal unknown control gains are k x∗ = −k̄ x = 3 and
kr∗ = −k̄r = −2.
The adaptive controller is designed as

u = k x (t) x + kr r (t) − θ (t) x 2

using the bi-objective optimal control modification method with the following pre-
dictor model:
 
x̂˙ = am x̂ + (a − am ) x + bλ̂ u (t − td ) + θ x 2 + ŵ

The bi-objective optimal control adaptive laws are

 
k̇ x = γx x e + νu λ̂bam−1 bλ̂
 
k̇r = γr r e + νu λ̂bam−1 bλ̂
   
θ̇ = −γθ x 2 e + νu λ̂bam−1 + e p − η u + 2θ x 2 λ̂b + ŵ am−1 bλ̂

    
λ̂˙ = −γλ u + θ x 2 e p − η u + 2θ x 2 λ̂b + ŵ am−1 b
   
ŵ˙ = −γw e p − η u + 2θ x 2 λ̂b + ŵ am−1

The initial conditions are k x (0) = k̄ x , kr (0) = k̄r , θ (0) = 0, λ̂ (0) = 1,

ŵ (0) = 0. The adaptive gains are chosen to be γx = γr = γθ = γλ = γw = 10, and
the modification parameters are chosen to be ν = 0.1 and η = 0.01.
The closed-loop response with r (t) = sin t for t ∈ [0, 60] is shown in Fig. 9.25.
It can be seen that x (t) eventually tracks xm (t), but the two signals are initially
180o out of phase due to the control reversal. The signal x̂ (t) approximates x (t)
very well after 10 s. Overall, the bi-objective optimal control modification method
demonstrates a good tracking performance.
304 9 Robust Adaptive Control

x, xm
0

−1
x x
m
−2
0 10 20 30 40 50 60
t

0
<
x, x

−2

x x^
−4
0 10 20 30 40 50 60
t

Fig. 9.25 Closed-loop response with bi-objective optimal control modification

10
kx

−10
0 10 20 30 40 50 60
t

5
kr

−5
0 10 20 30 40 50 60
t

0
θ

−5
0 10 20 30 40 50 60
t

Fig. 9.26 Adaptive parameter convergence of bi-objective optimal control modification

The adaptive parameters k x (t), kr (t), and θ (t) are shown in Fig. 9.26. These
adaptive parameters show a parameter convergence to their corresponding ideal val-
ues. The parameter convergence is also facilitated by having a persistently exciting
reference command signal r (t) = sin t.
The estimates of the control input effectiveness λ̂ (t) and disturbance ŵ (t) are
shown in Fig. 9.27. It can be seen that λ̂ (t) converges nicely to the ideal value of
−1. Because of the parameter convergence of λ̂ (t), the bi-objective optimal control
modification adaptive laws remain stable in light of the sign reversal of the control
input effectiveness. Without accurate estimation of λ̂ (t), it will be a challenge for
any adaptive control method to be able to maintain stability. It is also noted that this
parameter convergence is achieved even with a time delay of 0.1 s. The bi-objective
optimal control modification remains stable up to a time delay of 0.108 s.
For comparison, we redesign the adaptive controller with the optimal control
modification with the predictor model according to Sect. 9.10. The optimal control
modification adaptive laws are
9.11 Bi-Objective Optimal Control Modification for Systems … 305

0
<

λ
−5

−10
0 10 20 30 40 50 60
t

0
<
w, w

−2

w ^
w
−4
0 10 20 30 40 50 60
t

Fig. 9.27 Estimation of control input effectiveness and disturbance with bi-objective optimal con-
trol modification
 
k̇ x = γx x e + νxk x λ̂bam−1 bλ̂
 
k̇r = γr r e + νr kr λ̂bam−1 bλ̂
 
θ̇ = −γθ x 2 e p − ηx 2 θ λ̂bam−1 bλ̂

   
λ̂˙ = −γλ u + θ x 2 e p − η u + θ x 2 λ̂bam−1 b

 
ŵ˙ = −γw e p − ηŵam−1

Keeping the same initial conditions, it can be seen that the optimal control modi-
fication performs worse than the bi-objective optimal control modification as shown
in Figs. 9.28, 9.29 and 9.30. Even though the predictor model does an excellent job
of estimating x (t), the tracking performance is much worse than that with the bi-
objective optimal control modification. The adaptive parameters also suffers from
a poor parameter convergence with the optimal control modification as shown in
Figs. 9.29 and 9.30.
By setting ν = 0 and η = 0, one can recover the standard MRAC adaptive
laws with the predictor model. Comparing robustness of the three controllers, the
optimal control modification has the largest time-delay margin of 0.116 s, whereas
the standard MRAC has the smallest time-delay margin of 0.082 s.
306 9 Robust Adaptive Control

x, xm
0

−1

−2
0 10 20 30 40 50 60
t

1
<
x, x

−1
x x^
−2
0 10 20 30 40 50 60
t

Fig. 9.28 Closed-loop response with optimal control modification

10
x

0
k

−10
0 10 20 30 40 50 60
t

10
r

0
k

−10
0 10 20 30 40 50 60
t

0
θ

−1
0 10 20 30 40 50 60
t

Fig. 9.29 Adaptive parameter convergence of optimal control modification

< −1
λ

−2

−3
0 10 20 30 40 50 60
t

0.5

0
<
w, w

−0.5

−1
w ^
w
−1.5
0 10 20 30 40 50 60
t

Fig. 9.30 Estimation of control input effectiveness and disturbance with optimal control modifi-
cation
9.12 Adaptive Control for Singularly Perturbed Systems … 307

9.12 Adaptive Control for Singularly Perturbed Systems

with First-Order Slow Actuator Dynamics

When there exists a timescale separation between a fast plant and a slow actuator
which prevents the plant from following a reference model, a singular perturbation
approach can be used to separate the time scale between the plant and actuator
and then modify the adaptive law to account for the slow actuator in the singularly
perturbed system [32]. The singular perturbation approach transforms the original
system into a reduced-order system in slow time. A model matching condition is
applied to the reduced-order system and the reference model in the transformed slow
time coordinate that results in changes in the actuator command to accommodate
the slow actuator dynamics. The resulting control signal can then track the reference
model better than if the actuator command is not modified.
Consider a linear plant as
 
ẋ = Ax + B u + Θ ∗ x (9.505)

where x (t) ∈ Rn is a state vector, u (t) ∈ Rn is a control vector, A ∈ Rn × Rn and

B ∈ Rn×m with m ≥ n are known matrices such that the pair (A, B) is controllable
and furthermore A is Hurwitz, and Θ ∗ ∈ Rm × Rn is an unknown constant matrix.
The controller u (t) is subject to slow first-order actuator dynamics

u̇ = εG (u − u c ) (9.506)

where u c (t) ∈ Rm is an actuator command vector, ε is a positive constant, and

G ∈ Rm×m is a known Hurwitz matrix.
The objective is to design the controller u (t) that enables the plant to follow a
reference model
ẋm = Am xm + Bm r (9.507)

where Am ∈ Rn × Rn is Hurwitz and known, Bm ∈ Rn × Rq is also known, and

r (t) ∈ Rq ∈ L∞ is a bounded command vector.
If the actuator dynamics are sufficiently fast relative to the reference model dynam-
ics, that is, ε G  Am , then the effect of actuator dynamics may be negligible.
This is generally the case for most applications whereby the actuator dynamics typi-
cally are several times faster than the plant dynamics to be controlled. The robustness
issues with unmodeled actuator dynamics can then be avoided. Then, we design a
controller u (t) to follow an actuator command as

u c = K x∗ x + K r∗r − u ad (9.508)

where K x∗ ∈ Rm × Rn and K r∗ ∈ Rm × Rq are known nominal gain matrices, and

u ad (t) ∈ Rm is an adaptive signal.
308 9 Robust Adaptive Control

The control gain matrices K x∗ and K r∗ can be chosen to satisfy the model matching
conditions A + B K x∗ = Am and B K r∗ = Bm , and the adaptive signal u ad (t) is chosen
to be
u ad = Θ  Φ (x) (9.509)

where Θ (t) is an estimate of Θ ∗ .

Defining the tracking error as e (t) = xm (t) − x (t), then the tracking error
equation becomes
ė = Am e + B Θ̃  x (9.510)

where Θ̃ (t) = Θ (t) − Θ ∗ is the estimation error.

In the case of slow actuator dynamics when ε  1 is a small parameter and
ε G  A, then x (t) has “fast” dynamics and u (t) has “slow” dynamics. This
will lead to robustness issues. To decouple the fast state x (t) and slow state u (t),
we perform a timescale separation by applying the singular perturbation method.
Toward that end, we consider a slow time transformation

τ = εt (9.511)

where τ is a slow time variable.

Then, the plant and actuator models are transformed into a singularly perturbed
system as
dx  
ε = Ax + B u + Θ ∗ x (9.512)
dτ
du
= G (u − u c ) (9.513)
dτ
The Tikhonov’s theorem can be used to approximate the solution of the singularly
perturbed system with the solution of a “reduced-order” system by setting ε = 0 [33].
Then, x (u, ε) is on a fast manifold. The fast dynamics result in x (u, ε) tending to
its asymptotic solution in near zero time as ε → 0. Thus, the reduced-order system
is given by
d x0  
ε = Ax0 + B u 0 + Θ ∗ x0 = 0 (9.514)
dτ
du 0
= G (u 0 − u c ) (9.515)
dτ
where x0 (τ ) and u 0 (τ ) are the “outer” solution of the singularly perturbed system.
The term “outer” is in connection with the concept of “inner” or “boundary layer”
and “outer” solutions which have the origin in the boundary layer theory due to
Prandtl. The “inner” or “boundary layer” solution for this system is obtained from
 
ẋi = Axi + B u i + Θ ∗ xi (9.516)
9.12 Adaptive Control for Singularly Perturbed Systems … 309

u̇ i = εG (u i − u c ) = 0 (9.517)

The solution is then expressed as

x (t) = x0 (t) + xi (t) − x M AE (t) (9.518)

where x M AE (t) is a correction term by a matched asymptotic expansion method

applied to both the inner and outer solutions [34]. The outer solution is in fact the
asymptotic solution of the original system as t → ∞.
The solution of x0 (τ ) is obtained as
 −1
x0 = − A + BΘ ∗ Bu 0 (9.519)

Differentiating Eq. (9.519) with respect to the slow time variable and then substi-
tuting the actuator model into the result yield

d x0  −1
= − A + BΘ ∗ BG (u 0 − u c ) (9.520)
dτ
From Eq. (9.514), we have

u 0 = − B̄ −1 Ax0 − Θ ∗ x0 (9.521)
 −1
where B̄ −1 = B  B B  is the right pseudo-inverse of B.
Hence, we obtain the following reduced-order plant model constrained by the
slow actuator dynamics:

d x0  −1  
= A + BΘ ∗ BG B̄ −1 Ax0 + Θ ∗ x0 + u c (9.522)
dτ
Using the matrix inversion lemma, we obtain
 −1  −1 ∗ −1
A + BΘ ∗ = A−1 − A−1 B I + Θ ∗ A−1 B Θ A (9.523)
 −1 ∗ −1
Let Ψ ∗ = A−1 B I + Θ ∗ A−1 B Θ A . Then,

d x0    
= A−1 BG B̄ −1 Ax0 + −Ψ ∗ BG B̄ −1 A + A−1 − Ψ ∗ BGΘ ∗ x0
dτ  
+ A−1 − Ψ ∗ BGu c (9.524)

We will consider the asymptotic solution of the singularly perturbed system. In

effect, the inner solution is neglected so that x (t) = x0 (t). The reduced-order model
is expressed as
310 9 Robust Adaptive Control

dx  
= As x + Bs Λ u c + Θs∗ x (9.525)
dτ
 
where As = A−1 BG B̄ −1 A, Bs = A−1 BG, Bs Λ = A−1 − Ψ ∗ BG, and
 
Bs ΛΘs∗ = −Ψ ∗ BG B̄ −1 A+ A−1 − Ψ ∗ BGΘ ∗ .
If As is Hurwitz and Θs∗ = 0, then the Tikhonov’s theorem guarantees that the
reduced solution with ε > 0 converges to the solution of the original system as
ε → 0 [33].
Because of the slow actuator dynamics, the time scale of the response of the
plant cannot exceed that of the actuator. Thus, if the reference model is faster than
the actuator model, the tracking error cannot be guaranteed to be small even with
adaptive control due to the model mismatch. A possible solution is to revise the
reference model to match the actuator-constrained plant model, or alternatively to
redesign the actuator command to reduce the tracking error.
In slow time, the reference model is expressed as

d xm 1
= (Am xm + Bm r ) (9.526)
dτ ε
We make the following choice for the actuator command signal:

u c = K x x + K r r − u ad (9.527)

where K x and K r are nominal control gains for the normal plant without the slow
actuator dynamics and u ad (t) is of the form

u ad = ΔK x (t) x + ΔK r (t) r − Θs (t) x = −Ω  (t) Φ (x, r ) (9.528)

   
where Ω  (t) = Θs (t) − ΔK x (t) −ΔK r (t) and Φ (x, r ) = x  r  .
We assume that there exist ideal control gain matrices ΔK x∗ and ΔK r∗ that satisfy
the following model matching conditions:

1  
Am = As + Bs Λ K x + ΔK x∗ (9.529)
ε
1  
Bm = Bs Λ K r + ΔK r∗ (9.530)
ε
Then, the closed-loop system becomes

dx 1
= (Am xm + Bm r ) − Bs ΛΩ̃  Φ (x, r ) (9.531)
dτ ε

where Ω̃s (t) = Ωs (t) − Ωs∗ .

The tracking error equation in slow time is obtained as
9.12 Adaptive Control for Singularly Perturbed Systems … 311

de d xm dx 1
= − = Am e + ε Bs ΛΩ̃  Φ (x, r ) (9.532)
dτ dτ dτ ε

Since Am is Hurwitz and if Ω̃ (t) is bounded, then the Tikhonov’s theorem guar-
antees that the reduced solution with ε > 0 converges to the solution of the original
system as ε → 0.
Transforming into the real time, the tracking error equation becomes

ė = Am e + ε Bs ΛΩ̃  Φ (x, r ) (9.533)

If Λ is a diagonal matrix of known sign, then the optimal control modification

adaptive law in real time is given by
 
Ω̇ = −εΓ Φ (x, r ) e PsgnΛ − ενΦ  (x, r ) Ω Bs P A−1
m Bs (9.534)

However, this assumption is not always guaranteed since Λ can be of mixed sign.
The bi-objective optimal control modification adaptive law can be used to estimate
Λ using the following predictor model:
 
x̂˙ = Am x̂ + (ε As − Am ) x + ε Bs Λ̂ u c + Θs x (9.535)

Then, the bi-objective optimal control modification adaptive laws are expressed
as

Ω̇ = −εΓΩ Φ (x, r ) e P + ενu  Λ̂ B  P A−1 
m + ep W
  
−εη u + 2Θs Φ (x) Λ̂ Bs W A−1 m Bs Λ̂ (9.536)
   
Θ̇s = −εΓΘs x ep W − εη u + 2Θs Φ (x) Λ̂ Bs W A−1
m Bs Λ̂ (9.537)

    
Λ̂˙  = −εΓΛ u + Θs Φ (x) ep W − εη u + 2Θs Φ (x) Λ̂ Bs W A−1
m Bs
(9.538)
where e p (t) = x̂ (t) − x (t).

Example 9.18 Consider the following simple scalar system:

ẋ = ax + bu + θ ∗ x + w (t)

with slow first-order actuator dynamics

u̇ = εg (u − u c )

where a < 0, g < 0, ε > 0, |εg| < |a|, and w (t) is a disturbance signal.
The reference model is
312 9 Robust Adaptive Control

ẋm = am xm + bm r

where am < 0.
The actuator command for the slow actuator dynamics is designed as

u c = k x x + kr r − Ω  Φ (x, r )
 
where k x = amb−a , kr = bbm , and Φ (x, r ) = x r .
Note that, if the actuator dynamics are fast, then the actuator command is

u c = k x x + kr r − θ x

The optimal control modification adaptive law for the slow actuator dynamics is
 
Ω̇ = −εΓ Φ (x, r ) esgnλ − ενΦ  (x, r ) Ωbs am−1 bs

where bs = bg
a
and λ = a
a+bθ ∗
, and for the fast actuator dynamics is
 
θ̇ = −Γ x e − νxθ bam−1 b

If a and g are nominally of the same order of magnitude, then we note that, for
the slow actuator dynamics, the effective adaptive gain is reduced by ε for a similar
performance as that for the fast actuator dynamics.
For the numerical example, we let a = −1, b = 1, θ ∗ = 0.1, g = −1, ε = 0.1,
am = −5, bm = 1, r (t) = sin t, and w (t) = 0.05 sin 10t. The responses due to the
standard MRAC adaptive law and optimal control modification adaptive law with the
singular perturbation approach are shown in Fig. 9.31. The response for the standard
MRAC exhibits more initial transients than that for the optimal control modification
using the same adaptive gain.
Figure 9.32 is the plot of the control input and actuator command with the singular
perturbation approach. As can be seen, the actuator command signal is quite large
relative to the control input. This is due to the fact that the actuator dynamics are
slow so a large actuator command does not translate into the same control input in
a finite time. The effectiveness of the optimal control modification is demonstrated
by the reduced amplitude of oscillations in the control input significantly over that
due to the standard MRAC.
Figure 9.33 shows the responses due to the unmodified actuator command for the
fast actuator dynamics. As can be seen, the control input is insufficient to allow the
plant to follow the reference model even with adaptive control.
9.12 Adaptive Control for Singularly Perturbed Systems … 313

1
MRAC Γ=1000
0.5

x, xm
0

−0.5
x xm
−1
0 10 20 30 40 50 60
t

1
Optimal Control Modification Γ=1000 ν=0.5
0.5
m
x, x

−0.5
x xm
−1
0 10 20 30 40 50 60
t

Fig. 9.31 Responses for slow actuator dynamics with MRAC and optimal control modification

40
MRAC Γ=1000
20
u, uc

−20
u u
m
−40
0 10 20 30 40 50 60

40
Optimal Control Modification Γ=1000 ν=0.5
20
u, uc

−20
u u
m
−40
0 10 20 30 40 50 60

Fig. 9.32 Control and actuator command for slow actuator dynamics with MRAC and optimal
control modification

0.4
MRAC Γ=1000
0.2
x, xm

−0.2
x x
m
−0.4
0 10 20 30 40 50 60
t

0.4
Optimal Control Modification Γ=1000 ν=0.5
0.2
x, xm

−0.2
x x
m
−0.4
0 10 20 30 40 50 60
t

Fig. 9.33 Responses for slow actuator dynamics due to unmodified actuator command
314 9 Robust Adaptive Control

9.13 Optimal Control Modification for Linear Uncertain

Systems with Unmodeled Dynamics

Consider the system

ẋ = Ax + Bu + Δ (x, z, u) (9.539)

ż = f (x, z, u) (9.540)

y = Cx (9.541)

where z (t) is an internal state vector, Δ (x, z, u) is the plant model error that is
unknown and not accounted for, ż (t) is the unmodeled dynamics, and y (t) is the
plant output vector.
When the tracking error based on only the output signal y (t) is used for MRAC,
such a class of adaptive control is called output feedback adaptive control. If a
stable first-order reference model ym (t) is specified for the output y (t) to track, then
model-reference adaptive control is feasible if the plant transfer function satisfies a
certain property called strictly positive real (SPR) condition. Therefore, an essential
requirement for output feedback adaptive control is the SPR condition that the transfer
function must satisfy. In Chap. 8, we see that the standard MRAC is unstable for non-
minimum phase plants (see Example 8.3). Non-minimum phase systems do not have
SPR transfer functions. Therefore, output feedback adaptive control of systems with
non-SPR transfer functions can be challenging.
The strictly positive real condition is defined as follows:
Definition 9.1 Consider a proper SISO transfer function

Z (s)
G (s) = (9.542)
R (s)

where Z (s) and R (s) are polynomials of degrees m and n, respectively, with n > m.
The relative degree is defined as n − m. If the relative degree of G (s) is not greater
than 1 and the real part of G (s) is at least positive for all σ ≥ 0 where s = σ + jω,
that is,  (G (s)) ≥ 0 for σ ≥ 0, then the transfer function is said to be positive real
(PR). Furthermore, if G (s − ε) is positive real for some ε > 0, then G (s) is strictly
positive real (SPR) [30].
Example 9.19 • Consider the transfer function in Example 8.2

s+1
G (s) =
s2 + 5s + 6

Expressing in the complex variables σ and ω, we get

 3   
σ + 6σ 2 + σ ω2 + 11σ + 4ω2 + 6 + jω −σ 2 − 2σ − ω2 + 1
G (s) =  2
σ 2 + 5σ − ω2 + 6 + ω2 (2σ + 5)2
9.13 Optimal Control Modification for Linear Uncertain Systems … 315

We see that  ((G (s))) > 0 for all σ ≥ 0 and ω ≥ 0. If we let σ = −ε and
ω = 0, then  (G (s)) ≥ 0 for 0 < ε ≤ 1. So, G (s) is SPR and is a minimum
phase, stable transfer function with relative degree 1.
• Consider the transfer function
 3   
s σ + 5σ 2 + σ ω2 + 6σ + 5ω2 + jω −σ 2 − ω2 + 6
G (s) = 2 =  2
s + 5s + 6 σ 2 + 5σ − ω2 + 6 + ω2 (2σ + 5)2

We see that  (G (s)) ≥ 0 for all σ ≥ 0 and ω ≥ 0. So, G (s) is only PR but not
SPR. Note that G (s) is a stable transfer function with a zero on the jω axis.
• The transfer function in Example 8.2

s−1
G (s) =
+ 5s + 6
s2
 3   
σ + 4σ 2 + σ ω2 + σ + 6ω2 − 6 + jω −σ 2 + 2σ − ω2 + 11
=  2
σ 2 + 5σ − ω2 + 6 + ω2 (2σ + 5)2

is not SPR since  (G (s)) can be negative for some values of σ ≥ 0 and ω ≥ 0.
Note that G (s) is non-minimum phase.
• The transfer function
s+1
G (s) =
s 2 − 5s + 6
 3   
σ − 4σ 2 + σ ω2 + σ − 6ω2 + 6 + jω −σ 2 − 2σ − ω2 + 11
=  2
σ 2 − 5σ − ω2 + 6 + ω2 (2σ − 5)2

is not SPR since  (G (s)) can be negative for some values of σ ≥ 0 and ω ≥ 0.
Note that G (s) is unstable.
• The transfer function
1
G (s) =
s 2 + 5s + 6
 2 
σ + 5σ − ω2 + 6 − jω (2σ + 5)
= 2
σ 2 + 5σ − ω2 + 6 + ω2 (2σ + 5)2

is not SPR since  (G (s)) can be negative for some values of σ ≥ 0 and ω ≥ 0.
Note that G (s) has relative degree 2.

Research in robust adaptive control was motivated by instability phenomena of

adaptive control. In fact, instability of adaptive control in the early 1960s which
contributed to the crash of one of the NASA X-15 hypersonic vehicles caused a
great deal of concern about the viability of adaptive control. Rohrs et al. investigated
316 9 Robust Adaptive Control

various instability mechanisms of adaptive control due to unmodeled dynamics [1].

The Rohrs counterexample demonstrates the weakness of MRAC in its lack of ro-
bustness. Let us revisit the topic of unmodeled dynamics as discussed in Sect. 8.4
[20].
The open-loop transfer function of a first-order SISO plant with a second-order
unmodeled actuator dynamics is given by

y (s) bωn2
=   (9.543)
u (s) (s − a) s 2 + 2ζ ωn s + ωn2

where a < 0 is unknown, b is unknown, but b > 0 is known, and ζ and ωn are
known.
The system has relative degree 3 and therefore is not SPR.
The transfer function of the reference model is specified as

ym (s) bm
= (9.544)
r (s) s − am

The reference model is SPR with relative degree 1. Since the relative degree of
the reference model is less than the relative degree of the plant, perfect tracking is not
possible. Adaptive control of systems with relative degrees greater than 1 is generally
much more difficult to handle since the model reference cannot be chosen to be SPR
[30].
The controller is given by

u = k y (t) y + kr (t) r (9.545)

k̇ y = γx ye (9.546)

k̇r = γr r e (9.547)

where e (t) = ym (t) − y (t).

k y (t) and kr (t) are initialized with k y (0) = −1 and kr (0) = 1.
The reference command signal is given by

r = r0 sin ω0 t (9.548)
√
where r0 = 1 and ω0 = 30 rad/s is the frequency at which the closed-loop transfer
function has zero phase margin.
The closed-loop system is unstable as shown in Fig. 9.34.
The underlying cause of instability is the lack of robustness of the closed-loop
system. Changing either the initial condition of k y (t) and or the frequency in the
reference command signal can result in stabilization of the closed-loop system if it
has a sufficient phase margin.
9.13 Optimal Control Modification for Linear Uncertain Systems … 317

The optimal control modification adaptive law can be designed to handle linear
systems with unmodeled dynamics by utilizing the linear asymptotic property.
Suppose the open-loop plant is expressed in general by the following transfer
function:
y (s) Z p (s)
 W p (s) = k p (9.549)
u (s) R p (s)

where k p is a high-frequency gain, and Z p (s) and R p (s) are monic Hurwitz poly-
nomials of degrees m p and n p , respectively, and n p − m p > 0 is the relative degree
of the plant.
The reference model is given by a transfer function

ym (s) Z m (s)
 Wm (s) = km (9.550)
r (s) Rm (s)

where km is a high-frequency gain, and Z m (s) and Rm (s) are monic Hurwitz poly-
nomials of degrees m m and n m , respectively, and n m − m m > 0 is the relative degree
of the reference model.
Let n p − m p > n m − m m . So the SPR condition is no longer possible to ensure
tracking of the reference model. Stability of the closed-loop system cannot also be
guaranteed with the standard MRAC.

1 5

0.5
0
0
u
x

−5
−0.5

−1 −10
0 5 10 0 5 10
t t

−0.5 4

−1
3
y

−1.5
k
k

2
−2

−2.5 1
0 5 10 0 5 10
t t

Fig. 9.34 Instability of MRAC by Rohrs counterexample

Suppose the adaptive controller is redesigned with the optimal control modifica-
tion as
u = k y y + kr r (9.551)

where  
k̇ y = γ y ye − νy 2 k y (9.552)
 
k̇r = γr r e − νr 2 kr (9.553)
318 9 Robust Adaptive Control

Using the linear asymptotic property of the optimal control modification, the
asymptotic value of the adaptive controller u can be computed as γ y → ∞ and
γr → ∞ in the limit. Then,
2ym − 2y
ū = (9.554)
ν
The asymptotic closed-loop transfer function can now be computed as

ȳ 2W p (s) Wm (s) 2km k p Z p (s) Z m (s)

= =   (9.555)
r ν + 2W p (s) Rm (s) ν R p (s) + 2k p Z p (s)

By a suitable selection of the modification parameter ν, the asymptotic closed-

loop transfer function can be designed to have the closed-loop stability. Once the
modification parameter ν is chosen, the adaptation rates γ y and γr can be selected
to be any reasonable values without compromising the closed-loop stability of the
adaptive laws.
Referring back to the first-order plant with the second-order unmodeled actuator
dynamics, the adaptive controller asymptotically tends to

2ym − 2y 2bm r 2y
ū = = − (9.556)
ν ν (s − am ) ν

as γ y → ∞ and γr → ∞.
Then, the asymptotic closed-loop transfer function is obtained as

ȳ 2bωn2 bm
 G (s) =    (9.557)
r 2bωn2
ν (s − am ) (s − a) s 2 + 2ζ ωn s + ωn2 + ν

Note that the closed-loop transfer function has a relative degree 4 while the transfer
function of the reference model has a relative degree 1. This prevents the output y (t)
from tracking ym (t).
The characteristic equation of the transfer function G (s) with an input time delay
is computed as

  2bωn2 −td s
s 3 + (2ζ ωn − a) s 2 + ωn2 − 2aζ ωn s − aωn2 + e =0 (9.558)
ν
Substituting s = jω yields

  2bωn2
− jω3 −(2ζ ωn − a) ω2 + ωn2 − 2aζ ωn jω−aωn2 + (cos ωtd − j sin ωtd ) = 0
ν
(9.559)
This results in two frequency equations
9.13 Optimal Control Modification for Linear Uncertain Systems … 319

2bωn2
− (2ζ ωn − a) ω2 − aωn2 + cos ωtd = 0 (9.560)
ν

  2bωn2
− ω3 + ωn2 − 2aζ ωn ω − sin ωtd = 0 (9.561)
ν
We then obtain the cross-over frequency and phase margin as functions of the
modification parameter ν in the following expressions:
! "
 2  4  2  2 2 4b2
ω + a + 4ζ ωn − 2ωn ω + ωn + 4a ζ − 2a ωn ω + a − 2 ωn4 = 0
6 2 2 2 2 2 2 2
ν
  2  (9.562)
−ω 3
+ ω − 2aζ ωn ω
φ = ωtd = tan−1 n
(9.563)
(2ζ ωn − a) ω2 + aωn2
<
Note that the cross-over frequency at zero phase margin ω0 = ωn2 − 2aζ ωn is
still the same as that without modification (see Sect. 8.4). However, this cross-over
frequency corresponds to a modification parameter νmin . By choosing a modification
ν > νmin that meets a certain phase margin requirement, the closed-loop adaptive
system can be made robustly stable.

Example 9.20 Consider Example 8.5 with the following parameters: a = −1, b =
1, am = −2, bm = 2, ωn = 5 rad/s,√ and ζ = 0.5. The cross-over frequency at zero
phase margin is computed to ω0 = 30 rad/s. The cross-over frequency and phase
margin as a function of the modification parameter ν are shown in Fig. 9.35.
The modification parameter ν corresponding to zero phase margin and 90o phase
margin are νmin = ν0 = 0.3226 and νmax = ν90o = 0.9482, respectively. So, the
adaptive systems with the unmodeled dynamics can be stabilized with the selection of
the modification parameter ν between these two values. The< cross-over frequencies
√
margin and 90o phase margin are ω0 = ωn2 − 2aζ ωn = 30
at the zero phase 7
7
aω2
rad/s and ω90o = − 2ζ ωn n−a = 25 6
rad/s, respectively. So, robustness is achieved
with the optimal control modification by reducing the cross-over frequency of the
closed-loop system.
Suppose a 45o phase margin is selected. Then, the cross-over frequency can be
computed from Eq. (9.563) as the root of the following equation:
 
ω3 + (2ζ ωn − a) ω2 + 2aζ ωn − ωn2 ω + aωn2 = 0

which yields ω45o = 3.7568 rad/s. Then, the modification parameter ν is computed
from Eq. (9.562) to be ν45o = 0.5924.
The asymptotic closed-loop transfer function with ν45o = 0.5924 is computed to
be
100
ν45o
G (s) =  
s 4 + 8s 3 + 42s 2 + 85 + 50
ν45o
s + 50 + 100
ν
320 9 Robust Adaptive Control

If r (t) is a constant command signal, then in the limit as t → ∞, G (0) =

2
2+ν45o
= 0.7715 as compared to the steady-state value of the transfer function of the
(s)
reference model yrm(s) = 1. Thus, tracking of the reference model is not achieved,
but in return the closed-loop adaptive system is robustly stable.

ω, rad/sec 15

10

0
0 0.2 0.4 0.6 0.8 1
ν

100

50
φ, deg

−50

−100
0 0.2 0.4 0.6 0.8 1
ν

Fig. 9.35 Asymptotic phase margin and cross-over frequency as functions of ν

For the simulations, the time step is chosen to be 1 × 10−6 sec to permit the
adaptation rates γx and γr to be set at 1√× 106 to simulate fast adaptation. The
reference command signal is r (t) = sin 30t. The asymptotic response ȳ (t) and
control signal ū (t) evaluated analytically agree very well with the simulation results
of y (t) and u (t), respectively, as shown in Fig. 9.36. The control signal u (t) exhibits
sharp spikes due to the adaptive gain kr (t) but otherwise tracks the asymptotic control
signal ū (t) very well. Thus, the linear asymptotic property of the optimal control
modification is demonstrated to be able to facilitate the stability analysis of linear
uncertain systems with unmodeled dynamics, time delay, or non-minimum phase
behaviors. In this study, the optimal control modification can handle systems with
relative degrees greater than 1 with relative ease.

1 _ 4 _
y y u u
0.5 2
|
u, u
y, y

|
0 0

−0.5 −2

−1 −4
0 5 10 0 5 10
t t

100 100

50 50
y

kr
k

0 0

−50 −50
0 5 10 0 5 10
t t

Fig. 9.36 Closed-loop response with optimal control modification (ν = 0.5924, γx = γr = 106 )
9.13 Optimal Control Modification for Linear Uncertain Systems … 321

Example 9.21 Consider the Rohrs’ counterexample problem in Example 8.6 [1].
The cross-over frequency and phase margin as a function of the modification para-
meter ν are shown in Fig. 9.37. The cross-over frequency 7 at the zero phase margin
√
and 90 phase margin are ω0 = 259 rad/s and ω90o = 229
o
31
rad/s, respectively.
The corresponding modification parameters are ν0 = 0.1174 and ν90o = 1.3394.
We select ν45o = 0.4256 corresponding to a 45o phase margin and a cross-over
frequency ω45o = 7.5156 rad/s for the optimal control modification adaptive laws.
The closed-loop response is shown in Fig. 9.38. The closed-loop adaptive system is
robustly stable.
All the three robust modification schemes, namely σ modification, e modification,
and optimal control modification exhibit minimum values of the modification para-
meters at which the plant in the Rohrs counterexample begins to stabilize. The σ and
e modification parameters can be found by trial-and-error. In contrast, the modifica-
tion parameter ν is found analytically by taking advantage of the linear asymptotic
property of the optimal control modification method.
30
ω, rad/sec

20

10

0
0 0.2 0.4 0.6 0.8 1
ν

100

50
φ, deg

−50
0 0.2 0.4 0.6 0.8 1
ν

Fig. 9.37 Asymptotic phase margin and cross-over frequency of Rohrs counterexample with
optimal control modification

0.4 4

0.3
2
0.2
u
y

0
0.1

0 −2
0 50 100 0 50 100
t t

0.5 1.5

0 1
y

r
k
k

−0.5 0.5

−1 0
0 50 100 0 50 100
t t

Fig. 9.38 Closed-loop response of Rohrs counterexample with optimal control modification
(ν = 0.4256, γx = γr = 1)
322 9 Robust Adaptive Control

9.14 Adaptive Control of Non-Minimum Phase Plants with

Relative Degree 1

Non-minimum phase plants generally pose more difficulty to model-reference adap-

tive control than minimum phase plants. Output feedback adaptive control generally
relies on the SPR property to ensure stability. Non-minimum phase plants do not
possess SPR transfer functions. Therefore, adaptive control for non-minimum phase
plants cannot be designed using the SPR property. The difficulty with adaptive control
design for non-minimum phase plants is due to the ideal property of model-reference
adaptive control which attempts to seek asymptotic tracking. This results in a pole-
zero cancellation in the right half plane for non-minimum phase systems which leads
to unbounded signals.
The linear asymptotic property of the optimal control modification can be used for
adaptive control of non-minimum phase plants. By modifying the standard MRAC,
the pole-zero cancellation in the right half plane can be prevented. This then results
in bounded tracking as opposed to asymptotic tracking. Closed-loop stability can be
obtained by a judicious choice of the modification parameter ν. In this section, we
will demonstrate an adaptive control method for non-minimum phase plants with a
relative degree 1 using the optimal control modification method [20].
Consider the following SISO plant:

ẋ = ax + bu + gz (9.564)

ż = hz + lx + mu (9.565)

y=x (9.566)

where z (t) is the unmeasurable state with internal dynamics, a unknown but all the
other parameters are known, and h < 0.
The objective is to design an output feedback adaptive controller to track the
following reference model:

Z m (s) bm r
ym = Wm (s) r = km r= (9.567)
Rm (s) s − am

where am < 0 and km = bm .

The transfer functions of the plant is expressed as

y Z p (s) b (s − h) + gm
= W p (s) = k p = (9.568)
u R p (s) (s − a) (s − h) − gl

where k p = b.
Note that Wm (s) is SPR with a relative degree 1. The plant also has a relative
degree 1 and is assumed to be stable. So, R p (s) = (s − a) (s − h) − gl is Hurwitz.
9.14 Adaptive Control of Non-Minimum Phase Plants … 323

The ideal output feedback adaptive controller is designed to be of the form

b−1 θ2∗ y ∗ b−1 θ3∗ u ∗

u ∗ = −b−1 θ1∗ y ∗ − − + b−1 bm r (9.569)
s−λ s−λ

where λ < 0 is a chosen parameter, and θ1∗ , θ2∗ , and θ3∗ are unknown constants.
u ∗ (s) is obtained explicitly as
 
b−1 −θ1∗ (s − λ) − θ2∗ y ∗ + b−1 bm (s − λ) r
u∗ = (9.570)
s − λ + b−1 θ3∗

Then, the ideal output y ∗ (s) is written as

 
Z p (s) −θ1∗ (s − λ) − θ2∗ y ∗ + bm (s − λ) r
∗
y = (9.571)
R p (s) s − λ + b−1 θ3∗

We consider the following cases:

9.14.1 Minimum Phase Plant

If the plant is minimum phase, then Z p (s) is Hurwitz.

Then, the pole-zero cancellation can take place in the left half plane. Therefore,
it follows that

Z p (s) = s − λ + b−1 θ3∗ (9.572)

This results in
θ3∗ = b (λ − h) + gm (9.573)

The ideal controller u ∗ (s) then becomes

 
∗ b−1 −θ1∗ (s − λ) − θ2∗ y ∗ + b−1 bm (s − λ) r
u = (9.574)
Z p (s)

The ideal output y ∗ (s) is then obtained as

 ∗ 
∗ −θ1 (s − λ) − θ2∗ y ∗ + bm (s − λ) r
y = (9.575)
(s − a) (s − h) − gl

This results in the following ideal closed-loop transfer function:

y∗ bm (s − λ)
= 2   (9.576)
r s − a + h − θ1∗ s + ah − gl − λθ1∗ + θ2∗
324 9 Robust Adaptive Control

We want the ideal closed-loop plant to track the reference model. So, the ideal
closed-loop transfer function must be equal to the reference model transfer function
Wm (s). Thus,

bm (s − λ) bm
  = (9.577)
s 2 − a + h − θ1∗ s + ah − gl − λθ1∗ + θ2∗ s − am

This leads to the following model matching conditions:

a + h − θ1∗ = λ + am (9.578)

ah − gl − λθ1∗ + θ2∗ = λam (9.579)

θ1∗ and θ2∗ are then obtained as

θ1∗ = a − am + h − λ (9.580)

θ2∗ = gl − ah + λ (a + h − λ) (9.581)

The adaptive controller is now established as

b−1 θ2 y b−1 θ3 u
u = −b−1 θ1 y − − + b−1 bm r (9.582)
s−λ s−λ

where θ1 (t), θ2 (t), and θ3 (t) are the estimates of θ1∗ , θ2∗ , and θ3∗ , respectively.
Let θ̃1 (t) = θ1 (t) − θ1∗ , θ̃2 (t) = θ2 (t) − θ2∗ , and θ̃3 (t) = θ3 (t) − θ3∗ . Then, the
output y (s) is expressed as
5 6
Wm (s) θ̃2 y θ̃3 u
y = Wm (s) r − θ̃1 y + + (9.583)
bm s−λ s−λ

Define the tracking error as e (t) = ym (t)−y (t). Then, the tracking error equation
is obtained as
ė = am e + Θ̃  Φ (t) (9.584)
   
where Θ̃ (t) = θ̃1 (t) θ̃2 (t) θ̃3 (t) and Φ (t) = φ1 (t) φ2 (t) φ3 (t) .
The adaptive law is given by

Θ̇ = −Γ Φ (t) e (9.585)

where
φ1 = y (9.586)

φ̇2 = λφ2 + y (9.587)

9.14 Adaptive Control of Non-Minimum Phase Plants … 325

φ̇3 = λφ3 + u (9.588)

Therefore, all signals are bounded, and the tracking error is asymptotically stable.

9.14.2 Non-Minimum Phase Plant

For non-minimum phase plants, because the standard MRAC attempts to seek as-
ymptotic tracking by performing a pole-zero cancellation in the right half plane, the
resulting adaptive controller will become unbounded. Thus, if the adaptive law can
be modified to seek only bounded tracking instead of asymptotic tracking, then this
would prevent a pole-zero cancellation in the right half plane. The plant then can be
stabilized.
For non-minimum phase plants, we consider two possible adaptive controllers.
1. We use the same adaptive controller in Eq. (9.582), but with the optimal control
modification adaptive law
 
Θ̇ = −Γ Φ (t) e − νΦ  (t) Θam−1 (9.589)

By invoking the linear asymptotic property of the optimal control modification

as Γ → ∞, we get
am (ym − y)
Θ  Φ (t) = (9.590)
ν
Then, the asymptotic linear controller tends to

am y [νbm − am Wm (s)] r
u= + (9.591)
νb νb
Comparing the ideal controller with the asymptotic linear controller, we see that
the adaptive controller does not attempt to cancel the unstable zero of W p (s)
since it has a stable pole at s = am due to the term Wm (s). Otherwise, the pole-
zero cancellation would take place in the right half plane since Z p (s) is unstable.
Therefore, the stability of the adaptive controller is no longer affected by the
non-minimum phase behavior of W p (s). The stability of the closed-loop plant is
then determined by a proper selection of the modification parameter ν such that
ν > 0 and the closed-loop transfer function is stable.
The asymptotic closed-loop transfer function is expressed as

y W p (s) [νbm − am Wm (s)]

= (9.592)
r νb − am W p (s)

As Γ → ∞, the equilibrium value of y (t) tends to

326 9 Robust Adaptive Control

W p (0) [νbm − am Wm (0)]

ȳ = r (9.593)
νb − am W p (0)

Thus, the tracking is bounded regardless whether or not W p (s) is minimum

phase. The closed-loop plant therefore is robustly stable with the optimal control
modification. However, poor tracking will result if ν is too large to guarantee the
closed-loop stability.
2. We use a simple adaptive controller for the first-order SISO plant as if the non-
minimum phase dynamics do not exist with z (t) = 0

u (t) = k y (t) y + kr r (9.594)

where kr = b−1 bm and k y (t) is computed by the optimal control modification

adaptive law  
k̇ y = γ y y e + νyk y bam−1 b (9.595)

By invoking the linear asymptotic property of the optimal control modification,

we get
am (y − ym )
ky y = (9.596)
νb
Then, the asymptotic linear controller tends to

am y [νbm − am Wm (s)] r
u= + (9.597)
νb νb
Note that this asymptotic linear controller is the same as that with the first adaptive
controller. Thus, even though both adaptive controllers are different, the closed-
loop plants for both controllers behave the same in the limit.
We now will formalize the proof of the optimal control modification for the non-
minimum phase plant with relative degree 1 described by Eqs. (9.564) and (9.565)
with a unknown but all the other parameters are known and h < 0. The adaptive
controller is given by Eqs. (9.594) and (9.595).
The closed-loop plant becomes

ẏ = am y + bm r + bk̃ y y + gz (9.598)
 
ż = hz + l + mk ∗y y + m k̃ y y + mkr r (9.599)

where k̃ y (t) = k y (t) − k ∗y and k ∗y = amb−a .

The tracking error equation is given by

ė = am e − bk̃ y y − gz (9.600)

We define the reference internal state dynamics as

9.14 Adaptive Control of Non-Minimum Phase Plants … 327
 
ż m = hz m + l + mk ∗y ym + mkr r (9.601)

Let ε (t) = z m (t) − z (t) be the internal state tracking error. Then, the internal
state tracking error equation is given by
 
ε̇ = hε + l + mk ∗y e − m k̃ y y (9.602)

Proof Choose a Lyapunov candidate function

 
V e, ε, k̃ y = αe2 + βε2 + αγ y−1 k̃ 2y (9.603)

 α >0 and β > 0.

where
V̇ e, ε, k̃ y is evaluated as
   
V̇ e, ε, k̃ y =2αam e2 − 2αg (z m − ε) e + 2βhε 2 + 2β l + mk ∗y eε − 2βm k̃ y yε
 
−1
+ 2ανb2 am k̃ y k̃ y + k ∗y y 2 ≤ −2α |am | e2 + 2α |g| z m  e
    
   
− 2β |h| ε2 + 2 αg + β l + mk ∗y  e ε + 2β |m| y k̃ y  ε
 −1   2  −1   ∗   
− 2ανb2 am  y2  
k̃ y  + 2ανb2 am  k  y2  
k̃ y  (9.604)
y

b2
Using the inequality 2 a b ≤ δ 2 a2 + δ2
, we get
 
V̇ e, ε, k̃ y ≤ −2α |am | e2 + 2α |g| z m  e − 2β |h| ε2
! "
   2 ε2

+ αg + β l + mk y ∗ 
δ1 e + 2
2
δ1
!  2 ε2 "
 
+ β |m| δ22 y2 k̃ y  + 2
δ2
   2
 
− 2ανb2 am−1  y2 k̃ y 
    
 
+ 2ανb2 am−1  k ∗y  y2 k̃ y  (9.605)

   2
 
Note that the negative definite term −2ανb2 am−1  y2 k̃ y  of the optimal
control modification can be made to dominate the positive definite term
 2  
 
β |m| δ22 y2 k̃ y  to enable V̇ e, ε, k̃ y to be negative definite.
   |αg+β (l+mk ∗y )|
Let c1 = 2α |am | − αg + β l + mk ∗y  δ12 , c2 = 2β |h| − δ12
− β|m|
δ22
,
 
2  −1  ανb2 |am−1 ||k ∗y |
c3 = 2ανb am − β |m| δ2 , and c4 =
2
c3
. Then,
328 9 Robust Adaptive Control

    2
 
V̇ e, ε, k̃ y ≤ −c1 e2 + 2α |g| z m  e − c2 ε2 − c3 y2 k̃ y  − c4 + c3 c42 y2
(9.606)
We note that y ≤ e + 2 e ym  + ym  . The ultimate boundsof ym
2 2 2

and zm  can be  −1 

 shown  −1 to be y∗m ≤−1c y r0 and z m  ≤ cz r0 where c y = am bm2 ,
 −1
cz = h mkr + h     
l + mk y am bm , and r0 = r . Let c5 = c1 − c3 c4 ,
c = ( α|g|cz +c3 c42 c y )r0
, and c = c c2 + c c2 c2 r 2 . Then,
6 c5 7 5 6 3 4 y 0

    2
 
V̇ e, ε, k̃ y ≤ −c5 (e − c6 )2 − c2 ε2 − c3 y2 k̃ y  − c4 + c7 (9.607)

ν, α, β, δ1 , and δ2 are
 chosen such that c2 > 0, c3 > 0, and c5 > 0. Then, it
follows that V̇ e, ε, k̃ y ≤ 0 if
$
c7
e ≥ c6 + =p (9.608)
c5
$
c7
ε ≥ =q (9.609)
c2
  
  c7
k̃ y  ≥ c4 +  2 = κ (9.610)
c3 p + c y r0

Thus, the closed-loop non-minimum phase system is stable with output feedback
adaptive control with the optimal control modification. The ultimate bound of e is
then obtained as
$
β
e ≤ p 2 + q 2 + γ y−1 κ 2 (9.611)
α

It should be noted that the aforementioned output feedback adaptive control de-
sign is feasible only if the knowledge of the plant transfer function is mostly known
with the exception of the parameter a. This requirement imposes a severe restriction
on the design of output feedback adaptive control for non-minimum phase plants
since in many cases such knowledge may not exist if the plants are uncertain. An-
other drawback with this approach is that the desired tracking performance is not
guaranteed. In practical situations, this poor tracking performance can be a major
issue.

Example 9.22 Consider the system

ẋ = ax + bu + gz
9.14 Adaptive Control of Non-Minimum Phase Plants … 329

ż = −z + u

y=x

where a < 0 is unknown but a = −2 for simulation purposes, and b = 1 is known.

The reference model is given by the transfer function Wm (s) with am = −1 and
bm = 1.
The open-loop transfer function is

s+1+g
W p (s) =
(s − a) (s + 1)

The system is minimum phase if g > −1. Consider a minimum phase plant with
g = 2. Then, we design the output feedback adaptive controller according to Eq.
(9.582). Let λ = −1, Γ = I , and r (t) = 1. Figure 9.39 shows the response of the
closed-loop plant which tracks the reference model very well in the limit as t → ∞.
Consider a non-minimum phase plant with g = −2. We use both adaptive con-
trollers given by Eqs. (9.582) and (9.594).
The asymptotic closed-loop transfer function is given by

y (s − 1) [νbm (s − am ) − am bm ]
=
r (s − am ) [ν (s − a) (s + 1) − am (s − 1)]

The steady-state closed-loop transfer function is equal to

ȳ bm (ν + 1)
=
r νa − am

For a = −2 and am = −1, the transfer function is stable for ν > 0.5. Choose
ν = 2 and γ y = 1. The steady-state closed-loop output is equal to ȳ (t) = −1.
Figure 9.40 shows the stable closed-loop responses of the non-minimum phase plant
with both the adaptive controllers. The adaptive controller 1 with the adaptive pa-
rameter Θ (t) has a faster response than the adaptive controller 2 with the adaptive
parameter k y (t). Both adaptive controllers tend to the same steady-state closed-loop
response.
The closed-loop plant tracks the reference model very poorly as expected, but
the response is stable and tends to an equilibrium value ȳ = −1 which agrees with
the analytical result from the steady-state closed-loop transfer function. The poor
tracking performance is typical of a non-minimum phase system with the output
generally exhibiting an opposite response to the input.

As can be seen from Example 9.22, while the optimal control modification can
stabilize a non-minimum phase plant, the tracking performance is quite unaccept-
able. The issue at hand is the requirement for a non-minimum phase plant to track
330 9 Robust Adaptive Control

1.5

m
y, y
0.5
y
ym
0
0 5 10 15 20
t

0.5
u

0
0 5 10 15 20
t

Fig. 9.39 Closed-loop response of minimum phase plant with MRAC

0.5
m
y, y

0
y − Controller 1
−0.5 y − Controller 2
y
m
−1
0 5 10 15 20
t

1.5

1
u

0.5 u − Controller 1
u − Controller 2
0
0 5 10 15 20
t

Fig. 9.40 Closed-loop response of non-minimum phase plant with optimal control modification

a minimum phase reference model with the same relative degree. This is a very de-
manding and perhaps unrealistic requirement. As a consequence, MRAC attempts to
seek asymptotic tracking which results in an unstable pole-zero cancellation. If the
reference model could be redesigned so that the unstable pole-zero cancellation can-
not occur while the tracking performance can still be met, then the output feedback
adaptive control design would be more acceptable. One such approach is to design
an observer state feedback adaptive control using the Luenberger observer design.
Consider a MIMO system as
      
ẋ1 A11 A12 x1 B1
ẋ = Ax + Bu ⇔ = + u (9.612)
ẋ2 A21 A22 x2 B2
 
  x1
y = C x = C1 0 = C1 x1 (9.613)
x2
9.14 Adaptive Control of Non-Minimum Phase Plants … 331

where x1 (t) ∈ R p , x2 (t) ∈ Rn− p , u (t) ∈ Rm , y (t) ∈ R p with p > m, A11 ∈ R p× p

is unknown, A12 ∈ R p×(n− p) and A21 ∈ R(n− p)× p are known, A22 ∈ R(n− p)×(n− p) is
known and assumed to be Hurwitz, B1 ∈ R p×m and B2 ∈ R(n− p)×m are known, and
C1 ∈ R p× p is known and has full rank. We assume the pair (A, B) is controllable
and the pair (A, C) is observable. This partitioned form of the plant can be used
advantageously to design an output feedback adaptive control. The system expressed
in this form is sufficiently general for many practical applications. For example,
systems with unmodeled dynamics could be expressed in this form with x1 (t) being
the plant state variable and x2 (t) being the unmodeled state variable.
The Kalman-Yakubovich lemma [30] can be used to determine if the transfer
function matrix G (s) = C (s I − A)−1 B is SPR. If A is Hurwitz and there exist
P = P  ∈ Rn×n > 0 and Q = Q  > 0 ∈ Rr ×n such that

P A + A P = −Q (9.614)

B P = C (9.615)

then G (s) is SPR.

We assume that A = A0 + ΔA where A0 is known and ΔA is a small un-
known perturbation of A0 due to the uncertainty in A11 . The transfer function
G (s) = C (s I − A)−1 B is assumed to be non-SPR. The Luenberger observer design
constructs an observer state-space model of the plant as
 
x̂˙ = Â x̂ + L y − ŷ + Bu (9.616)

where x̂ (t) is the observer state which estimates the plant state x (t), Â (t) is the
estimate of A, ŷ (t) = C x̂ (t) is the observer output which estimates the plant output
y (t), and L is the Kalman filter gain matrix computed using A0 .
A full-state feedback controller can be designed to enable the output y (t) to track a
reference command signal r (t). For example, we can use the LQR method to design
the full-state feedback controller using the following cost function for tracking a
constant reference command signal r (t):
,
1 tf  
J = lim (C x − r ) Q (C x − r ) + u  Ru dt (9.617)
t f →∞ 2 0

Then, the control gains can be computed as

K x∗ = −R −1 B  W (9.618)
 −1 
K r = −R −1 B  A − W B R −1 B  C Q (9.619)

where W is the solution of the Riccati equation

332 9 Robust Adaptive Control

W A + A W − W B R −1 B  W + C  QC = 0 (9.620)

Then, a reference model is constructed from this LQR design with Am = A+ B K x∗

and Bm = B K r .
If A is unknown, then we can design the following adaptive controller:

u = K x (t) x̂ + K r r (9.621)

where the observer state x̂ (t) replaces the plant state.

The tracking error is defined as e (t) = xm (t) − x̂ (t). We also define the state
estimation error as e p (t) = x (t) − x̂ (t). Then, the error equations are

ė = Am e − LCe p − Ã x̂ − B K̃ x x̂ (9.622)

 
ė p = A p + ΔA e p − Ã x̂ (9.623)

where à (t) =  (t) − A, K̃ x (t) = K x (t) − K x∗ , A p = A0 − LC, and A p + ΔA is

Hurwitz by a suitable choice of L.
Note that the state estimation error signal e p (t) is generally not available, but for
the class of MIMO systems under consideration, it can be constructed. Since C1 is
invertible, then x1 (t) can be constructed from y (t). Let z (t) = x2 (t) be the internal
state, then z (t) can be computed from the following equation:

ż = A22 z + A21 C1−1 y + B2 u (9.624)

with z (0) = z 0 and A21 , A22 , and B2 known.

 
Then, the plant state can be constructed as x (t) = y  (t) C1− z  (t) .
A stable adaptation of the non-SPR plant of Eqs. (9.612) and (9.613) can be
achieved by the optimal control modification adaptive laws based on the observer
state x̂ (t) and the constructed state x (t) from the output y (t) and the internal state
z (t)  
K̇ x = Γx x̂ e P + ν x̂  K x B  P A−1
m B (9.625)
 
Â˙  = Γ A x̂ e P + ep W + η x̂  Â P A−1
m (9.626)

where P = P  > 0 and W = W  > 0 are solutions to the Lyapunov equations

P Am + A
m P = −Q (9.627)

W A p + Ap W = −R (9.628)

with Q = Q  > 0 and R = R  > 0.

9.14 Adaptive Control of Non-Minimum Phase Plants … 333

The stability of the observer output feedback adaptive control using the optimal
control modification is provided as follows:
Proof Choose a Lyapunov candidate function
     
V e, e p , K̃ x , Ã = e Pe + ep W e p + trace K̃ x Γx−1 K̃ x + trace ÃΓ A−1 Ã
  (9.629)
Then, V̇ e, e p , K̃ x , Ã is evaluated as
 
V̇ e, e p , K̃ x , Ã = − e Qe − ep R̄e p − 2e P LCe p + 2ν x̂  K x B  P A−1
m B K̃ x x̂
   
+ ην x̂  Â P A−1   + 2c3 e e p 
2
m à x̂ ≤ −c1 e − c2 e p
2

 2   2  2
 
− νc4 x̂   K̃ x  − c5 + νc4 c52 x̂ 
 2   2  2
 
− ηc6 x̂   Ã − c7 + ηc6 c72 x̂  (9.630)

 
where R̄ = R − W ΔA − ΔA W , c1 = λmin (Q), c2 = λmin R̄ , c3 = P LC,
  m B
 K x∗ B  P A−1  
c4 = λmin B  A− −1
m Q Am B , c5 = c4
, c6 = λmin A− −1
m Q Am , c7 =
m 
ΔA P A−1
.
c6
 2
We utilize the inequality 2 a b ≤ a2 + b 2
and also note that x̂  ≤

e + 2 e xm  + xm  . Then, V̇ e, e p , K̃ x , Ã is bounded by
2 2

    
V̇ e, e p , K̃ x , Ã ≤ − c1 − c3 − νc4 c52 − ηc6 c72 e2 + 2 νc4 c52
  2  2   2  2
 
+ηc6 c72 e xm  − (c2 − c3 ) e p  − νc4 x̂   K̃ x  − c5 − ηc6 x̂ 
  2  
 
 Ã − c7 + νc4 c52 + ηc6 c72 xm 2 (9.631)

Note that the ultimate bound of xm  can be expressed as xm  ≤ cx r0 . Let
c8 = c1 − c3 − νc4 c52 − ηc6 c72 , c9 = ( 4 5 c8 6 7 ) x 0 , c10 = c2 − c3 , and c11 =
νc c2 +ηc c2 c r

   
c8 c92 + νc4 c52 + ηc6 c72 cx2 r02 . Then, V̇ e, e p , K̃ x , Ã ≤ 0 outside the compact set

;
S = e (t) ∈ Rn , e p (t) ∈ Rn , K̃ x (t) ∈ Rn×m , Ã (t) ∈ Rn×n : c8 (e − c9 )2
  #
 2  2 
 
2  2 
 
2
+ c10 e p  + νc4 x̂   K̃ x  − c5 + ηc6 x̂   Ã − c7 ≤ c11 (9.632)

by choosing L, Q, R, ν, and η appropriately such that c8 > 0 and c10 > 0.

334 9 Robust Adaptive Control

By setting ν = 0 and η = 0, werecover the standard  MRAC. Then, it can be

shown by Barbalat’s lemma that V̇ e, e p , K̃ x , Ã is uniformly continuous since
 
V̈ e, e p , K̃ x , Ã is bounded. It follows that e (t) → 0 and e p → 0 as t → ∞.

Note that if the reference model is not derived from the ideal controller of the
non-SPR plant, the standard MRAC will not be able to stabilize the plant. On the
other hand, the optimal control modification adaptive law can handle the mismatch
between the plant and the reference model. Therefore, for a stable adaptation using
the standard MRAC for output feedback adaptive control, the reference model must
be established from the ideal controller design of the non-SPR plant. The mismatch
between the reference model and the non-SPR plant causes MRAC to continue to
seek a high-gain control in order to achieve asymptotic tracking. This would lead to
instability.
Suppose a reference model is specified as

ẋm = A∗m xm + Bm r (9.633)

where A∗m is not established from the non-SPR plant. Then, the model matching
condition cannot be satisfied since there exists no solution of K x∗ . To show this, we
see that suppose K x∗ exists and can be solved using the pseudo-inverse of B with
m < p < n as
 −1   ∗ 
K x∗ = B  B B Am − A (9.634)

But  −1   ∗ 
A + B K x∗ = A + B B  B B Am − A = A∗m (9.635)

For the standard MRAC, the tracking error equation in the presence of the
mismatch between the reference model and the non-SPR plant is established as
 
ė = A∗m e + A∗m − Am x̂ − LCe p − Ã x̂ − B K̃ x x̂ (9.636)

Because the optimal control modification only seeks bounded tracking, so the
model matching condition is not satisfied. Using the linear asymptotic property of
the optimal control modification, the asymptotic value of K x (t) and  (t) can be
computed from Eqs. (9.625) and (9.626) by letting Γ → ∞. Then, K x (t) → K̄ x
and  (t) → Ā for a constant reference command signal r (t). From the linear
asymptotic property, we get

1   ∗− −1 
K̄ x x̂ = − B Am P B B Pe (9.637)
ν
1  
Ā x̂ = − P −1 A∗ Pe + W e p (9.638)
η m
9.14 Adaptive Control of Non-Minimum Phase Plants … 335

where P = P  > 0 now solves the Lyapunov equation

P A∗m + A∗
m P = −Q (9.639)

We redefine the estimation errors as K̃ x (t) = K x (t) − K̄ x and à (t) =  (t) − Ā.
Then, the error equations are established as
 
1   ∗− −1  1 −1 ∗
ė = A∗m + B B Am P B B P + P Am P e
ν η
! "
1 −1 ∗
− LC − P Am W e p + A∗m x̂ − Ã x̂ − B K̃ x x̂ (9.640)
η
! "
1 −1 ∗ 1 −1 ∗
ė p = P Am Pe + A p + P Am W e p + A x̂ − Δ Ã x̂ (9.641)
η η

The standard MRAC results in instability due to the mismatch between the refer-
ence model and the non-SPR plant if A + B K x∗ = Am = A∗m and P Am + A m P ≮ 0,
whereas a stable adaptation of the plant can be achieved with the optimal control
modification.
Proof Choose thesame Lyapunov  candidate function in Eq. (9.627). Then, for the
standard MRAC, V̇ e, e p , K̃ x , Ã is evaluated as
     ∗ 
V̇ e, e p , K̃ x , Ã = e P A∗m + A∗ 
m P e + e P Am − Am x̂
 
+x̂  A∗ 
m − Am Pe
−2e P LCe p − ep R̄e p (9.642)

Upon substituting x̂ (t) = xm (t) − e (t), we get

   
V̇ e, e p , K̃ x , Ã = e P Am + A   ∗   
m P e + 2e P Am − Am x m − 2e P LCe p − e p R̄e p
(9.643)

Note that P Am + A
m P is not necessarily negative definite. Therefore,
      ∗   2 
V̇ e, e p , K̃ x , Ã ≤  P Am + A  
m P e + P Am − Am
2  e + xm 2
  2   2
+ c3 e2 + e p  − c2 e p  (9.644)
 
Thus, V̇ e, e p , K̃ x , Ã  0. Therefore, the tracking error is unbounded and the
closed-loop system is unstable.
336 9 Robust Adaptive Control
 
On the other hand, for the optimal control modification, V̇ e, e p , K̃ x , Ã is
evaluated as
  ;  −1  −  #
1 1
V̇ e, e p , K̃ x , Ã = e − Q + P B B  A∗− m P B + B  ∗−
A m P B B 
P e
η ν
 
1  ∗ 
+ 2e P A∗m xm − 2e P LC − Am + P A∗m P −1 W + A W e p
η
   
1
+ e
p − R̄ + W P
−1 ∗
Am + A∗m P −1 W e p + 2e p W Ax m
η
+ 2ν x̂  K x B  P A−1   −1
m B K̃ x x̂ + ην x̂ Δ Â P Am Δ Ã x̂ (9.645)

Note that P −1 A∗ ∗ −1

m + Am P = −P −1 Q P −1 . Then,
   2    
V̇ e, e p , K̃ x , Ã ≤ −c12 e2 − c13 e p  + 2c14 e e p  + 2c15 e + 2c16 e p 
 2   2  2  2   2  2
   
−νc4 x̂   K̃ x  − c5 + νc4 c52 x̂  − ηc6 x̂   Ã − c7 + ηc6 c72 x̂ 
(9.646)
  −1   ∗− −  
where c12 = λmin η1 Q − ν1 P B B  A∗− m PB + B Am P B B P ,
   
 
c13 = λmin R̄ − η1 W P −1 Q P −1 W , c14 = P LC + η1 Q P −1 W + A W , c15 =
 
 P A∗  cx r0 , c16 = W A cx r0 , and c4,5,6,7 are defined previously.
m
 2
Further simplification using x̂  ≤ e2 + 2 e xm  + xm 2 leads to
    2  2
V̇ e, e p , K̃ x , Ã ≤ −c17 (e − c18 )2 − c19 e p  − c20 − νc4 x̂ 
  2  2   2
   
 K̃ x  − c5 − ηc6 x̂   Ã − c7 + c21 (9.647)

where c17 = c12 − c14 − νc4 c52 − ηc6 c72 , c18 = 15 ( 4 5c17 6 7 ) x 0 , c19 = c13 − c14 ,
c + νc c2 +ηc c2 c r
 
c20 = cc16 , and c21 = c17 c18
2
+ c19 c20
2
+ νc4 c52 + ηc6 c72 cx2 r02 .
19  
Choose L, Q, R, ν, and η such that c17 > 0 and c19 > 0. Then, V̇ e, e p , K̃ x , Ã ≤ 0
outside a compact set. Therefore, the closed-loop system is uniformly ultimately
bounded with the optimal control modification.

Using the linear asymptotic property of the optimal control modification,
 L, Q,ν
and η can also be chosen such that the closed-loop matrix formed by ẋ (t) , x̂˙ (t)
9.14 Adaptive Control of Non-Minimum Phase Plants … 337
⎡  −1  ⎤
A 1
BB  A∗− PB B P
⎢ 1 −1 ∗
ν  m −1  ⎥
Ac = ⎣ LC − η P Am W −LC + ν1 B B  A∗−
m PB B P ⎦ (9.648)
+ η1 P −1 A∗
m (P + W )

is Hurwitz.
To demonstrate the observer state feedback adaptive control, we return to Example
9.22.

Example 9.23 The non-minimum phase plant in Example 9.22 is defined by the
following matrices:
   
a g b  
A= , B= ,C= 10
0 −1 1

where a = −2, b = 1, and g = −2.

The Kalman filter gain is computed as
 
0.4641
L=
−0.2679

Using Q = 100 and R = 1, the LQR control gains are obtained as

 
K x∗ = −2.7327 −5.4654 , kr = −9.8058

Choose Γx = I , Γ A = 0.1I , Q = I , and R = I . Figure 9.41 shows the closed-

loop response with the standard MRAC of the plant output y (t) which tracks the
redesigned reference model very well. Notice that the reference model now has a
non-minimum phase behavior as evidenced by the reversal in the initial response.
All the control and adaptive parameter signals are bounded.
Suppose, instead of the reference model computed from the LQR, we use the ideal
minimum phase first-order reference model described by the following matrices:
   
am 0 b
A∗m = , B= m
0 −1 0

Figure 9.42 shows the unstable closed-loop response with the standard MRAC
which no longer tracks the reference model. The adaptive parameters are unbounded
and drifting as t → ∞.
The optimal control modification is then used with ν = 0.13 selected. Figure 9.43
shows the closed-loop response with the optimal control modification which is able
to track the ideal reference model very well. All the control and adaptive parameter
signals are bounded.

338 9 Robust Adaptive Control

Fig. 9.41 Closed-loop output response to LQR non-minimum phase reference model with observer
state feedback MRAC

Fig. 9.42 Closed-loop output response to ideal minimum phase reference model with observer
state feedback MRAC

Example 9.23 illustrates the significance of the reference model in adaptive con-
trol of non-minimum phase plants. If the reference model can be redesigned to be
non-minimum phase, then the standard MRAC can achieve asymptotic tracking, al-
beit with a non-minimum phase behavior. Otherwise, even with the observer state
feedback adaptive control design, if the reference model is minimum phase, instabil-
ity will still result with the standard MRAC. On the other hand, the optimal control
modification can produce bounded tracking for both types of reference models.
We have simplified the analysis of the Luenberger observer state feedback adaptive
control. There are some excellent textbooks that cover in-depth the subject of output
feedback adaptive control. The reader is referred to the textbooks by Ioannu [5] and
by Narendra and Annaswamy [35]. The textbook by Lavretsky and Wise presents a
9.14 Adaptive Control of Non-Minimum Phase Plants … 339

Fig. 9.43 Closed-loop output response to ideal minimum phase reference model with optimal
control modification

Luenberger observer state feedback adaptive control method based on squaring up

the control input matrix to satisfy the SPR condition for MIMO systems [36]. This
section illustrates the general challenges with adaptive control design for systems
with only output information.

9.15 Summary

Robust adaptive control is a well-researched topic. The primary objective of robust

adaptive control is to improve robustness of adaptive control systems to the parame-
ter drift, time delay, unmodeled dynamics, and other destabilizing effects. Various
robust modification schemes have been developed to ensure boundedness of adaptive
parameters in the last several decades. The principle of robust modification is based
on two central themes: (1) limiting adaptive parameters and (2) adding damping
mechanisms to model-reference adaptive control. The robustness issues with the pa-
rameter drift, non-minimum phase behaviors, time delay, unmodeled dynamics, and
fast adaptation are largely ameliorated with these robust modification schemes, but
are not entirely eliminated if the nature of the uncertainty is not completely known.
The standard techniques for robust adaptive control include the dead-zone method,
the projection method, the σ modification, and the e modification. Normalization and
covariance adaptive gain adjustments are two other techniques common to all adap-
tive control methods. More recent methods include the optimal control modification,
the adaptive loop recovery modification, and the L1 adaptive control, and many oth-
ers which have been developed in the recent years. These new methods add more
richness to this fertile field of endeavor. In particular, the optimal control modifi-
cation is an adaptive optimal control method that seeks to minimize a L2 tracking
error norm bounded away from the origin to provide improved robustness. The linear
asymptotic property of the optimal control modification for linear uncertain plants
340 9 Robust Adaptive Control

results in asymptotic linear plants in the limit which can be exploited in adaptive
control system design and analysis using many existing linear control techniques.
For systems with input uncertainty, the bi-objective optimal control modification
provides an adaptation mechanism based on both the tracking error and the predictor
error using a predictor model that approximates the plant dynamics. For systems with
slow first-order actuator dynamics, a singular perturbation method is developed by
scaling the adaptive law to allow the plant to track a reference model. An output feed-
back adaptive control approach based on the optimal control modification method
is shown to be able to stabilize first-order plants with unmodeled dynamics having
relative degree greater than 1 by taking advantage of the linear asymptotic prop-
erty. This approach is extended to non-minimum phase plants with relative degree 1.
For these non-minimum phase plants, the optimal control modification adaptive law
can be designed by a suitable selection of the modification parameter to ensure the
closed-loop stability by preventing an unstable pole-zero cancellation typically oc-
curred with the standard MRAC. The potential degraded tracking performance of the
output feedback adaptive control for non-minimum phase systems can be addressed
by observer state feedback adaptive control using the Luenberger observer design.

9.16 Exercises

1. Consider a time-delay second-order SISO system

ÿ + 2ζ ωn ẏ + ωn2 y = bu (t − td )

where b = 1, td = 13 s, and ζ and ωn are unknown but their actual values are
−0.5 and 1 rad/s, respectively.
The system is designed to track a second-order reference model

ÿm + 2ζm ωm ẏm + ωm2 ym = bm r (t)

where ζm = 0.5, ωm = 2 rad/s, bm = 4, and r (t) = 1, with an adaptive controller

u = K x (t) x + kr r
   
where x (t) = y (t) ẏ (t) and K x (t) = k p (t) kd (t) .
a. Calculate the fixed-gain values of k pmin and kdmin to achieve a phase margin
of 60o and a time-delay margin of 1/3 s.
b. Define a convex set described by an ellipse that contains k p (t) and kd (t)
! "2 ! "2
  kp kd
g k p , kd = + −1≤0
a b
9.16 Exercises 341

where a and b are to be determined from k pmin and kdmin . Design a projection
method for the adaptive controller to ensure robustness in the presence of
time delay. Write down the adaptive law. Implement the adaptive controller
in Simulink using the following information: y (0) = 0, ẏ (0) = 0, K x (0) =
0, and Γx = 0.2I with a time step Δt = 0.001 s. Plot the time histories of
y (t), u (t), k p (t), and kd (t) for t ∈ [0, 600] s. What happens when the
projection method is removed from the adaptive law?
2. Implement in Simulink the σ modification and e modification for the Rohrs’
counterexample with the reference command

r = 0.3 + 1.85 sin 16.1t

using the same initial conditions k y (0) and kr (0) with γx = γr = 1, σ = 0.2,
and μ = 0.2 and Δt = 0.001 s. Plot the time histories of y (t), u (t), k y (t),
and kr (t) for t ∈ [0, 100] s. Experiment with different values of σ and μ and
determine by trial-and-error the values of σ and μ at which the system begins
to stabilize.
3. Consider a first-order SISO system

ẋ = ax + bu + w

where a is unknown, b is known, and w is an unknown disturbance.

To prevent the parameter drift, the σ modification is used in an adaptive regulator
design
u = k x (x) x
 
k̇ x = −γx x 2 b + σ k x

Suppose x (t) is a sinusoidal response where x (t) = sin t.

a. Derive the general time-varying disturbance w (t) that produces the given
response x (t) in terms of a, b, γx , σ , and k x (0). Let a = 1, b = 1, γx = 10,
σ = 0.1, x (0) = 0, and k x (0) = 0. Express w (t).
b. Implement in Simulink the control system with a time step Δt = 0.001 sec.
Plot the time histories of x (t), u (t), w (t), and k x (t) for t ∈ [0, 20] sec.
c. Repeat part (b) with the standard MRAC by setting σ = 0. Does the system
exhibit the parameter drift?
4. Consider a linear system
ẋ = Ax + Bu

y = Cx

Design a reference model for tracking the output y (t) with a constant refer-
ence command r (t) using the optimal control approach and the following cost
342 9 Robust Adaptive Control

function:
,
1 tf  
J = lim (C x − r ) Q (C x − r ) + u  Ru dt
t f →∞ 2 0

Derive the expressions for the optimal control gain matrices K x and K r for the
closed-loop system
ẋ = (A + B K x ) x + B K r r

Given    
1 2 2
ẋ = + u
1 −1 1
 
y= 10

r = sin t − 2 cos 4t − 2e−t sin2 4t

Implement in Simulink the control system. Let Q = q and R = q1 . Determine

7 :
t
a suitable value of q, K x , and K r such that t1f 0 f (y − r )2 dt ≤ 0.05 for
 
t ∈ [0, 10] s. Initialize with x (0) = −2 1 . Plot the time histories of y (t)
and r (t) on the same plot, and e (t) = y (t) − r (t).
5. Consider a time delay second-order SISO system

ÿ − ẏ + y = u (t − td )

where td = 0.1 s is a time delay.

The unstable open-loop plant is stabilized with an adaptive controller

u = Kx x
   
where x (t) = y (t) ẏ (t) ∈ R2 and K x (t) = k p (t) kd (t) , to achieve an
ideal reference model
ÿm + 6 ẏm + ym = 0

a. Express the optimal control modification adaptive law for K x (t). Let Γ →
∞ and Q = I , calculate the equilibrium values of K x (t) as a function of
the modification parameter ν.
b. Determine numerically the value of the modification parameter ν to achieve
the maximum time-delay margin to within 0.001. Compute the equilibrium
values of K x (t) corresponding to this modification parameter ν. Implement
the adaptive controller in Simulink with this modification parameter using
the following information: Γ = 10I , y (0) = 1, ẏ (0) = 0, and K x (0) = 0
with a time step Δt = 0.001 s. Plot the time histories of x (t), u (t), and
K x (t) for t ∈ [0, 10] s.
9.16 Exercises 343

c. Increase the adaptation rate to Γ = 10000I . Repeat the simulations with a

time step Δt = 0.0001 s. Compare the steady-state values of K x (t) at 10 s
with those results computed in part (b).

6. Consider a first-order SISO plant as

 
ẋ = ax + b u + θ ∗ x + w

with a = −1, b = 1, θ ∗ = 2, and

w = cos t + 4 sin t − 4e−t sin t − (cos 2t + 2 sin 2t) sin t

This disturbance will cause a parameter drift when the standard MRAC is used
in a regulator design.
An adaptive controller is designed as

u = kr r − θ (t) x − ŵ (t)

to enable the plant to follow a reference model

ẋm = am xm + bm r

where am = −2, bm = 2, and r (t) = 1.

a. Calculate kr . Express the adaptive loop recovery modification adaptive laws
for θ (t) and ŵ (t) using a modification parameter η = 0.1.
b. Implement the adaptive controller in Simulink using the following informa-
tion: x (0) = 0, θ (0) = 0, d̂ (0) = 0, and γ = γd = 100 with a time step
Δt = 0.001 s. Plot the time histories of x (t), u (t), θ (t), and d (t) and d̂ (t)
together on the same plot for t ∈ [0, 100] s.
7. Consider a second-order SISO plant

ÿ + 2ζ ωn ẏ + ωn2 y = bu (t − td )

where ζ = −0.5 and ωn = 1 rad/s are unknown, b = 1 is known, and td is a

known time delay.
Design an adaptive controller using the normalized MRAC without the projection
method to allow the plant to follow a reference model

ÿm + 2ζm ωm ẏm + ωm2 ym = bm r (t)

where ζm = 3, ωm = 1, bm = 1, and r (t) = r0 sin t.

344 9 Robust Adaptive Control

a. Implement the adaptive controller in Simulink using the following infor-

mation: td = 0, x (0) = 0, K x (0) = 0, and Γx = 100I with a time step
Δt = 0.001 s for the standard MRAC by setting R = 0 with r0 = 1 and
r0 = 100. Plot the time histories of y (t) and ym (t), e1 (t) = ym (t) − y (t),
u (t), and K x (t) for t ∈ [0, 100] s. Comment on the effect of the amplitude
of the reference command signal on MRAC.
b. Repeat part (a) for the normalized MRAC with R = I and r0 = 100 for
td = 0 and td = 0.1 s. Comment on the effect of normalization on the
amplitude of the reference command signal and time delay.
8. For the Rohrs’ counterexample, design a standard MRAC with the covariance
adjustment method without the projection method.

a. Implement the adaptive controller in Simulink using the following infor-

mation: y (0) = 0, k y (0) = −0.65, kr (0) = 1.14, γ y (0) = γr (0) = 1,
and η = 5 with a time step Δt = 0.01 s. Plot the time histories of k y (t),
kr (t), γ y (t), and γr (t) for t ∈ [0, 100] s. Note: plot γ y (t) and γr (t) with
the logarithmic scale in the y axis for better visualization.
b. Repeat part (a) with t ∈ [0, 1000] s. Do k y (t) and kr (t) reach their equilib-
rium values or do they exhibit a parameter drift behavior?

9. Consider a first-order SISO plant

 
ẋ = ax + bλ u (t − td ) + θ ∗ φ (x) + w

where a = −1 and b = 1 are known, λ = −1 and θ ∗ = 0.5 are unknown,

but the sign of λ is known, φ (x) = x 2 , td = 0.1 s is a known time delay, and
w (t) = 0.02 + 0.01 cos 2t.
The reference model is given by

ẋm = am xm + bm r

where am = −2, bm = 2, and r (t) = sin t.

a. Design an adaptive controller using the standard tracking error-based opti-
mal control modification method. Express the adaptive laws.
b. Implement the adaptive controller in Simulink using the following informa-
tion: x (0) = k x (0) = kr (0) = θ (0) = 0 and γx = γr = γθ = 20 with a
time step Δt = 0.001 sec for the standard MRAC with ν = 0 and for the
optimal control modification with ν = 0.2. Plot the time histories of x (t)
and xm (t) on the same plot, u (t), k x (t), kr (t), and θ (t) for t ∈ [0, 60] sec.
9.16 Exercises 345

10. For Exercise 9.9, suppose λ is completely unknown.

a. Design an adaptive controller using the bi-objective optimal control modi-
fication method. Express the adaptive laws.
b. Implement the adaptive controller in Simulink using γλ = γw = 20, η = 0,
and the rest of the information in Exercise 9.9 along with the initial conditions
λ̂ (t) = 1 and ŵ (t) = 0. Plot the time histories of x (t) and xm (t) on the
same plot, u (t), k x (t), kr (t), θ (t), λ̂ (t) and ŵ (t) and w on the same plot
for t ∈ [0, 60] s.
c. Comment on the results of Exercises 9.9 and 9.10. Which method seems to
work better?

References

1. Rohrs, C. E., Valavani, L., Athans, M., & Stein, G. (1985). Robustness of continuous-time
adaptive control algorithms in the presence of unmodeled dynamics. IEEE Transactions on
Automatic Control, AC-30(9), 881–889.
2. Ioannou, P., & Kokotovic, P. (1984). Instability analysis and improvement of robustness of
adaptive control. Automatica, 20(5), 583–594.
3. Ioannu, P. A. and Sun, J., (1996). Robust Adaptive Control, Prentice-Hall, Inc., 1996
4. Narendra, K. S., & Annaswamy, A. M. (1987). A new adaptive law for robust adaptation
without persistent excitation. IEEE Transactions on Automatic Control, AC-32(2), 134–145.
5. Ioannu, P. A., & Sun, J. (1996). Robust adaptive control. Upper Saddle River: Prentice-Hall,
Inc.
6. Calise, A. J., & Yucelen, T. (2012). Adaptive loop transfer recovery. AIAA Journal of Guidance,
Control, and Dynamics, 35(3), 807–815.
7. Lavretsky, E. (2009). Combined/composite model reference adaptive control. IEEE Transac-
tions on Automatic Control, 54(11), 2692–2697.
8. Hovakimyan, N., & Cao, C. (2010). L1 Adaptive control theory: Guaranteed robustness with
fast adaptation. Philadelphia: Society for Industrial and Applied Mathematics.
9. Nguyen, N. (2012). Optimal control modification for robust adaptive control with large adaptive
gain. Systems and Control Letters, 61(2012), 485–494.
10. Yucelen, T., & Calise, A. J. (2011). Derivative-free model reference adaptive control. AIAA
Journal of Guidance, Control, and Dynamics, 34(4), 933–950.
11. Nguyen, N., Krishnakumar, K., & Boskovic, J. (2008). An optimal control modification to
model-reference adaptive control for fast adaptation. In AIAA Guidance, Navigation, and Con-
trol Conference, AIAA 2008-7283, August 2008.
12. Campbell, S., Kaneshige, J., Nguyen, N., & Krishnakumar, K. (2010). An adaptive control
simulation study using pilot handling qualities evaluations. In AIAA Guidance, Navigation,
and Control Conference, AIAA-2010-8013, August 2010.
13. Campbell, S., Kaneshige, J., Nguyen, N., & Krishnakumar, K. (2010). Implementation and
evaluation of multiple adaptive control technologies for a generic transport aircraft simulation.
In AIAA Infotech@Aerospace Conference, AIAA-2010-3322, April 2010.
346 9 Robust Adaptive Control

14. Hanson, C., Johnson, M., Schaefer, J., Nguyen, N., & Burken, J. (2011). Handling qualities
evaluations of low complexity model reference adaptive controllers for reduced pitch and roll
damping scenarios. In AIAA Guidance, Navigation, and Control Conference, AIAA-2011-6607,
August 2011.
15. Hanson, C., Schaefer, J., Johnson, M., & Nguyen, N. Design of Low Complexity Model Ref-
erence Adaptive Controllers, NASA-TM-215972.
16. Nguyen, N., Hanson, C., Burken, J., & Schaefer, J. (2016). Normalized optimal control modi-
fication and flight experiments on NASA F/A-18 Aircraft. AIAA Journal of Guidance, Control,
and Dynamics.
17. Schaefer, J., Hanson, C., Johnson, M., & Nguyen, N. (2011). Handling qualities of model
reference adaptive controllers with varying complexity for pitch-roll coupled failures. AIAA
Guidance, Navigation, and Control Conference, AIAA-2011-6453, August 2011.
18. Bryson, A. E., & Ho, Y. C. (1979). Applied optimal control: Optimization, estimation, and
control. John Wiley & Sons, Inc.
19. Bryson, A. E., & Ho, Y. C. (1979). Applied optimal control: Optimization, estimation, and
control. New Jersey: Wiley Inc.
20. Nguyen, N. (2013). Adaptive control for linear uncertain systems with unmodeled dynam-
ics revisited via optimal control modification. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2013-4988, August 2013.
21. Nguyen, N. (2010). Asymptotic linearity of optimal control modification adaptive law with
analytical stability margins. In AIAA Infotech@Aerospace Conference, AIAA-2010-3301, April
2010.
22. Calise, A. J., Yucelen, T., Muse, J., & Yang, B. (2009). A loop recovery method for adaptive
control. In AIAA Guidance, Navigation, and Control Conference, AIAA-2009-5967, August
2009.
23. Cao, C., & Hovakimyan, N. (2007). Guaranteed transient performance with L1 adaptive con-
troller for systems with unknown time-varying parameters and bounded disturbances: Part I.
In American Control Conference, July 2007.
24. Cao, C., & Hovakimyan, N. (2008). Design and analysis of a novel L1 adaptive control ar-
chitecture with guaranteed transient performance. IEEE Transactions on Automatic Control,
53(2), 586–591.
25. Nguyen, N., Burken, J., & Hanson, C. (2011). Optimal control modification adaptive law with
covariance adaptive gain adjustment and normalization. In AIAA Guidance, Navigation, and
Control Conference, AIAA-2011-6606, August 2011.
26. Nguyen, N. (2014). Multi-objective optimal control modification adaptive control method for
systems with input and unmatched uncertainties. In AIAA Guidance, Navigation, and Control
Conference, AIAA-2014-0454, January 2014.
27. Nguyen, N. and Balakrishnan, S. N. (2014). Bi-objective optimal control modification adaptive
control for systems with input uncertainty. IEEE/CAA Journal of Automatica Sinica, 1(4), 423–
434.
28. Nguyen, N. (2012). Bi-objective optimal control modification adaptive control for systems with
input uncertainty. In AIAA Guidance, Navigation, and Control Conference, AIAA-2012-4615,
August 2012.
29. Nguyen, N. (2013). Bi-objective optimal control modification adaptive law for unmatched
uncertain systems. In AIAA Guidance, Navigation, and Control Conference, AIAA-2013-4613,
August 2013.
30. Slotine, J.-J., & Li, W. (1991). Applied nonlinear control. Upper Saddle River: Prentice-Hall,
Inc.
31. Nguyen, N., Krishnakumar, K., Kaneshige, J., & Nespeca, P. (2008). Flight dynamics model-
ing and hybrid adaptive control of damaged asymmetric aircraft. AIAA Journal of Guidance,
Control, and Dynamics, 31(3), 751–764.
32. Nguyen, N., Ishihara, A., Stepanyan, V., & Boskovic, J. (2009). Optimal control modification
for robust adaptation of singularly perturbed systems with slow actuators. In AIAA Guidance,
Navigation, and Control Conference, AIAA-2009-5615, August 2009.
References 347

33. Kokotovic, P., Khalil, H., & O’Reilly, J. (1987). Singular perturbation methods in control:
analysis and design. Philadelphia: Society for Industrial and Applied Mathematics.
34. Ardema, M. (1981). Computational singular perturbation method for dynamical systems. AIAA
Journal of Guidance, Control, and Dynamics, 14, 661–663.
35. Narendra, K. S., & Annaswamy, (2005). Stable adaptive systems. New York: Dover Publica-
tions.
36. Lavretsky, E., & Wise, K. (2012). Robust and adaptive control. Berlin: Springer.
Chapter 10
Aerospace Applications

Abstract This chapter presents several adaptive control applications with a partic-
ular focus on aerospace flight control applications. Two relatively simple pendulum
applications are used to illustrate a nonlinear dynamic inversion adaptive control
design method for tracking a linear reference model. The rest of the chapter presents
several adaptive flight control applications for rigid aircraft and flexible aircraft.
The chapter concludes with an application of the optimal control modification to a
F-18 aircraft model. The flight control applications for rigid aircraft include the σ
modification, e modification, bi-objective optimal control modification, least-squares
adaptive control, neural network adaptive control, and hybrid adaptive control which
combines model-reference adaptive control with a recursive least-squares parameter
identification method. An adaptive control design for a flexible aircraft is presented
using a combined adaptive law with the optimal control modification and adaptive
loop recovery modification to suppress the dynamics of the aircraft flexible modes.
An adaptive linear quadratic gaussian (LQG) design based on the optimal control
modification for flutter suppression is presented. By taking advantage of the linear
asymptotic property, the adaptive flutter suppression can be designed to achieve a
closed-loop stability with output measurements.

Adaptive control is a potentially promising technology that can improve performance

and stability of a conventional fixed-gain controller. In recent years, adaptive con-
trol has been receiving a significant amount of attention. In aerospace applications,
adaptive control technology has been demonstrated successfully in both unmanned
aircraft and man-rated aircraft arenas. During 2002 to 2006, NASA conducted the
intelligent flight control system (IFCS) program to demonstrate the ability of a neu-
ral network intelligent flight control system onboard a NF-15B research aircraft (tail
number 837), shown in Fig. 10.1 for enhanced aircraft flight control performance
under faults and failures [1, 2]. The flight demonstration program was a joint collab-
oration between NASA Ames Research Center, NASA Armstrong Flight Research
Center, the Boeing Company, and other organizations. The intelligent flight control
system was based on the e modification [3] with the sigma-pi neural network based
on the work by Calise [4]. The first phase of the flight test program revealed that
the intelligent flight control system did not perform as well as it was in simulations.

© Springer International Publishing AG 2018 349

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0_10
350 10 Aerospace Applications

In some instances, the intelligent flight control system generated large commands
that caused load limit excursions. A subsequent modification to simplify the neu-
ral network demonstrated an improvement in the intelligent flight control system
in simulations. However, during the final phase of the flight test experiment, mixed
results were obtained. Lateral pilot-induced oscillations were experienced during an
in-flight simulated stabilator failure. Consequently, a further study was conducted at
the conclusion of the IFCS program in 2006. It was noted during this study that the
normalization method performed well in simulations.

Fig. 10.1 NASA NF-15B research aircraft (tail number 837)

During 2010 to 2011, NASA conducted a follow-on flight test program of adaptive
control onboard a F/A-18A aircraft (tail number 853) shown in Fig. 10.2. The adaptive
controllers were a simplified MRAC with Φ (x) = x and the optimal control modifi-
cation with and without normalization [5–8]. Prior to the flight test program, several
adaptive controllers were evaluated by eight NASA test pilots in the Advanced Flight
Concept Simulator at NASA Ames Research Center in 2009 [9, 10]. The evaluation
study showed that the optimal control modification performed well across different
failure scenarios [7, 9]. NASA Armstrong Flight Research Center conducted a fur-
ther investigation of the optimal control modification and simulation experiments in
a F-18 flight simulator [11]. The simulation results supported the improved perfor-
mance with the optimal control modification in a simplified adaptive control design.
In addition, the normalization method was also implemented in conjunction with the
optimal control modification based on the previous study during the IFCS program
[12]. The flight test program showed improvements in the adaptation process under
most failure scenarios [5–8].
The use of adaptive control potentially could help reduce the development cost
of production flight control systems for aerospace vehicles. Typically, in aerospace
production systems, accurate plant models must be developed to ensure high con-
fidence in a flight control design. The model development of a typical aerospace
system in general can be quite expensive since extensive experimental validation in
wind tunnel facilities and in flight is often required. Adaptive control may permit
the use of less accurate plant models since the adaptation process can accommodate
10 Aerospace Applications 351

Fig. 10.2 NASA F/A-18A research aircraft (tail number 853)

the uncertainty in the plants. For this to happen, certification barriers for adaptive
control must be overcome.
In aerospace applications, the ability of an adaptive controller to modify a pre-
designed control system is viewed as a strength and a weakness at the same time.
On the one hand, the premise of being able to accommodate system failures is a
major advantage of adaptive control. On the other hand, potential problems with
adaptive control exist with regard to robustness. Therefore, in spite of the many
recent advances in adaptive control, currently there are no adaptive controllers that
have yet been certified for mission-critical or safety-critical systems. This is not to
say that adaptive control is generally not considered as a viable option. Quite the
contrary, in a number of major aerospace systems, adaptive control has been viewed
as an option in the recent years.
The development of certifiable adaptive flight control systems represents a major
challenge to overcome. Adaptive control will never become part of the future unless it
can be proven that it is is highly safe and reliable. Technical challenges are continued
to be addressed by the research community.
To demonstrate the potential benefits of adaptive control, this chapter presents
some example adaptive control applications with a particular focus on aerospace
flight control. The learning objectives of this chapter are:
• To be able to design adaptive controllers for a variety of applications;
• To be able to design nonlinear dynamic inversion adaptive controllers for nonlinear
systems;
• To develop a familiarity with various aircraft applications and a basic understand-
ing of aircraft flight control;
• To learn how to design adaptive controllers for systems with partial-state informa-
tion using the Kalman filter; and
• To be able to combine both direct MRAC and indirect least-squares parameter
estimation to improve effectiveness of adaptive controllers.

10.1 Inverted Pendulum

An inverted pendulum, shown in Fig. 10.3, is an unstable nonlinear system that is

fairly analogous to a launch vehicle. When the pendulum is in equilibrium at the top,
352 10 Aerospace Applications

a slight perturbation will cause the pendulum to swing down to the stable equilibrium
in the vertically downward position.

Fig. 10.3 Inverted pendulum

Consider the equation of motion of a simple inverted pendulum

1 1
m L 2 θ̈ − mgL sin θ + cθ̇ = u (t − td ) (10.1)
3 2
where m is the mass of the pendulum, L is the length, g is the gravity constant, c is the
damping coefficient which is assumed to be unknown, θ (t) is the angular position
of the pendulum, u (t) is the control input which represents the motor torque, and td
is a time delay which represents the motor actuator dynamics.
 
Let x1 (t) = θ (t), x2 (t) = θ̇ (t), and x (t) = x1 (t) x2 (t) . Then,

ẋ = f (x) + B [u (t − td ) − cx2 ] (10.2)

where  
x2
f (x) = 3g (10.3)
2L
sin x1
 
0
B= 3 (10.4)
m L2

We want to design an adaptive controller that enables the closed-loop plant to

follow a reference model
ẋm = Am xm + Bm r (10.5)

Using the pseudo-inverse, the adaptive controller is obtained as

 −1 
u = B B B [Am x + Bm r − f (x)] + ĉ (t) x2 (10.6)

The delay-free closed-loop plant is described by

ẋ = Am x + Bm r + B c̃x2 (10.7)
10.1 Inverted Pendulum 353

The delay-free tracking error equation is expressed as

ė = Am e − B c̃x2 (10.8)

The standard MRAC adaptive law for ĉ (t) is then given by

ĉ˙ = γ x2 e P B (10.9)

However, in the presence of time delay, the adaptive law may be modified to
include the optimal control modification as
 
ĉ˙ = γ x2 e P B + νx22 ĉB  P A−1
m B (10.10)

ν can be chosen to guarantee stability in the presence of the time delay by using
the linear asymptotic property of the optimal control modification which yields

1   − −1 
B ĉx2 = − B Am P B Pe (10.11)
ν
as γ → ∞.
The closed-loop plant with the time delay is expressed as

ẋ = f (x) − f (x (t − td )) + Am x (t − td ) + Bm r (t − td ) + B ĉ (t − td ) x2 (t − td ) − Bcx2
(10.12)

The nonlinear term f (x) does create a problem since linear stability analysis
cannot be used. However, consider the case when td is small, then
 
x2 − x2 (t − td )
f (x) − f (x (t − td )) = 3g (10.13)
2L
3g
sin x1 − 2L sin x1 (t − td )

If td is small, then using a first-order finite-difference method, we obtain

sin x1 − sin x1 (t − td ) d sin x1

≈ = ẋ1 cos x1 = x2 cos x1 (10.14)
td dt

Therefore,  
x2 − x2 (t − td )
f (x) − f (x (t − td )) ≈ 3g (10.15)
t x cos x1
2L d 2

The worst case for linear stability is when cos x1 = 1. So,

 
1   − −1 
ẋ = A (td ) x − A (0) x (t − td ) + Am + B Am P B P x (t − td ) + Bm r (t − td )
ν
1   − −1 
− B Am P B P xm (t − td ) − Bcx2 (10.16)
ν
354 10 Aerospace Applications

where  
0 1
A (t) = 3g (10.17)
0 2L t

It is noted that this approach is a bounded linear approximation for stability

analysis. It is not always possible to bound a nonlinear system with a linear system
for stability analysis. In general, the Lyapunov stability theory for time delay systems
must be used if a system is nonlinear. However, the Lyapunov stability theory for
nonlinear time delay systems can be very challenging and generally produces very
conservative results [13].

Example 10.1 Let m = 0.1775 slug, L = 2 ft, c = 0.2 slug-ft2 /s, td = 0.05 s, and
θ (0) = θ̇ (0) = 0. The reference model is specified by

θ̈m + 2ζm ωm θ̇m + ωm2 θm = ωm2 r

where ζm = 0.5, ωm = 2, and r (t) = 15◦ .

The response of the closed-loop system with the optimal control modification
using γ = 100 and ν = 0.5 is shown in Fig. 10.4.

20 20
.
.
, m , deg/sec

15
, m , deg

10 m

10
. 0
5 .
0 -10
0 5 10 0 5 10
t, sec t, sec

1 0.1

0
c, lb-ft-sec

0
u, lb-ft

-0.1
-1 <
-0.2

-2 -0.3
0 5 10 0 5 10
t, sec t, sec

Fig. 10.4 Closed-loop response with optimal control modification

10.2 Double Pendulum in Robotic Applications

Robotic arms are usually modeled as a double linkage pendulum as shown in

Fig. 10.5. The control objective of the double pendulum is to achieve a certain con-
figuration described by the angular positions of the linkages.
The equations of motion of a double pendulum are given by
10.2 Double Pendulum in Robotic Applications 355

1 1 1
(m 1 + 3m 2 ) L 21 θ̈1 + m 2 L 1 L 2 θ̈2 cos (θ2 − θ1 ) − m 2 L 1 L 2 θ22 sin (θ2 − θ1 )
3 2 2
1
+ (m 1 + 2m 2 ) gL 1 sin θ1 + (c1 + c2 ) θ̇1 − c2 θ̇2 = u 1 (t − td ) + u 2 (t − td )
2
(10.18)

Fig. 10.5 Double pendulum

1 1 1
m 2 L 22 θ̈2 + m 2 L 1 L 2 θ̈1 cos (θ2 − θ1 ) + m 2 L 1 L 2 θ12 sin (θ2 − θ1 )
3 2 2
1
+ m 2 gL 2 sin θ2 + c2 θ̇2 = u 2 (t − td ) (10.19)
2
where m 1 and m 2 are the masses of the linkages, L 1 and L 2 are the lengths, c1 and c2
are the friction coefficients at the joints which are assumed to be unknown, g is the
gravity constant, θ1 (t) and θ2 (t) are the angular positions of the double pendulum,
u 1 (t) and u 2 (t) are the control variables which represent the motor torques at the
joints, and td is the time delay due to the motor actuator dynamics.
These equations can be recast as

 
(m 1 +3m 2 )L 21 m 2 L 1 L 2 cos(θ2 −θ1 )
3 2
θ̈1
m 2 L 1 L 2 cos(θ2 −θ1 ) m 2 L 22 θ̈2
 2
 3

p(x1 )


m 2 L 1 L 2 θ22 sin(θ2 −θ1 )
− (m 1 +2m 22)gL 1 sin θ1
= 2 2
m L L θ sin(θ −θ )
− 2 1 2 12 2 1 − m 2 gL22 sin θ2
  
f (x1 )
     
11 u 1 (t − td ) −c1 − c2 c2 θ̇1
+ + (10.20)
01 u 2 (t − td ) 0 −c2 θ̇2
     
C D
356 10 Aerospace Applications

   
Let x1 (t) = θ1 (t) θ2 (t) , x2 (t) = θ̇1 (t) θ̇2 (t) , and u (t) =
 
u 1 (t) u 2 (t) . Then, the equations of motion become

ẋ2 = p −1 (x1 ) f (x1 ) + p −1 (x1 ) Cu (t − td ) + p −1 (x1 ) Dx2 (10.21)

We want to design an adaptive controller that enables the closed-loop plant to

follow a reference model
      
ẋm 1 0 I xm 1 0
= + r (10.22)
ẋm 2 K p Kd xm 2 −K p
     
Am Bm

where K p and K d are the proportional and derivative control gain matrices such that
 
−ωm2 1 0  
Kp = = diag −ωm2 1 , −ωm2 2 (10.23)
0 −ωm2 2
 
−2ζm 1 ωm 1 0  
Kd = = diag −2ζm 1 ωm 1 , −2ζm 2 ωm 2 (10.24)
0 −2ζm 2 ωm 2

The adaptive controller can be found by inverting the delay-free equations of

motion with td = 0 such that

K p x1 + K d x2 − K p r = ẋ2 = p −1 (x1 ) f (x1 ) + p −1 (x1 ) Cu + p −1 (x1 ) D̂ (t) x2

(10.25)
This yields
 
u = C −1 p (x1 ) K p x1 + K d x2 − K p r − C −1 f (x1 ) − C −1 D̂ (t) x2 (10.26)

The delay-free closed-loop plant is then described by

ẋ2 = K p x1 + K d x2 − K p r − p −1 (x1 ) D̃ (t) x2 (10.27)

The term p −1 (x1 ) D̃ (t) x2 (t) can be evaluated as


m 2 L 22 2 −θ1 )
−1 1 − m 2 L 1 L 2 cos(θ
p (x1 ) D̃ (t) x2 = 3 2
(m 1 +3m 2 )L 21
det p (x1 ) − m 2 L 1 L 2 cos(θ2 −θ1 )
2 3
  
−c̃1 − c̃2 c̃2 θ̇1
0 −c̃2 θ̇2


m L (c̃ +c̃ )
2
m 2 L 22 c̃2
− 2 2 31 2 0 m 2 L 1 L 2 c̃2
= 3
(m +3m )L 2 c̃
2
0 − 1 3 2 1 2 m 2 L 1 L 22(c̃1 +c̃2 ) − m 2 L 12L 2 c̃2
  
Θ̃ 
10.2 Double Pendulum in Robotic Applications 357
⎡ θ̇1
⎤
det p(x1 )
⎢ θ̇2 ⎥
⎢ ⎥
⎢ det p(x1 ) ⎥ = Θ̃  Φ (x1 , x2 ) (10.28)
⎢ θ̇1 cos(θ2 −θ1 ) ⎥
⎣ det p(x1 ) ⎦
θ̇2 cos(θ2 −θ1 )
det p(x1 )
  
Φ(x1 ,x2 )

Then, this implies

D̂ (t) x2 = p (x1 ) Θ  Φ (x1 , x2 ) (10.29)

So, the adaptive controller becomes

 
u = C −1 p (x1 ) K p x1 + K d x2 − K p r − C −1 f (x1 ) − C −1 p (x1 ) Θ  Φ (x1 , x2 )
(10.30)
Thus, the closed-loop plant can be expressed as
        
ẋ1 0 I x1 0 0
= + r− Θ̃  Φ (x1 , x2 ) (10.31)
ẋ2 K p Kd x2 −K p I
       
Am Bm B

 
Let e (t) = xm1 (t) − x1 (t) xm2 (t) − x2 (t) , then the tracking error equation
is obtained as
ė = Am e + B Θ̃  Φ (x1 , x2 ) (10.32)

where ⎡ ⎤
0 0 1 0
⎢ 0 0 0 1 ⎥
Am = ⎢
⎣ −ωm2
⎥
⎦ (10.33)
1
0 −2ζm 1 ωm 1 0
0 −ωm 2 2
0 −2ζm 2 ωm 2
⎡ ⎤
0 0
⎢0 0⎥
B=⎢
⎣1
⎥ (10.34)
0⎦
0 1

Due to the presence of the time delay which represents the motor actuator dynamics,
the adaptive law for Θ (t) should include a robust modification scheme or the pro-
jection method. For example, the adaptive law for Θ (t) with the σ modification is
given by  
Θ̇ = −Γ Φ (x1 , x2 ) e P B + σ Θ (10.35)
358 10 Aerospace Applications

150 300

θ , θm , deg

θ , θ , deg
100 200

2
m
50 100

2
θ1 θm θ2 θm
1 2

0 0
0 5 10 0 5 10
t, sec t, sec

200 15

10
0
u, lb−ft

Θ
5
−200
0
u1 u2
−400 −5
0 5 10 0 5 10
t, sec t, sec

Fig. 10.6 Closed-loop response with MRAC

Example 10.2 Let m 1 = m 2 = 0.1775 slug, L 1 = L 2 = 2 ft, c1 = c2 = 0.2 slug-

ft2 /s, td = 0.005 s, and θ1 (0) = θ̇1 (0) = θ2 (0) = θ̇2 (0) = 0. The reference model
 
is specified by ωm 1 = ωm 2 = 2 rad/s, ζm 1 = ζm 2 = 0.5, and r (t) = 90◦ 180◦
rad. Let Γ = 10I and σ = 1.

The responses of the closed-loop system with the standard MRAC and σ modifi-
cation are shown in Figs. 10.6 and 10.7. The closed-loop response with the standard
MRAC is very poor due to the time delay. The adaptive control signals are very large,
and the adaptive parameters appear to be drifting. In other words, the closed-loop
system is on the verge of instability. On the other hand, the σ modification produces
a very good closed-loop response. The adaptive control signals exhibit some ini-
tial high-frequency oscillations but then settle to steady-state values. The adaptive
parameters all converge to zero.

150 300
θ , θm , deg

θ2, θm , deg

100 200
1

50 100
1

θ1 θm θ2 θm
1 2

0 0
0 5 10 0 5 10
t, sec t, sec

40 2

1
20
u, lb−ft

0
0
−1
u1 u2
−20 −2
0 5 10 0 5 10
t, sec t, sec

Fig. 10.7 Closed-loop response with σ modification

10.3 Adaptive Control of Aircraft Longitudinal Dynamics 359

10.3 Adaptive Control of Aircraft Longitudinal Dynamics

The motion of an aircraft is unconstrained in a three-dimensional space. Therefore,

aircraft dynamics possess all six degrees of freedom involving translational and
angular velocities in the roll, pitch, and yaw axes. The combined motion of aircraft
can be decomposed into a symmetric or longitudinal motion in the pitch axis and
asymmetric or lateral-directional motion in roll and yaw axes.

Fig. 10.8 Aircraft longitudinal dynamics

The longitudinal dynamics of an aircraft are described by the airspeed V , the angle
of attack α, and the pitch rate q as shown in Fig. 10.8. There are two longitudinal
modes: the phugoid mode and the short-period mode. The short-period mode involves
the angle of attack and pitch rate that affect the aircraft longitudinal stability.
The equations of motion for a symmetric flight are given by [14]

mV α̇ = −C L q̄ S − T sin α + mg cos (θ − α) + mV q (10.36)

I yy q̇ = Cm q̄ S c̄ + T z e (10.37)

where m (t) is the aircraft mass, q̄ (t) is the dynamic pressure, S is the wing reference
area, c̄ is the mean aerodynamic chord, T (t) is the engine thrust, V (t) is the airspeed,
α (t) is the angle of attack, θ (t) is the pitch attitude, q (t) = θ̇ (t) is the pitch rate,
I yy (t) is the mass moment of inertia about the aircraft pitch axis, C L (t) is the lift
coefficient, and Cm (t) is the pitching moment coefficient about the aircraft center
of gravity. Note that the aircraft mass and inertia properties as well as aerodynamic
coefficients are time-varying due to the fuel burn.
The motion can be decomposed into a trim motion and a perturbed motion. The
trim motion is a motion when the aircraft is in an equilibrium for a steady-level flight
for which V (t) = V̄ , α (t) = ᾱ, and q (t) = q̄ = 0. The perturbed motion is a small
amplitude motion about the trim condition. Because the amplitudes are assumed to
be small, linearization of the equations of motion can be performed to derive the
linear equations of motion for the perturbed motion. The equations of motion are
linearized as
360 10 Aerospace Applications
 
α̇ c̄ q c̄
m V̄ α̇ = − C L α α + C L α̇ + CLq + C L δe δe q̄ S + mg sin γ̄ (α − θ ) + m V̄ q
2 V̄ 2 V̄
(10.38)
 
α̇ c̄ q c̄
I yy q̇ = Cm α α + Cm α̇ + Cm q + Cm δe δe q̄ S c̄ (10.39)
2 V̄ 2 V̄

where γ̄ = θ̄ − ᾱ is the trim flight path angle; δe (t) is the elevator control surface
deflection; C L α , C L α̇ , C L q , and C L δe are stability and control derivatives of C L ; and
Cm α , Cm α̇ , Cm q , and Cm δe are stability and control derivatives of Cm .
Assuming γ̄ = 0 and neglecting C L α̇ and C L q which are generally small, then the
equations of motion can be written as
⎡ ⎤ ⎡ Zα ⎤⎡ ⎤
α̇ V̄
0 1 α
⎣ θ̇ ⎦ = ⎣ 0 0 1 ⎦⎣θ ⎦
q̇ Mα + Mα̇V̄Z α 0 Mq + Mα̇ q
        
ẋ A x
⎛ ⎞
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
Z δe ⎜ α ⎟ wα
⎢ V̄ ⎥⎜⎜  
∗ ⎣ ⎦⎟
⎟
+ λ⎣ 0 δ (t −
⎦ ⎜e  d t ) + θ ∗
0 θ θ + ⎣ wθ ⎦
   q ⎟
α q
M Z
Mδe + α̇V̄ δe ⎝ ⎠ wq
u(t−td )
   Θ ∗   
B x
(10.40)

C L α q̄ S C L q̄ S
where Z α = − m
and Z δe = δme are the normal force derivatives (positive
C q̄ S c̄ Cm q̄ S c̄ Cm α̇ q̄ S c̄2 Cm q q̄ S c̄2
down); Mα = mIαyy , Mδe = δIeyy , Mα̇ = 2I , and Mq = 2I are the
yy V̄ yy V̄
∗ ∗
pitching moment derivatives (positive nose up); θα and θq are the uncertainty in
α (t) and q (t) due to failures; λ is the uncertainty in the control effectiveness of the
elevator control surface; and td is a time delay which represents the elevator control
surface actuator dynamics.
Suppose an adaptive pitch attitude controller is to be designed to enable the pitch
attitude to follow a reference model specified as

θ̈m + 2ζm ωm θ̇m + ωm2 θm = ωm2 r (10.41)

For perfect tracking, the pitch attitude should have the same dynamics. Therefore,

θ̈ = −2ζm ωm q − ωm2 (θ − r ) (10.42)

The delay-free pitch rate equation is written as

   
Mα̇ Z α   Mα̇ Z δe  
θ̈ − Mα + α − Mq + Mα̇ θ̇ = Mδe + δe + Θ ∗ x (10.43)
V̄ V̄
10.3 Adaptive Control of Aircraft Longitudinal Dynamics 361

   
where Θ ∗ = θα∗ 0 θq∗ and x (t) = α (t) θ (t) q (t) .
Therefore, the elevator control surface deflection δe (t) can be designed by invert-
ing the pitch rate equation. This yields

δe = kα α + kθ (θ − r ) + kq q − Θ  (t) x (10.44)

where
Mα + Mα̇ Z α
kα = − V̄
Mα̇ Z δe
(10.45)
Mδe + V̄

ωm2
kθ = − Mα̇ Z δe
(10.46)
Mδe + V̄

2ζm ωm + Mq + Mα̇
kq = − Mα̇ Z δe
(10.47)
Mδe + V̄

Thus, the attitude controller is a proportional-integral (PI) controller of the pitch

rate, where k p is the proportional gain and kθ is the integral gain. The feedback gain
kα in α (t) is designed to cancel out the angle-of-attack dynamic in the pitch rate
equation.
Using the nominal controller with no adaptation, i.e., Θ (t) = 0, the full-state
reference model can be established as
⎡ ⎤ ⎡ Z α +Z δ kα Z δ kθ Z δe kq ⎤ ⎡ ⎤ ⎡ Z δe ωm2
⎤
α̇m e e
1 + α m −
⎦ ⎣ θm ⎦ + ⎢
V̄ V̄ V̄ Mδe V̄ +Mα̇ Z δe ⎥
⎣ θ̇m ⎦ = ⎣ 0 0 1 ⎣ 0 ⎦ r (10.48)
q̇m 0 −ωm −2ζm ωm
2 q m ωm 2
           
ẋm Am xm Bm

The delay-free tracking error equation is

ė = Am e + B Θ̃  x (10.49)

The optimal control modification can be used for the adaptive law to take advan-
tage of the linear asymptotic property that allows the modification parameter ν to
be computed to guarantee stability of the closed-loop system in the presence of time
delay.
The adaptive controller is designed using the bi-objective optimal control modi-
fication adaptive laws as follows:

u = K x x + kr r + u ad (10.50)

where
u ad = ΔK x (t) x + Δkr (t) r − Θ  (t) x (10.51)
362 10 Aerospace Applications

Then, ΔK x (t), Δkr (t), and Θ  (t) are computed from the following bi-objective
optimal control modification adaptive laws:

Δ K̇ x = Γ K x x e P + νu ad

λ̂B  P A−1
m B λ̂ (10.52)

Δk̇r = γkr r e P + νu ad

λ̂B  P A−1
m B λ̂ (10.53)

  ! 

Θ̇ = −ΓΘ x e P + νu ad

λ̂B  P A−1
m + e 
p P − η u + 2Θ  Φ (x) λ̂B  + ŵ P A−1
m B λ̂
(10.54)

  "  #
λ̂˙ = −γλ u + Θ  x ep P − η u + 2Θ  Φ (x) λ̂B  + ŵ P A−1
m B
 " # (10.55)

ŵ˙  = −γw ep P − η u + 2Θ  Φ (x) λ̂B  + ŵ P A−1
m (10.56)

Example 10.3 The short-period mode dynamics of a transport aircraft at Mach 0.8
and 30,000 ft are described by
 
ẋ = Ax + Bλ u (t − td ) + Θ ∗ x

where λ = 0.5 represents an uncertainty in the effectiveness of the elevator control

surface, td = 0.05 s is the time delay which the elevator actuator dynamics, and the
matrices A, B, and Θ ∗ are
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
−0.7018 0 0.9761 −0.0573 0.5
A=⎣ 0 0 1 ⎦, B = ⎣ 0 ⎦ , Θ∗ = ⎣ 0 ⎦
−2.6923 0 −0.7322 −3.5352 −0.4

The reference model is specified by ζm = 0.75 and ωm = 1.5 rad/s to give

a desired handling quality. A nominal controller is designed with kα = −0.7616,
kθ = −kr = 0.6365, and kq = 0.4293. The closed-loop eigenvalues of the nominal
plant are −0.6582 and −1.2750 ± 0.7902i. The reference model is then established
from the closed-loop nominal plant with
⎡ ⎤ ⎡ ⎤
−0.6582 −0.0365 0.9466 0.0365
Am = ⎣ 0 0 1 ⎦ , Bm = ⎣ 0 ⎦
0 −2.2500 −2.5500 2.2500

The parametric uncertainty Θ ∗ and the control input uncertainty λ result in a

short-period mode damping ratio of 0.2679, which is 34% less than the nominal
short-period mode damping ratio of 0.4045.
10.3 Adaptive Control of Aircraft Longitudinal Dynamics 363

The adaptive flight control architecture for the bi-objective optimal control mod-
ification is illustrated in Fig. 10.9 [15].
Figure 10.10 shows the aircraft response due to the baseline controller. With no
adaptation, the closed-loop plant does not track the reference model well.
Figure 10.11 is the plot of the aircraft response with the standard MRAC for the
adaptation rates Γx = ΓΘ = 50I and γr = 50. The attitude angle command tracking
has improved considerably. However, there are large initial transients in the pitch
rate response as well as high-frequency oscillations.
Figure 10.12 shows the aircraft response with the bi-objective MRAC for the
adaptation rates Γx = ΓΘ = 50I and γr = γλ = γw = 50 by setting ν = η = 0 in the
bi-objective optimal control modification adaptive laws. The closed-loop becomes
unstable after 14 s. The instability of the adaptive laws is consistent with the theory
which shows that η cannot be zero when an external disturbance w (t) exists due to
the term c8 in the stability theorem. Moreover, it is also consistent with the MRAC
theory which establishes that the standard MRAC generally exhibits a parameter drift
in the presence of a disturbance. To prevent a parameter drift, the disturbance estimate
ŵ (t) must be bounded by setting η > 0. If the disturbance w (t) is not estimated by
setting γw = 0, then the bi-objective MRAC is stable as shown in Fig. 10.13 since
the term c8 becomes bounded when the adaptive law for ŵ (t) is not present. The
bi-objective MRAC with γw = 0 has a very similar tracking performance to that with
the standard MRAC.
Figure 10.14 shows the aircraft response with the bi-objective optimal control
modification for the same adaptation rates with ν = η = 0.4. The closed-loop
response with the bi-objective optimal control modification is significantly improved
with a very good tracking performance and no high-frequency oscillations.

Fig. 10.9 Adaptive flight control architecture for bi-objective optimal control modification
364 10 Aerospace Applications

3
α α
m

α, deg
0

−3
0 5 10 15 20 25 30 35 40
t, sec
5
θ θm

θ, deg
0

−5
0 5 10 15 20 25 30 35 40
t, sec
5
qm
q, deg/sec

q
0

−5
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.10 Aircraft response with baseline controller

3
α αm
α, deg

−3
0 5 10 15 20 25 30 35 40
t, sec
5
θ θ
m
θ, deg

−5
0 5 10 15 20 25 30 35 40
t, sec
5
qm
q, deg/sec

q
0

−5
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.11 Aircraft response with standard MRAC

3
α αm
α, deg

−3
0 5 10 15 20 25 30 35 40
t, sec
5
θ θ
m
θ, deg

−5
0 5 10 15 20 25 30 35 40
t, sec
5
qm
q, deg/sec

q
0

−5
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.12 Aircraft response with bi-objective MRAC (γw > 0)

10.3 Adaptive Control of Aircraft Longitudinal Dynamics 365

3
α α
m

α, deg
0

−3
0 5 10 15 20 25 30 35 40
t, sec
5
θ θm

θ, deg
0

−5
0 5 10 15 20 25 30 35 40
t, sec
5
qm
q, deg/sec

q
0

−5
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.13 Aircraft response with bi-objective MRAC (γw = 0)

3
α α
m
α, deg

−3
0 5 10 15 20 25 30 35 40
t, sec
5
θ θm
θ, deg

−5
0 5 10 15 20 25 30 35 40
t, sec
5
qm
q, deg/sec

q
0

−5
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.14 Aircraft response with bi-objective optimal control modification

Figure 10.15 compares the elevator deflections produced by all the various con-
trollers. The elevator deflections with the standard MRAC and bi-objective MRAC
with γw = 0 exhibit significant high-frequency oscillations and high amplitudes.
The elevator deflection with the bi-objective MRAC with γw > 0 is in full satura-
tion before the controller goes unstable at 14 s. In contrast, the bi-objective optimal
control modification produces a well-behaved control signal for the elevator deflec-
tion with no discernible saturation or high-frequency oscillations. The amplitude of
the elevator deflection produced by the bi-objective optimal control modification is
roughly double that for the baseline controller to account for the 50% reduction in
the control effectiveness of the elevator.
366 10 Aerospace Applications

δ , deg
0

e
Baseline
−5
0 5 10 15 20 25 30 35 40
t, sec

δ , deg
0

e
MRAC
−5
0 5 10 15 20 25 30 35 40
t, sec

5
δ , deg

0 Bi−Objective MRAC
e

−5
0 5 10 15 20 25 30 35 40
t, sec

5
δ , deg

0
e

Bi−Objective MRAC, γ =0
−5 w
0 5 10 15 20 25 30 35 40
t, sec

5
δ , deg

0
e

Bi−Objective Optimal Control Modification

−5
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.15 Elevator deflections

10.4 Recursive Least-Squares and Neural Network Pitch

Attitude Adaptive Flight Control

Consider the short-period dynamics of an aircraft with a matched unstructured uncer-

tainty f (α) as a function of the angle of attack due to nonlinear aerodynamics [16]
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ Z δe ⎤
α̇ Zα
ū
0 1 α ū
⎣ θ̇ ⎦ = ⎣ 0 0 1 ⎦⎣θ ⎦ + ⎣ 0 ⎦ [δe + f (α)]
q̇ Mα + Mα̇ūZ α 0 Mq + Mα̇ q M Z
Mδe + α̇ū δe
(10.57)
A pitch attitude controller is designed to track a desired second-order pitch attitude
dynamics. The pitch rate equation is written as
   
Mα̇ Z α   Mα̇ Z δe
θ̈ − Mα + α − Mq + Mα̇ θ̇ = Mδe + [δe + f (α)] (10.58)
ū ū

The elevator input is designed with the following proportional-integral (PI) control
law:
δe = kα α + kθ (θ − r ) + kq q − fˆ (α) (10.59)
10.4 Recursive Least-Squares and Neural Network Pitch … 367
   
where x (t) = α (t) θ (t) q (t) , K x = kα kθ kq , and fˆ (α) is the function
approximation of the unstructured uncertainty f (α).
The plant modeling error is computed as ε (t) = ẋd (t) − ẋ (t) = Am x (t) +
Bm r (t) − ẋ (t), where ẋ (t) is estimated by a backward finite-difference method.
The uncertainty is modeled with the first four terms of the Chebyshev orthogonal
polynomials
   
fˆ (α) = Θ  Φ (α) = θ0 + θ1 α + θ2 2α 2 − 1 + θ3 4α 3 − 3α (10.60)

Θ (t) is estimated by both the least-squares gradient adaptive law

 −1
Θ̇ = −Γ Φ (α) ε B B  B (10.61)

and the RLS adaptive laws

 −1
Θ̇ = −RΦ (α) ε B B  B (10.62)

Ṙ = −η RΦ (α) Φ  (α) R (10.63)

Note that the adaptive laws are implemented with the scaling of the B matrix by
 
−1
B B to keep the RLS parameter η < 2 (see Sect. 7.3.2).
As discussed in Sect. 6.6, the signal ẋ (t) may not be necessarily available and
therefore needs to be estimated. The predictor model of the plant can be implemented
to estimate the signal ẋ (t) as follows:
 
x̂˙ = Am x̂ + (A − Am ) x + B u + Θ  Φ (α) (10.64)

Then, the predictor-based least-squares gradient and RLS adaptive laws are given
by
 −1
Θ̇ = −Γ Φ (α) εp B B  B (10.65)
 −1
Θ̇ = −RΦ (α) εp B B  B (10.66)

where ε p (t) = ẋd (t) − x̂˙ (t).

For comparison, the least-squares gradient adaptive controller is replaced by the
standard MRAC using the same Chebyshev orthogonal polynomials

Θ̇ = −Γ Φ (α) e P B (10.67)

where e (t) = xm (t) − x (t).

In addition, instead of using the Chebyshev orthogonal polynomials, a two-layer
neural network with a sigmoidal activation function is used to approximate the
unstructured uncertainty as
368 10 Aerospace Applications
 
fˆ (α) = Θ  σ W  ᾱ (10.68)
 
where ᾱ (t) = 1 α (t) and σ () is the sigmoidal activation function.
The neural network adaptive controller is specified by the following adaptive laws:
 
Θ̇ = −ΓΘ Φ W  ᾱ e P B (10.69)
  
Ẇ = −ΓW ᾱe P BV  σ W  ᾱ (10.70)
 
where Θ  (t) = V0 (t) V  (t) .

Example 10.4 The numerical model of the short-period dynamics for a transport
aircraft at Mach 0.8 and 30,000 ft is given by
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
α̇ −0.7018 0 0.9761 α −0.0573
⎣ θ̇ ⎦ = ⎣ 0 0 1 ⎦⎣θ ⎦ + ⎣ 0 ⎦ [δe + f (α)]
q̇ −2.6923 0 −0.7322 q −3.5352

For simulation purposes, the unstructured uncertainty that represents nonlinear

aerodynamics is described by

f (α) = 0.1 cos α 3 − 0.2 sin 10α + 0.05e−α

2

The least-squares gradient and RLS adaptive controllers are implemented with
η = 0.2 and Γ = R (0) = I . The aircraft longitudinal responses with the nominal
controller and with the least-squares gradient and RLS adaptive controllers are shown
in Figs. 10.16, 10.17 and 10.18.

20
α, deg

0
α αm

−20
0 10 20 30 40 50 60
t, sec

50
θ θm
θ, deg

−50
0 10 20 30 40 50 60
t, sec

50
q, deg/sec

q qm
0

−50
0 10 20 30 40 50 60
t, sec

Fig. 10.16 Aircraft response with nominal controller

10.4 Recursive Least-Squares and Neural Network Pitch … 369

The aircraft response with the nominal controller is seen to be extremely poor with
the maximum angle of attack of almost 14o which is well into stall. The benefit of
the least-squares gradient adaptive control is clearly demonstrated by the simulation
results which show a very good tracking performance. However, the RLS adaptive
control does not perform well. This could be due to the slow parameter convergence of
the RLS adaptive control which can improve stability robustness of adaptive control
in the presence of time delay or unmodeled dynamics at the expense of the tracking
performance.

5
α, deg

0
α α
m
−5
0 10 20 30 40 50 60
t, sec
5
θ θ
m
θ, deg

−5
0 10 20 30 40 50 60
t, sec
10
q, deg/sec

q qm
0

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.17 Aircraft response with least-squares gradient adaptive control

5
α, deg

0
α αm

−5
0 10 20 30 40 50 60
t, sec
10
θ θ
m
θ, deg

−10
0 10 20 30 40 50 60
t, sec
10
qm
q, deg/sec

q
0

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.18 Aircraft response with RLS adaptive control (η = 0.2)

370 10 Aerospace Applications

α, deg
0
α αm

−5
0 10 20 30 40 50 60
t, sec
10
θ θm

θ, deg
0

−10
0 10 20 30 40 50 60
t, sec
10
qm
q, deg/sec

q
0

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.19 Aircraft response with predictor-based least-squares gradient adaptive control

5
α, deg

0
α αm

−5
0 10 20 30 40 50 60
t, sec
10
θ θ
m
θ, deg

−10
0 10 20 30 40 50 60
t, sec
10
q
q, deg/sec

q m

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.20 Aircraft response with predictor model-based RLS adaptive control (η = 0.2)

Figures 10.19 and 10.20 illustrate the use of the predictor-based plant modeling
error ε p (t) in the least-squares gradient and RLS adaptive control instead of the
true plant modeling error ε (t). The predictor-based least-squares gradient adaptive
controllers provide a very good tracking performance as compared to its counterpart
using the true plant modeling error. The pitch rate response exhibits some small
initial high-frequency oscillations. The predictor-based RLS adaptive controller is
observed to provide much better tracking after 20 s than the original RLS adaptive
controller, but overall both the RLS adaptive controllers perform worse than the least-
squares gradient adaptive controllers. The results indicate that the predictor model
can provide a good estimation of the signal ẋ (t) without differentiation which can
introduce noise in practical applications.
10.4 Recursive Least-Squares and Neural Network Pitch … 371

α, deg
0
α αm

−5
0 10 20 30 40 50 60
t, sec
10
θ θ
m

θ, deg
0

−10
0 10 20 30 40 50 60
t, sec
10
q
q, deg/sec

q m

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.21 Aircraft response with MRAC

5
α, deg

0
α αm

−5
0 10 20 30 40 50 60
t, sec
10
θ θm
θ, deg

−10
0 10 20 30 40 50 60
t, sec
10
q
q, deg/sec

q m

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.22 Aircraft response with neural network MRAC

The aircraft responses with MRAC (Γ = ΓΘ = ΓW = 10I ) using the Chebyshev

orthogonal polynomial and the neural network are shown in Figs. 10.21 and 10.22.
The responses in Figs. 10.21 and 10.22 both exhibit initial high-frequency oscilla-
tions, but the subsequent tracking performance is very good. The neural network
MRAC has much more pronounced high-frequency oscillations which are due to the
initialization of the neural network weights with random numbers that range between
−1 and 1. As a result, a saturation in the elevator is encountered.
The elevator commands for all four adaptive controllers are shown in Fig. 10.23.
Both the MRAC and neural network MRAC produce high-frequency control input
signals in the first 10 s. This is generally undesirable and therefore should be avoided.
372 10 Aerospace Applications

0 0
LS Gradient RLS

−5 −5

δ , deg

δ deg
e
e
−10 −10

−15 −15
0 20 40 60 0 20 40 60
t, sec t, sec

0 20
MRAC Neural Network MRAC
−5 10
δ , deg

δ , deg
−10 0
e

e
−15 −10

−20 −20
0 20 40 60 0 20 40 60
t, sec t, sec

Fig. 10.23 Elevator deflections

To illustrate the issue of robustness and show that the RLS adaptive controller
is actually better able to handle time delay or unmodeled dynamics than the least-
squares gradient adaptive controller or MRAC, a numerical evidence of the time-
delay margin is computed for each of the four adaptive controllers using a time step
of 0.001 s. The results are shown in Table 10.1.

Table 10.1 Estimated time-delay margins

Adaptive law Numerical evidence of time delay Margin
Least-squares gradient 0.073 s
Predictor-based least-squares gradient 0.032 s
RLS with η = 0.2 0.269 s
Predictor-based RLS with η = 0.2 0.103 s
MRAC 0.020 s
Neural network MRAC 0.046 s

The RLS adaptive controller has the best time-delay margin than the other three
adaptive controllers. The standard MRAC has very poor robustness which is a well-
known fact [17]. Generally, the standard MRAC has to be modified to improve
robustness using the projection method or various modification techniques, such
as the σ modification [18], e modification [3], optimal control modification [19],
and adaptive loop recovery modification [20]. The time-delay margin for the neural
network MRAC varies with the initialization of the neural network weights. It should
be noted that both the predictor-based least-squares adaptive controllers suffer a
significant reduction in the time-delay margin by a factor of more than 2.
10.4 Recursive Least-Squares and Neural Network Pitch … 373

The aircraft responses due to a 0.020 s time delay for the four adaptive controllers
are shown in Figs. 10.24, 10.25, 10.26, and 10.27. As shown, the least-squares gra-
dient adaptive controller maintains a very good tracking performance with a 0.020 s
time delay. The MRAC and neural network MRAC exhibit high-frequency oscilla-
tions. The RLS adaptive controller exhibits low-frequency transients as it is much
more robust than the other three adaptive controllers. Thus, overall, the least-squares
gradient adaptive controller seems to perform the best among all of the adaptive
controllers.

5
α, deg

0
α α
m
−5
0 10 20 30 40 50 60
t, sec
5
θ θ
m
θ, deg

−5
0 10 20 30 40 50 60
t, sec
10
qm
q, deg/sec

q
0

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.24 Aircraft response with least-squares gradient with 0.020 s time delay

10
α, deg

0
α αm

−10
0 10 20 30 40 50 60
t, sec
10
θ θm
θ, deg

−10
0 10 20 30 40 50 60
t, sec
10
q
q, deg/sec

q m

−10
0 10 20 30 40 50 60
t, sec

Fig. 10.25 Aircraft response with RLS with 0.020 s time delay (η = 0.2)
374 10 Aerospace Applications

10

α, deg
0
α αm

−10
0 10 20 30 40 50 60
t, sec
10
θ θ
m

θ, deg
0

−10
0 10 20 30 40 50 60
t, sec
100
q
q, deg/sec

q m

−100
0 10 20 30 40 50 60
t, sec

Fig. 10.26 Aircraft response with MRAC with 0.020 s time delay

10
α, deg

0
α α
m
−10
0 10 20 30 40 50 60
t, sec
10
θ θ
m
θ, deg

−10
0 10 20 30 40 50 60
t, sec
50
qm
q, deg/sec

q
0

−50
0 10 20 30 40 50 60
t, sec

Fig. 10.27 Aircraft response with neural network MRAC with 0.020 s time delay

10.5 Adaptive Control of Flexible Aircraft

Lightweight aircraft design has received a considerable attention in recent years as

a means for improving cruise efficiency. Reducing aircraft weight results in lower
lift requirement which directly translates into lower drag, hence reduced engine
thrust requirement during cruise. The use of lightweight materials such as advanced
composite materials has been adopted by airframe manufacturers in many current
and future aircraft. Modern lightweight materials can provide less structural rigid-
ity while maintaining the required load-carrying capacity. As structural flexibility
increases, aeroelastic interactions with aerodynamic forces and moments become an
increasingly important consideration in aircraft design. Understanding aeroelastic
effects can improve the prediction of aircraft aerodynamic performance and pro-
vide an insight into how to design an aerodynamically efficient flexible airframe to
10.5 Adaptive Control of Flexible Aircraft 375

reduce the fuel consumption. Structural flexibility of airframes can also cause signif-
icant aeroelastic interactions that can degrade vehicle stability margins, potentially
leading to degraded flying qualities. There exists a trade-off between the desire of
having lightweight, flexible structures for weight savings and the need for maintain-
ing sufficient robust stability margins in consideration of flight dynamics, stability,
and control.
It is of interest to examine the effect of aeroelasticity on the short-period mode
of an aircraft. The aeroelastic effect on an aircraft is assumed to be contributed only
by the flexible wing structures. Aeroelastic contributions by the fuselage and tail
empennage are assumed to be negligible. For simplicity, only the first symmetric
bending mode (1B) and first symmetric torsion mode (1T) are considered. Quasi-
steady aerodynamics assumption is invoked to simplify the model. The coupled
aeroelastic flight dynamic model of an aircraft that accounts for interactions between
wing bending and torsion on aircraft performance and stability in the pitch axis can
be expressed in the following state-space form [21]:
⎡  C L q c̄
CLw
⎡ ⎤ − 1 CL 1 mV∞
q∞ S − 2V∞ − 1 1
α̇ ⎢  m αα α  C m αα   C m αα c̄

⎢ q̇ ⎥ ⎢ 1 m qα
Cm α + m αα 1 c̄
mq m qα C L q c̄
+ m αα 2V∞ − mV ∞ 1 m w1
+
m qα C L w1
⎢ ⎥ ⎢ CLα
⎢ ẇ1 ⎥ ⎢
m qq m qq 2V∞ q∞ S m qq c̄ m αα c̄
⎢ ⎥ ⎢ 0 0 0
⎢ θ̇1 ⎥ = ⎢
⎢ ⎥ ⎢ 0 0 0
⎣ ẅ1 ⎦ ⎢ ⎢
⎣ − m w1 w h w1 α − m w1 w h w1 q − m w1 w kw1 w1
θ̈1 1 1 1 1 1 1
− mθ 1
h θ1 α − mθ 1
h θ1 q − m θ1 θ kθ1 w1
1 θ1 1 θ1 1 1
C L ẇ C L c̄ ⎤
θ̇1
− m1ααC L θ1 − 1 1
− 1 ⎡ ⎤
  m αα V̄∞   C mc̄αα 2V̄∞ C c̄  ⎥ α
Cm ẇ m qα C L ẇ1 m θ̇ m qα L θ̇1 ⎥
1 m qα
Cm θ1 + m αα C L θ1 1 1
+ m αα 1 1
+ m αα 2V̄ ⎥ ⎢ q⎥
m qq m qq V̄∞ V̄∞ m qq 2 V̄∞ ⎥⎢ ⎥
∞
⎥⎢w ⎥
0 1 0 ⎥⎢ ⎥
⎥⎢ θ ⎥
0 0 1 ⎥⎢ ⎥
⎥ ⎣ ẇ ⎦
− m w1 w kw1 θ1 − m w1 w cw1 w1 − m w1 w cw1 θ1 ⎥
1 1 1 1 1 1
⎦ θ̇
− mθ 1
k θ1 θ1 − m θ θ c θ 1 w1 1
− m θ1 θ cθ1 θ1
1 θ1 1 1 1 1
⎡ ⎤
− 1 C − 1 C
 m αα L δe  m αα L δ
⎢ 1 m qα m qα ⎥
⎢ m qq Cm δe + m αα C L δe 1
Cm δ + m αα CLδ ⎥
⎢ m qq ⎥ 
⎢ 0 0 ⎥ δe
+⎢
⎢
⎥
⎥ δf
⎢ 0 0 ⎥
⎢ 1 ⎥
⎣ 0 g
m w1 w1 w1 δ1 ⎦
0 − m w1 w gθ1 δ1
1 1
(10.71)

where δ f (t) is a symmetric flap control surface deflection on the wing and m αα , m qα ,
and m qq are defined as
m V̄∞ C L α̇ c̄
m αα = + (10.72)
q∞ S 2 V̄∞

Cm α̇ c̄
m qα = − (10.73)
2 V̄∞
376 10 Aerospace Applications

IY Y
m qq = (10.74)
q∞ S c̄

The subscripts w1 and θ1 denote the first bending and first torsion quantities,
respectively. The symbols m, c, and k denote the generalized mass, generalized
damping which includes both structural and aerodynamic damping, and generalized
stiffness, respectively. The subscripts α and q denote the angle of attack and pitch
rate, respectively. The symbols h and g denote the generalized forces acting on the
wing structure due to the rigid aircraft states α (t) and q (t) and due to the symmetric
flap control surface deflection δ f (t), respectively.

Example 10.5 Consider a flexible wing aircraft at a midpoint cruise condition of

Mach 0.8 and 30,000 ft with 50% fuel remaining. For the configuration with 50%
fuel remaining and assuming a structural damping of ζ1 = 0.01, the A matrix is
given by
⎡ ⎤
−8.0134 × 10−1 9.6574 × 10−1 1.2608 × 10−2 5.0966 × 10−1 5.4634 × 10−4 −2.4249 × 10−3
⎢ −2.4526 × 100 −9.1468 × 10−1 4.6020 × 10−2 2.1726 × 100 3.5165 × 10 −3 −6.2222 × 10 ⎥
−2
⎢ ⎥
⎢ ⎥
⎢ 0 0 0 0 1 0 ⎥
A=⎢ ⎥
⎢ 0 0 0 0 0 1 ⎥
⎢ ⎥
⎣ 1.4285 × 103 1.5869 × 101 −3.1602 × 101 −1.4029 × 103 −2.4360 × 10 0 5.2088 × 10 ⎦
0
−3.9282 × 102 −1.8923 × 100 5.6931 × 100 −2.8028 × 102 3.2271 × 10 −1 −6.1484 × 10 0

The eigenvalues of the short-period mode of the rigid aircraft can be computed
from the 2 × 2 upper left matrix partition of the A matrix. These eigenvalues are
stable as shown
λ S P = −0.8580 ± 1.5380i

The eigenvalues of the 4 × 4 lower right matrix partition corresponding to the 1B

and 1 T modes are also stable as shown

λ1B = −2.0955 ± 8.2006i

λ1T = −2.1967 ± 15.1755i

The eigenvalues of the flexible aircraft are also stable, but with a reduced damping
in the 1 T mode, as shown

λ S P = −0.5077 ± 0.5229i

λ1B = −3.1878 ± 8.3789i

λ1T = −1.4547 ± 15.1728i

The computed frequencies and damping ratios of the short-period mode, and the
1B and 1 T modes with 50% fuel remaining are shown in Table 10.2. The short-period
mode can be seen to be significantly affected by the aeroelastic coupling of the rigid
10.5 Adaptive Control of Flexible Aircraft 377

aircraft flight dynamics with the 1B and 1 T modes. The frequency of the coupled
short-period mode is significant reduced, but the damping increases.

Table 10.2 Frequencies and damping ratios at Mach 0.8 and 30,000 ft for flexible aircraft
Mode Short period 1B 1T
Uncoupled frequency, 1.761 8.4641 15.3337
rad/s
Coupled frequency, 0.7288 8.9648 15.2424
rad/s
Uncoupled damping 0.4872 0.2476 0.1433
ratio
Coupled damping ratio 0.6966 0.3556 0.0954

The frequencies and damping ratios as a function of the airspeed at the same
altitude of 30,000 ft are shown in Figs. 10.28 and 10.29. Generally, the frequencies
of the short-period mode and 1B mode increase with increasing the airspeed, while
the frequency of the 1 T mode decreases precipitously with increasing the airspeed.
The divergence speed is the airspeed at which the torsion modal frequency becomes
zero.
The damping ratios for both the short-period mode and 1B mode generally increase
with increasing airspeed. The damping ratio for the 1 T mode increases with increas-
ing the airspeed up to Mach 0.7 and thereafter decreases rapidly. The flutter speed is
the airspeed at which the damping ratio of any aeroelastic modes becomes zero. It is
apparent that the 1 T mode would exhibit a zero damping at a flutter speed of about
Mach 0.85. The low damping ratio of the 1 T mode can be a problem for aircraft sta-
bility. Uncertainty in aeroelastic properties can adversely affect the performance and
stability of a flexible aircraft. Active feedback control can potentially help improve
the stability margins of the aeroelastic modes.

25
Short−Period Mode
1st Bending Mode
1st Torsion Mode
20
Frequency, rad/sec

15

10

0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Mach Number

Fig. 10.28 Frequencies of flexible aircraft modes

378 10 Aerospace Applications

0.8
Short−Period Mode
0.7 1st Bending Mode
1st Torsion Mode
0.6

0.5

Damping Ratio
0.4

0.3

0.2

0.1

−0.1

−0.2
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Mach Number

Fig. 10.29 Damping ratios of flexible aircraft modes

Consider a linearized model of a flexible aircraft with a matched uncertainty

 
ẋ = Ax + B u + Θ ∗ Φ (xr ) (10.75)

xr = C x (10.76)

where x (t) ∈ Rn is a state vector that is composed of a rigid aircraft state vector
xr (t) ∈ Rnr and an elastic wing state vector xe (t) ∈ Rn e =n−nr , u (t) ∈ Rm is a
control vector, A ∈ Rn × Rn and B ∈ Rn × Rm are constant and known matrices,
Θ ∗ ∈ R p×m is a constant and unknown matrix that represents a matched parametric
uncertainty in the rigid aircraft state, and Φ (xr ) ∈ R p is a vector of known regressor
functions of the rigid aircraft states.
Assuming that there is a sufficient frequency separation between the “slow” rigid
aircraft dynamics and “fast” elastic wing dynamics, then the fast and slow dynamics
can be decoupled using the standard singular perturbation method. The fast dynamics
of the elastic wing modes are assumed to approach the equilibrium solution infinitely
fast. Therefore, the rigid aircraft dynamics with approximately zero-order elastic
wing dynamics can be obtained by setting ẋe (t) = ε (x), where ε (x) is a small
parameter [22]. Thus,
      
ẋr Arr Ar e xr Br  
= + u + Θ ∗ Φ (xr ) (10.77)
ε Aer Aee xe Be
10.5 Adaptive Control of Flexible Aircraft 379

Then, the elastic wing dynamics are approximated by

 
xe = A−1 −1 −1 ∗
ee ε (x) − Aee Aer xr − Aee Be u + Θ Φ (xr ) (10.78)

Substituting xe (t) into the rigid aircraft dynamics yields

 
ẋr = A p xr + B p u + Θ ∗ Φ (xr ) + Δ (x) (10.79)

where
A p = Arr − Ar e A−1
ee Aer (10.80)

B p = Br − Ar e A−1
ee Be (10.81)

Δ (x) = Ar e A−1
ee ε (x) (10.82)

The term Δ (x) represents the effect of unmodeled dynamics of the elastic wing
modes. The reduced-order plant matrix A p is assumed to be Hurwitz. Otherwise, an
output feedback adaptive control design would be necessary.
The objective is to design an adaptive control that enables the rigid aircraft state
vector xr (t) to tracks a reference model

ẋm = Am xm + Bm r (10.83)

where Am ∈ Rnr ×Rnr is a known Hurwitz matrix, Bm ∈ Rnr ×Rr is a known matrix,
and r (t) ∈ Rr is a piecewise continuous bounded reference command vector.
The adaptive controller is designed with

u = K x xr + K r r − Θ  Φ (xr ) (10.84)

where Θ (t) is an estimate of Θ ∗ and it is assumed that K x and K r can be found such
that the following model matching conditions are satisfied:

A p + B p K x = Am (10.85)

B p K r = Bm (10.86)

Defining the tracking error as e (t) = xm (t) − xr (t), then the tracking error
equation becomes
ė = Am e + B Θ̃  Φ (xr ) − Δ (x) (10.87)

where Θ̃ (t) = Θ (t) − Θ ∗ is the estimation error.

380 10 Aerospace Applications

Because of the presence of unmodeled dynamics, the standard model-reference

adaptive law that adjusts Θ (t) which is given by

Θ̇ = −Γ Φ (xr ) e P B (10.88)

is not robust. As the adaptive gain Γ increases, the adaptive law becomes increasingly
sensitive to the unmodeled dynamics Δ (x) that can lead to instability [17].
To improve robustness to unmodeled dynamics, we use the optimal control modifi-
cation adaptive law to estimate the unknown parameter Θ ∗ [19]. The optimal control
modification adaptive law is given by
 
Θ̇ = −Γ Φ (xr ) e P B − νΦ (xr ) Φ  (xr ) Θ B  P A−1
m B (10.89)

where Γ = Γ  > 0 ∈ R p × R p is the adaptation rate matrix, ν > 0 ∈ R is the

modification parameter, and P is the solution of the Lyapunov equation

P Am + A
m P = −Q (10.90)

Alternatively, the adaptive loop recovery modification adaptive law [20] can be
used to adjust Θ (t) as
 
 dΦ (xr ) dΦ  (xr )
Θ̇ = −Γ Φ (xr ) e P B + η Θ (10.91)
d xr d xr

where η > 0 ∈ R is the modification parameter.

Both the optimal control modification and the adaptive loop recovery modification
adaptive laws can also be blended together in a combined adaptive law as follows:


dΦ (xr ) dΦ  (xr )
Θ̇ = −Γ Φ (xr ) e P B − νΦ (xr ) Φ  (xr ) Θ B  P A−1
m B+η Θ
d xr d xr
(10.92)

Example 10.6 The reduced-order model of the flexible aircraft in Example 10.5
that retains only the rigid aircraft states is given by
        
α̇ −0.2187 0.9720 α −0.0651   α
= + δe + θα∗ θq∗
q̇ −0.4053 −0.8913 q −3.5277 q
   
Δα (x) f (t)
+ + α
Δq (x) f q (t)

where θ ∗ = 0.4 and θq∗ = −0.3071 represent a parametric uncertainty corresponding

to a short-period frequency of 1.3247 rad/s and a damping ratio of 0.0199 which is
close to neutral pitch stability.
10.5 Adaptive Control of Flexible Aircraft 381

The time-varying functions f α (t) and f q (t) are disturbances due to a moderate
vertical wind gust modeled by the Dryden turbulence model [14] with a vertical
velocity amplitude of about 10 ft/s and a pitch rate amplitude of 1.5 deg/s as shown
in Fig. 10.30.
A desired reference model of the pitch attitude is given by

θ̈m + 2ζ ωm θ̇m + ωm2 θm = ωn2 r

where ζm = 0.85 and ωm = 1.5 rad/s are chosen to give a desired handling charac-
teristic.    
Let xr (t) = α (t) θ (t) q (t) , u (t) = δe (t), and Θ ∗ = θα∗ 0 θq∗ .
A nominal
  as ū (t) = K x x (t) + kr r (t) where K x =
controller is  designed
− b13 a31 ωn2 2ζ ωn + a33 = − 0.1149 0.6378 0.4702 and kr = b13 ωn2 =
−0.6378. The closed-loop eigenvalues are −0.2112 and −1.2750 ± 0.7902i. The
nominal closed-loop plant is then chosen to be the reference model as
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
α̇m −0.2112 −0.0415 0.9414 αm 0.0415
⎣ θ̇m ⎦ = ⎣ 0 0 1 ⎦ ⎣ θm ⎦ + ⎣ 0 ⎦ r
q̇m 0 −2.2500 −2.5500 qm 2.2500
           
ẋm Am xm Bm

0.5
w /V , deg

0
∞ a

−0.5

−1
0 20 40 60 80 100
t, sec

1
qa, deg/sec

−1

−2
0 20 40 60 80 100
t, sec

Fig. 10.30 Vertical wind gust model

An adaptation rate Γ = 100I is used for the adaptive controller with the input
 
function chosen as Φ (xr ) = 1 α θ q whereby the bias input is needed to handle
the time-varying wind gust disturbances.
For the optimal control modification, the modification parameter is set to ν =
0.2. For the adaptive loop recovery modification, the modification parameter is set
to η = 0.2. Also, the Jacobian of the input function dΦ (xr ) /d xr is simply an
382 10 Aerospace Applications

identity matrix, thereby making the adaptive loop recovery modification effectively
a σ modification adaptive law [18].
A pitch attitude doublet is commanded. The response of the flexible aircraft with-
out adaptive control is shown in Fig. 10.31. It is clear that the aircraft response does
not track well with the reference model. Moreover, the high-frequency oscillations
in the pitch rate response due to the aeroelastic interactions with the bending and
torsion modes are highly objectionable from the pilot handing quality perspective.

5
α α
m
α, deg

−5
0 20 40 60 80 100
t, sec
5
θ θm
θ, deg

−5
0 20 40 60 80 100
t, sec
5
qm
q, deg/sec

q
0

−5
0 20 40 60 80 100
t, sec

Fig. 10.31 Longitudinal response of flexible aircraft with no adaptive control

5
α α
m
α, deg

−5
0 20 40 60 80 100
t, sec
5
θ θ
m
θ, deg

−5
0 20 40 60 80 100
t, sec
5
qm
q, deg/sec

q
0

−5
0 20 40 60 80 100
t, sec

Fig. 10.32 Longitudinal response of flexible aircraft with standard MRAC (Γ = 100I )

Using the standard model-reference adaptive control (MRAC) by setting ν = η = 0,

the pitch attitude tracking is much improved as shown in Fig. 10.32. However, the
initial transient in the pitch rate is quite large and is characterized by high-frequency
oscillations in the pitch rate response.
10.5 Adaptive Control of Flexible Aircraft 383

5
α α
m

α, deg
0

−5
0 20 40 60 80 100
t, sec
5
θ θ
m

θ, deg
0

−5
0 20 40 60 80 100
t, sec
5
q
q, deg/sec

q m
0

−5
0 20 40 60 80 100
t, sec

Fig. 10.33 Longitudinal response of flexible aircraft with optimal control modification (OCM)
(Γ = 100I , ν = 0.2)

5
α α
m
α, deg

−5
0 20 40 60 80 100
t, sec
5
θ θm
θ, deg

−5
0 20 40 60 80 100
t, sec
5
q
q, deg/sec

q m
0

−5
0 20 40 60 80 100
t, sec

Fig. 10.34 Longitudinal response of flexible aircraft with adaptive loop recovery modification
(ALR) (Γ = 100I , η = 0.2)

In contrast, Fig. 10.33 shows that the optimal control modification adaptive law
is able to suppress the large initial transient of the pitch rate and the amplitude of
the high-frequency oscillations. The response of the aircraft due to the adaptive loop
recovery modification adaptive law as shown in Fig. 10.34 is very similar to that with
the optimal control modification adaptive law.
The aeroelastic wing tip bending deflection and torsional twist are shown in
Figs. 10.35 and 10.36 for the four different controllers: baseline nominal controller,
standard MRAC, optimal control modification, and adaptive loop recovery modi-
fication. The aeroelastic wing is modeled to be rather flexible to demonstrate the
aeroelastic effects on adaptive control. The rigid aircraft pitch attitude command
and wind gust result in a bending deflection amplitude of 5 ft and a torsional twist
amplitude of about 3 deg at the wing tip. The aeroelastic deflections are quite signif-
icant since the flight condition at Mach 0.8 is approaching the flutter speed at Mach
384 10 Aerospace Applications

0.85. It is noted that the standard MRAC results in a very large initial transient of
the torsional twist. This large torsional twist is clearly not realistic and in practice
would result in excessive wing loading and wing stall. These effects are not taken
into account in the simulations. Nonetheless, this illustrates the undesirable behavior
of the standard MRAC in the flight control implementation for flexible aircraft.

10 10
Wing Tip Deflection, ft

Wing Tip Deflection, ft

5 5

0 0

−5 −5
Baseline MRAC Γ=100
−10 −10
0 50 100 0 50 100
t, sec t, sec

10 10
Wing Tip Deflection, ft

5 Wing Tip Deflection, ft 5

0 0

−5 −5
OCM Γ=100 ν=0.2 ALR, Γ=100 η=0.2
−10 −10
0 50 100 0 50 100
t, sec t, sec

Fig. 10.35 Wing tip deflection of first bending mode

Figure 10.37 is the plot of the elevator deflections for the four controllers. The
standard MRAC produces a significant control saturation during the initial transient.
This saturation leads to an undesirable rigid aircraft response and aeroelastic deflec-
tions. Both the optimal control modification and adaptive loop recovery modification
adaptive laws produce quite similar elevator deflections. Also, it is observed that the
elevator deflection for the baseline controller has a lower-frequency content than
those corresponding to the optimal control modification and adaptive loop recovery
modification adaptive laws. This indicates that the adaptive controllers effectively
account for the aeroelastic wing dynamics in the elevator deflection commands to
suppress the responses of the aeroelastic wing modes, whereas the baseline controller
has no knowledge of the effect of the aeroelastic wing dynamics.
This study shows that adaptive control can be used to accommodate uncertainty for
flexible aircraft. The effect of aeroelasticity is captured in the reduced-order model
as unmodeled dynamics. The results show that the standard MRAC is neither suffi-
ciently robust nor able to produce well-behaved control signals. Excessive torsional
twist and control saturation due to the standard MRAC are noted. Both the optimal
control modification and adaptive loop recovery modification adaptive laws are seen
to be more effective in reducing the tracking error while maintaining the aeroelastic
deflections to within reasonable levels.
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 385

4 10

Wing Tip Twist, deg

Wing Tip Twist, deg

2
0
0
−10
−2
Baseline MRAC Γ=100
−4 −20
0 50 100 0 50 100
t, sec t, sec

4 4
Wing Tip Twist, deg

Wing Tip Twist, deg

2 2

0 0

−2 −2
OCM Γ=100 ν=0.2 ALR, Γ=100 η=0.2
−4 −4
0 50 100 0 50 100
t, sec t, sec

Fig. 10.36 Wing tip twist of first torsion mode

4 20

2 10
δ , deg

δe, deg

0 0
e

−2 −10
Baseline MRAC Γ=100
−4 −20
0 50 100 0 50 100
t, sec t, sec

10 10

5 5
δe, deg

δe, deg

0 0

−5 −5
OCM Γ=100 ν=0.2 ALR, Γ=100 η=0.2
−10 −10
0 50 100 0 50 100
t, sec t, sec

Fig. 10.37 Elevator deflection

10.6 Adaptive Linear Quadratic Gaussian Flutter

Suppression Control

Flutter is a structural dynamic phenomenon in aircraft flight dynamics that mani-

fests itself as an aeroelastic instability which can result in structural failures. Air-
craft are designed to ensure that their flutter boundaries are far away from their
flight envelopes. As modern aircraft begin to employ lightweight materials in air-
craft structures for improved fuel efficiency, increased aeroelastic interactions and
reduced flutter margins become a definite possibility. Active flutter suppression con-
trol is a structural feedback control that suppresses structural vibration and increases
aeroelastic stability of aircraft structures. This study illustrates an application of
adaptive augmentation Linear Quadratic Gaussian (LQG) control for flutter suppres-
sion for a new type of aircraft flight control surfaces. Optimal control modification
adaptive law will be used as the adaptive augmentation controller.
386 10 Aerospace Applications

In the recent years, NASA has developed a new type of aircraft flight control sur-
faces called the Variable Camber Continuous Trailing Edge Flap (VCCTEF) [23–25].
The VCCTEF, illustrated in Fig. 10.38, is a wing shaping control device designed to
reshape a flexible aircraft wing in-flight to improve the aerodynamic performance
while suppressing any adverse aeroelastic interactions. It employs three chordwise
flap segments to provide a variable camber to change the wing shape for increas-
ing the aerodynamic performance. The flap is also made up of individual spanwise
sections which enable different flap settings at each flap spanwise position. This
results in the ability to control the wing shape as a function of the wing span, thereby
resulting in a change to the wing twist to establish the best lift-to-drag ratio at any
aircraft gross weight or mission segment. The individual spanwise flap sections are
connected with a flexible elastomer material to form a continuous trailing edge with
no flap gaps in the wing planform for drag and noise reduction purposes.

Fig. 10.38 Variable camber continuous trailing edge flap concept

Consider a notional aircraft equipped with highly flexible wings. The flexibility
of modern transport wings can cause a reduction in the flutter margin which can
compromise the stability of an aircraft. A flutter analysis is conducted to examine
the effect of the increased flexibility of a conventional aircraft wing based on a NASA
Generic Transport Model (GTM) [26]. The baseline stiffness of the GTM wing is
reduced by 50%. This configuration is referred to as the flexible wing GTM. The
flutter analysis computes the aeroelastic frequencies and damping ratios over a range
of the airspeed as shown in Figs. 10.39 and 10.40 [27]. When the damping ratio of
an aeroelastic mode encounters a zero crossing, a flutter speed is noted. The lowest
flutter speed establishes a flutter boundary.
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 387

Flexible Wing GTM 80% Fuel Anti−Symmetric Mode @ 35K ft

14

12

10

Frequency, Hz
8

0
0 0.5 1 1.5
Mach @ 35K ft

Fig. 10.39 Natural frequencies of antisymmetric modes of flexible wing GTM

Flexible Wing GTM 80% Fuel Anti−Symmetric Mode @ 35K ft

1

0.8

0.6
Aerodynamic Damping

0.4

0.2

−0.2

−0.4
0 0.5 1 1.5
Mach @ 35K ft

Fig. 10.40 Damping ratios of antisymmetric modes of flexible wing GTM

Table 10.3 shows the flutter speed prediction at 35,000 ft for the stiff wing GTM
and the flexible wing GTM. The critical flutter mode is an antisymmetric mode. The
flutter boundary is significantly reduced by 31% with the flexible wing GTM.
The FAA (Federal Aviation Administration) requires the aircraft certification to
demonstrate a flutter margin of at least 15% above the dive speed which is normally
determined from flight testing. For a maximum operating Mach 0.8, the dive speed
may be estimated to be about 15% over the maximum operating Mach, or Mach 0.92.
Thus, the flutter clearance for this notional aircraft would require a minimum flutter
speed of Mach 1.06 at 35,000 ft. The stiff wing GTM meets this flutter clearance,
but the flexible wing GTM does not. To increase the flutter speed, active flutter sup-
pression is an option. Currently, active flutter suppression has not been certified for
transport aircraft, but this situation may change as the FAA has begun to investigate
388 10 Aerospace Applications

certification requirements for active flutter suppression control for commercial trans-
port aircraft.

Table 10.3 Flutter speed prediction

Symmetric mode Antisymmetric mode
Flutter Mach @ 35K ft (Stiff 1.358 1.310
Wing)
Flutter Mach @ 35K ft 0.938 0.925
(Flexible Wing)

Consider a general aeroservoelastic (ASE) state-space model of the form

⎡ ⎤
       δ
η̇ A11 A12 η B11 B12 B13 ⎣ ⎦
= + δ̇ (10.93)
μ̇ A21 A22 μ B21 B22 B23
δ̈

y = C1 η + C2 μ + D1 δ + D2 δ̇ + D3 δ̈ (10.94)

where μ (t) is a fast state vector that contains the generalized aeroelastic states
for high-frequency aeroelastic modes, η (t) is a slow state vector that contains the
generalized aeroelastic states for lower-frequency aeroelastic modes as well as the
rigid aircraft states, δ (t) is the control surface deflection vector, and y (t) is the output
vector.
Note that the aeroelastic states are dependent on the velocity and acceleration of
the control surface deflection vector δ (t). This dependency is usually neglected in
rigid aircraft flight dynamics, but can be significant in aeroservoelasticity especially
for flutter suppression.
The quality of fast and slow states can be examined by the eigenvalues of the
partitioned matrices A11 and A22 . Since μ (t) is a fast state vector, then we can write
ε
A11
<
A22
, where ε is a small quantity.
In general, an ASE state-space model contains both the rigid aircraft modes, which
usually have low frequencies, and the aeroelastic modes which are at much higher
frequencies than the rigid aircraft modes. Flutter modes are usually associated with
those aeroelastic modes in the low-frequency range. As a result, high-frequency
aeroelastic modes normally do not participate in a flutter response. In a control
design, it is usually expedient to remove high-frequency modes in order to simplify a
controller design. By removing the high-frequency aeroelastic modes, the ASE state-
space model is reduced in the model order. We will employ the singular perturbation
method to reduce the order of an ASE state-space model [28]. The advantage of the
singular perturbation method is that the physical modes are properly retained for a
flutter suppression control design.
Using the singular perturbation approach, the fast and slow dynamics can be
decoupled. To decouple the fast and slow states, we perform a time scale separation
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 389

by applying the singular perturbation method. Toward that end, we consider the usual
slow or stretched time transformation

τ = εt (10.95)

where τ is a slow time variable.

Then, the fast and slow state-space models are transformed into a singularly
perturbed system as

η̇ = A11 η + A12 μ + B11 δ + B12 δ̇ + B13 δ̈ (10.96)

dμ
ε = A21 η + A22 μ + B21 δ + B22 δ̇ + B23 δ̈ (10.97)
dτ
The Tikhonov’s theorem can be used to approximate the solution of the singularly
perturbed system with the solution of a “reduced-order” system by setting ε = 0 [22].
Thus, the reduced-order system is given by

η̇0 = A11 η0 + A12 μ0 + B11 δ + B12 δ̇ + B13 δ̈ (10.98)

A21 η0 + A22 μ0 + B21 δ + B22 δ̇ + B23 δ̈ = 0 (10.99)

where η0 (τ ) and μ0 (τ ) are the “outer” solution of the singularly perturbed system.
The term “outer” is in connection with the concept of “inner” or “boundary layer”
and “outer” solutions which have the origin in the boundary layer theory due to
Prandtl. The “inner” or “boundary layer” solution for this system is obtained from

A11 ηi + A12 μi + B11 δ + B12 δ̇ + B13 δ̈ = 0 (10.100)

μ̇i = A21 ηi + A22 μi + B21 δ + B22 δ̇ + B23 δ̈ (10.101)

The solution is then expressed as

η = η0 + ηi − η M AE (10.102)

μ = μ0 + μi − μ M AE (10.103)

where η M AE (t) and μ M AE (t) are correction terms by a matched asymptotic expan-
sion method applied to both the inner and outer solutions [29]. The outer solution is
in fact the asymptotic solution of the original system as t → ∞.
Since the asymptotic behavior of a closed-loop system is an important consider-
ation for stability implication, the outer solution of the singularly perturbed system
is of significant importance. Thus, we obtain the outer solution as the reduced-order
model using only the outer solution of the slow state vector η0 (t) as
390 10 Aerospace Applications
⎛ ⎞
⎜ ⎟
η̇0 = ⎝ A11 − A12 A−1 A21 ⎠ η0
  22 
Ā11

⎡δ⎤
B11 − A12 A−1 −1 −1
22 B21 B12 − A12 A22 B22 B13 − A21 A22 B23 ⎣ ⎦
+        δ̇ + Δη̇0
B̄11 B̄12 B̄13 δ̈
(10.104)

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
y = ⎝C1 − C2 A−1 A21 ⎠ η0 + ⎝ D1 − C2 A−1 B21 ⎠ δ + ⎝ D2 − C2 A−1 B22 ⎠ δ̇
  22    22    22 
C̄1 D̄1 D̄2
⎛ ⎞
⎜ ⎟
+ ⎝ D3 − C2 A−1 B23 ⎠ δ̈ + Δy (10.105)
  22 
D̄3

We will use the slow state vector η0 (t) as the approximation of the actual slow
state vector η (t).
Now, consider a simplified second-order actuator model

δ̈ + 2ζ ωn δ̇ + ωn2 δ = ωn2 δc (10.106)

The state-space representation then becomes

⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤
η̇ Ā11 B̄11 − B̄13 ωn2 B̄12 − 2 B̄13 ζ ωn η B̄13 ωn2 Δη̇
⎣ δ̇ ⎦ = ⎣ 0 0 I ⎦ ⎣ δ ⎦ + ⎣ 0 ⎦ δc + ⎣ 0 ⎦
 
δ̈ 0 −ωn2
−2ζ ωn δ̇ ω2 u 0
          n    
ẋ A x B Δẋ
⎡ ⎤ (10.107)
  η
y = C̄1 D̄1 − D̄3 ωn2 D̄2 − 2 D̄3 ζ ωn ⎣ δ ⎦ + D̄3 ωn2 δc +Δy (10.108)
   δ̇     
u
C    D
x

which can be expressed in the canonical form

ẋ = Ax + Bu + Δ (10.109)

y = C x + Du (10.110)

where Δ is the residual high-order dynamics of the high-frequency aeroelastic modes.

10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 391

Example 10.7 To illustrate the model reduction, the ASE state-space models with
22 modes at various flight conditions are developed for the flexible wing GTM roll
dynamics coupled to the wing antisymmetric modes [30]. The open-loop ASE model
has two unstable modes at Mach 0.86 and an altitude of 10,000 ft. Using the model
reduction method above, it can be shown in Table 10.4 that a reduced-order model
using only the first 8 modes can capture all the unstable modes and approximate the
first 6 modes of the higher-order ASE model quite well [28].

Table 10.4 Antisymmetric modes of ESAC wing @ Mach 0.86 and Altitude 10,000 ft
Mode n = 6 n=7 n=8 n = 22 (Full)
Rigid −2.7392 −2.7395 −2.7385 −2.7385
1 2.7294 ± 19.8683i 2.7512 ± 19.8529i 2.7804 ± 19.8561i 2.7842 ± 19.8513i
2 −0.1553 ± 24.3565i −0.1557 ± 24.3562i −0.1547 ± 24.3553i −0.1549 ± 24.3552i
3 −6.3434 ± 24.0892i −6.3272 ± 24.0739i −6.4220 ± 23.9949i −6.4174 ± 23.9920i
4 −0.3902 ± 37.1580i −0.3782 ± 37.1461i 0.0571 ± 37.4423i 0.0584 ± 37.4846i
5 −20.0160 ± −20.2813 ± −20.4217 ± −20.4833 ±
32.3722i 32.3013i 32.4999i 32.5445i


Suppose we introduce an uncertainty into the nominal ASE state-space model as

ẋ = (A + ΔA) x + (B + ΔB) u (10.111)

y = C x + Du (10.112)

where A ∈ Rn × Rn , B ∈ Rn × Rm , C ∈ R p × Rn , D ∈ R p × Rm with p ≥ m,
ΔA = δ A A, and ΔB = δ B B are known perturbations of the A and B matrices, and
δ A and δ B are assumed to be small parameters that represent multiplicative model
variations.
Observer state feedback control based on output measurements is generally used
for flutter suppression control since the full-state information is not measurable. The
idea is to design an adaptive augmentation controller that is robust to the perturbation
in the nominal plant model due to the model variation. Toward that end, we assume
the pair (A, B) is controllable, and the pair (A, C) is observable.
In general, output feedback adaptive control of MIMO systems can be quite dif-
ficult. The SPR condition for a system transfer function matrix is required as is the
case with a SISO system for stable adaptive control. In general, the system transfer
function matrix from u (t) to y (t) can be non-square, hence non-SPR, if the number
of inputs and the number of outputs are different. The Kalman–Yakubovich lemma
[31] can be used to determine whether the transfer function G (s) = C (s I − A)−1 B
is SPR by the following conditions:
392 10 Aerospace Applications

P A + A P = −Q (10.113)

P B = C (10.114)

for any P = P  > 0 and Q = Q  > 0.

Equation (10.114) leads to the following necessary and sufficient condition for a
SPR transfer function matrix [32]:

B  P B = B C  = C B > 0 (10.115)

Therefore, the condition C B > 0 requires C B to be a symmetric positive-definite

square matrix. This condition requires as a minimum that rank (C B) = m. This
implies that the number of inputs is equal to number of outputs. If the numbers of
inputs and outputs are not equal, we will assume that the number of outputs is greater
than the number of inputs m. This condition will lead to an easier output feedback
adaptive control design by employing various methods of squaring up a non-square
matrix [33, 34]. These methods are used in a number of adaptive output feedback
control approaches such as the observer state feedback adaptive control method by
Lavretsky and Wise [35].
As shown in Sects. 9.13 and 9.14, the optimal control modification method can
be used to control non-SPR SISO systems with non-minimum phase behaviors or
with relative degree greater than 1. We will extend this approach to MIMO sys-
tems by taking advantage of the linear asymptotic property of the optimal control
modification.
The LQG method is a standard technique for control design of systems with output
or partial-state information. A Luenberger state observer is constructed to estimate
the nominal plant model using the Kalman filter optimal estimation method as
 
x̂˙ = A x̂ + L y − ŷ + Bu (10.116)

where x̂ (t) is the observer state vector, L is the Kalman filter gain, and ŷ (t) is the
observer output given by
ŷ = C x̂ + Du (10.117)

A loop transfer recovery technique can be applied to improved the closed-loop

stability margin of the observer system.
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 393

We employ the separation principle of control and estimation by invoking an

assumption that a full-state feedback design provides a stabilizing controller for the
observer system.
Consider an ideal full-state observer controller
 
u ∗ = K x∗ x̂ + K y∗ y − ŷ + K r∗r (10.118)

The ideal closed-loop observer model is then expressed as

   
x̂˙ = A x̂ + L y − ŷ + B K x∗ x̂ + B K y∗ y − ŷ + B K r∗r (10.119)

The objective is to design a closed-loop full-state observer model that track a

reference model
ẋm = Am xm + Bm r (10.120)

where
A + B K x∗ = Am (10.121)

B K r∗ = Bm (10.122)

B K y∗ = −L (10.123)

Then, the adaptive controller is designed as

 
u = K x (t) x̂ + K y (t) y − ŷ + K r (t) r (10.124)

Let K̃ x (t) = K x (t)− K x∗ , K̃ y (t) = K y (t)− K y∗ , and K̃ r (t) = K r (t)− K r∗ be the

estimation errors of K x (t), K y (t), and K r (t), respectively. Then, the closed-loop
plant is written as
 
x̂˙ = Am x̂ + Bm r + B K̃ x x̂ + B K̃ y y − ŷ + B K̃ r r (10.125)

Let e p (t) = x (t) − x̂ (t) be the state estimation error. Then, the state estimation
error equation is computed as
   
ė p = A p e p + ΔA e p + x̂ + ΔB K x x̂ + K y Ce p + K r r (10.126)

where A p = A − LC is Hurwitz.
394 10 Aerospace Applications

Let e (t) = xm (t) − x̂ (t) be the tracking error, then the tracking equation is
expressed as
ė = Am e − B K̃ x x̂ − B K̃ y Ce p − B K̃ r r (10.127)

The adaptive laws for K x (t), K y (t), and K r (t) using the optimal control modi-
fication [28] are given by
 
K̇ x = Γx x̂ e P + νx x̂  K x B  P A−1
m B (10.128)
   
K̇ y = Γ y y − ŷ e P + ν y y − ŷ K y B  P A−1
m B (10.129)

 
K̇ r = Γr r e P + νr r  K r B  P A−1
m B (10.130)

with the initial conditions K x (0) = K̄ x , K y (0) = K̄ y , and K r (0) = K̄ r , where

Γx = Γx > 0, Γ y = Γ y > 0, and Γr = Γr > 0 are the adaptation rate matrices;
νx > 0 ν y > 0, and νr > 0 are the modification parameters; and P = P  > 0 is the
solution of the Lyapunov equation

P Am + A
mP + Q = 0 (10.131)

where Q = Q  > 0.
As discussed in Sect. 9.14, if the reference model is SPR and the plant is non-
SPR, then an adaptive control design with the standard MRAC is infeasible unless
the reference model is modified to be non-SPR from the ideal control design for
the non-SPR plant. The optimal control modification has been shown to be able
to handle non-SPR plants with either SPR or non-SPR reference models. For an
adaptive augmentation design, the robust baseline controller can also serve to reduce
the sensitivity of the closed-loop plant to the non-minimum phase behavior if a robust
modification or the projection method is used.
Consider an adaptive augmentation regulator design with the following adaptive
controller:  
u = K̄ x x̂ + ΔK x x̂ + K y y − ŷ (10.132)

where K̄ x is obtained from the Linear Quadratic Regulator (LQR) design of the
full-state equation, and the adaptive laws for ΔK x (t) and K y (t) are given by
 
Δ K̇ x = −Γx x̂ x̂  P − νx ΔK x B  P A−1
m B (10.133)
   
K̇ y = −Γ y y − ŷ x̂  P − ν y y − ŷ K y B  P A−1
m B (10.134)

The adaptive laws can be shown to be stable as follows:

Proof Choose a Lyapunov candidate function

10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 395
  
V x̂, e p , Δ K̃ x , K̃ y = x̂  P x̂ + e −1  −1 
p W e p + trace Δ K̃ x Γx Δ K̃ x + trace K̃ y Γ y K̃ y
(10.135)
where
W A p + Ap W = −R < 0 (10.136)

Then, evaluating V̇ x̂, e p , Δ K̃ x , K̃ y yields

V̇ x̂, e p , Δ K̃ x , K̃ y
   
= −x̂  Q x̂ − e  
p Re p + 2e p W ΔA e p + x̂ + 2e p W ΔB K̄ x x̂ + ΔK x x̂ + K y Ce p
+ 2νx x̂  ΔK x B  P A−1     −1
m BΔ K̃ x x̂ + 2ν y e p C K y B P Am B K̃ y Ce p
= −x̂  Q x̂ − e    − −1
p Re p − νx x̂ Δ K̃ x B Am Q Am BΔ K̃ x x̂
− ν y e    − −1  ∗  −1
p C K̃ y B Am Q Am B K̃ y Ce p + 2νx x̂ ΔK x B P Am BΔ K̃ x x̂
+ 2ν y e  ∗  −1
p C K y B P Am B K̃ y Ce p
⎛ ⎞
⎜ ⎟
+ 2e ⎜ ∗ ⎟
p W ⎝ΔA + ΔB K y C ⎠ e p
  
ΔA p
⎛ ⎞
⎜ ∗⎟
+ 2e  
p W ⎝ΔA + ΔB K̄ x + ΔBΔK x ⎠ x̂ + 2e p W ΔBΔ K̃ x x̂ + 2e p W ΔB K̃ y Ce p
  
ΔAm
(10.137)

V̇ x̂, e p , Δ K̃ x , K̃ y is bounded by

 $ $2 $ $2 $ $
$2
$
V̇ x̂, e p , Δ K̃ x , K̃ y ≤ − c1 $x̂ $ − νx c2 $x̂ $ $Δ K̃ x $
$ $2 $$ $ $ $2
$
+ 2νx c2 c3 $x̂ $ $Δ K̃ x $ − (c4 − 2c7 ) $e p $
$ $2 $ $ $2
$ $ $2 $ $ $
$ $ $$ $
− ν y c5 $e p $ $ K̃ y $ + 2ν y c5 c6 $e p $ $ K̃ y $ + 2c8 $x̂ $ $e p $
$ $$ $$ $
$
$ $ $2 $ $ $
$
+ 2c9 $x̂ $ $e p $ $Δ K̃ x $ + 2c10 $e p $ $ K̃ y $ (10.138)

 
B  P A−1
m B

ΔK x

∗
where c1 = λmin (Q), c2 = λmin B  A− −1
m Q Am B , c3 = c2
, c4 =
m B

K y

B  P A−1 ∗ $ $
λmin (R), c5 = c2
C
, c6 =
2
c5
, c7 = $W ΔA p $, c8 =
W ΔAm
,
c9 =
W ΔB
, and c10 = c9
C
.
Using the inequality 2
a

b
≤
a
2 +
b
2 , we obtain
396 10 Aerospace Applications

 $ $2 $ $2 $$
$2
$
V̇ x̂, e p , Δ K̃ x , K̃ y ≤ − (c1 − c8 ) $x̂ $ − (νx c2 − c9 ) $x̂ $ $Δ K̃ x $
$ $2 $$ $ $ $2
$
+ 2νx c2 c3 $x̂ $ $Δ K̃ x $ − (c4 − 2c7 − c8 − c9 ) $e p $
$ $2 $ $ $2
$   $ $2 $$ $
$
− ν y c5 $e p $ $ K̃ y $ + 2 ν y c5 c6 + c10 $e p $ $ K̃ y $
(10.139)

With further simplification, we get

% &
 νx2 c22 c32 $ $2
V̇ x̂, e p , Δ K̃ x , K̃ y ≤ − c1 − c8 − $x̂ $
νx c2 − c9
 
$ $2 $ $
$
$ νx c2 c3 2
− (νx c2 − c9 ) $x̂ $ $Δ K̃ x $ −
νx c2 − c9

 2 
ν y c5 c6 + c10 $ $2
− c4 − 2c7 − c8 − c9 − $e p $
ν y c5
 2
$ $2 $ $
$ $2 ν y c5 c6 + c10
− ν y c5 $e p $ $ K̃ y $ − (10.140)
ν y c5

We choose Q, R, νx , and ν y to satisfy the following inequalities:

νx2 c22 c32

c1 − c8 − >0 (10.141)
νx c2 − c9

νx c2 − c9 > 0 (10.142)
 2
ν y c5 c6 + c10
c4 − 2c7 − c8 − c9 − >0 (10.143)
ν y c5

Then, V̇ x̂, e p , Δ K̃ x , K̃ y ≤ 0. Therefore, the adaptive regulator design is stable.

Note that this adaptive augmentation controller is designed to tolerate the uncer-
tainty due to the plant model variation, but does not attempt to accommodate the
uncertainty by canceling the uncertainty. The goal is to enable the adaptive control
design to have a sufficient robustness due to the plant model variation. In the event
the plant model variation is zero, then c7 = 0, c8 = 0, c9 = 0, and c10 = 0. This
implies νx ≤ ccc1 2 and ν y ≤ ccc4 2 . On the other hand, if the plant model variation is
2 3 5 6
large, it is possible that there are no suitable values of νx and ν y that can stabilize the
plant.
The Lyapunov stability analysis shows how the modification parameters νx and
ν y could be selected to ensure a stable adaptation. As discussed in Sect. 9.5.3, the
Lyapunov stability analysis often results in a high degree of conservatism that could
render the selection of the modification parameters infeasible. Therefore, in practice,
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 397

the linear asymptotic property of the optimal control modification can be invoked to
find suitable values of νx and ν y to ensure the closed-loop stability. In the limit as
Γ → ∞, the following adaptive parameters tend to:

1   − −1 
BΔK x → B B Am P B B P (10.144)
νx

1   − −1 
BKy y → B B Am P B B P x̂ (10.145)
νy

Using this linear asymptotic property, it is possible to estimate stability margins of

the closed-loop system in the limit as shown in Sects. 9.13 and 9.14. The closed-loop
system then tends to an asymptotic linear system in the limit as
   −1
1 1
ẋ = (1 + δ A ) Ax +(1 + δ B ) B K̄ x x̂ + + (1 + δ B ) B B  A−
m PB B  P x̂
νx νy
(10.146)
   
1 1  −1 
x̂˙ = A − LC + B K̄ x + + B B  A−
m PB B P x̂ + LC x
νx νy
(10.147)
Thus, stability margins can be determined for the closed-loop plant matrix which
is given by
⎡   −1  ⎤
(1 + δ A ) A (1 + δ B ) B K̄ x ++ ν1y (1 + δ B ) B B  A−
1
νx m PB B P
Ac = ⎣    −1
⎦
LC A − LC + B K̄ x + ν1x + ν1y B B  A− m P B B 
P
(10.148)
Then, νx and ν y can be chosen to provide the closed-loop stability.
It is noted that K y (t) accounts for the model variation in the Luenberger state
observer design. Alternatively, K y (t) could be eliminated if the observer design
is sufficiently robust using the loop transfer recovery method. Then, the observer
feedback adaptive control with the optimal control modification comes down to
selecting a suitable value of νx to ensure the closed-loop stability.

Example 10.8 The ASE state-space model in Example 10.7 contains 45 states,
64 outputs, and 4 inputs. The states include the aircraft roll rate p (t) and two
generalized states for each of the 22 aeroelastic modes. The outputs include the
aircraft airspeed V (t), angle of attack α (t), sideslip angle β (t), aircraft angu-
   
lar rates p (t) q (t) r (t) , aircraft position x (t) y (t) h (t) , aircraft attitude
   
φ (t) θ (t) ψ (t) , accelerations in the three axes N x , N y , Nz at the forward and
aft locations of the wing tip, and the four hinge moments of the control surfaces.
The control surfaces are the four outboard third camber segments of the VCCTEF.
All these surfaces are blended together to provide a single input to address the
relative constraints imposed by the elastomer material. For the flutter suppression
398 10 Aerospace Applications

control design, only the two Nz acceleration measurements are used as the outputs.
Figure 10.41 illustrates the input and output locations [28].

All the control surfaces are not entirely independent in their motions due to the
physical constraints imposed by the elastomer material. This material has certain
displacement and rate limits. Thus, any adjacent control surfaces will also have
relative displacement and rate limits. These limits are in addition to the normal
position and rate limits that each of the control surfaces is subjected to. Thus, these
relative constraints can cause challenges in a control design of this system.
Consider the following relative displacement and rate constraints:

|δi+1 − δi | ≤ Δδ (10.149)
' '
'δ̇i+1 − δ̇i ' ≤ Δδ̇ (10.150)

where i = 1, 2, 3.
For the VCCTEF design, the relative motion between any adjacent flap sections
is allowed to be within two degrees. The rate constraint imposed by the elastomer
material is not yet defined and thus is assumed to be large. The actuator dynamics
are modeled as a second-order system. This actuator model is highly simplified since
it does not take into account the hinge moment which is a function of the states and
the dynamics of the elastomer material in the overall actuator model.
To address the relative displacement limit, a concept of virtual control is introduced
[36]. The control surface deflections are described by a shape function. This shape
function can be any reasonable shape function with a smooth and gradual slope.
One simple function is a linear function. The control surface deflections are then
parametrized as a linear function
iδv
δi = (10.151)
4
where i = 1, 2, 3, 4 such that δ1 (t) is the inboard flap and δ4 (t) is the outboard flap,
and δv (t) is the virtual control surface deflection.
Since the inboard flap section δ1 (t) cannot deflect more than two degrees relative
to the stationary flap adjacent to it, then δv (t) ≤ 8 deg. Also, the outboard flap
deflection δ4 (t) is the same as the virtual control surface deflection. Thus, one can
think that the outboard flap δ4 (t) is a master control input, while the other three
control surfaces are slave control inputs since their motions are dependent on the
master control input.
Thus, the virtual control derivatives are computed as

(
4
i Bi jk
B jk = (10.152)
i=1
4

where B jki is the control derivative of mode jth with respect to the displacement
(k = 1), velocity (k = 2), and acceleration (k = 3) of flap section ith.
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 399

Fig. 10.41 Flutter suppression control surfaces and accelerometer locations

The simulation is conducted with only the reduced-order aeroservoelastic state-

space model for a flight condition at Mach 0.86 and an altitude of 10,000 ft. There
are two unstable aeroelastic modes: mode 1 and mode 4, as shown in Table 10.4.
Process noise and sensor noise are introduced to simulate the structural response to
atmospheric turbulence. The baseline full-state feedback controller is designed with
a LQR controller tuned to give good performance. A LQG output feedback controller
is then designed using the ideal full-state feedback gain. The adaptive augmentation
controller is then turned on. The adaptation rate matrices and modification parameters
are selected to be Γx = Γ y = 1 and ηx = η y = 0.1.
The root locus of the open-loop transfer functions between the accelerometers and
the virtual control are shown in Figs. 10.42 and 10.43. It is shown that the individual
transfer functions have two unstable poles corresponding to the two unstable modes
and zeros on the right half plane. However, the transmission zeros of the MIMO
system transfer function actually are stable.
An initial roll rate of 0.1 rad/s is specified. Figures 10.44, 10.45, and 10.46 show the
responses of the roll rate, the generalized displacement of mode 1, and the generalized
displacement of mode 4, respectively, without the process and sensors noises for the
baseline full-state feedback LQR controller and the output feedback LQG controller
with and without the adaptive augmentation controller. The full-state feedback LQR
controller performs much better than the output feedback LQG controller with and
without the adaptive augmentation controller using only the two accelerometers at
the wing tip. The LQG controller does seem to need further tuning. The adaptive
augmentation controller causes an increase in the overshoot of the responses due to
the output feedback LQG controller as well as injects high-frequency contents into
the modal responses. Nonetheless, all the controllers are able to suppress the two
unstable aeroelastic modes.
The effects of process and sensor noises are examined. Figures 10.47, 10.48, and
10.49 show the responses of the roll rate, the generalized displacement of mode
1, and the generalized displacement of mode 4, respectively, with the process and
400 10 Aerospace Applications

Root Locus of Open−Loop Forward Accelerometer

150

100

50

jω (seconds−1)
0

−50

−100

−150
−50 −40 −30 −20 −10 0 10 20 30 40 50
−1
σ (seconds )

Fig. 10.42 Root locus of open-loop transfer function of forward accelerometer

Root Locus of Open−Loop Aft Accelerometer

500

400

300

200
jω (seconds−1)

100

−100

−200

−300

−400

−500
−100 −80 −60 −40 −20 0 20 40 60 80 100
−1
σ (seconds )

Fig. 10.43 Root locus of open-loop transfer function of aft accelerometer

sensor noises for the baseline full-state feedback LQR controller and the output
feedback LQG controller with and without the adaptive augmentation controller. All
the controllers are able to maintain good performance in the presence of the process
and sensor noises.
Figures 10.50 and 10.51 show the root locus plots of the closed-loop transfer
functions of the accelerometers with the adaptive augmentation controller using the
final gain matrix. As can be seen, the closed-loop transfer functions are completely
stable. Figures 10.52 and 10.53 are the frequency response plots for the open-loop and
closed-loop transfer functions of the accelerometers. The largest frequency response
is due to mode 4. The closed-loop frequency response is significantly less than the
open-loop frequency response, indicating the effectiveness of the aeroelastic mode
suppressioncontroller.
Next, the plant model variation is introduced into the aeroservoelastic state-space
model by specifying ΔA = 0.05A and ΔB = −0.1B. The process and sensor
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 401

0.12
LQR
0.1 LQG
LQG + Adaptive
0.08

0.06

Roll Rate p, rad/sec

0.04

0.02

−0.02

−0.04

−0.06

−0.08
0 2 4 6 8 10
t, sec

Fig. 10.44 Roll rate response for LQR and LQG controllers without process and sensor noises

7
LQR
6 LQG
LQG + Adaptive
5
Mode 1 Generalized Displacement

−1

−2

−3
0 2 4 6 8 10
t, sec

Fig. 10.45 Mode 1 generalized displacement response for LQR and LQG controllers without
process and sensor noises

1.5
LQR
LQG
LQG + Adaptive
1
Mode 4 Generalized Displacement

0.5

−0.5

−1
0 2 4 6 8 10
t, sec

Fig. 10.46 Mode 4 generalized displacement response for LQR and LQG controllers without
process and sensor noises
402 10 Aerospace Applications

0.12
LQR
0.1 LQG
LQG + Adaptive
0.08

0.06

Roll Rate p, rad/sec

0.04

0.02

−0.02

−0.04

−0.06

−0.08
0 2 4 6 8 10
t, sec

Fig. 10.47 Roll rate response for LQR and LQG controllers with process and sensor noises

7
LQR
6 LQG
LQG + Adaptive
5
Mode 1 Generalized Displacement

−1

−2

−3
0 2 4 6 8 10
t, sec

Fig. 10.48 Mode 1 generalized displacement response for LQR and LQG controllers with process
and sensor noises
1.5
LQR
LQG
LQG + Adaptive
1
Mode 4 Generalized Displacement

0.5

−0.5

−1

−1.5
0 2 4 6 8 10
t, sec

Fig. 10.49 Mode 4 generalized displacement response for LQR and LQG controllers with process
and sensor noises
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 403

Root Locus of Closed−Loop Forward Accelerometer

250

200

150

100

jω (seconds−1)
50

−50

−100

−150

−200

−250
−50 −40 −30 −20 −10 0 10 20 30 40 50
−1
σ (seconds )

Fig. 10.50 Root locus of closed-loop transfer function of forward accelerometer

Root Locus of Closed−Loop Aft Accelerometer

500

400

300

200
jω (seconds−1)

100

−100

−200

−300

−400

−500
−100 −80 −60 −40 −20 0 20 40 60 80 100
σ (seconds−1)

Fig. 10.51 Root locus of closed-loop transfer function of aft accelerometer

Bode Plot of Forward Accelerometer

Open−Loop
50 Closed−Loop
Magnitude (dB)

−50
1080
900 Open−Loop
Closed−Loop
Phase (deg)

720
540
360
180
0
−180
−2 0 2 4
10 10 10 10
ω (rad/s)

Fig. 10.52 Frequency response of forward accelerometer

404 10 Aerospace Applications

Bode Plot of Aft Accelerometer

Open−Loop
50 Closed−Loop

Magnitude (dB)
0

−50
1080
Open−Loop
900 Closed−Loop
Phase (deg)

720
540
360
180
0
−2 0 2 4
10 10 10 10
ω (rad/s)

Fig. 10.53 Frequency response of aft accelerometer

noises are also included. The output feedback LQG controller without the adaptive
augmentation controller is unstable as shown in Figs. 10.54, 10.55, and 10.56. On the
other hand, the adaptive augmentation controller is able to stabilize the aeroelastic
modes in the presence of the plant model variation. The closed-loop plant matrix
without the adaptive augmentation controller is in fact unstable. Stability of the
closed-loop plant is restored in the presence of the adaptive augmentation controller.

0.15
LQG
LQG + Adaptive

0.1
Roll Rate p, rad/sec

0.05

−0.05

−0.1
0 2 4 6 8 10
t, sec

Fig. 10.54 Roll rate response with and without adaptive augmentation controller
10.6 Adaptive Linear Quadratic Gaussian Flutter Suppression Control 405

20
LQG
LQG + Adaptive
15

Mode 1 Generalized Displacement

10

−5

−10

−15

−20
0 2 4 6 8 10
t, sec

Fig. 10.55 Mode 1 generalized displacement response with and without adaptive augmentation
controller

2.5
LQG
2 LQG + Adaptive

1.5
Mode 4 Generalized Displacement

0.5

−0.5

−1

−1.5

−2

−2.5
0 2 4 6 8 10
t, sec

Fig. 10.56 Mode 4 generalized displacement response with and without adaptive augmentation
controller

Figure 10.57 is the plot of the time history of the virtual control command for
the output feedback LQG controller with and without the adaptive augmentation
controller. The largest amplitude of the stabilizing virtual control command for the
adaptive augmentation controller is 6.22o . The linear mapping between the virtual
control command and the physical control commands results in 1.56o which meets
the physical constraint of 2o on the relative deflection of the VCCTEF.
406 10 Aerospace Applications

6
LQG
LQG + Adaptive
4

Virtual Control Command δ, deg

2

−2

−4

−6

−8
0 2 4 6 8 10
t, sec

Fig. 10.57 Virtual control command with and without adaptive augmentation controller

10.7 Adaptive Flight Control

Flight control is an important aircraft system that provides many critical functions to
enable safe operations of an aircraft. A pilot controls an aircraft by moving a control
input device such as a control yoke, rudder pedal, or side stick. A pilot command
signal is sent to a flight control system which processes the signal further and converts
it into an actuator command signal. This signal is then sent to an actuator system that
drives the position of a flight control surface such as the aileron, elevator, or rudder.
This flight control feature is called a pilot command tracking task. Another important
flight control feature is to enhance aircraft stability via a stability augmentation
system (SAS). The SAS provides the additional damping to the rigid aircraft modes
by feedback control. A typical SAS on an aircraft is a yaw damper.
Connections between control input devices to flight control surfaces can be made
via mechanical linkages which exist in older aircraft. Hydraulic flight control systems
use hydraulic fluid circuits in conjunction with partial mechanical flight control sys-
tems to drive flight control surfaces. Hydraulic flight control systems are still found
in many aircraft in operation nowadays. Modern aircraft usually employ fly-by-wire
(FBW) flight control systems which replace mechanical systems with electronics. A
pilot command is converted into an electrical signal transmitted to a flight control
computer for processing. The flight control computer computes actuator command
signals and transmits these signals to hydraulic actuators to drive flight control sur-
faces.
A typical flight control system consists of an inner-loop stability augmentation
system and an outer-loop autopilot system as illustrated in the Fig. 10.58. The com-
mand input into a flight control system can be accomplished via a pilot’s direct
mechanical input or an autopilot. When engaged, the autopilot generates an attitude
command of the attitude angles or an altitude command for a selected autopilot task
such as cruise or landing. This command is processed as a feedforward signal and
then is summed with the SAS to form a control signal. The SAS uses feedback of
10.7 Adaptive Flight Control 407

aircraft’s sensed angular rate variables, such as the roll rate, pitch rate, and yaw rate
to provide the additional damping to the aircraft dynamics. The control signal is
transmitted to flight control actuators that drive flight control surfaces to change the
aircraft dynamics.

Fig. 10.58 Flight control system

Most conventional flight control systems utilize extensive gain-scheduling

schemes via table-lookups in order to achieve desired handling qualities through-
out a flight envelope. While this approach has proven to be very successful, the
development process can be expensive and often results in aircraft-specific imple-
mentations. In rare occurrences when a failure or damage occurs to an aircraft, a
conventional flight control system may no longer perform pilot command tracking
tasks as intended since the aircraft flight dynamics may deviate from its design char-
acteristics substantially, thereby causing a degradation in the performance of the
flight control system. Adaptive flight control can offer a possibility for restoring the
performance of a flight control system in an event of failure or damage to an aircraft.
The linearized equations of the angular motion for an aircraft with failures or
damage can be expressed in general as [19]

ẋ = (C + ΔC) x + (D + ΔD) u + (E + ΔE) z (10.153)

 
where x (t) = p (t) q (t) r (t) is an inner-loop state vector comprising the roll,
 
pitch, and yaw rates; u (t) = δa (t) δe (t) δr (t) is a control vector
comprising the aileron, elevator, and rudder control surface deflections;
 
z (t) = φ (t) α (t) β (t) V (t) h (t) θ (t) is an outer-loop state vector comprising
the bank angle, angle of attack, sideslip angle, airspeed, altitude, and pitch angle;
C ∈ R3 × R3 , D ∈ R3 × R3 , and E ∈ R3 × R6 are the nominal aircraft plant matrices
for the inner-loop dynamics; and ΔC, ΔD, and ΔE are the changes in the inner-loop
plant matrices due to failures or damage.
In general, the inner-loop state variables; namely, the angular rates have faster
dynamics than the outer-loop variables. This frequency separation allows the inner-
loop dynamics to be decoupled from the outer-loop dynamics in a flight control
design. This can simplify the design significantly. The outer-loop dynamics are
described by the following equation:
408 10 Aerospace Applications

ż = (F + ΔF) z + (G + ΔG) u + (H + ΔH ) x (10.154)

where F ∈ R6 × R6 , G ∈ R6 × R3 , and H ∈ R6 × R3 are the nominal outer-loop

aircraft plant matrices; and ΔF, ΔG, and ΔH are the changes in the outer-loop plant
matrices due to failures or damage.
Consider the following inner-loop rate command adaptive flight control archi-
tecture as shown in Fig. 10.59. The control architecture comprises: (1) a reference
model that translates a rate command r (t) into a reference acceleration command
ẋm (t), (2) a proportional-integral (PI) feedback control for stability augmentation
and tracking, (3) a dynamic inversion controller that computes actuator commands
u c (t) to achieve a desired acceleration command ẋd (t), and (4) an adaptive control
augmentation that computes an adaptive signal u ad (t) that augments the reference
acceleration command to form a desired acceleration command for the dynamic
inversion controller. The adaptive control augmentation is the standard MRAC with
any robust modification schemes, such as the projection method, σ modification, e
modification, optimal control modification, and adaptive loop recovery modification
with or without the normalization and covariance adjustment methods.

Fig. 10.59 Adaptive flight control architecture

For rate command attitude hold (ACAH) flight control, a first-order reference
model is used to filter the rate command r (t) into a reference acceleration command
ẋm (t) as
ẋm = −Ω (xm − r ) (10.155)
 
where Ω = diag ω p , ωq , ωr is a diagonal matrix of the reference model frequencies.
The nominal controller is then specified as
) t
ū = K p (xm − x) + K i (xm − x) dτ (10.156)
0
 
 K2 p =2 diag
where  2ζ p ω p , 2ζq ωq , 2ζr ωr is the proportional gain matrix and K i =
diag ω p , ωq , ωr2 is the integral gain matrix [37].
For transport aircraft, typical values of the reference model frequencies ω p , ωq ,
and ωr are 3.5, 2.5, and 2.6, respectively; and typical values of the reference model
damping ratios ζ p , ζq , and ζr are √12 .
10.7 Adaptive Flight Control 409

The desired acceleration is computed as

ẋd = ẋm + ū − u ad (10.157)

The adaptive signal u ad (t) is designed to estimate the uncertainty in the aircraft
plant dynamics.
The desired acceleration is the ideal response of the nominal aircraft plant dynam-
ics. Therefore, the ideal aircraft dynamics can be expressed as

ẋd = C x + Du + E z (10.158)

The dynamic inversion controller then computes the actuator command from the
closed-loop plant model of the ideal aircraft dynamics [37]

u c = D −1 (ẋd − C x − E z) (10.159)

Substituting u c (t) for u (t) into the aircraft model, then the closed-loop plant
model of the aircraft is obtained as

ẋ = ẋm + ū − u ad + ΔC x + ΔDu c + ΔE z (10.160)

   
Let Θ ∗ = ΔC ΔD ΔE and Φ (x, u c , z) = x  (t) u  
c (t) z (t) . Then, the
adaptive signal u ad (t) can be designed as
⎤ ⎡
 x 
u ad = ΔĈ (t) x + Δ D̂ (t) u c + Δ Ê (t) z = ΔĈ Δ D̂ Δ Ê ⎣ u c ⎦ = Θ  (t) Φ (x, u c , z)
z
(10.161)

Then,
) t
ẋ = ẋm + K p (xm − x) + K i (xm − x) dτ − Θ̃  Φ (x, u c , z) (10.162)
0

where Θ̃ (t) = Θ (t) − Θ ∗ is the estimation error of the unknown matrices.

The tracking error equation is expressed as
) t
ẋm − ẋ = −K p (xm − x) − K i (xm − x) dτ + Θ̃  Φ (x, u c , z) (10.163)
0

*t   
Let e (t) = 0 xm (t) − x  (t) dτ xm (t) − x  (t) . Then,

ė = Am e + B Θ̃  Φ (x, u c , z) (10.164)

where
410 10 Aerospace Applications
 
0 I
Am = (10.165)
−K i −K p
 
0
B= (10.166)
I

Then, the adaptive law can be designed with any robust adaptive control scheme
such as the e modification
 $ $ 
Θ̇ = −Γ Φ (x, u c , z) e P B + μ $e P B $ Θ (10.167)

Alternatively, if the optimal control modification is used, then

 
Θ̇ = −Γ Φ (x, u c , z) e P − νΦ  (x, u c , z) Θ B  P A−1
m B (10.168)

Note that actuator dynamics can affect the closed-loop stability of an aircraft.
Therefore, a suitable selection of the modification parameter should be made to ensure
sufficient robustness. An actuator system must have a higher frequency bandwidth
than the frequency of a given mode in consideration. Roughly speaking, the frequency
bandwidth of an actuator should be at least several times greater than the highest
frequency of the aircraft modes. For example, the short-period mode is about 2.5 rad/s
and then the frequency of the actuator for the elevator could be about 10 times larger
or 25 rad/s. The damping ratio should be sufficiently large to provide a well-damped
control surface deflection.
For example, a second-order actuator system is modeled as

ü + 2ζa ωa u̇ + ωa2 u = ωa2 u c (10.169)

 
where ωa  max ω p , ωq , ωr .
The response of an actuator is subject to both position and rate limits. Since flight
control surfaces are generally of a flap-type design, as the flap deflection increases
beyond a certain limit, the performance of the flap becomes degraded due to flow
separation. This would result in nonlinear aerodynamics. Therefore, flight control
surface positions are generally designed to operate within a linear range. Typically,
for the elevator and aileron, the position limits are maintained between ±20o . For
the rudder, the position limit is smaller since the rudder is a powerful flight control
surface that can generate a significant structural loading on the vertical tail at a high
rudder deflection. The rudder position limit is also scheduled with the altitude and
airspeed. As the airspeed and altitude increase, the rudder position limit becomes
smaller.
Typically, the rate limit of an actuator system is established by the design of
the actuator. A typical rate limit is about 60o /s for transport aircraft. If an actuator
is operating normally, a pilot command would translate into an expected response.
However, when a rate limit occurs due to a degraded actuator performance, the
actuator would not produce a sufficiently fast response to meet the pilot command.
10.7 Adaptive Flight Control 411

Consequently, the pilot would tend to apply a greater command to try to overcome
the lag in the actuator dynamics. This actuator lag and the pilot positive feedback can
result in an adverse consequence known as pilot-induced oscillations (PIO), which
could lead to a catastrophic loss of control.

Example 10.9 Adaptive flight control is an effective method for controlling a dam-
aged aircraft. A transport aircraft with wing damage is shown in Fig. 10.60. The
wing damage is modeled as a 28% loss of the left wing [37]. Since the damage is
asymmetric, the motion of the damaged aircraft is fully coupled in the roll, pitch,
and yaw axes.

Fig. 10.60 Transport aircraft with left wing damaged

A level flight condition of Mach 0.6 at 15,000 ft is selected. The remaining right
aileron control surface is the only roll control effector available. The reference model
is specified
√ by ω p = 2.0 rad/s, ωq = 1.5 rad/s, ωr = 1.0 rad/s, and ζ p = ζq = ζr =
1/ 2.
The pilot pitch rate command is simulated with a series of ramp input longitu-
dinal stick command doublets, corresponding to the reference pitch angle ±3.81o
from trim. The tracking performance of the nominal controller without adaptation
is compared against the standard MRAC, the e modification with μ = 0.1, and the
optimal control modification with ν = 0.1. The adaptive gains are selected to be
as large as possible within the numerical stability limit of the adaptive laws. This
results in Γ = 60 for the standard MRAC, Γ = 800 for the e modification, and
Γ = 2580 for the optimal control modification. Thus, it can be seen that the opti-
mal control modification can tolerate a much larger adaption rate than the standard
MRAC. This large adaptation rate allows the optimal control modification to achieve
fast adaptation to better adapt to uncertainty than the standard MRAC.
The aircraft angular rate responses are shown in Fig. 10.61. The nominal controller
without adaptation produces a rather poor tracking of the reference pitch rate. Both the
standard MRAC and e modification improve the tracking performance significantly.
412 10 Aerospace Applications

The optimal control modification (OCM) is able to provide a better tracking than
both the standard MRAC and e modification. Due to the asymmetric wing damage,
the roll axis is most affected. With the nominal controller, there is a significant roll
rate of as high as 20o /s. Both the standard MRAC and e modification reduce the
maximum amplitude of the roll rate to about 10o /s. The optimal control modification
further reduces the roll rate to a maximum value of about 4o /s. All the three adaptive
controllers significantly reduce the yaw rate to a very low level. The e modification
is observed to perform slightly better than the standard MRAC and optimal control
modification in the yaw axis.

20 20 20 20
p, deg/sec

p, deg/sec

p, deg/sec

p, deg/sec
0 0 0 0

Nom. Ref. MRAC Ref. e Mod. Ref. OCM Ref.

−20 −20 −20 −20
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec
5
2 2 2
q, deg/sec

q, deg/sec

q, deg/sec

q, deg/sec
0 0 0 0

−2 Nom. Ref. −2 MRAC Ref. e Mod. Ref. −2 OCM Ref.

−5
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec

0.4 0.4 0.4 0.4

r, deg/sec

r, deg/sec

r, deg/sec

r, deg/sec
0.2 0.2 0.2 0.2
0 0 0 0
−0.2 Nom. Ref. −0.2 MRAC Ref. −0.2 e Mod. Ref. −0.2 OCM Ref.
−0.4 −0.4 −0.4 −0.4
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec

Fig. 10.61 Closed-loop responses of roll, pitch, and yaw rates

The aircraft attitude responses are shown in Fig. 10.62. When there is no adapta-
tion, the reference pitch attitude could not be followed accurately. With adaptive con-
trol, the tracking is much improved and the optimal control modification produces
the pitch angle tracking better than both the standard MRAC and e modification.
Without adaptation, as expected the damaged aircraft exhibits a rather severe roll
behavior with a large bank angle between −30o and 20o . Both the standard MRAC
and e modification reduce the bank angle significantly. However, the optimal control
modification shows a drastic improvement in the arrest of the roll motion with the
bank angle maintained very close to the trim value. All the three adaptive controllers
produce similar angle-of-attack responses. With the optimal control modification,
the sideslip angle is reduced to near zero, while both the standard MRAC and e
modification still produce considerable sideslip angle responses.
10.7 Adaptive Flight Control 413

15 15 15 15
Nom. Ref. MRAC Ref. e Mod. Ref. OCM Ref.
10 10 10 10

θ, deg

θ. deg

θ, deg

θ, deg
5 5 5 5

0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec
20 20 20 20

0 0 0 0
φ, deg

φ, deg

φ, deg

φ, deg
−20 −20 −20 −20
Nom. MRAC e Mod. OCM
−40 −40 −40 −40
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec
10 10 10 10
α, deg

α, deg

α, deg

α, deg
5 5 5 5

Nom. MRAC e Mod. OCM

0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec
5 5 5 5
β, deg

β, deg

β, deg

β, deg
0 0 0 0

Nom. MRAC e Mod. OCM

−5 −5 −5 −5
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
t, sec t, sec t, sec t, sec

Fig. 10.62 Closed-loop responses of pitch angle, bank angle, angle of attack, and sideslip angle

The control surface deflections are shown in Fig. 10.63. Because of the wing
damage, the damaged aircraft has to be trimmed with a rather large aileron deflection.
This causes the roll control authority to be severely limited. Therefore, a roll control
saturation is present in all cases. The elevator deflection is nominally similar for all
the four controllers and is well within its control authority. The rudder deflection
produced by the baseline controller is quite significant. Generally, it is desired to
keep the rudder deflection as small as possible in normal operation. Typically, the
rudder deflection limit is reduced as the airspeed and altitude increase. Both the
standard MRAC and e modification reduce the baseline rudder deflection to some
extent, but the optimal control modification is observed to produce the smallest rudder
deflection.

40 40 40 40

30 30 30 30
δa, deg

δa, deg

δ , deg

δ , deg
a

20 20 20 20
Nom. MRAC e Mod. OCM
10 10 10 10
0 20 40 0 20 40 0 20 40 0 20 40
t, sec t, sec t, sec t, sec
2 2 2 2

0 0 0 0
δe, deg

δe, deg

δe, deg

δ , deg
e

−2 −2 −2 −2
Nom. MRAC e Mod. OCM
−4 −4 −4 −4
0 20 40 0 20 40 0 20 40 0 20 40
t, sec t, sec t, sec t, sec
5 5 5 5

0 0 0 0
δr, deg

δ , deg

δ , deg

δ , deg
r

−5 −5 −5 −5
Nom. MRAC e Mod. OCM
−10 −10 −10 −10
0 20 40 0 20 40 0 20 40 0 20 40
t, sec t, sec t, sec t, sec

Fig. 10.63 Closed-loop responses of aileron, elevator, and rudder deflections

414 10 Aerospace Applications

10.8 Hybrid Adaptive Flight Control

In the adaptive flight control design in Sect. 10.7, MRAC is used to estimate the
plant uncertainty due to ΔC, ΔD, and ΔE. However, it is well known that MRAC
is designed to drive the tracking error to zero, but is not designed with the primary
objective of parameter estimation. Least-squares methods on the hand are well-
known techniques for parameter estimation. It is possible to combine both the least-
squares parameter estimation and MRAC into a control architecture. This approach
for flight control is called hybrid adaptive flight control [37].

Fig. 10.64 Hybrid adaptive flight control system

The hybrid adaptive flight control architecture is shown in Fig. 10.64. The archi-
tecture includes an indirect adaptive control that uses a least-squares parameter esti-
mation technique to estimate the true plant model of the aircraft in conjunction with
a direct MRAC designed to reduce the tracking error.
In the hybrid adaptive flight control approach, the dynamic inversion controller
computes the actuator command from a predictor or estimated plant model of the
aircraft dynamics rather than the ideal plant model. The estimated plant model is
expressed as
  
x̂˙ = C + ΔĈ x + D + Δ D̂ u + E + Δ Ê z (10.170)

Setting the desired acceleration ẋd (t) equal to x̂˙ (t), the actuator command is
computed from the estimated plant model by the dynamic inversion controller as
10.8 Hybrid Adaptive Flight Control 415

 −1   
u c = D + Δ D̂ ẋd − C + ΔĈ x − E + Δ Ê z (10.171)

The actuator command now depends on the estimates of the plant matrices ΔC,
ΔD, and ΔD which are computed from the indirect adaptive control using a least-
squares parameter estimation method. The plant modeling error ε (t) is computed by
subtracting the plant model from the estimated plant model as
  
ε = ẋd −ẋ = ΔĈ − ΔC x+ Δ D̂ − ΔD u c + Δ Ê − ΔE z = Θ̃  Φ (x, u c , z)
(10.172)
The recursive least-squares (RLS) indirect adaptive laws for computing Θ (t) are
given by
Θ̇ = −RΦ (x, u c , z) ε (10.173)

Ṙ = −RΦ (x, u c , z) Φ  (x, u c , z) R (10.174)

where R (t) = R  (t) > 0 is the covariance matrix that acts as a time-varying
adaptation rate matrix.
Alternatively, the least-squares gradient indirect adaptive law can be used by
setting R (t) to be a constant.
The matrices ΔĈ (t), Δ D̂ (t), and Δ Ê (t) are computed from Θ (t) at each time
step and then are used in the dynamic inversion controller to compute the actuator
command.
Using Eq. (10.157), the closed-loop plant is expressed as
) t
ẋ = ẋm + K p (xm − x) + K i (xm − x) dτ − u ad − ε (10.175)
0

where u ad (t) is a direct adaptive control signal computed as

u ad = ΔΘ  Φ (x, u c , z) (10.176)

where ΔΘ (t) is the residual estimate of the unknown plant matrices.

The tracking error equation is expressed as

ė = Am e + B (u ad + ε) (10.177)

The direct MRAC law for computing ΔΘ (t) can be selected to be any robust
modification scheme such as the e modification given by
 $ ' 
ΔΘ̇ = −Γ Φ (x, u c , z) e P B + μ $e P B ' ΔΘ (10.178)

The hybrid adaptive flight control thus uses the RLS indirect adaptive control to
estimate the plant uncertainty. The goal of the RLS indirect adaptive control is to
416 10 Aerospace Applications

drive the plant modeling error to zero. This information is then used in the dynamic
inversion controller to compute the actuator command. The control signal from the
actuator command produces a closed-loop aircraft response which is then compared
with the reference model to form the tracking error. However, because the plant mod-
eling error is reduced by the RLS indirect adaptive control, the resulting tracking error
will be smaller than if the RLS indirect adaptive control is not present. The MRAC
law further drives the residual tracking error to zero. Because the residual tracking
error is small, the adaptation rate for the direct MRAC can be set to a small value to
improve robustness. The RLS parameter estimation method is inherently robust, so,
together with the direct MRAC, the hybrid adaptive flight control architecture can
provide a very effective and robust adaptation mechanism.

Example 10.10 For the same wing-damage aircraft simulation in Example 10.9,
the hybrid adaptive flight control is implemented using the RLS indirect adaptive
control in conjunction with the direct adaptive control using the e modification. The
covariance matrix is initialized with R (0) = 100I . The closed-loop response of the
damaged aircraft due to the hybrid adaptive flight control is shown in Fig. 10.65.
The closed-loop response is exceedingly well-behaved. Both the roll and yaw rates
are much smaller than those with the e modification alone. The pitch rate and pitch
attitude both track the reference model extremely well. The bank angle and sideslip
angle are reduced close to zero as shown in Fig. 10.66. The saturation in the aileron is
still present, but the rudder deflection is much smaller than that with the e modification
alone as shown in Fig. 10.67. Thus, the performance of the hybrid adaptive flight
control is much better than the e modification alone.

10
p, deg/sec

0
p p
m
−10
0 5 10 15 20 25 30 35 40
t, sec

5
q, deg/sec

0
q qm
−5
0 5 10 15 20 25 30 35 40
t, sec

0.2
r, deg/sec

0
r r
m
−0.2
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.65 Closed-loop responses of roll, pitch, and yaw rates

10.9 Adaptive Flight Control for F-18 Aircraft with Optimal Control Modification 417

10 2

8 0

θ, deg

φ, deg
6 −2

4 −4
θ θ
m
2 −6
0 20 40 0 20 40
t, sec t, sec

10 0.5

8
0
α, deg

β, deg
6
−0.5
4

2 −1
0 20 40 0 20 40
t, sec t, sec

Fig. 10.66 Closed-loop responses of pitch angle, bank angle, angle of attack, and sideslip angle

40
δ , deg

20
a

0
0 5 10 15 20 25 30 35 40
t, sec
5
δe, deg

−5
0 5 10 15 20 25 30 35 40
t, sec
−1
δ , deg

−1.5
r

−2
0 5 10 15 20 25 30 35 40
t, sec

Fig. 10.67 Closed-loop responses of aileron, elevator, and rudder deflections

10.9 Adaptive Flight Control for F-18 Aircraft with

Optimal Control Modification

Adaptive flight control can be used to provide consistent handling qualities and
restore stability of aircraft under off-nominal flight conditions such as those due to
failures or damage. Suppose an aircraft plant is described by

ẋ = A11 x + A12 z + B1 u + f 1 (x, z) (10.179)

ż = A21 x + A22 z + B2 u + f 2 (x, z) (10.180)

418 10 Aerospace Applications

where Ai j and Bi , i = 1, 2, j = 1, 2 are nominal aircraft matrices which are

 
assumed to be known; x (t) = p (t) q (t) r (t) is an inner-loop state vector of
 
roll, pitch, and yaw rates; z (t) = Δφ (t) Δα (t) Δβ (t) ΔV (t) Δh (t) Δθ (t)
is an outer-loop state vector of aircraft attitude angles, airspeed, and altitude;
 
u (t) = Δδa (t) Δδe (t) Δδr (t) is a control vector of aileron, elevator, and rudder
deflections; and f i (x, z), i = 1, 2 is an unstructured uncertainty due to off-nominal
events which can be approximated as

f i (x, z) = Θi∗ Φ (x, z) + δ (x, z) (10.181)

where Θi∗ is an unknown, constant ideal weight matrix, and Φ (x, z) is the input
regressor function vector chosen to be
 
Φ (x, z) = x  px  q x  r x  z  u  (x, z) (10.182)

The inner-loop rate feedback control is designed to improve aircraft rate response
characteristics such as the short-period mode and Dutch roll mode. The reference
model is a second-order model that specifies desired handling qualities with good
damping and natural frequency characteristics as
 2 
s + 2ζ p ω p s + ω2p φm = g p δlat (10.183)
 2 
s + 2ζq ωq s + ωq2 θm = gq δlon (10.184)
 2 
s + 2ζr ωr s + ωr2 βm = gr δr ud (10.185)

where φm (t), θm (t), and βm (t) are the reference bank, pitch, and sideslip angles;
ω p , ωq , and ωr are the natural frequencies for desired handling qualities in the roll,
pitch, and yaw axes; ζ p , ζq , and ζr are the desired damping ratios; δlat (t), δlon (t), and
δr ud (t) are the lateral stick input, longitudinal stick input, and rudder pedal input;
and g p , gq , and gr are the input gains.
Let pm (t) = φ̇m (t), qm (t) = θ̇m (t), and rm (t) = −β̇m (t), then the reference
model can be represented as
) t
ẋm = −K p xm − K i xm (τ ) dτ + Gr (10.186)
0

   
where xm (t) = pm (t) qm (t) rm (t) , K p = diag 2ζ p ω p , 2ζq ωq , 2ζr ωr , K i = diag
 2 2 2    
ω p , ωq , ωr = Ω 2 , G = diag g p , gq , gr , and r (t) = δlat (t) δlon (t) δr ud (t) .
For the roll axis, the reference model could also be a first-order model
 
s + ω p pm = g p δlat (10.187)
10.9 Adaptive Flight Control for F-18 Aircraft with Optimal Control Modification 419

Assuming the pair (A11 , B1 ) is controllable and the outer-loop state vector z (t)
is stabilizable, then the nominal PI feedback controller, defined by ū (t), is given by
) t
ū = K p (xm − x) + K i [xm (τ ) − x (τ )] dτ (10.188)
0

and the adaptive controller, defined by u ad (t), is given by

u ad = Θ1 Φ (x, z) (10.189)

Assuming B1 is invertible, then the dynamic inversion controller is computed as

u = B1−1 (ẋm − A11 x − A12 z + ū − u ad ) (10.190)

In a more general case when the control vector has more inputs than the number
of states to be controlled, then an optimal control allocation strategy using a pseudo-
inverse method is used to compute the dynamic inversion controller as
 −1
u = B1 B1 B1 (ẋm − A11 x − A12 z + ū − u ad ) (10.191)
*t   
Let e (t) = 0 xm (t) − x  (t) dτ xm (t) − x  (t) be the tracking error, then
the tracking error equation is given by
 
ė = Am e + B Θ1 Φ (x, z) − f 1 (x, z) (10.192)

where  
0 I
Am = (10.193)
−K i −K p
 
0
B= (10.194)
I

Let Q = 2cI , where c > 0 is a weighting constant and I is an identity matrix,

then it can be shown that
 −1 
K i K p + K −1
p (K i + I ) K i−1 
P=c >0 (10.195)
K i−1 K −1
p I + K i−1
 −1 
PB = c K i  (10.196)
K p I + K i−1
−1

−2
B  P A−1
m B = −cK i <0 (10.197)

Then, the optimal control modification adaptive law for a nominal PI feedback
controller is specified by
420 10 Aerospace Applications
 
Θ̇1 = −Γ Φ (x, z) e P B + cνΦ (x, z) Φ  (x, z) Θ1 K i−2 (10.198)

which can also be expressed as

 
Θ̇1 = −Γ Φ (x, z) e P B + cνΦ (x, z) Φ  (x, z) Θ1 Ω −4 (10.199)

Suppose the tracking error equation is of a proportional-derivative (PD) type

ë = −K d ė − K p e + Θ̃2 Φ (x, z) (10.200)

where e (t) = xm (t) − x (t)

Then, the optimal control modification adaptive law for a nominal PD feedback
controller is specified by
 
Θ̇2 = −Γ Φ (x, z) e P B + cνΦ (x, z) Φ  (x, z) Θ2 K −2
p (10.201)

where  
K −1 
PB = c p
(10.202)
K d−1 I + K −1
p

Furthermore, if the tracking error equation is of a proportional type

ė = −K p e + Θ̃3 Φ (x, z) (10.203)

then the optimal control modification adaptive law for a nominal proportional con-
troller is specified by
 
Θ̇3 = −cΓ Φ (x, z) e K −1  −2
p + νΦ (x, z) Φ (x, z) Θ3 K p (10.204)

An adaptive flight control design is implemented on an F-18 aircraft model of

NASA F/A-18 research aircraft (tail number 853) as shown in Fig. 10.2, with both
the standard baseline dynamic inversion controller and the adaptive controller with
the optimal control modification adaptive law [11]. The flight condition is a test point
of Mach 0.5 and an altitude of 15,000 ft. All of the pilot inputs used in the simulations
are from pre-selected piloted stick inputs for comparison purposes. A time delay of
0.01 s is injected at the actuators to account for the actuator lag.
The first simulation scenario is called an A matrix failure which is emulated by a
center of gravity (CG) shift so as to reduce the amplitude of the pitch stability deriva-
tive Cm α which is normally negative for a stable aircraft in pitch. Figure 10.68 shows
the responses of the aircraft. The failure is inserted at 13 s. Upon the failure insertion,
the closed-loop plant with the nominal controller can no longer track the reference
model. Excessive overshoots in the pitch rate and angle of attack are observed. With
the adaptive controller, the pitch rate tracks the reference pitch rate command quite
well and the overshoots are significantly reduced.
10.9 Adaptive Flight Control for F-18 Aircraft with Optimal Control Modification 421

Figure 10.69 shows the tracking errors in the roll, pitch, and yaw axes along with
the adaptation weights. With the adaptive controller, the roll rate tracking error is
essentially reduced to zero. Both the pitch rate tracking error and yaw rate tracking
error are reduced to no more than about 2◦ /s with the adaptive controller. The weights
exhibit some large initial transients but then converge to their steady-state values
quickly after 28 s.
Figure 10.70 shows the control surface deflections with the nominal controller and
the adaptive controller. The actuator models are high-fidelity fourth-order models
with time delays. The control surface deflections with the adaptive controller are
well-behaved and smaller in amplitudes than the control surface deflections with the
nominal controller.

Fig. 10.68 Longitudinal states due to an A matrix failure (Cm α shift at 13 s) with and without
optimal control modification

Fig. 10.69 Tracking errors due to an A matrix failure (Cm α shift at 13 s) with and without optimal
control modification
422 10 Aerospace Applications

Fig. 10.70 Control surfaces due to an A matrix failure (Cm α shift at 13 s) with and without optimal
control modification

The second simulation scenario is called a B matrix failure which is emulated by a

jammed left stabilator at +2.5o from trim. Figure 10.71 and 10.72 show the responses
of the aircraft in the pitch, roll, and yaw axes. The failure is inserted at 13 s. Upon the
failure insertion, the closed-loop plant with the nominal controller is unable to follow
the reference model. The pitch rate can longer track the reference pitch rate command.
The roll and yaw rates are quite large. With the adaptive controller, the pitch, roll,
and yaw rates are significantly reduced. The bank angle is also significantly reduced
from 12◦ with the nominal controller to about 5◦ . The sideslip angle is reduced from
8◦ to 4◦ .

Fig. 10.71 Longitudinal states due to a B matrix failure (Stabilator Jammed at 2.5o at 13 s) with
and without optimal control modification
10.9 Adaptive Flight Control for F-18 Aircraft with Optimal Control Modification 423

Fig. 10.72 Lateral-directional states due to a B matrix failure (Stabilator Jammed at 2.5o at 13 s
with and without optimal control modification

Figure 10.73 shows the tracking errors in the roll, pitch, and yaw axes along with
the adaptation weights. The roll rate tracking error and pitch rate tracking error are
generally smaller with the adaptive controller than with the nominal controller. The
yaw rate tracking error is about the same for both the nominal controller and the
adaptive controller. The weights converge after about 20 s.

Fig. 10.73 Tracking errors due to a B matrix failure (Stabilator Jammed at 2.5o at 13 s) with and
without optimal control modification

The optimal control modification adaptive control is shown to be able to achieve

fast adaptation with a good damping characteristic. The simulation scenario with
the A matrix failure is conducted with the adaptation rate changed from Γ = 0.5
to Γ = 50 while keeping the modification parameter ν = 1 constant. The failure
424 10 Aerospace Applications

is inserted at 2 s. Figure 10.74 shows the pitch rate tracking error for Γ = 0.5 and
Γ = 50. Increasing the adaptation rate Γ from 0.5 to 50 results in a significantly
smaller tracking error. This is as expected. The response is nicely damped with
Γ = 50 without any high-frequency oscillations. The weights converge faster with
increasing the adaptation rate Γ . Thus, the simulation shows that a larger adaptation
rate can produce better tracking with the optimal control modification without causing
any high-frequency oscillations.

Fig. 10.74 Pitch rate error and weights due to an A-matrix failure (Cm α Shift at 2 s) with optimal
control modification (Γ = 0.5 and Γ = 50 with fixed ν = 1)

The modification parameter ν provides the damping to the adaptive controller. The
same simulation scenario with the A matrix failure is conducted with the modification
parameter changed from ν = 0.25 to ν = 1 while keeping the adaptation rate Γ = 5
constant. Figure 10.75 shows the pitch rate tracking errors for ν = 0.25 and ν = 1.
Increasing the modification parameter ν from 0.25 to 1 adds more damping to the
adaptive controller. As a result, the pitch rate tracking error is well-damped with
ν = 1 without any oscillations. All the weights converge faster with increasing the
modification parameter ν. Thus, this simulation shows that a larger modification
parameter can help improved the performance of the adaptive controller.
This F-18 aircraft simulation study validates the effectiveness of adaptive control
in general and the optimal control modification in particular in improving the flight
control performance of a degraded aircraft. The study coupled with a piloted high-
fidelity F-18 flight simulator experiment established the basis for a flight test program
on NASA F/A-18 research aircraft (tail number 853) in 2010 and 2011.
10.10 Summary 425

Fig. 10.75 Pitch rate error and weights due to an A-matrix failure (Cm α shift at 2 s) with optimal
control modification (fixed Γ = 5 with ν = 0.25 and ν = 1)

10.10 Summary

Adaptive control enjoys many applications in aerospace. It has a long history dated
back to the 1950s with the first adaptive control deployed on NASA X-15 hyper-
sonic vehicles. Applications of adaptive control in other domains are numerous.
Adaptive flight control applications for aircraft longitudinal dynamics using the
bi-objective optimal control modification and least-squares adaptive control with
Chebyshev polynomial approximation and neural networks are illustrated. Adap-
tive flight control for flexible aircraft with aeroelastic interactions using the optimal
control modification and adaptive loop recovery modification is presented. An adap-
tive LQG control method is developed for a flutter suppression of a flexible wing.
A general adaptive flight control architecture based on a dynamic inversion with
adaptive augmentation control is outlined. Robust modification schemes based on
σ modification, e modification, and optimal control modification are used in the
adaptive augmentation control to improve robustness. Hybrid adaptive flight control
is a robust adaptive flight control architecture that blends both direct MRAC and
indirect adaptive control using least-squares parameter identification together. This
technique can significantly improve the performance of a flight control system under
off-nominal operation of an impaired or damaged aircraft. Finally, an adaptive flight
control design for a F/A-18A NASA aircraft based on the optimal control modifica-
tion is implemented in simulations. This implementation was later used in the actual
flight testing in 2010 and 2011.
426 10 Aerospace Applications

10.11 Exercises

1. Consider the equation of motion of an inverted pendulum

1 1
m L 2 θ̈ − mgL sin θ + cθ̇ = u (t − td )
3 2
a. Expand sin θ using the Taylor series expansion about θ (t) = 0 for the first
two terms. Then, express the equation of motion in the form of
 
ẋ = Ax + B u (t − td ) + Θ ∗ Φ (x)
 
where x1 (t) = θ (t), x2 (t) = θ̇ (t), x (t) = x1 (t) x2 (t) , Φ (x) is com-
prised of the function in the nonlinear term of the Taylor series expansion of
sin θ and the function in the damping term, and Θ ∗ is a vector of parameters
associated with Φ (x) which are assumed to be unknown.
b. Let m = 0.1775 slug, L = 2 ft, c = 0.2 slug-ft2 /s, td = 0.05 sec, and θ (0) =
θ̇ (0) = 0. Using the equation of motion in part (a), design an adaptive
controller using the optimal control modification to enable the closed-loop
plant to track a reference model specified by

θ̈m + 2ζm ωm θ̇m + ωm2 θm = ωm2 r

π
where ζm = 0.5, ωm = 2, and r = 12 . Calculate K x and kr .
c. Implementthe adaptive
 controller in Simulink using the nonlinear plant with
Θ  (0) = θ1∗ 0 and a time step Δt = 0.001 s for the standard MRAC with
Γ = 100 and the optimal control modification with Γ = 100 and ν = 0.5.
Plot the time histories of x (t) and xm (t) on the same plot, u (t), and Θ (t) for
t ∈ [0, 10] s. Compare the closed-loop response with the optimal control
modification to that in Example 10.1. Does the linear nominal controller
design in this problem appear to work as well as the nonlinear nominal
controller design in Example 10.1?

2. Implement a longitudinal dynamic model of an aircraft.

 
ẋ = Ax + B u (t − td ) + Θ ∗ x
   
where x (t) = α (t) θ (t) q (t) , u (t) = δe (t), and Θ ∗ = θα∗ 0 θq∗ , with
the following information: V̄ = 795.6251 ft/s, γ̄ = 0, Z α = −642.7855 ft/s2 ,
Z δe = −55.3518 ft/s2 , Mα = −5.4898 s−2 , Mδe = −4.1983 s−2 , Mq = −0.6649
s−1 , Mα̇ = −0.2084 s−1 , θα∗ = 0.5, θq∗ = −0.5, and td = 0.1 s.
a. Design an adaptive pitch attitude controller using the optimal control mod-
ification to enable the closed-loop plant to follow a second-order reference
model of the pitch attitude specified by ζm = √12 and ωm = 2 rad/s. Express
10.11 Exercises 427

the adaptive controller with the feedback gain values and the reference
model.
b. Implement the adaptive controller in Simulink using the following infor-
mation: x (0) = 0 and Θ (0) = 0 with a time step Δt = 0.01 s for: 1) the
nominal controller, 2) the standard MRAC with Γ = 500, and 3) the optimal
control modification with Γ = 500 and ν = 0.5. The reference command
signal r (t) is a pitch attitude doublet specified in the following plot.
6

2
r, deg

−2

−4

−6
0 5 10 15 20 25 30
t, sec

For each controller, plot the time histories of each of the elements of x (t)
and xm (t) on the same plot, and u (t) for t ∈ [0, 30] s. Plot in units of deg
for α (t), θ (t), and δe (t), and deg/s for q (t). Comment on the simulation
results.

References

1. Bosworth, J., & Williams-Hayes, P. S. (2007). Flight test results from the NF-15B IFCS project
with adaptation to a simulated stabilator failure. In AIAA Infotech@Aerospace Conference,
AIAA-2007-2818, May 2007.
2. Williams-Hayes, P. S. Flight test implementation of a second generation intelligent flight control
system. In: NASA TM-2005-213669.
3. Narendra, K. S., & Annaswamy, A. M. (1987). A new adaptive law for robust adaptation
without persistent excitation. IEEE Transactions on Automatic Control, AC-32(2), 134–145.
4. Calise, A. J., & Rysdyk, R. T. (1998). Nonlinear adaptive flight control using neural networks.
IEEE Control System Magazine, 18(6):14–25.
5. Hanson, C., Johnson, M., Schaefer, J., Nguyen, N., & Burken, J. (2011). Handling qualities
evaluations of low complexity model reference adaptive controllers for reduced pitch and roll
damping scenarios. In AIAA Guidance, Navigation, and Control Conference, AIAA-2011-6607,
August 2011.
6. Hanson, C., Schaefer, J., Johnson, M., & Nguyen, N. Design of low complexity model reference
adaptive controllers. In: NASA-TM-215972.
7. Nguyen, N., Hanson, C., Burken, J., & Schaefer, J. (2016). Normalized optimal control modifi-
cation and flight experiments on NASA F/A-18 aircraft. In: AIAA Journal of Guidance, Control,
and Dynamics.
8. Schaefer, J., Hanson, C., Johnson, M., & Nguyen, N. (2011). Handling qualities of model
reference adaptive controllers with varying complexity for pitch-roll coupled failures. In AIAA
Guidance, Navigation, and Control Conference AIAA-2011-6453, August 2011.
428 10 Aerospace Applications

9. Campbell, S., Kaneshige, J., Nguyen, N., & Krishnakumar, K. (2010) An adaptive control
simulation study using pilot handling qualities evaluations. In AIAA Guidance, Navigation,
and Control Conference, AIAA-2010-8013, August 2010.
10. Campbell, S., Kaneshige, J., Nguyen, N., & Krishnakumar, K. (2010). Implementation and
evaluation of multiple adaptive control technologies for a generic transport aircraft simulation.
In AIAA Infotech@Aerospace Conference, AIAA-2010-3322, April 2010.
11. Burken, J., Nguyen, N., & Griffin, B. (2010). Adaptive flight control design with optimal control
modification for f- 18 aircraft model. In AIAA Infotech@Aerospace Conference, AIAA-2010-
3364, April 2010.
12. Nguyen, N., Burken, J., & Hanson, C. (2011). Optimal control modification adaptive law with
covariance adaptive gain adjustment and normalization. In AIAA Guidance, Navigation, and
Control Conference, AIAA-2011-6606, August 2011.
13. Nguyen, N., & Summers, E. (2011). On time delay margin estimation for adaptive control and
robust modification adaptive laws. In AIAA Guidance, Navigation, and Control Conference,
AIAA-2011-6438, August 2011.
14. Nelson, R. C. (1989). Flight stability and automatic control. New York: McGraw-Hill
15. Nguyen, N., & Balakrishnan, S. N. (2014). Bi-objective optimal control modification adaptive
control for systems with input uncertainty. IEEE/CAA Journal of Automatica Sinica, 1(4),
423–434.
16. Nguyen, N. (2013). Least-squares model reference adaptive control with chebyshev orthogonal
polynomial approximation. AIAA Journal of Aerospace Information Systems, 10(6), 268–286.
17. Rohrs, C. E., Valavani, L., Athans, M., & Stein, G. (1985). Robustness of continuous-time
adaptive control algorithms in the presence of unmodeled dynamics. IEEE Transactions on
Automatic Control, AC-30(9), 881–889.
18. Ioannou, P., & Kokotovic, P. (1984). Instability analysis and improvement of robustness of
adaptive control. Automatica, 20(5), 583–594.
19. Nguyen, N. (2012). Optimal control modification for robust adaptive control with large adaptive
gain. Systems & Control Letters, 61(2012), 485–494.
20. Calise, A. J., & Yucelen, T. (2012). Adaptive loop transfer recovery. AIAA Journal of Guidance,
Control, and Dynamics, 35(3), 807–815.
21. Nguyen, N., Tuzcu, I., Yucelen, T., & Calise, A. (2011). Longitudinal dynamics and adaptive
control application for an aeroelastic generic transport model. In AIAA Atmospheric Flight
Mechanics Conference, AIAA-2011-6319, August 2011.
22. Kokotovic, P., Khalil, H., & O’Reilly, J. (1987). Singular perturbation methods in control:
analysis and design. Philadelphia: Society for Industrial and Applied Mathematics.
23. Nguyen, N. (2010). Elastically Shaped Future Air Vehicle Concept,” NASA Innovation Fund
Award 2010 Report, October 2010, Submitted to NASA Innovative Partnerships Program.
http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/20110023698.pdf.
24. Nguyen, N., Trinh, K., Reynolds, K., Kless, J., Aftosmis, M., Urnes, J., & Ippolito, C., Elas-
tically shaped wing optimization and aircraft concept for improved cruise efficiency. In AIAA
Aerospace Sciences Meeting, AIAA-2013-0141, January 2013.
25. Urnes, J., Nguyen, N., Ippolito, C., Totah, J., Trinh, K., & Ting, E., A mission adaptive variable
camber flap control system to optimize high lift and cruise lift to drag ratios of future N+3
transport aircraft. In AIAA Aerospace Sciences Meeting, AIAA-2013-0214, January 2013.
26. Jordan, T., Langford, W., Belcastro, C., Foster, J., Shah, G., Howland, G., et al. (2004). Devel-
opment of a dynamically scaled generic transport model testbed for flight research experiments.
In AUVSI Unmanned Unlimited, Arlington, VA.
27. Nguyen, N., Ting, E., Nguyen, D., & Trinh, K., Flutter analysis of mission-adaptive wing with
variable camber continuous trailing edge flap. In 55th AIAA/ASME/ASCE/AHS/ASC Structures,
Structural Dynamics, and Materials Conference, AIAA-2014-0839, January 2014.
28. Nguyen, N., Swei, S., & Ting, E. (2015). Adaptive linear quadratic gaussian optimal control
modification for flutter suppression of adaptive wing. In AIAA Infotech@Aerospace Confer-
ence, AIAA 2015-0118, January 2015.
References 429

29. Ardema, M. (1981). Computational singular perturbation method for dynamical systems. AIAA
Journal of Guidance, Control, and Dynamics, 14, 661–663.
30. Dykman, J., Truong, H., & Urnes, J. (2015). Active control for elastic wing structure dynamic
modes. In 56th AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Con-
ference, AIAA-2015-1842, January 2015.
31. Slotine, J.-J., & Li, W. (1991) Applied Nonlinear Control. Englewood Cliffs: Prentice-Hall,
Inc.
32. Hsu, L., Teixeira, M. C. M., Costa, R. R., & Assunc̃ao, E. (2011). Necessary and sufficient con-
dition for generalized passivity, passification and application to multivariable adaptive systems.
In 18th International Federation of Automatic Control World Congress, August 2011.
33. Misra, P. (1993). Numerical algorithms for squaring-up non- square systems part II: general
case. In American Control Conference, June 1993.
34. Qu, Z., Wiese, D., Annaswamy, A. M., & Lavretsky, E. (2014). Squaring-up method in the
presence of transmission zeros. In 19th International Federation of Automatic Control World
Congress, August 2014.
35. Lavretsky, E., & Wise, K. (2012). Robust and adaptive control. Berlin: Springer.
36. Nguyen, N., & Urnes, J. (2012). Aeroelastic modeling of elastically shaped aircraft concept via
wing shaping control for drag reduction. In: AIAA Atmospheric Flight Mechanics Conference,
AIAA-2012-4642, August 2012.
37. Nguyen, N., Krishnakumar, K., Kaneshige, J., & Nespeca, P. (2008). Flight dynamics model-
ing and hybrid adaptive control of damaged asymmetric aircraft. AIAA Journal of Guidance,
Control, and Dynamics, 31(3), 751–764.
Suggested Exam Questions

1. For the following systems, determine the equilibrium points. Use the Lyapunov’s
direct method to determine the type of Lyapunov stability for each of the equi-
librium points. Determine all the invariant sets and the values of the Lyapunov
function on the sets. If an equilibrium point is stable, conclude if it is asymptot-
ically stable, and if so, show whether or not it is also exponentially stable.
a.      
ẋ1 (x2 − x1 ) x12 + x22 − 1 
=
ẋ2 − (x1 + x2 ) x12 + x22 − 1

b.    
ẋ1 x22 − x2
=
ẋ2 −x1 − x2 + 1

c.    
ẋ1 x2
=
ẋ2 −x1 − (1 + sin x1 ) x2

2. Linearize the systems in problem 2 and determine the types of equilibrium points.
Plot phase portraits.
3. Given the following system
    
ẋ1 −2 + sin2 x1 1 − sin x1 cos x2 x1
=
ẋ2 −1 + sin x1 cos x2 −2 − cos2 x2 x2

Determine the stability of this system about the origin using the Lyapunov can-
didate function
1
V (x) = x  x
2
If asymptotically stable, determine if the origin is exponentially stable and find
 
the rate of convergence of x, where x (t) = x1 (t) x2 (t) .

© Springer International Publishing AG 2018 431

N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0
432 Suggested Exam Questions

4. Given a linear system

ẋ = Ax + Bh (t)
 
where x (t) = x1 (t) x2 (t) ∈ R2 and
   
−1 2 1  
A= , B= , h (t) = 1 + e−t (sin t + cos t)
−4 −2 1

a. Compute P that solves the Lyapunov equation

P A + A P = −I

and also compute the eigenvalues of P to verify that P is positive definite.

b. Use the following Lyapunov candidate function

V (x) = x  P x

to compute V̇ (x). Establish an upper bound on V̇ (x) in terms of x, and

then determine a lower bound on x that satisfies V̇ (x) ≤ 0 using the L∞
norm and the Cauchy–Schwartz inequality

C D ≤ C D

c. Find an analytical solution of an upper bound of the Lyapunov function

V (t) as an explicit function of t from V̇ (x) in part (b), given V (0) = 2, by
utilizing the following relationship for a positive definite function

λmin (P) x2 ≤ V (x) = x  P x ≤ λmax (P) x2

and the following variable transformation


W (t) = V (t)

d. Find the ultimate bound of x by finding the limit of V (t) as t → ∞. If

an ultimate bound exists, then the solution of x (t) is uniformly ultimately
bounded.

5. Consider a first-order nonlinear SISO system with a matched uncertainty

 
ẋ = ax + b u + θ ∗ x 2

where a is unknown but b is known, and θ ∗ is unknown.

A reference model is specified by

ẋm = am xm + bm r
Suggested Exam Questions 433

where am < 0 and bm are known, and r (t) is a bounded command signal.
a. Design a direct adaptive controller that enables the plant output x (t) to track
the reference model signal xm (t). Show by Lyapunov stability analysis that
the tracking error is asymptotically stable; i.e., e (t) → 0 as t → ∞.
b. Implement the adaptive controller in Simulink, given b = 2, am = −1,
bm = 1, and r (t) = sin t. For adaptation rates, use γx = 1 and γ = 1.
For simulation purposes, assume a = 1 and θ ∗ = 0.2 for the unknown
parameters. Plot e (t), x (t), xm (t), u (t), and θ (t) for t ∈ [0, 50].
c. Repeat part (b) for γx = 10 and γ = 10. Plot the same sets of data as in
part (b). Comment on the simulation results for parts (b) and (c) regarding
the tracking of the reference model, the quality of the signal in terms of the
relative frequency content, and the convergence of k x (t) and θ (t) as the
adaptation rates increase.
d. Repeat part (b) for r (t) = 1 (t) where 1 (t) is the unit-step function. Plot
the same sets of data as in part (b). Comment on the convergence of k x (t)
and θ (t) to the ideal values k x∗ and θ ∗ .
6. Given a first-order nonlinear system

ẋ = ax + Bu + cx 2
 
where x (t) ∈ R, u (t) ∈ R2 , a is an unknown constant, B = 1 2 is known,
and c is an unknown constant.
The reference model is specified as

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t.

Express the system in the form of a matched uncertainty
 
ẋ = ax + B u + Θ ∗ Φ (x)

Determine K x∗ , K r∗ , and Θ ∗ . Write down the adaptive laws for K x (t) and Θ (t).
Implement the controller in Simulink. Use γx = γΘ = 1. Assume all initial
conditions to be zero and a = 1, c = 0.2 for simulation purpose. Plot e (t), x (t)
versus xm (t), K x (t), and Θ (t) for t ∈ [0, 40].
7. The symmetric sigmoidal function

1 − e−x
σ (x) =
1 + e−x

can be used to model a control actuator saturation, which frequently exists in

real systems.
434 Suggested Exam Questions

0.8

0.6

0.4

0.2
B(u)u

−0.2

−0.4

−0.6

−0.8

−1
−10 −5 0 5 10
u

Saturation occurs when a control actuator ceases to be effective. When there are
more inputs than commands, a control allocation strategy should be developed
to allocate redundant control effectors in an optimal manner so as to produce an
output that tracks a command. Define y (u) as the output of a control allocator
as  
y = V σ W u

where y (u) ∈ Rn , V ∈ Rm × Rn , W ∈ R p × Rm , and u ∈ R p , p ≥ n.

V can be used to specify a saturation limit, while V  W  plays the role of
a nonlinear B (u) matrix. Develop an optimal control allocation strategy by
computing the gradient of the following cost function with respect to u, i.e., ∇ Ju

1 
J (u) =  
2
where  = y − r and r ∈ Rn are a command vector for which an optimal control
vector u is to be found to minimize the cost function.
Given r = 1 and    
0.75 1.2 0.8
V = , W =
0.5 0.5 1.5

Write a MATLAB code to compute u using the steepest descent method with
an adaptation rate ε = 0.1 and a number of iteration of n = 1000. Indicate the
final value of u and plot u.
8. Adaptive control can be used for disturbance rejection. Disturbances are usually
time signals that may have multiple frequency contents. Unlike unstructured
uncertainty in the form of an unknown function f (x), an unknown function of
Suggested Exam Questions 435

time f (t) should be approximated by a bounded function. This prevents adaptive

signals from blowing up in time. Both the sigmoidal and radial basis functions
are bounded functions, but a polynomial function is not. Consider a first-order
system with an unknown disturbance

ẋ = ax + b [u + f (t)]

where a and f (t) are unknown, but b = 2. For simulation purpose, a = 1 and
f (t) = 0.1 sin 2.4t − 0.3 cos 5.1t + 0.2 sin 0.7t.
The reference model is given by

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = sin t.

Implement in Simulink a direct adaptive control using the least-squares gradient
method to approximate f (t) by a sigmoidal neural network with Θ (t) ∈ R5 ,
W (t) ∈ R2 × R4 using the activation function σ (x) = 1+e1 −x . Write down
the neural net adaptive laws for k x (t), Θ (t), and W (t). All initial neural net
weights are randomized between 0 and 1. The initial condition for k x (t) is zero.
Use Γx = 10I . Plot e (t),  (t), x (t) versus xm (t) with disturbance rejection,
x (t) versus xm (t) without disturbance rejection, k x (t), Θ (t), and W (t) for
t ∈ [0, 40].
9. Given the following plant

ẋ = −2x − z + u + w

ż = −3z + 4u

y=x

where x (t) is the plant output, z (t) is an internal state, and w (t) = 1 is a
constant disturbance.

a. If a linear controller u (t) = k x x (t) is used, where k x is constant, express

the transfer function from w (t) to x (t). Find all values of k x for which the
closed-loop plant is stable.
b. Find the equilibrium state x̄ as a function of k x from part (a). Suppose an
adaptive regulator controller is designed with the σ modification

u = k x (t) x
 
k̇ x = −γ x 2 + σ k x

Find the minimum value of the modification parameter σmin to within 0.01
by finding the roots of a polynomial in terms of k̄ x for which one or more
roots satisfy the values of k x in part (a). Calculate k̄ x and x̄.
436 Suggested Exam Questions

c. Implement the adaptive controller in Simulink with σ = σmin − 0.05 and

σ = 0.5 using the following information: x (0) = 0, z (0) = 0, k x (0) = 0,
and γ = 10 using a time step Δt = 0.001 s. Plot the time histories of
x (t) and θ (t) for t ∈ [0, 10] s for both values of σ . Comment on the two
responses. Calculate k¯x and x̄ for σ = 0.5 analytically and compare them
with the simulation results.

10. Given a first-order SISO system with a matched uncertainty

 
ẋ = ax + b u + θ ∗ x 2

subject to x (0) = x0 , where a = 1 and b = 1 are known, and θ ∗ = 2 is

unknown.
An adaptive controller is designed using the optimal control modification adap-
tive law to enable the plant to follow a reference model

ẋm = am xm + bm r

where am = −1, bm = 1, and r (t) = 1

The adaptive controller is given by

u = k x x + kr r − θ (t) x 2

a. Express the closed-loop system with the nominal (non-adaptive) controller

u = k x x in terms of the reference model parameters am and bm . Determine
whether or not the closed-loop system with the nominal controller is uncon-
ditionally (globally) stable by explicitly integrating the plant model to find
the solution of x (t). If the closed-loop plant is not globally stable, find the
stability condition imposed on x0 .
b. Express the optimal control modification adaptive law for θ (t). Use
Sect. 9.5.3 to estimate the limiting value of the modification parameter
νmax to within 0.001. If applicable, express ϕ (x , xm  , ν, θ ∗ ). Then,
νmax can be found by trial and error to be the largest value for which
ϕ (x , xm  , νmax , θ ∗ ) = 0 such that x > xm . Express the ulti-

mate bound of e and θ̃ as a function x, ν, and γ . Evaluate them
for γ = 500.
c. Implement the adaptive controller in Simulink with MRAC for which ν = 0
and the optimal control modification with ν = νmax determined from part
(b) using the following information: x (0) = 1, θ (0) = 0, and γ = 500
with a time step Δt = 0.001 s. Plot the time histories of x (t), u (t), and
θ (t) for t ∈ [0, 10] s for both MRAC and the optimal control modification.
Comment on the responses of the two adaptive controllers and compare the

maximum tracking error e and maximum parameter estimation error θ̃
due to the optimal control modification to those determined from part (b).
Suggested Exam Questions 437

11. Consider the equation of motion of an inverted pendulum constrained to move

horizontally by a control force u (t)

1
 1 1 1
m L 2 4 − 3 cos2 θ θ̈ − mgL sin θ + m L 2 θ̇ 2 sin 2θ + cθ̇ = L cos θu (t − td )
12 2 8 2

where m is the mass of the pendulum, L is the length, g is the gravity constant, c
is the damping coefficient which is assumed to be unknown, θ (t) is the angular
position, u (t) is the control input which represents the horizontal force at point
O, and td is a time delay which represents the motor actuator dynamics.
 
a. Let x1 (t) = θ (t), x2 (t) = θ̇ (t), and x (t) = x1 (t) x2 (t) . Derive the
expressions for the nonlinear dynamic inversion adaptive controller and the
σ modification adaptive law to estimate the unknown coefficient c in order
to enable the closed-loop plant to track a reference model specified by

θ̈m + 2ζm ωm θ̇m + ωm2 θm = ωm2 r

which can be expressed in general as

ẋm = Am xm + Bm r

b. Given m = 0.1775 slug, g = 32.174 ft/s, L = 2 ft, c = 0.2 slug-ft2 /s,

π
ζm = 0.75, ωm = 2, and r = 12 sin 2t. Implement the adaptive controller
in Simulink with the following information: x (0) = 0, ĉ (0) = 0, γ = 100,
and a time step Δt = 0.001 s for the following cases: (1) the standard
MRAC with td = 0, (2) the standard MRAC with td = 0.01 s, and (3) the
σ modification with σ = 0.1. For each case, plot the time histories of x (t)
and xm (t) on the same plot, u (t), and ĉ (t) for t ∈ [0, 10] s. Plot in the units
of deg for x1 (t), deg/s for x2 (t), lb for u (t), lb-ft-sec for ĉ (t).
438 Suggested Exam Questions

12. Given a longitudinal dynamic model of an aircraft with a matched uncertainty

   Zα
 
α̇ 1 α
= V̄
q̇ Mα + Mα̇V̄Z α Mq + Mα̇ q
   
Z δe
 ∗ ∗ θ
+ V̄
M Z δe (t − td ) + θα θq
Mδe + α̇V̄ δe q

with the following information: V̄ = 795.6251 ft/s, γ̄ = 0, Z α = −642.7855

ft/s2 , Z δe = −55.3518 ft/s2 , Mα = −5.4898 sec−2 , Mδe = −4.1983 s−2 , Mq =
−0.6649 sec−1 , Mα̇ = −0.2084 s−1 , θα∗ = −5.4, θq∗ = −0.3, and td = 0.01 sec.
a. Design a nominal proportional–integral control
 t
δe = k p α + ki (α − r ) dτ + kq q
0

by finding the general expressions and the numerical values for k p , ki , and
kq to enable the aircraft to track a reference model of the angle of attack

α̈m + 2ζm ωm α̇m + ωm2 αm = ωm2 r

where ζm = 0.75
 t and ωm = 1.5 rad/s.
b. Let z (t) = 0 (α (t) − r (t)) dτ , provide the general expression and the
numerical value for the reference model of the aircraft as

ẋm = Am xm + Bm r
 
where x (t) = z (t) α (t) q (t) .
 
c. Let Θ ∗ = 0 θα∗ θq∗ . Design an adaptive angle-of-attack controller using
the optimal control modification to enable the closed-loop plant to track
the reference model. Express the adaptive controller and the adaptive law.
Given Q = 100I , select the modification parameter to guarantee stability of
the closed-loop plant by using the linear asymptotic property of the optimal
control modification and the following formulas to compute the crossover
frequency and time delay margin for MIMO systems. Plot ν versus td for
ν ∈ [0, 5] and determine ν to within 0.01 for td = 0.01 s
For a general time delay system

ẋ = Ax + Bu (t − td )

with a linear controller

u = Kx x

the crossover frequency and time delay margin can be estimated as

Suggested Exam Questions 439

ω = μ (− j A) + B K x 

1 μ (A)
td = cos−1
ω μ (−B K x )

where μ̄ is a matrix measure quantity defined as

 
C + C∗
μ (C) = max λi
1≤i≤n 2

for a general complex-value matrix C with its conjugate transpose C ∗ .

d. Implement the adaptive controller in Simulink using the following informa-
tion: x (0) = 0, Θ (0) = 0, Γ = 1000I , and ν determined from part (c)
with a time step Δt = 0.001 s. The reference command signal r (t) is a pitch
attitude doublet specified in the following plot.
6

2
r, deg

−2

−4

−6
0 5 10 15 20 25 30
t, sec
Plot the time histories of each of the elements of x (t) and xm (t) on the same
plot, and u (t) for t ∈ [0, 30] s. Plot in the units of deg–sec for z (t), deg for
α (t) and δe (t), and deg/s for q (t).

13. Given a longitudinal dynamic model of a damaged aircraft

   
Zα Z δe
α̇ = + ΔAαα α + q + + ΔBα δe (t − td )
V̄ V̄
 
Mα̇ Z α  
q̇ = Mα + + ΔAqα α + Mq + Mα̇ + ΔAqq q
V̄
 
Mα̇ Z δe
+ Mδe + + ΔBq δe (t − td )
V̄
440 Suggested Exam Questions

a. Design an ACAH hybrid adaptive flight controller for the pitch axis to enable
the aircraft to follow a reference model

q̇m = −ωq (qm − r )

by providing the expressions for the hybrid adaptive controller, the least-
squares gradient parameter estimation of ΔAqα , ΔAqq , ΔBq , and the optimal
control modification adaptive law to handle the residual tracking error.
b. Implement the hybrid adaptive flight controller in Simulink using the same
aircraft parameters from Exam Problem 12 and the following additional
information: td = 0.02 s, ζq = 0.75, ωq = 2.5 rad/s, ΔAαα = 0.1616/s,
ΔAqα = 2.1286 /s2 , ΔAqq = 0.5240 /s, ΔBα = −0.0557 /s, ΔBq =
−2.5103 /s2 , α (0) = 0, q (0) = 0, Δ Âqα (0) = 0, Δ Âαα (0) = 0,
Δ B̂q (0) = 0, R = 1000I , Γ = 1000I , and ν = 0.1 with a time step
Δt = 0.001 s. The reference command signal r (t) is a pitch rate doublet
specified in the following plot.
3

1
r, deg/sec

−1

−2

−3
0 5 10 15 20 25 30
t, sec

c. Simulate three cases: (1) nominal controller, (2) only direct MRAC, and (3)
hybrid adaptive control with both direct MRAC and indirect least-squares
gradient adaptive control. For each case, plot the time histories of each of
the elements of α (t), θ (t) , q (t) and qm (t) on the same plot, and u (t) for
t ∈ [0, 30] s. In addition, plot Θ (t) for case 2; and Δ Âqα (t), Δ Âqq (t),
Δ B̂q (t) for case 3. Plot in the units of deg for α (t), θ (t), and δe (t), and
deg/s for q (t).
Index

A Pitch Attitude Control, 362–364

Actuator Dynamics Amplitude Saturation, 20, 21
Robustness Issues, 196, 200 Autonomous Systems, 18–21, 47–67
Slow Actuator, 307–313
Adaptive Flight Control
Dynamic Inversion Control, 355, 359, B
409 Barbalat’s Lemma, 74–79
e Modification, 350–354 Barbashin–Krasovskii Theorem, 61, 62
F-18 Aircraft, 358–365 Bi-Objective Optimal Control Modification
Hybrid, 350–354 Adaptive Law, 290–307
Inner Loop, 407, 408 Aircraft Adaptive Control, 310–312,
Optimal Control Modification, 350–354, 363–368
359–365 Singularly Perturbed Systems, 309, 312
Outer Loop, 408
Rate Command Attitude Hold, 349
Stability Augmentation System, 406 C
Adaptive Linear Quadratic Gaussian Cauchy Theorem, 38, 39
Adaptive Flutter Suppression, 337–344 Chebyshev Orthogonal Polynomial Approx-
Adaptive Law, 332, 394, 410 imation, 153, 154, 367, 371
Adaptive Loop Recovery Modification Compact Set, 37
Adaptive Law, 261–263, 265, 266 Controllability, 111
Combined with Optimal Control Modifi- Covariance Adjustment, 280–285
cation, 381, 383–385 Cross-Over Frequency, 194, 197, 200, 201,
Flexible Aircraft Adaptive Control, 384, 319
386–388
Adjoint Equation, 236, 239, 241, 242, 294,
295 D
Aircraft Adaptive Control Dead-Zone Method, 212–213
Adaptive Flight Control, 408–427 Dynamic Inversion Control, 411, 413, 421
Bi-Objective Optimal Control Modifica-
tion, 363–368
e Modification, 411–416 E
Least-Squares Adaptive Control, 369 e Modification
Longitudinal Dynamics, 360–364 Adaptive Flight Control, 410–413
Neural Network Adaptive Control, 370 Adaptive Law, 228–233
Optimal Control Modification, 411–416, Hybrid Adaptive Flight Control, 415
421–427 Energy Function, 52, 53
© Springer International Publishing AG 2018 441
N.T. Nguyen, Model-Reference Adaptive Control, Advanced Textbooks
in Control and Signal Processing, https://doi.org/10.1007/978-3-319-56393-0
442 Index

Equilibrium, 23 Gradient Method, 130

Error Signals Recursive, 134–137
Approximation, 126 Regression, 126–128
Plant Modeling, 138, 139, 369 Least-Squares Gradient Adaptive Control
Predictor, 287 Aircraft Adaptive Control, 369
State Estimation, 129, 130 Combined with Model-Reference Adap-
Tracking, 88, 89 tive Control, 133–141
Direct Adaptive Control, 168–173
Direct Adaptive Control with Neural
F Network, 173–175
F-18 Aircraft Adaptive Flight Control Indirect Adaptive Control, 137, 139
Flight Validation, 353 Predictor-Based, 369
Optimal Control Modification, 417, 424 Limit Cycle, 28
Fast Adaptation, 200, 201 Linearization, 22, 24
Finite Escape Time, 27 Lipschitz Condition
Flexible Aircraft Adaptive Control, 382–386 Global, 39
Adaptive Flutter Suppression, 387–408 Local, 40
Adaptive Linear Quadratic Gaussian, Luenberger Observer Design, 330, 394
391–408 Lyapunov Equation
Adaptive Loop Recovery Modification, Algebraic, 59, 68, 101, 242, 294
382–386 Differential, 67–69, 242
Coupled Aeroelastic Flight Dynamic Lyapunov Function
Model, 376–380 Candidate, 56
Flutter Speed, 380, 387–391 Decrescent, 71
Model Reduction, 391–394 Definition, 55, 56
Optimal Control Modification, 382–386, Lyapunov’s Direct Method
397–399 Autonomous Systems, 52–57
Singular Perturbation, 381, 391–394 Non-Autonomous Systems, 70
Flight Validation, 4, 5, 351, 353 Lyapunov Theorem
Function Approximation, 151–159 Exponential Stability, 60
Local Stability, 57, 60

G
Generic Transport Model M
Damaged Aircraft, 412–419 Model Matching Condition, 90, 100, 111
Flexible Aircraft, 377–380, 382 Model-Reference Adaptive Control
Longitudinal Dynamics, 364, 370 Composition, 85–90
Direct, 85, 89–96, 98–107, 109–118
Indirect, 95–99, 105–110, 117–120
H Neural Network, 179–183, 369–370
Hamiltonian Function, 235, 239, 241, 242, Nonlinear Integral Control, 197–204
294 Overview, 83–85
Hybrid Adaptive Flight Control, 414, 416 Robustness Issues, 186–204
Time-Delay Margin, 257

K
Kalman-Yakubovich Lemma, 331, 394 N
Negative Definiteness, 24–27
Neural Network
L Aircraft Adaptive Control, 370
L1 Adaptive Control, 267–274 Learning Law, 156–158
Lagrange Multiplier, 235 Radial Basis Function, 155–157
LaSalle’s Invariant Set Theorem, 63–66 Sigmoidal Function, 155–158, 369–370
Least-Squares Non-Autonomous Systems, 18–20, 69–79
Index 443

Nonlinear Systems, 17–28 Projection Method, 97–99, 110, 119, 213–

Non-Minimum Phase Systems 221
Optimal Control Modification, 322–332
Robustness Issues, 190–193
Norm R
Matrix, 34–38 Radial Unbounded Functions, 60–63
Vector, 31–35 Rate Limiting, 20
Normalization, 274–281 Rate of Convergence, 51, 52, 133
Recursive Least-Squares Adaptive Control
Aircraft Adaptive Control, 369
O Combined with Model-Reference Adap-
Optimal Control, 234–241 tive Control, 141–143
Optimal Control Allocation, 421 Direct Adaptive Control, 160–169
Optimal Control Modification Hybrid Adaptive Flight Control, 417
Adaptive Flight Control, 411–416 Indirect Adaptive Control, 137–139,
Adaptive Flutter Suppression, 396–399 143–145
Predictor-Based, 369
Adaptive Law, 234–254
Reference Model
F-18 Aircraft, 421–427
Model Mismatch, 333–343
Flexible Aircraft Adaptive Control, 382–
Overview, 88–90
386
Relative Degree, 314–318
Linear Asymptotic Property, 253–261,
Riccati Equation
317–321, 354–356, 398–399
Algebraic, 240
Output Feedback of Non-SPR and Non-
Differential, 240, 332
Minimum Phase Systems, 314–343
Robust Adaptive Control, 209, 211
Singularly Perturbed Systems, 310–311
Rohrs’ Counterexample
Systems with Control Input Uncertainty,
Optimal Control Modification Redesign,
285–291
320–321
Output Feedback Adaptive Control
Robustness, 199, 200
Observer-Based Design, 329–343, 393–
399
Single-Input-Single-Output Plant of Rel- S
ative Degree 1, 322–332 σ Modification, 220–228
Switching, 227, 228
Singular Perturbation
P Adaptive Control, 307–314
Parameter Convergence Flexible Aircraft, 381–382, 390–394
Combined Least-Squares Model- Inner Solution, 391–393
Reference Adaptive Control, 145 Outer Solution, 307–310
Least-Squares, 131–135 Tikhonov’s Theorem, 308, 391–393
Model-Reference Adaptive Control, Spectral Radius, 36, 37
104–105, 177–178 Stability, 47–79
Parameter Drift, 177–178 Asymptotic, 50–52, 60–63
Parameter Identification, 125–148 Autonomous Systems, 47–67
Pendulum Definition, 49
Double, 357–361 Exponential, 51–53, 59–61
Inverted, 354–356 Global, 60–63
Persistent Excitation, 131–135 Local, 24–27, 56–59
Phase Margin, 197, 198, 200, 202, 319 Lyapunov, 49
Phase Plane Analysis, 24–27 Non-Autonomous Systems, 69–79
Polynomial Approximation, 151–153 Uniform, 70, 71
Pontryagin’s Minimum Principle, 236 Steepest Descent, 130, 238
Positive Definiteness, 40–43 Strictly Positive Real, 314–318
Predictor Model, 147, 287, 292 Sweep Method, 239, 242
444 Index

T Unmatched, 87
Time-Delay Margin, 202, 203, 257, 258 Unstructured, 86, 176
Time-Delay Systems, 193–195 Uniform Boundedness, 71
Trace Operator, 35, 108 Uniform Continuity, 74–76
Tranversality Condition, 236, 239–241 Uniform Ultimate Boundedness, 72–75
Two-Point Boundary Value Problem, 236, Unmodeled Dynamics
237 Definition, 86, 95
Robustness Issues, 195–200
U
Uncertainty
Control Input, 88, 286, 291 V
Matched, 86, 87, 90 Van der Pol Oscillator, 28
Parametric, 85–87 Verification and Validation, 6–9
Structured, 86, 90 Virtual Control, 399–400