This action might not be possible to undo. Are you sure you want to continue?
Engineering
Power System Design and Applications for Alternative Energy Sources FERRY, AKERS, AND GREENEICH Ultra Large Scale Integrated Microelectronics HA YKIN, En. Selected Topics in Signal Processing JoHNSON Lectures on Adaptive Parameter Estimation MILUTINOVIC, En Microprocessor Design for GaAs Technology QuACKENBUSH, BARNWELL III, AND CLEMENTS Objective Measures of Speech Quality RoFFEL, VERMEER, AND CHIN Simulation and Implementation of SelfTuning Controllers SASTRY AND BonsoN Adaptive Control: Stability, Convergence, and Robustness
DENNO
ADAPTIVE CONTROL
PRENTICE HALL INFORMATION AND SYSTEM SCIENCES SERIES
Thomas Kailath, Editor
Stability, Convergence, and Robustness
Shankar Sastry
University qf California at Berkeley
ANDERSON & MOORE AsTROM & WITTENMARK GARDNER GOODWIN & SIN GRAY & DAVISSON HAYKIN JAIN JOHNSON KAILATH KuNG KUNG, WHITEHOUSE, & KAILATH, EDS. LTUNG MACOVSKI MELSA & SAGE SASTRY & BoDSON SPILKER WILLIAMS
Optimal Filtering ComputerControlled Systems: Theory and Design Statistical Spectral Analysis: A Nonprobabilistic Theory Adaptive Filtering, Prediction, and Control Random Processes: A Mathematical Approach for Engineers Adaptive Filter Theory Fundamentals of Digital Image Processing Lectures on Adaptive Parameter Estimation Linear Systems VLSI Array Processors VLSI and Modern Signal Processing System Identification: Theory for the User Medical Imaging Systems An Introduction to Probability and Stochastic Processes Adaptive Control: Stability, Convergence, and Robustness Digital Communications by Satellite Designing Digital Filters
Marc Bodson
Carnegie Mellon University
Prentice Hall, Englewood Cliffs, New Jersey 07632
Library of Congress CataloginginPublication Data
Sastry, Shankar (Sosale Shankara) Adaptive control : stability, convergence, and robustness I Shankar Sastry and Marc Bodson. p. cm.(Prentice Hall information and system sciences series) (Prentice Hall advanced reference series) Bibliography: p. Includes index. ISBN 0130043265 I. Adaptive control systems. I. Bodson, Marc. II. Title. Ill. Series. TJ217.S27 1989 8828940 629.8 '36<lc 19 CIP
PrenticeHall Advanced Reference Series PrenticeHall Information and System Sciences Series
To our parents
•
© 1989 by PrenticeHall, Inc.
= =
A Division of Simon & Schuster Englewood Cliffs, New Jersey 07632
All rights reserved. No part of this book may be reproduced, in any form or by any means, without permission in writing from the publisher.
Printed in the United States of America 10 9 8 7 6 5 4 3 2 1
ISBN 0130043265
PrenticeHall International (UK) Limited, London PrenticeHall of Australia Pty. Limited, Sydney PrenticeHall Canada Inc., Toronto PrenticeHall Hispanoamericana, S.A., Mexico PrenticeHall of India Private Limited, New Delhi PrenticeHall of Japan, Inc., Tokyo Simon & Schuster Asia Pte. Ltd., Sinxapore Editora PrenticeHall do Brasil, Ltda., Rio de Janeiro
CONTENTS
Preface Chapter 0 Introduction 0.1 Identification and Adaptive Control 0.2 Approaches to Adaptive Control 0.2.1 Gain Scheduling 0.2.2 Model Reference Adaptive Systems 0.2.3 Self Tuning Regulators 0.2.4 Stochastic Adaptive Control 0.3 A Simple Example Chapter 1.1 1.2 1.3 1.4 1 Preliminaries Notation LP Spaces, Norms Positive Definite Matrices Stability of Dynamic Systems 1.4.1 Differential Equations 1.4.2 Stability Definitions 1.4.3 Lyapunov Stability Theory 1.5 Exponential Stability Theorems 1.5.1 Exponential Stability of Nonlinear Systems 1.5.2 Exponential Stability of Linear TimeVarying Systems 1.5.3 Exponential Stability of Linear Time Invariant Systems 1.6 Generalized Harmonic Analysis
xv 1 1 4 4 5 8 9 11
17 17 18 19 20 20 23 25 28 28
32 38 39
2 LeastSquares Algorithms 2.1 Output Error SchemeLinearized Equations 4.1 Model Reference Adaptive Control Problem 3.5.3 Stability of the Identifier 2.4 Averaging TheoryTwoTime Scales 4.3 Regressor Vector Filtering 5.2 StabilityOutput Error Direct Adaptive Control 3.7 Frequency Domain Conditions for Parameter Convergence 2.2 Mixed Time Scales 4.7.5.1 Instability Theorems Using Averaging 5.3 The Usage of Prior Information 5.2 Strictly Positive Real Error Equation and Identification Algorithms 2.6 Conclusions Chapter 5.3 5 Robustness Structured and Unstructured Uncertainty The Rohrs Examples Robustness of Adaptive Algorithms with Persistency of Excitation 5.1 Exponential Convergence and Robustness 5.7.3 Adaptive Control Schemes 3.6 Methods for Improving RobustnessQualitative Discussion 5.6.8 Exponential Parameter Convergence 3.2 Specification of the Closed Loop Control ObjectiveChoice of Control Model and of Reference Input 5.3 StabilityIndirect Adaptive Control 45 45 52 53 57 58 61 63 63 66 69 71 76 76 82 85 90 90 95 98 99 99 103 104 110 111 3.6. 7.3 Exponential Convergence of the Gradient Algorithms with SPR Error Equations 2.1 5.2 Partial Parameter Convergence 2.3.5.5 Analysis of the Model Reference Adaptive Control System 3.6 Model Reference IdentifiersSPR Error Equation 2.6.3.3.1 Gradient Algorithms 2.0 2.3.1 Model Reference Identifiers 2.6 Useful Lemmas 3.1 Deadzone and Relative Deadzone 5.7.3.1 Input Error Direct Adaptive Control 3.3. 7.3.3 Application to Identification 4.4 Time Variation of Parameters 5. 7 Robustness via Update Law Modifications 5.1 StabilityInput Error Direct Adaptive Control 3.4.7.4 Heuristic Analysis of the Rohrs Examples 5.8 Conclusions Chapter 3 Adaptive Control 3.7.4.4 Properties of the Identification AlgorithmsIdentifier Stability 2.4.2 Leakage Term (aModification) 5.3.5 Adaptive Pole Placement Control 3.6.5.3 2 Identification Introduction Identification Problem Identifier Structure Linear Error Equation and Identification Algorithms 2.1 Examples of Averaging Analysis 4.1 Parameter Convergence 2.3 Indirect Adaptive Control 3.1 Robust Identification Schemes 5.2 Robustness of an Adaptive Control Scheme 5.6.0 Introduction 4.2 Output Error Direct Adaptive Control 3. 7.4 Slow Adaptation.8 Conclusions 154 156 158 158 159 166 175 179 183 186 187 188 192 202 207 209 209 215 219 221 225 231 236 236 241 248 248 118 123 127 129 130 132 138 142 142 149 151 250 250 251 251 251 253 253 254 254 .4 The Stability Problem in Adaptive Control 3.0 Introduction 3.1 2.4 Alternate Model Reference Schemes 3.X XI Chapter 2.9 Conclusions Chapter 4 Parameter Convergence Using Averaging Techniques 4.3 Input Error Scheme 4. 7 Stability Proofs 3.2 Application to the Output Error Scheme 5.6. Averaging and Hybrid Update Laws 5.2 Controller Structure 3.2 5.5.7.2 Output Error SchemeNonlinear Equations 4.5 Persistent Excitation and Exponential Parameter Convergence 2. 5 Averaging Analysis of Slow Drift Instability 5.4.5 Applications to Adaptive Control 4.4.1 Separated Time Scales 4.1 Gradient Algorithms 2.6.3.2 LeastSquares Algorithms 2.2 2.2 Averaging TheoryOneTime Scale 4.
2 Effect of Unmodeled Dynamics 6.3.1 SingleInput SingleOutput.1 The SingleInput SingleOutput Case 7.1 Factorization of Transfer Function Matrices 6.5 Alternate Schemes 6.2.3.2 Indirect Adaptive Pole Placement 6. Relative Degree One Case 7.1 General Conclusions 8.1 Basic Theory 7.3 Model Reference Control for Nonlinear Systems 7.2.3.2 Preliminaries 6.2.2 Extensions to Higher Relative Degree SISO Systems 7.2.1.1 Use of Prior Information 6.2 Global Stability of Indirect Adaptive Control Schemes 6.4 Conclusions Chapter 7 Adaptive Control of a Class of Nonlinear Systems 7.1 Introduction 6.2.3 Indirect Adaptive StabilizationThe Factorization Approach 6.3.3.3 Adaptive Control of Linearizable Minimum Phase Systems 7.2.3 Adaptive Control of MIMO Systems Decouplable by Static State Feedback 7.2.1 Indirect Adaptive Control Scheme 6.3.Xll Xlll Chapter 6 Advanced Topics in Identification and Adaptive Control 6.2 Linearizing Control for a Class of Nonlinear SystemsA Review 7.3.3.3 Model Reference Adaptive ControlController Structure 6.2.2 Minimum Phase Nonlinear Systems 7.2 Future Research Appendix References Index 324 324 325 331 359 373 274 277 277 278 278 282 286 290 292 293 294 294 295 295 299 300 305 307 309 309 312 320 322 .4 Conclusions 257 257 257 263 266 267 272 Chapter 8 Conclusions 8.2.2.2.3.2.3.1 Introduction 7.4 Model Reference Adaptive ControlInput Error Scheme 6.2 The MultiInput MultiOutput Case 7.1.2 Interactor Matrix and Hermite Form 6.3 Multivariable Adaptive Control 6.1 Identification of Partially Known Systems 6.
techniques of analysis and new directions of research in adaptive systems. Such a treatment is particularly timely. evaluation and testing of the algorithms. Our objective has been to give a clear. we present the deterministic theory of identification and adaptive control. For the most part the focus is on linear. Current research in the adaptive control of multiinput. conceptual presentation of adaptive methods. multioutput linear systems and a XV . in a concise and unified fashion. Several successful applications have been reported and the last ten years have seen an impressive growth in the availability of commercial adaptive controllers. In this book.PREFACE The objective of this book is to give. There have been many theoretical successes. The presentation includes the algorithms. limitations to future growth can hardly be expected to be computational. Indeed. Adaptive control has been the subject of active research for over three decades now. the major results. but rather from a lack of a fundamental understanding of the methodologies for the design. including the development of rigorous proofs of stability and an understanding of the dynamical properties of adaptive schemes. their dynamical properties and tools for analysisincluding the recently introduced averaging techniques. continuous time. to enable a critical evaluation of these techniques and suggest avenues of further development. singleinput singleoutput systems. given the rapid advances in microprocessor and multiprocessor technology which make it possible to implement the fairly complicated nonlinear and time varying control laws associated with adaptive control.
In Chapter 2. but not least. We are especially thankful to those who collaborated with us in research. including NASA (Grant NAG243). Bob Bitmead. for their diligent typing and layout of the manuscript in the presence of uncertainty. Some background in basic control systems and in linear systems theory at the graduate level is assumed. we would like to express our sincere appreciation to Nirmala and Cecilia for their patience and encouragement. at the Electronics Research Laboratory at Berkeley. We also acknowledge Bernard Goodwin of Prentice Hall for his friendly management of this enterprise and Elaine Lynch for coordinating production matters. in the Massachussetts Institute of Technology. They have helped us broaden our view and understanding of the field. Soura Dasgupta. Niklas Nordstrom. the IBM Corporation (Faculty Development Award). Andy Packard. and we wish them good luck. they are presented in such a way as to enable their transcription to the discrete time case. Chapter 1 is a chapter of mathematical pre! iminaries containing most of the key stability results used later in the book. thanks to the hospitality of Sanjoy Mitter. His advice and support from the beginning are gratefully acknowledged. Bob Narendra. The persistently exciting inputs of students at Berkeley and Carnegie Mellon have helped us refine much of the material of this book.. Brian Anderson and Petar Kokotovic offered excellent critical comments that were extremely helpful both in our research. Rick Johnson. Iven Mareels. A brief outline of the book is as follows: Chapter 0 is a brief historical overview of adaptive control and identification. We also thank Karl Astrom and Steve Morse for their enthusiasm about adaptive control. and deserve our deepest gratitude. we give a self contained presentation of averaging techniques and we analyze the rates of convergence of the schemes of Chapters 2 and 3. we develop several adaptive identification algorithms along with their stability and convergence properties. Graham Goodwin. Last. We are indebted to many colleagues for stimulating discussions at conferences. and we thank Karl Astrom for providing us with a copy of this software package. Berkeley. Bob Kosut. Charles Rohrs. Brad Paden and Tim Salcudean. and an introduction to various approaches. Background in stability theory for nonlinear systems is desirable. Brad Riedle. explicitly linearizable by state feedback and their adaptive control using the techniques of Chapter 3. and contributed to this work: Erwei Bai. Dorothee NormandCyrot. Michael Athans. Stephen Boyd. Special thanks are due to George Meyer and Jagdish Chandra: their continuous support of our research made this book possible. Chapter 6 covers some advanced topics: the use of prior information in adaptive identification schemes. We have used the book as a text several times for a onesemester graduate course at the University of California at Berkeley and at CarnegieMellon University. and in the revisions of the manuscript.XVI Preface Preface XVII class of nonlinear systems is also covered. Despite distance. workshops and meetings. Romeo Ortega. how to analyze their potential instability using averaging techniques and how to make the schemes more robust. indirect adaptive control as an extension of robust nonadaptive control and multivariable adaptive control.tems. Chapter 7 gives a brief introduction to the control of a class of nonlinear systems. Cynthia Bilbrey and Craig Louis. Fathi Salam and Lena Valavam. Ca!Ui>rnia Shankar Sastry Marc Bodson . but the presentation is mostly selfcontained. Michel de Mathelin. The figures were drafted by Osvaldo Garcia. We wish to express our gratitude to Carol Block and Tim Burns. We acknowledge the support of several organizations. our families have been a source of continuous support. and the National Science Foundation (Grant ECS881 0 145). Acknowledgments It is a pleasure to acknowledge the contributions of the people who helped us in the writing of this book. Ping Hsu. Chapter 3 is a corresponding development for model reference adaptive control. LiChen Fu. Jim Krause. Sanjoy Mitter. Although continuous time algorithms occupy the bulk of our interest. We are also grateful for the logistical support received from the administration of the Department of Electrical Engineering and Computer Sciences at the University of California at Berkeley and of the Electrical and Computer Engineering Department of Carnegie Mellon University. dissertation of the second author on which this book is based. Many of them have now adapted to new environments. This book is intended to introduce researchers and practitioners to the current theory of adaptive control.D. Chapter 8 concludes with some of our suggestions about the areas of future exploration. Alberto lsidori. Rogelio LozanoLeal. We are especially appreciative of the detailed and thoughtful reviews given by Charles Desoer of the original Ph. Laurent Praly. In Chapter 4. Ed Kamen. Jeff Mason. Simulations were executed using the package SIMNON. Petros Ioannou. Chapter 5 deals with robustness properties of the adaptive schemes. and for fruitful interactions. We would particularly like to mention Anu Annaswamy. Part of our work was also done at the Laboratory for Information and Decision Syc. the Army Research Office (Grant DAAG 2985K0072).
. Then. system identification could be aimed at determining if the plant to be controlled is linear or nonlinear. and only parameters of a fixed type of model need to be determined. we will thus be limiting ourselves to parametric system identification. the two steps will be taken separately.CHAPTER 0 INTRODUCTION 0. and some classes of multi variable and nonlinear plants. finite or infinite dimensional. Abstractly. Often. pragmatically. It is thus intuitive to aggregate system identification and control. Here we will restrict our attention to finite dimensional. We will see adaptive control. and has continuous or discrete event dynamics. Control design techniques then need to be augmented with an identification technique aimed at obtaining a progressively better understanding of the plant to be controlled. and parametric adaptive control. singleinput singleoutput linear plants. the plant to be controlled is too complex and the basic physical processes in it are not fully understood.1 IDENTIFICATION AND ADAPTIVE CONTROL Most current techniques for designing control systems are based on a good understanding of the plant under study and its environment. the primary step of system identification (structural identification) has already been taken. If the system identification is recursivethat is the plant model is periodically updated on the basis of previous estimates and new dataidentification and control may be performed concurrently. However. in a number of instances. Implicitly. as a direct aggregation of a (nonadaptive) control methodology with some form of recursive system identification.
robot manipulators. Parks [ 1966] found a way of redesigning the update laws proposed in the 1950s for model reference schemes so as to be able to prove convergence of his controller. Goodwin & Sin [1984].I. The design techniques for adaptive systems are studied and analyzed in theory for unknown but fixed (that is. gain scheduling based on some auxiliary measurements of airspeed was adopted. Mareels. Lyapunov's stability theory was firmly established as a tool for proving convergence in adaptive control schemes. & Valavani [1980] and Morse [1980]. However. In practice.T. Finally. robot manipulators. Several schemes of selfadjustment of the controller parameters were proposed. singleinput. An attempt to link the signal processing viewpoint .2 Introduction Section 0. System identification (offline) was thoroughly researched and understood. input output (Popov hyperstability based) proofs appeared in Egardt [ 1979] and Landau [ 1979]. The 1960s marked an important time in the development of control theory and adaptive control in particular. then. Adaptive control. it was frequently observed that a single constant gain controller would not suffice. In the 1950s. the 1980s have proven to be a time of critical examination and evaluation of the accomplishments to date.. automatic adjustment is needed. and are contained in the textbook by Goodwin & Sin [1984]. Ljung & Soderstrom [1983] and Ljung [ 1987]. which is an important part of adaptive control are by Eykhoff [1974]. typically highfrequency parasitic modes that were neglected to limit the complexity of the controller. All this flood of research and development is bearing fruit and the industrial use of adaptive control is growing. Overview of the Literature Research in adaptive control has a long and vigorous history. Egardt [1979]. Stochastic control made giant strides with the understanding of dynamic programming. and were verified to perform well under certain conditions. For a number of simple PID (proportional + integral + derivative) controllers in process control. Kokotovic. power systems. singleoutput plant and the usage of these parameter estimates to update an optimal linear quadratic controller. & Riedle [1986]. Johnson. complete proofs of stability for several adaptive schemes appeared. they are applied to slowly timevarying and unknown plants. Praly. Given the firtu. it was motivated by the problem of designing autopilots for aircraft operating at a wide range of speeds and altitudes. this period was marked by the culmination of the analytical efforts of the past twenty years. Adaptive control has a rich and varied literature and it is impossible to do justice to all the manifold publications on the subject. While the object of a good fixedgain controller was to build an autopilot which was insensitive to these (large) parameter variations. With this scheme in place several rudimentary model reference schemes were also attemptedthe goal in this scheme was to build a selfadjusting controller which yielded a closed loop transfer function matching a prescribed reference model. and they vary with time. Thus.1 Identification and Adaptive Control 3 Applications of such systems arise in several contexts: advanced flight control systems for aircraft or spacecraft. Lin. Further. Polderman ( 1988] and Caines [ 1988]. Kosut. Recent books dealing with the theory of adaptive control are by Landau [ 1979]. Consequently. In the late 1970s. State space (Lyapunov based) proofs of stability for model reference adaptive schemes appeared in the work of Narendra. The implementation of the ::omplicated nonlinear laws inherent in adaptive control has been greatly facilitated by the boom in microelectronics and today. Ioannou & Kokotovic [ 1984]. due to Bellman and others. Learning schemes proposed by Tsypkin. Ramadge. It is a tribute to the vitality of the field that there are a large number of fairly recent books and monographs. Anderson. In the 1970s. when the number of parameters is larger than three or four. process control. Stability proofs in the discrete time deterministic and stochastic case (due to Goodwin. Goqdwin & Payne [1977]. Kalman [ 1958] put on a firm analytical footing the concept of a general selftuning controller with explicit identification of the parameters of a linear. Feldbaum and others (see Tsypkin [1971] and [1973]) were shown to have roots in a single unified framework of recursive equations. is a technique of applying some system identification technique to obtain a model of the process and its environment from inputoutput experiments and using this model to design a controller. The parameters of the controller are adjusted during the operation of the plant as the amount of data available for plant identification increases. This sparked a flood of research into the robustness of adaptive algorithms: a reexamination of whether or not adaptive controllers were at least as good as fixed gain controllers. and others. Kumar and Varaiya [ 1986]. the development of tools for the analysis of the transient behavior of the adaptive algorithms and attempts at implementing the algorithms on practical systems (reactors. rule. time invariant) plants. & Caines [ 1980]) also appeared at this time. owing to the culmination of determined efforts by several teams of researchers. Some recent books on recursive estimation. Bitmead. this is often done manually. one can talk in terms of custom adaptive controller chips. and ship steering systems to mention only a few). analytical footing of the work to this point. It was first pointed out by Rohrs and coworkers [ 1982] that the assumptions under which stability of adaptive schemes had been proven were very sensitive to the presence of unmodeled dynamics. such as the sensitivity rules and the socalled M.
It is convenient especially if the plant dynamics depend in a wellknown fashion on a relatively few easily measurable variables. many of the existing results concern either algorithms. and the effectiveness of the design was corroborated by simulation. and a great deal more needs to be understood about the dynamic behavior of adaptive systems. two adaptive control schemes other than gain scheduling were proposed to compensate for changes in aircraft dynamics: a series. This. and in books edited by Narendra & Monopoli [ 1980] and Unbehauen [1980]. AUXILIARY MEASUREMENT Again in the context of flight control systems. Ninety controllers were designed. Airspeed sensors modified the control scheme of the helicopter in flight. the disadvantage of gain scheduling is that it is an openloop adaptation scheme. as was illustrated by the flight control system implemented on a CH47 helicopter. Consequently. In the example of flight control systems. our objective in this book is to address fundamental issues of stability. highgain scheme.1: Gain Scheduling Controller Figure 0. Although gain scheduling is extremely popular in practice.1 Gain Scheduling The advantage of gain scheduling is that the parameters can be changed quickly (as quickly as the auxiliary measurement) in response to changes in the plant dynamics. the dynamics depend in relatively simple fashion on the readily available dynamic pressurethat is the product of the air density and the relative velocity of the aircraft squared. It was introduced in particular in the context of flight control systems in the 1950s and 1960s. The idea is to find auxiliary process variables (other than the plant outputs used for feedback) that correlate well with the changes in process dynamics. and a linear interpolation between these controllers (linear in the horizontal and vertical flight velocities) was programmed onto a flight computer. 0.2.1. This is illustrated in Figure 0. The flight envelope of the helicopter was divided into ninety flight conditions corresponding to thirty discretized horizontal flight velocities and three vertical velocities. with no real "learning" or intelligence. Surveys of the applications of adaptive control are given in a book by Harris & Billings [ 1981 ]. As of the writing of this book. Further.2 shows a schematic of the series highgain scheme. we feel that there is a need for a "toolkit" of methods of analysis comparable to nonadaptive linear time invariant systems. It is then possible to compensate for plant parameter variations by changing the parameters of the regulator as functions of the auxiliary variables. and Narendra & Annaswamy are also nearing completion. and a parallel scheme.2 APPROACHES TO ADAPTIVE CONTROL 0.2 Model Reference Adaptive Systems One of the earliest and most intuitive approaches to adaptive control is gain scheduling. REFERENCE SIGNAL r PLANT OUTPUT Yp PLANT OUTPUT Yp Figure 0. two other books on adaptive control by Astrom & WittenmarY. Series HighGain Scheme Figure 0. Further. corresponding to each flight condition. has limited practical applications more than it should have. HighGain Scheme .2: Model Reference Adaptive ControlSeries. the extent of design required for its implementation can be enormous. 0. convergence and robustness.2 Approaches to Adaptive Control 5 with the adaptive control viewpoint is made in Johnson [ 1988]. Also.4 Introduction Section 0.2. we will review some common approaches to adaptive control systems and introduce the basic issues studied in this book with a simple example. we believe. In spite of the great wealth of literature. structures or specific applications. we hope to communicate our excitement about the problems and potential of adaptive control. In the remainder of the introduction.
The two axes were assumed decoupled for the purpose of control design. However. the gain is decreased. using differential and identical commands respectively.Ym to zero. The key problem in the scheme is to obtain an adjustment mechanism that drives the output error e0 = Yp. The limit cycle detector is typically just a rectifier and lowpass filter. The controller YM Reference Input Model Output eo Yp . so that the closedloop transfer function becomes close to 1. and the adaptive system attempts to make the plant output match the reference model output asymptotically. Due to the saturation. was used. and represents the sensitivity of the output error to variations in the controller parameter 0. tragically. until the onset of instability (a limit cycle) is detected. The idea behind the gradient update is to reduce by adjusting () along the direction of steepest descent. that is. The series highgain scheme is intuitive and simple: only one parameter is updated.3: Model Reference Adaptive ControlParallel Scheme can be thought of as having two loops: an inner or regulator loop that is an ordinary control loop consisting of the plant and regulator.1ce to this scheme is Osburn. the instability remained undetected. It is thus desired that the aircraft response. Ym· The simple analysis that goes into the scheme is as follows: consider P(s) to be the transfer function of the linear. time invariant plant and k the constant gain of the servo. Plant Output Controller Parameter 8 OUTER LOOP Figure 0. an ex¥erimental X15 aircraft flying this control system crashed in 1966 (cf. b) Noise in the frequ. If the limit cycle oscillations exceed some level. The roll and pitch axes were controlled by the right and left rear ailerons. c) Reference inputs may cause saturation due to the highgain.2 Approaches to Adaptive Control 7 The reference model represents a pilot's desired commandresponse characteristic. & Kezer [ 1961 ]. Indeed. the output Yp.2.2) where g is a positive constant called the adaptation gain. the gain is increased. Let the vector () contain the adjustable parameters of the controller.ansfer function is approximately 1 over the frequencies of interest. The interpretation of e 0 (0) is as follows: it is the output error (also a function of time) obtained by freezing the controller parameter at 0. Staff of the Flight Research Center [ 1971 ]).1) = 2geo(O) a ao (Yp(O)) (0. that is. that is dO dt e6 (0) a ao (e6 (O)) a 2geo(O) ao (eo(O)) . The gradient of e0(0) with respect to () is equal to the gradient of Yp with respect to 0. and an outer or adaptation loop that adjusts the parameters of the regulator in such a way as to drive the error between the model output and plant output to zero. and created aerodynamic forces too great for the aircraft to withstand. it has the following problems a) Oscillations are constantly present in the system. the following update.2. called the gradient update. owing partially to the saturation problems occurring in the highgain scheme. Parallel Scheme As in the series scheme. . matches the output of the reference model. The transfer function from Ym to y is k P(s )/ 1 + k P(s ).ency band of the limit cycle detector causes the gain to decrease well below the critical value. saturation of the actuators in the pitch axis caused the aircraft to lose controllability in the roll axis (since the ailerons were at maximum deflection). and leading to instability. d) Saturation may mask limit cycle oscillations. However. A block diagram is shown in Figure 0.g (0. Whitaker. allowing the gain to increase above the critical value. since Ym is indepenqent of 0.3.6 Introduction Section 0. the t. In the earliest applications of this scheme. The aim of the scheme is to let the gain k be as high as possible.Yp. An early refere. the desired performance of the closedloop system is specified through a reference model. so that Ym. Below this level. When the gain k is sufficiently large.
This is usually called a certainty equivalence principle. Through this book. was proposed. The controller is called self tuning. and a relationship between plant parameters and controller parameters.. and to the choice of identification scheme Reference Input r 1. where the parameters of the controller are updated directly to achieve the goal of model following. so that stable and provably convergent model reference schemes were obtained. James [ 1971 ]). the only real difference between them is that model reference schemes are direct adaptive control schemes.__ Plant Output yp Controller Parameter 8 Figure 0. It is easy to see that the inner or control loop of a self tuning regulator could be the same as the inner loop of a ~odel reference design.. in the same way as if these were the true parameters. The self tuning regulator. it can be thought of as having two loops: an inner loop consisting of a conventional controller. which replaced the unknown parameters by their estimates at time t. .. in contrast to the model reference schemes above.. The self tuning regulator is very flexible with respect to its choice of controller design methodology (linear quadratic.. it was observed that the M. The controller parameters are then obtained from the estimates of the plant parameters. The stability and convergence properties of model reference adaptive systems make them particularly attractive and will occupy a lot of our interest in this book. The resulting scheme is represented on Figure 0. one starts from a control design method for known plants. Again. At this point the socalled M. maximum likelihood. The model reference adaptive schemes update the controller parameters directly (no explicit estimate or identification of the plant parameters is made). Unfortunately. first identifies the plant parameters recursively.. Or. Empirically. minimum variance. or convergence of the output error to zero. with the sensitivity iJ Yp (0) I iJ 0 replaced by other functions. the model reference adaptive schemes can be seen as a special case of the self tuning regulators with an identity transformation between updated parameters and con~ troller parameters. it is not possible in general to prove closedloop stability. gainphase margin design. ). 0.2. we will distinguish between direct and indirect schemes rather than between model reference and self tuning algorithms. whereas self tuning regulators are indirect. The sensitivity function iJ Yp (0) I iJ 0 usually depends on the unknown plant parameters. and an outer loop consisting of an identifier and a design box (representing an online solution to a design problem for a system with known parameters) which adjust these controller parameters.8 Introduction Section 0. examples of instability could be obtained otherwise (cf.2 Approaches to Adaptive Control 9 Several problems were encountered in the usage of the gradient update.I. rule.... for schemes based on the M. The analysis of self tuning adaptive systems is however more complex than the ~nalysis of model reference schemes./. Parks [ 1966] found a way of redesigning adaptive systems using Lyapunov theory. An explicit separation between identificativn and control is assumed. . but with varying parameters. ..2). extended Kalman filtering.T. since it has the ability to tune its own parameters.. they are obtained using a recursive parameter identification algorithm. Direct and Indirect Adaptive Control While model reference adaptive controllers and self tuning regulators were introduced as different approaches. ).____.. T.4: Selftuning Controller (least squares.. rule performed well when the adaptation gain g and the magnitude of the reference input were small (a conclusion later confirmed analytically by Mareels et a/ [ 1986]). due primarily to the (usually nonlinear) transformation from identifier parameters to controller parameters. in other words.4.I. However. This design method is summarized by a controller structure.2. and then uses these estimates to update the controller parameters through some fixed transformation. Since the plant parameters are in fact unknown. The self tuning approach was originally proposed by Kalman [1958] and clarified by Astrom & Wittenmark [1973]. The update laws were similar to (0. and is consequently unavailable. rule.3 Self Tuning Regulators In this technique of adaptive control.T.
but its sign is known (say kP > 0). This property is referred to as dual control. however. Yet. the following interesting observations can be made of the optimal control law: in addition to driving the plant output to its desired value. linear system with transfer function (0.2 Approaches to Adaptive Control 11 0. time invariant. We consider a first order.4 Stochastic Control Approach Adaptive controller structures based on model reference or self tuning approaches are based on heuristic arguments. with the hyperstate approximated by the process state and process parameters estimate. and a criterion is formulated to minimize the expected value of a loss function. The control objective is to get the plant output to match a model output.1 Plant Output presence of this rich enough excitation may result in poor transient performance of the scheme. This can be done (in principle.5: "Generic" Stochastic Controller The estimator generates the conditional probability distribution of the state from the measurements: this distribution is called the hyperstate (usually belonging to an infinite dimensional vector space). When indeed they can be solved. While we will not explicitly study stochastic control in this book. and for the identification to be complete. the optimal controllers have the structure shown in Figure 0. This.1) s+a where a> 0 is known.2) Figure 0. We also discuss some of the adaptive schemes of the previous section in this context. 0. it would be appealing to obtain such structures from a unified theoretical framework. The system and its environment are described by a stochastic model. the foregoing tradeoff will be seen repeatedly: good adaptive control requires correct identification. In this section. at least) using stochastic control.10 Introduction Section 0.6: Simple Feedforward Controller Note that.5: an identifier (estimator) followed by a nonlinear feedback regulator. The Only gain compensationor feedforward controlis necessary. The design of the various . We will call(0. which is a scalar function of states and controls. the controller introduces probing signals which improve the identification and.3.3. we will introduce some of these problems with a simple example. displaying the tradeoff between learning and control performance. if kp were known. From some limited experience with stochastic control. The gain kp of the plant is unknown. Reference Slgnal_r_. The optimal regulator maintains a balance between the control activity for learning about the plant it is controlling and the activity for controlling the plant output to its desired value. 8 would logically be chosen to be 1 1kp. time invariant. even if the plant that they are controlling is linear. the controller signal has to be sufficiently rich to allow for the excitation of the plant dynamics. delicate technical problems. This leads to interesting and.3. namely a gain 8 at the plant input. as is shown on Figure 0. The self tuner may be thought of as an approximation of this controller. therefore future control. represents some cost in terms of control activity.6. It is usually very difficult to solve stochastic optimal control problems (a notable exception is the linear quadratic gaussian problem). where the reference model transfer function is M(s) = A Yp P(s) _!p_ 1 s+a (0. that is the value which realizes the output matching objective for all inputs.3 A SIMPLE EXAMPLE Adaptive control systems are difficult to analyze because they are nonlinear. time varying systems.2. Figure 0.3) the nominal value of the parameter 8.
3.3. Gain Scheduling Let v(t) e IR.14) o (0. it is easy to see that ayp(O) ao Note that since kp > 0. it would appear that e0 . Indeed.3.3. However.3. be some auxilia cy measurement that correlates in known fashion with kp.8) Subtracting (0. From Figure 0. Then. Having concluded that e0 .3. what can we say about 0? Does it indeed converge to 0* = 1lkp? The answer is that one can not conclude anything about the convergence of 0 to 0* without extra conditions on the reference input.6) Note that (0.T rule becomes 0 = geoYm g>O 0 yields v(e 0 .9) = Since we would like 0 to converge to the nominal value 0* define the parameter error as cJ> = o . Rule To apply the M.6) prescribes an update of the parameter 0 in the direction opposite to the "correlation" product of e0 and the model output Ym· Model Reference Adaptive Control Using the Lyapunov Redesign The control scheme is exactly as before.2.3.3.oo. This actually follows from further analysis. The plant and reference model are described by .3. the function v ( e0 .3.12) update law to be defined Model Reference Adaptive Control Using the M. The derivative of v along the trajectories of the error system (0.cJ>) = 2aecr ~ (0.T. we (0. we get.O*)r = Yp .14) and that e0 .ayP + kpOr aym + r = aym + kpO*r thereby guaranteeing that e6 + kpc/> 2 is decreasing along the trajectories of (0. The M. with e0 eo = a eo+ kp(O.jt) is similar in form to (0.T.0 as t . Since v(e 0 . cJ>) is decreasing and bounded below.12).o• 1 I kp . and cJ> are bounded.15) (0. say kp(t) = f(v(t)).7). provided that r is bounded (cf.oo.5) depends on the parameter kp which is unknown.6).0 as t .3. but the parameter update law is chosen to make a Lyapunov function decrease along the trajectories of the adaptive system (see Chapter 1 for an introduction to Lyapunov analysis).13) (0.3.3. we need to obtain ae 0 (0)IaO = ayp(O)IaO.c/>) = ecr + kpc/> 2 (0.7) (0.4) ae 0 + kpcJ>r (0. so that a Yp I a 0 is not available.3 A Simple Example 13 adaptive schemes proceeds as follows. if the reference input was a constant zero signal.I.3. (0. The Lyapunov redesign approach consists in finding an update law so that the Lyapunov function v(eo. Conditions for parameter convergence are important in adaptive control and will be. the sign of kp is known (kp > 0). Barbalat's lemma 1.Ym (0. so that we may merge the constant kp with the adaptation gain.¢ = 0.3.14). with the difference that e0 is correlated with r rather that Ym· An adaptation gain g may also be included in (0.3.1).studied in great detail. the gain scheduler chooses at timet O(t) = Note that since 0* is fixed (though unknown).3. with the understa ••ding that 0 is frozen.3. c/>) = ~ ( {.I.6.I.) = s+a I'· k pYm I (0. there would be no reason to expect 0 to converge to 0*.11) is decreasing along trajectories of the error system 1 f(v(t )) (0.12) is given by v ( e0 . cJ>) is a positive definite function.3.3. An answer to this question for the simple example will be given for the following indirect adaptive control scheme.12 Introduction Section 0. '"' (0.8) from (0. Note that (0.3.12) Choosing the update law We see immediately that the sensitivity function in (0.3. rule.3.10) .5) (0.
The difficulty with (0.3.Ji(t). which in turn depends on u and therefore on both 8 and r.3. we will assume that the control objective is still model matching. we define 1r• = kp.3. _P_ k (8 r) s+a 1r• w so that e.l I kp. as too.(0r) = .26) (0.3.3. The controller parameter is chosen following the certainty equivalence principle: since o• = l I kp and 1r* = kp. Figure 0. The update law is based on the identifier error e. Its derivative along the trajectories of the adaptive system is (0. L________ ______. It is easily seen that. !fw (0.1r• and O(t).(u) s+a s+a (0.26) is referred to as an identifiability condition and is much related to the socalled persistency of excitation that will be discussed in Chapter 2. with the same model as above. The hope is that. the update law causes a decreasing parameter error and all signals remain bounded.3. The update law now is an update law for the identifier parameter 1r(t ).18) is used in the actual implementation of the algorithm.19) Indirect Adaptive Control (Self Tuning) To be able to effectively compare the indirect scheme with the direct schemes given before.3 A Simple Example 15 (0. Define the identifier parameter error (0. (0. Therefore.26) is that it depends on w.16) and let 1 1 w = . The condition (0.3. e. Note that (0.20).7: A Simple Indirect Controller The identifier contains an identifier parameter 1r(t) that is an estimate· of the unknown plant parameter kp.Ji Figure 0.3.oo. since a > 0 is known. note that then . it excludes signals which tend to zero as t . The question of parameter convergence can be answered quite simply in this case.3.18) Equation (0.J I v = 1f2 (0.14 Introduction Section 0.3. Converting it into a condition on the exogenous reference input r(t) only is another of the problems which we will discuss in the following chapters. so that 1r(t).0.21) and let the Lyapunov function .24) I which may be explicitly integrated to get !f(t) = !f(O)exp(gjw 2(r)dr) 0 (0.20) Consider the update law ge.E~!~~~. .3.3.3.3. yielding the desired controller. and we proceed to derive one of them. := 1rW.3. 1r(t)kp.25) It is now easy to see that if I Jw 2(r)dr 0  00 as too (0.17) The signal w may be obtained by stable filtering of the input u. so that O(t)llkp. For the analysis.22) I w I I I I This Lyapunov function has the special form of the norm square of the identifier parameter error.23) Therefore.Yp (0. we let O(t) = l I 1r(t ). in particular.w g>O (0.21) represent the first order linear time varying system ~ = gw2. There are several possibilities at this point.3. 7 shows an indirect or self tuning type of model reference adaptive controller.
When u (t) is a function of time. Once feedback control is involved. and on nonstandard results. and 17 . Sometimes. CHAPTER 1 PRELIMINARIES This chapter introduces the notation used in this book. and Narendra. n(s) or simply n. 1. as well as some basic definitions and results. and this chapter will concentrate on the definitions used most often. operators. time invariant. since many connections exist between this work and their results. where if is both the ratios of polynomials in s and an operator in the Laplace transform domain. if (s) or if. Upper case letters are used to denote matrices. Polynomials in s will be denoted using lower case letters. The material is provided mostly for reference. u(s) denotes its Laplace transform. linear equations. & Valavani [1980]. and simply write u and u. Rational transfer functions of linear time invariant (LTI) systems will be denoted using upper case letters.16 Introduction The foregoing simple example showed that even when simple feedforward control of a linear. for example. or by the reader familiar with the results. we will drop the arguments. we may have if= n/d. or sets. Thus. It may be skipped in a first reading. The notation used in the adaptive systems literature varies widely. Lin. first order plant was involved. the analysis of the resulting closedloop dynamics could be involved: the equations were timevarying. the time domain and the Laplace transform domain will be mixed. the equations become nonlinear and time varying. and Vidyasagar [1978] for standard results. We elected to use a notation close to that of Narendra & Valavani [1978]. for example. We will refer to texts such as Desoer & Vidyasagar [1975].1 NOTATION Lower case letters are used to denote scalars or vectors. Without ambiguity.
2 If g. A square matrix A e R nx n is positive semidefinite if xT Ax ~ 0 for all x. then A is symmetric positive semidefinite.3. U is a unitary matrix).B ~ 0. The eigenvalues of a positive semidefinite matrix lie in the closed righthalf plane (RHP). then A +AT is symmetric positive semidefinite. so that this only establishes a partial order on symmetric matrices.oo). 0 I g I P. while those of a positive definite matrix lie in the open RHP. 17). is said to be positive if u ~ 0. since g. In particular.2) Direct from lemma 1. then f e Lp for all p e [ 1.3) an<i we say that u eLp when II uiiP exists.2. Note that a matrix can be neither positive semidefinite nor negative semidefinite. For example.5) When ( 1. and for functions of time.4) t > s and the extended Lp spaces are defined by Lpe = (/I for all s < 00 . Truncated (unctions are defined as II ull fsCt) f(t) 0 t ~ s ( 1. so that we can decompose A as (1. Corollary 1. The matrix A is negative semidefinite if A is positive semidefinite and for symmetric matrices.2.2 Lp SPACES.2. It is positive definite if. we can require xT A x ~ a for all x such that I xI = 1.2. xT Ax ~ axT x = a I x 1 2 for all x. if A ~ 0.andg e Lp.2. It is strictly positive if u > 0. The eigenvalues of a symmetric matrix are all real. We summarize here several facts that will be useful. while H(uv) is fi operating on the product u (t) v(t ).2 Lp Spaces.0 as t .2.0 as t .2. while uiiP = ( J I u(r)l Pdr)1/p 0 1. the notation is used for the Lp norm 00 Proof of Corollary 1.1 Barbalat's Lemma If f(t) is a uniformly continuous function. Lemma 1. Equivalently. p. if I e L 1 n L 00 . u(t) ~ a for all t. we have II u1 ll 00 = supl u(r)l TS. Recall that a scalar u. H(u)v is H(u) multiplied by v in the time domain.2. for some a> 0. or a function of time u (t ). Desoer & Vidyasagar [ 197 5]. We consider real matrices.1 cf.1) where U is the matrix of eigenvectors satisfying ur U = I (that is. Popov [1973] p. Conversely. a bounded function need not belong to L 1 • However. If A ~ 0 and A = A r. oo ). for some a> 0.oo. However. This is not even guaranteed iff is bounded.oo.1. 211.g bounded implies II for p e [ 1. Norms 19 parentheses will determine the sense to be made of an expression. or. and A is a diagonal matrix composed of the . The notation II II will be used to denote the induced norm of an operator. with f = that f is uniformly continuous. we write A ~ B if A . Such a matrix also h!ls n orthogonal eigenvectors. or u (t) ~ 0 for all t.6) For example.I A function f may belong to L 1 and not be bounded.forsomep e [1. oo] (cf. Also.1) Then g(t)O as too. ProofofLemma 1. Positive definite matrices are frequently found in work on adaptive systems. 1.2 (1.!s E Lp) ( 1.0 ( 1.3 POSITIVE DEFINITE MATRICES sup II ull OO = I u(t)l I :?. When p is omitted. e1 does not belong to L 00 . NORMS We denote by I x 1 the absolute value of x if x is a scalar and the euclidean norm of x if x is a vector. f e Lp does not imply that f.2.18 Preliminaries Chapter 1 Section 1.g e L 00 . iJ (u) or iJ u is the output of the LTI system iJ with input u. but e1 e L ooe u E L 00 e. note the following results.2. II A II = I xl sup ~ 1 I A xI (1. in particular the induced matrix norm Then f(t).2. denotes the L 2 norm. such that lim 100 Jf(r)dr 0 I exists and is finite.
From the theory of ordinary differential equations (cf.2) for all XJ.3) 1. that is. with I possibly depending on h. x) = A (t )x for some A(. with h arbitrary.4 • Stability of Dynamic Systems E 21 eigenvalues of A.1 Differential Equations This section is concerned with differential equations d the form . or timeinvariant. x 2 E Bh. then f is Lipschitz with constant /. and non autonomous. IR n. However.4. I does not depend on t. we can make the origin 0 an equilibrium point. 3. Xmax(A) I xl 2 (1. By this. following from ( 1. The constant I is called the Lipschitz constant. then it is continuous in x.4.4. • uniformly. We will always assume that f(t. if true for all x 0 in some ball Bh. it is known that f locally bounded. continuous with respect to t the solution x(t) of ( 1.4. if f has continuous and bounded partial derivatives in x. More formally. then ABalthough not necessarily symmetric. if true for all t 0 ~ 0. and A '12 = uTA'" u (I.2) is the square root matrix of A.IR n x n and nonlinear otherwise. while functions that are Lipschitz in any closed ball satisfy (1.1) satisfies Then ~ If .4. if A. Lipschitz Condition and Consequences The function f is said to be Lipschitz in x if. otherwise.4 STABILITY OF DYNAMIC SYSTEMS ( 1. Proposition 1.4) II A II and.20 Preliminaries Chapter 1 Section 1. • in any closed ball.1 ). so that if II Dz/11 : : . positive semidefinite matrices. This is of great notational help. x) = 0 for all t 0.4.4.3.X= f{t.5) 1. We define by Bh the closed ball of radius h centered at 0 in IRn. x 2 E IRn. positive semidefinite matrices.4. When A ~ 0. f is Lipschitz in x with constant I and is piecewise. if f does not depend on t. and we will assume henceforth that 0 is an equilibrium point of ( 1. and f locally Lipschitz in x imply the existence and uniqueness of the solutions of ( 1.1 ). B are symmetric.3. if true for all x 0 E Bh. when A is positive definite = Xmax<A) (1.4. with A = A '12 ·A '12 and (A '12 )T = A '12 • If A ~ 0 and B ~ 0.3) This simply follows from the fact that XT A X 2 2 and 1z 1 = zT z = I x 1 • We also have that = XT UTA U X = ZT Az (1.1) where x E IRn. if f(t.x) x(to) = Xo (1.4.1) is said to be autonomous.1pact set. This defines locally Lipschitz functions. t ~ 0. It is said to be linear if f(t. On the other hand.) : IR + .1 x = 0 is an equilibrium point of ( 1. then it is Lipschitz.3.3.2) for all XJ. Iff is Lipschitz in x. ] ax1 globally. or timevarying.2) for all Xt. By default. I.1) on some time interval (for as long as x E Bh). we mean that there are only a finite number of discontinuity points in any cor. the square root matrix A'12 is a diagonal matrix composed of the square roots of the eigenvalues of A. or positive semidefinitehas all eigenvalues real positive. x 2 E Bh. Definition Equilibrium Point x is called an equilibrium point of ( 1. is Xmin(A) I xl 2 ::::. The Lipschitz property is by default assumed to be satisfied uniformly. for some h > 0. properties will be true locally. Globally Lipschitz functions satisfy (1. The system defined by (1. we denote Dzf := [ af.x) is piecewise continuous with respect to t. there exists I ~ 0 such that ( 1. Properties will be said to be true: • locally. Another property of symmetric. then A + B ~ 0 but it is not true in general that A · B ~ 0. t ~ 0.4.1 ). Coddington & Levinson [ 19 55]).4. if true for all X o XT A X ::::.4.1 ). By translating the origin to an equilibrium point x 0 .
.4. More precisely.IR.4.).4) as long as x (t) remains in B h.1) starting from x 0 at t 0 • Definition l!niform Stability x = 0 is called a uniformly stable equilibrium point of ( 1. D Proposition 1. I xo I e/(1lo> The other inequality follows similarly from 0/7). the solutions will remain in Bh provided that the initial conditions are sufficiently small.4.12) and there exists a positive function s(t) such that (1..8) Solving ( 1.4.1.. T] Proof of Lemma 1.::: t 0 where x(t) is the solution of ( 1.4.5) I for all t e [0.4. if the trajectory x(t) remains close to 0 if the initial condition x 0 is close to 0. . for all t 0 t > 0.4.4. The proof is similar to the proof of proposition 1.4.4.4. Let T . Proof of Proposition 1.1) if. Or.1 implies that solutions starting inside Bh will remain inside Bh for at least a finite time interval.2 in Appendix. t) such that (t for all t .1 also says that x cannot tend to zero faster than exponentially. f globally Lipschitz implies that x e L 00 e.11) ( 1.1 ). 1. conversely.4.2 Stability Definitions Informally.::: 0. ocan be chosen independent of t 0 • Intuitively.4. and is left to the appendix. +.4.8) I for all t e [0. this definition captures the notion that the equilibrium point is not getting progressively less stable with time. T] llxl ~ dt d I I lxl ~ llxl ( 1.4.::: lx(t)l .).4.2 BellmanGronwall Lemma Let x(.4.4 Stability of Dynamic Systems 23 lxol /(llo).1 Note that Lemma 1.4.22 Preliminaries Chapter 1 e/(1lo) Section 1.4.::: 0 and . Proposition 1. 7) x(t) ~ u(O)e [ a(u)du I J a(u)du +I u(r)e' 0 dr (1. : : lxol (1. Stability is a very mild requirement for an equilibrium point.. given a time interval.) : IR. x = 0 is a stable equilibrium point.) is differentiable for all t e [0. it does not Ix(t)l >  lxol e/(llo) +I eI(ITls(r)dr Yt. there exists o 0. if. in the preceding definition. a(. The following lemma is an important result generalizing the wellknown BellmanGronwall lemma (Bellman [ 1943]).10) ~~lxl 2 1 so that 2lxll ~ lxll Then ( 1. u (. Also. In particular. If I I xl 2 = xT x implies that = x(t) ~ Ia(r)x(r)dr + u(t) 0 (1.9) Definition Stability in the Sense of Lyapunov x = 0 is called a stable equilibrium point of ( 1. + . we say (1. T] 2j xT ~ xI ~ 21 xl I ~ xI I x(t) ~I a(r)u(r)e' 0 Ja(u)du dr + u(t) (1.6) Since f is Lipschitz When u(.4.
1 ). To make this notion precise.1) without explicitly integrating the differential equation. such lxol < Global exponential stability means that ( 1. p. x 0 ) : IR+ x IR"IR+.1 ). Definition Positive Definite Functions A continuous function v(t.4. then the system will tend to its equilibrium. in the texts by Hahn [ 1967] and Vidyasagar [ 1978].4. is defined likewise.p.) equilibrium point of ( 1.s. Definition Asymptotic Stability x = 0 is called an asymptotically stable equilibrium point of ( 1.Xo) forallt ~. The constant a is called the rate of conver (a) (b) x x = = 0 is a stable equilibrium point of ( 1.4. Note that the speed of convergence is not quantified in the definitions of asymptotic stability. energy functions.O) = 0 x 0 e IR".p.) e K v(t.4. Global u. if it is asymptotically stable and lim lx(t)l = 0. but it is not true in general. we need to define exactly what we mean by a "measure of energy.1) if there exist m.4. is locally like an "energy function.a. Global asymptotic stability is defined as follows. for all A continuous function v(t.) e K).f) .) x = 0 is called a uniformly asymptotically stable (u.d.2). and some a(.1 ). E Bh.d. a > 0 such that the solution x (t) satisfies and v(t. there exists o(t 0 ).s. x) : IR+ x IR" .x) ~ a(lxl) An l. for some h > 0. More precisely.a. for all t 0 ~ 0. if (a) x = 0 is a uniformly stable equilibrium point of ( 1.to ~0 The previous definitions are local. A more complete development is available.a.f. if for some a(. 7).13) is satisfied for any x 0 E IR"." that is.5. Exponential stability in any closed ball is similar except that m and a may be functions of h.d. that is. Definition Global Asymptotic Stability x = 0 is called a globally asymptotically stable equilibrium point of (1. x) : IR + x IR" . if for all x 0 gence.s. Definition Class K Functions A function a (f) : IR + .) e K v(t.1).24 Preliminaries Chapter 1 Section 1.4. 0 is attractive. for instance.4.d.x 0 ) = 0 forallx 0 and lxol <o ==> lx(t)l :5: :~t~to. It will be shown that uniform asymptotic stability is equivalent to exponential stability for linear systems (see Section 1. strictly increasing.3 Lyapunov Stability Theory We now review some of the key concepts and results of Lyapunov stability theory for ordinary differential equations of the form ( 1. Definition Locally Positive Definite Functions that o ==:: lim lx(t)l 1+00 = 0 Definition Uniform Asymptotic Stability (u.) and are defined as follows. if it is continuous.f.IR + belongs to class K (denoted a(. such that lim 'Y(r.IR+ is called a positive definite function (p.IR + is called a locally positive definite function (l." Functions which are globally like "energy functions" are called positive definite functions (p. (b) the trajectory x(t) converges to 0 uniformly in t 0 . we first define class K functions (Hahn [ 1967].1 ).O)=O (1. In the following definition.4. t ~ t0 ~ 0.x)~a(ixl) for all X E IR n' t ~ 0 . The socalled Lyapunov second method enables one to determine the nature of stability of an equilibrium point of ( 1. For this. Exponential stability is assumed to be uniform with respect to t 0. and a (0) = 0. The method is basically a generalization of the idea that if some "measure of the energy" associated with a system is decreasing. Definition Exponential Stability.4. the convergence to zero is required to be at least exponential. since they concern neighborhoods of the equilibrium point.4 Stability of Dynamic Systems 25 require that trajectories starting close to the origin tend to the origin asymptotically. there exists o> 0 and a function 'Y (r.4.4. 1.f) if. That property is made precise in the following definition.13) and v(t.1 ). Rate of Convergence x = 0 is called an exponentially stable equilibrium point of ( 1.
26
F ·eliminaries
Chapter 1
Section 1.4
Conditions on v(t, x) l.p.d.f. l.p.d.f., decrescent J.p. d. f. l.p.d.f., decrescent p.d.f., decrescent
Stability of Dynamic Systems
Conditions on  v(t, x) ;:::: 0 locally ;:::: 0 locally l.p.d.f. l.p.d.f. p.d.f.
27
and the function a (p) oo as p  oo . In the definitions of l.p.d.f. and p.d.f. functions, the energy like functions are not bounded from above as t varies. This follows in the next definition. Definition Decrescent Function The function v(t ,x) is called decrescent, if there exists a function {3(.) e K, such that
Conclusions stable uniformly stable asymptotically stable uniformly asymptotically stable globally u.a.s
v(t,x)
~ {3(
II xll)
Proof of Theorem 1.4.3 cf. Vidyasagar [ 1978], p. 148 and after. Example We refer to Vidyasagar [ 1978] for examples of application of the theorems. An interesting example is the second order system
Examples Following are several examples of functions, and their membership in the various classes: v(t ,x) = Ixl 2 : p.d.f., decrescent
Xt
Xt(Xt+x}1)x2
(1.4.15)
v(t ,x) = xT P x, with P > 0: p.d.f., decrescent v(t,x) v(t, x) v(t,x)
= = =
(t + 1)lxl 2 : p.d.f. e 1 I xl 2 : decrescent sin 2 (1xl 2 ): l.p.d.f., decrescent
with the p.d.f.
v (xt>x2) = x[ +xi
The derivative of v is given by
(1.4.16)
Lyapunov Stability Theorems Generally speaking, the theorems state that when v(t, x) is a p.d.f., or an l.p.d.f., and dv I dt (t, x) ~ 0, then we can conclude the stability of the equilibrium point. The derivative of v is taken along the trajectories of ( 1.4.1 ); that is,
(1.4.17) so that  v is an J.p. d. f. (in Bh, where h < 1), establishing the local asymptotic stability of the origin. The origin is, in fact, not globally asymptotically stable: the system can be shown to have a circular limit cycle of radius I (by changing to polar coordinates). From ( 1.4.17), we can also show that for all x e Bh
dv(t,x) dt
I
=
(1. 4. 1>
av(t,x) + av(t,x) f(t,x) at ax
(1.4.14)
Theorem 1.4.3 Basic Theorems of Lyapunov Let v (t, x) be continuously differentiable: Then
v~
so that
2(!h 2 )v
(1.4.18)
v(t) ~ v(O)e 2 (1h 2 ) 1
lx(t)l ~ lx(O)I eOh
2 1 )
forallt;o::O
(1.4.19)
and, in fact, x = 0 is a locally exponentially stable equilibrium point. Note that the Lyapunov function is in fact the squared norm of the state, a situation that will often occur in the sequel.
28
Preliminaries
Chapter 1
Section 1.5
Exponential Stability Theorems
29
Comments The theorems of Lyapunov give sufficient conditions guaranteeing the stability of the system ( 1.4.1 ). It is a remarkable fact that the converse of theorem 1.4.3 is also true: for example, if an equilibrium point is stable, there exists an l.p.d.f. v(t, x) with v(t, x) :s; 0. The usefulness of theorem 1.4.3 and its converse is limited by the fact that there is no general (and computationally nonintensive) prescription for generating the Lyapunov functions. A significant exception to this concerns exponentially stable systems, which are the topic of the following section. 1.5 EXPONENTIAL STABILITY THEOREMS We will pay special attention to exponential stability for two reasons. When considering the convergence of adaptive algorithms, exponential stability means convergence, and the rate of convergence is a useful measure of how fast estimates converge to their nominal values. In Chapter 5, we will also observe that exponentially stable systems possess at least some tolerance to perturbations, and are therefore desirable in engineering applications. 1.5.1 Exponential Stability of Nonlinear Systems The following theorem will be useful in proving several results and relates exponential stability to the existence of a specific Lyapunov function. Theorem 1.5.1 Converse Theorem of Lyapunov
Comments Again, the derivative in (1.5.3) is a derivative taken along the trajectories of ( 1.5 .1 ), that is dv(t ,x) = av(t ,X) + av(t ,X) f(t ,x) (1.5.5) dt (I.s. 1> at ax This means that we consider x to be a function of t to calculate the derivative along the trajectories of ( 1. 5.1) passing through x at t. It does not require of x to be the solution x(t) of ( 1.5.1) starting at x(t 0 ). Theorem 1.5.1 can be found in Krasovskii [1963] p. 60, and Hahn [ 1967] p. 273. It is known as one of the converse theorems. The proof of the theorem is constructive: it provides an explicit Lyapunov function v (t , x ). This is a rather unusual circumstance, and makes the theorem particularly valuable. In the proof, we derive explicit values of the constants involved in ( 1.5.2)( 1.5.4). Proof of Theorem 1.5.1 (a) implies (b). (i) Denote by p(r, x, t) the solution at timer of (1.5.1) starting at x(t), t , and define
t+T
I
v (t ,
X)
=
J p (r , X , t) 2 dr
I 1
( 1.5.6)
I
Assume that /(t, x) : IR + x IR n  IR n has continuous and bounded first partial derivatives in x and is piecewise continuous in t for all x E Bh, t ~ 0. Then, the following statements are equivalent: (a) x = 0 is an exponentially stable equilibrium point of
where T > 0 will be defined in (ii). From the exponential stability and the Lipschitz condition m
lxl
ea(rt) ~ lp(r,x,t)l ~
lxl
el(rt)
( 1.5. 7)
and inequality ( 1.5.2) follows with
a 1 := (1e 21 T)/2l
.X = f(t ,x)
(b)
x(to)
= Xo
(1.5.1)
a 2 := m 2 (1e 2aT)/2a
( 1.5.8)
There exists a function v (t , x ), and some strictly positive constants h', a 1,a2,a3.a4, such that, for all x E Bh', t~O ( 1.5.2)
(ii) Differentiating ( 1.5.6) with respect to t, we obtain dv (t ,x) dt
I
p (t + T 'X' t) 1 2
I+T

I
p (t 'X' t) 1
X , t)
2
+
dv(t,x)l dt
J dt ( l,p(
I
d
T ,
I 2 ) dr
( 1.5.9)
:S:a3lx12
(1.5.1)
(1.5.3)
Note that d I dt is a derivative with respect to the initial time t and is taken along the trajectories of ( 1. 5.1) By definition of the solution p, p (T
'
( 1.5.4)
X (t
+ At) ' t + At) = p (T
'
X
(t ) ' t )
(1.5.10)
30
Preliminaries
Chapter 1
Section 1.5
Exponential Stability Theorems
31
for all ~ t, so that the term in the integral is identically zero over [t, t + T]. The second term in the righthand side of ( 1.5.9) is simply , 1xj 2 while the first is related to I xl 2 by the assumption of exponential stability. It follows that
and, using the exponential stability again, ( 1.5.14) becomes
Iav~;x) I : ; 2 J
I
I+ T
m
Jxl
e<ka)(TI)dT
(1.5.18)
dv~;x)
which is ( l.5 .4) if we define
::;; (lm2e2aT)Ixl2
(1.5.11)
a4
:= 2m(e<ka)T_l)l(ka)
(1.5.19)
Inequality (1.5.3) follows, provided that T > (I I a) In m and
..
av(t,X) ax;
=
a3
:= l  m2e2aT
X;,
(1.5.12)
Note that the function v (t ,x) is only defined for x E Bh' with h' = hIm, if we wish to guarantee that p (T , x , t) E Bh for all T 2. t .
(b) implies (a)
(iii) Differentiating ( 1.5.6) with respect to
we have
(1.5.13)
21r
I
This direction is straightforward, using only (1.5.2)(1.5.3), and we find
(1.5.20)
0
±pj(T,X,t)apj(T,X,t) dT }=1 ax;
Under the assumptions, the partial derivative of the solution with respect to the initial conditions satisfies
[fj(T ,p(T ,X,!)))
k=l axk
±aJ I
1
(1.5.14)
T,P(T,X,I)
(except possibly at points of discontinuity of j(T ,x )). Denote
Qij(T,X,t) := ap;(T,X,t)lax1 AiJ(x,t) aj;(t,x);ax1
(1.5.15)
Comments The Lyapunov function v (t , x) can be interpreted as an average of the squared norm of the state along the solutions of ( 1.5.1 ). This approach is actually the basis of exact proofs of exponential convergence presented in Sections 2.5 and 2.6 for identification algorithms. On the other hand, the approximate proofs presented in Chapter 4 rely on methods for averaging the differential system itself. Then the norm squared of the state itself becomes a Lyapunov function, from which the exponential convergence can be deduced. Theorem 1.5.1 is mostly useful to establish the existence of the Lyapunov function corresponding to exponentially stable systems. To establish exponential stability from a Lyapunov function, the following theorem will be more appropriate. Again, the derivative is to be taken along the trajectories of ( 1.5.1 ). Theorem I .5.2 Exponential Stability Theorem If there exists a function v (t , x }, and strictly positive constants a I> a2, a3, and o, such that for all x E Bh, t 2. 0
so that ( 1.5 .14) becomes
dT Q (T , X , t)
d
=
A {p (T , X , t ) , T ) · Q (T , X , t)
( 1.5.16)
T
Equation ( l. 5.16) defines Q (T , x , t ), when integrated from
=
t to
T= t+T, with initial conditions Q(t,x,t)=l. Thus, Q(T,x,t) is the state transition matrix (cf. Section 1.5.2) associated with the time varying matrix A (p (T ,x, t), T). By assumption, II A(. , .) II ::;; k for some k,
a1lxl 2
dt
::;; v(t,x)::;;
a2lxl 2
(1.5.21) ( 1.5.22)
d v(t,x(t))
so that
'
I :; 0
(1.5.1)
II
Q(T,x,t)ll ::;; ek<T1)
(1.5.17)
32
t+o
Preliminaries
Chapter 1
Section 1.5
Exponential Stability Theorems
33
j
t
d
dv(T,X(T))I
T
Definition StateTransition Matrix
dT :>; a 3 1x(t)l 2
(1.5.1)
( 1.5.23)
The statetransition matrix .P (t, t 0) e IR n xn associated with A (t) is, by definition, the unique solution of the matrix differential equation
dt(.P(t,to))
Then
x(t) converges exponentially to 0.
d
=
A(t)<P(t,to)
<P(to, to) = I
( 1.5.32)
Proof of Theorem 1.5.2 From (1.5.23)
v(t,x(t))v(t+o,x(t+o)) ~ (a 3 /a 2 )v(t,x(t))
(1.5.24)
for all t
~
0, so that
v(t + o, x(t + o)) :>; (1 a 3 /a 2) v(t,x(t))
Note that linear systems with A(.) e L ooe automatically satisfy the Lipschitz condition over any finite interval, so that the solutions of ( 1.5.31) and ( 1.5.32) are unique on any time interval. It is easy to verify that the solution of(1.5.31) is related to that of(1.5.32) through
x(t)
=
( 1.5.25)
<P(t, t 0 )x(to)
( 1.5.33)
for all t
~
0. From (1.5.22)
v (tt>x(t 1)) :>; v (t ,x(t))
for all t 1 e [t, t + o]
(1.5.26)
Choose fort the sequence t 0 , t 0 +o, t 0 +2o, ... so that v(t,x(t)) is bounded by a staircase v(t 0,x(t 0)), v(t 0 +o,x(t 0 +o)), ... where the steps are related in geometric progression through (1.5.24). It follows that
v(t ,x(t)) :>; mve
a(tto)
In particular, this expression shows that trajectories are "proportional" to the size of the initial conditions, so that local and global properties of LTV systems are identical. The statetransition matrix satisfies the socalled semigroup property (cf. Kailath [1980], p. 599):
<P (t , to)
=
<P (t , T) <P (T , t 0)
for all t
~ T ~
t0
( 1.5.34)
and its inverse is given by ( 1.5.35) A consequence is that
•
v(to,x(to))
( 1.5.27)
for all t
~
t0
~
0, where (1.5.28)
d <P(t, to) to
d
d (<P(to, t)) 1 d to
Similarly, (1.5.29) where
 <P (to, t)  1A (to) <P (t 0, t ) <P (t 0 , t )  <P (t , to) A Uo)
1
( 1.5.36)
m
0
a
1 = lb
The following propositions relate the stability properties of ( 1.5.31) to properties of the statetransition matrix.
I [
n
1 ] 1 a3/a2
( 1.5.30) Proposition 1.5.3 Uniform Asymptotic Stability of LTV Systems x = 0 is a uniformly asymptotically stable equilibrium point of ( 1.5.31) if and only if x = 0 is stable, which is guaranteed by sup ( sup II <P (t, to) II ) < oo
to:<:: 0
1.5.2 Exponential Stability of Linear TimeVarying Systems We now restrict our attention to linear timevarying (LTV) systems of the form
(1.5.37)
I :<:: 10
.X
where A (t)
Looe'
E
= A(t)x
x(to) = Xo
(1.5.31)
and x = 0 is attractive, which is gttaranteed by
IR n X n is a piecewise continuous function belonging to
II<P(t,to)IJ 0 as too
uniformly in t 0.
(
1.5.38)
t 0 + o) E IR.41) For all t ~ t 0 .5.4 direct from the expression of the solution ( 1.to+o) ~ {jii (1. some additional results on the stability of linear timevarying systems will now be established.5.5.nxn. x(t 0 ) can be reconstructed from the knowledge of y(. where x (t) is the solution of ( 1.n.3 and proposition 1.5.) using the expression to+6 x(to) < M [.33).5. Proposition 1. t 0 + (n + 1) T].5. while A(t) e IR.44) fails to be satisfied. We now show the converse.43) starting at x(t 0). A unique property of linear systems is the equivalence between uniform asymptotic stability and exponential stability.mxn.to+nT)II II <t>(to+nT.4.45) to Comments Note that.5.47) ( 1. We consider the linear timevarying system [C(t).5.39).5.5.A(t)] is called uniformly completely observable (UCO) if there exist strictly positive constants {j~tfh5. M [.5.5.5. Proof of Proposition 1. C(t) e IR. n X n is the socalled observability grammian t0 + 6 if and only if x = 0 is an exponentially stable equilibrium point of (1.46) is satisfied for all x(t 0).33). Proof of Proposition 1.46) 1 II <I>(to + T.5 That exponential stability implies uniform asymptotic stability is obvious from their definitions and in particular from proposition 1. Using the semigroup property recursively.BII x(to)l 2 (1. using (1. Definition Uniform Complete Observability (UCO) The system [C(t). Proposition 1.B2I x(to)l 2 ~ J I C(T)X(T) to 1 2 dT ~ . t 0 ~ 0.t 0 +5) = J <I>T(T.43) is satisfied uniformly for all to and complete because ( 1. A specific x(t 0) is observable on a specific time interval [t 0.t 0 +5).3 direct from the expression of the solution ( 1.5 Exponential and Uniform Asymptotic Stability where x(t) E IR. t 0 + o] if condition ( 1.42) which can easily be expressed in the form of (1.44) x = 0 is a uniformly asymptotically stable equilibrium point of ( 1.] n ::. to)ll < M for all t ~to~ 0 (1.t0)CT(T)y(T)dT to (1.40) .5.5.5.34 Preliminaries Chapter 1 Section 1.5 Exponential Stability Theorems 35 Proof of Proposition 1.5.t 0)dT (1.m. (1. if condition (1.5. Proposition 1.43) forallt ~t 0 ~0.to)ll :::: II <t>(t. (1.1 J <I>T(T. 0 On the other hand. such that.5.a(t to) ( 1. to)ll = me. there exists an integer n such that t e [to+ n T.5. such that N(t 0.46) shows that there exists an x(t 0) "# 0 such that the corresponding output • . n.t 0)CT(T)C(T)<l>(T.to)ll for all x(t 0) e IR.5. to)ll < 2 for all t 0 ~ 0 (1.5.41) II <t>(t. condition (1. rtt )/T 0 = N(to.5.5.31) if and only if for some m. as stated in the following theorem.5.5. for all t 0 ~ 0 {j2/ ~ N(to.5.31) where N (to. The observability is called uniform because ( 1.5.44) can be rewritten as t0 +6 II and <I>(t.40).5.46) is satisfied on that interval. are piecewise continuous functions (therefore belonging to L 00 e).4 Exponential Stability of LTV Systems x = 0 is an exponentially stable equilibrium point of ( 1. a > 0 Uniform Complete ObservabilityDefinition and Results Through the definition of uniform complete observability.5.33).39) y(t) ( 1.31). Then. y(t) E IR.3 implies that there exists M > 0 and T > 0.5.5.5.A(t)] defined by x(t) A(t)x(t) C(t)x(t) II <I> (t. together with (1..
trivial. we did not impose initial conditions in (1. To .48) show the upper inequality. t) :::.49) and (1.2ao I (1. t ) 1+0 ':t ~ [ I t +o J~ 20 T (T . (b) For all C(t) e IRm xn (with m arbitrary) such that the pair [C{t). Since the integral from t to oo is not less than from t to t + o.5. [ + 0) dr l~(t + 0 . This equation is a linear differential equation.51).5.49) is satisfied so that P(t) is well defined. t ) dr = ~ T(t + o.5.5. t + 26] r + 2o so that x (t 0 ) is not distinguishable from 0.5.'Y2I (b) implies (c) (c) implies (a).5.50) .56) Using the UCO property I+ I vdr I o +a {32I ~ J ~T(r.55) Define 00 so that .57) for all t ~ 0. ( 1.36 satisfies Preliminaries Chapter 1 Section 1.t)dr I (1. and using (1.5 Exponential Stability Theorems 37 t0 + o to I !y(r)i 2dr 0 ( 1.5. :::. {3 2 m 2e. t ) dr I 0.5. f32(1+m2e2ao+m2e4ao+···)I .5.49) ( 1.t)dr I ~ {31I ( 1. we divide the interval of integration in intervals of size o and sum the individual integrals.xT(t)A T(t)P(t)x(t) . the lower inequality in ( 1.5. Consider the Lyapunov function v(t . t + 0) C T ( T) C (T) ~( T ..51). but first note that by differentiating (1. so that the solution is unique for given initial conditions.50).52) Exponential convergence follows fr'om theorem 1. such ti. On the interval [t + 0. t ) C T( r) C ( r) ~ (r .xT(t)P(t)A(t)x(t). P(t) = I ~T(r.X) = XT(t)P(t)x(t) Proof of Theorem 1.53) where we used the UCO and exponential stability assumptions.36).51) We will show that (1. (1.v .P(t) for all t (c) ~ A T(t)P(t) + P(t)A (t) + CT(t) C(t) I ~ T(r . t ) C T(r) C (r) ~ (r . there exists a symmetric P(t) e IRnxn and some 'YI> ')' 2 >0. However.at (1.5.5.5.5.52).50).6 Exponential Stability of LTV Systems The following statements are equivalent: (a) x = 0 is an exponentially stable equilibrium point of ( 1.50) are satisfied.54) For some C(t) e IRmxn (with m arbitrary) such that the pair [C(t).t)CT(r)C(r)~(r.2. such that 'Y2I ~ P(t) ~ 'YI I I ~ T(r . there exists a symmetric P(t) e IRnxn. Therefore 00 (1.5. it follows that P(t) satisfies the linear differential equation (1.5. By the UCO assumption t (o (1. Theorem 1.5.5.6 (a) implies (b) (1.t)CT(r)C(r)~(r.5.5. and some 'YI> ')' 2 > 0.5.5. and P(O) is in fact given by (1.5.A(t)] is UCO.50) follows directly from ( 1.A{t)] is UCO.31 ).5.xT(t)P(t)x(t) xT(t) CT(t) C(t)x(t) ).
we provide them in this section. Proof of Theorem 1.7 Lyapunov Lemma The following statements are equivalent: (a) All eigenvalues of A lie in the open lefthalf plane (b) x = 0 is an exponentially stable equilibrium point of (1. Vidyasagar [1978]. and the upper bound of = I eA T I (I .7 (a) implies (b).A] is observable.5. Using both expressions in (1. m x n (with m arbitrary). This follows from the fact that the transition matrix is given by the exponential matrix (cf.5. they become "asymptotically" timeinvariant.5.5.5.62) X =A X x(to) = Xo ( 1. when C integration tends to infinity. Boyd & Sastry [1983] and [1986] illustrated its application to adaptive systems.3 Exponential Stability of Linear Time Invariant Systems When the system ( 1.60) (b) implies (c).5. (d) implies (b) as in theorem 1. all properties are uniform (in t 0). S(t)>O for all t >0.to) = eA(IIo) ( 1.5 .A] is observable. (d) For some C e JR.59). 171) ~(t . since ejwt is an eigenfunction of such systems.5.5. m x n (with m arbitrary). but. Let S(t) I eArrcrceArdr 0 I . Specifically. such that the pair [ C . A] is observable. 1.n x n satisfying ( 1. there exists a symmetric positive definite P e JR. local and global stability properties are identical.64) (c) implies (d) trivial.6 GENERALIZED HARMONIC ANALYSIS An appropriate tool for the study of stable linear timeinvariant systems is Fourier transforms or harmonic analysis.u) c T c eA (I .38 Preliminaries Chapter 1 Section 1.5. we consider the system Clearly S(t)=ST(t).58) (c) where . 0 = I. Since the proofs of the yarious lemmas used in this book are neither difficult nor long. Theorem 1. We stated a more general version of the Lyapunov lemma partly to show the intimate connection between exponential stability and observability for linear systems. We will be concerned with the analysis of algorithms for identification and adaptive control of linear timeinvariant systems.5.5. For these systems.61) S(t) = eATicTceAI = ATS(t)+S(t)A +CTc (1. such that the pair [C .u) d/ (j'"' (1.5.5 Exponential Stability Theorems 39 It is interesting to note that the Lyapunov function is identical to the function used in theorem 1. even further simplification of the stability criterion results.nxn satisfying(J.61) 0 1. The theory has been known since the early part of the century and was developed in Wiener's Generalized Harmonic Analysis (Wiener [1930]).58) and using the exponential stability of A lim S(t) 1+00 = 0 = AT p + p A + cr c (1.65) Then. and since [C.P lim S(t) 1+00 (1.5. and since it is timeinvariant. under certain conditions.5.1. pp.31) is timeinvariant. and leading to the socalled Lyapunov equation: ( 1. generalized harmonic analysis is a useful tool of analysis. 0 For all C e JR.6. The overall adaptive systems are timevarying systems.5. there exists a symmetric positive definite p e JR.5.63) Note that since the system is linear.59) Remark The original version of the Lyapunov lemma is stated with Q positive definite replacing cr C (which is only semidefinite). the observability hypothesis is trivial.
w0 • If u (t) is periodic. Since Ru is real.6.. .6.6. while Im (Su) is an antisymmetric matrix.6. For example. ~ k cr< ~ J u(u)uT(u+t1 t. for all t 1 . that is.1 Define the scalar valued function v (t) by k v (t) : = ~ cr u (t + t. There is a strong analogy between ( 1.40 Preliminaries Chapter 1 Section 1.1) lo in which instance. On the other hand. and Im ( Su) is an odd function of w. . S[(dw) = s:(dw) . we find that (1.6. for example.6. In particular..6.9) also shows that Re (Su) is a symmetric matrix. and C1.)du)c1 I.nxn (1.6. w) (w From (1. and .. its spectral measure has point masses concentrated at countably many points. In other words.n k and 00 J Su(dw) = 211' Ru(O) < ( 1. . Proposition 1.) i =I ( 1.6.wo. and is equal to R~loch (t) for almost all w. 2wo. Widder [ 1971]) that Ru is the Fourier transform of a positive semidefinite matrix Su(dw) of bounded measures. the limit of ( 1.. . for a wide sense stationary ergodic process u(t.6.+ T ( 1.1) shows that Ru (t) = R[ (. so that (1. that is.6.6.6) lim Too ~ to+ T I J u(r)uT(t +r) dr e IR.j ~ ( 1. Ru(t) = 1 11' 2 00 f eiwt Su(dw) 00 (1.6. it follows (see.t) 0 :::.6. .6.2) follows.6) is positive. .t.) Cj ~ 0 i. with period 21r/w 0.) CJ i .9) oo ~ cr Ru (tJ . .2w 0 . then u is said to have a spectral line at frequency w0 • Su(dw) has point mass (a delta function) at w0 and . consider the scalar signal .6. Ck e IR. ( 1. when u is a wide sense stationary stochastic process.)uT(r+tj)dr)Cj J l. T k JI v (r) I 0 2 dr T (1.6.J = I Equations ( 1.5) as Too is k ~ cr Ru(tj. ~ cr( T (u(r+t.J =I b Therefore.6. then Su(dw) has point mass at wo.4) .6.6 Generalized Harmonic Analysis 41 Definition Stationarity. .7) oo here denotes a sample point of the underlying probability space). Indeed. uniformly in t 0 Ru(t) := Since u has an autocovariance. the limit Ru(t) is called the autocovariance of u. if u has a sinusoidal component at frequency w0 .1) and R~loch(t) = E [u(r)uT(t +r)].1) and ( 1.4). if u (t) is almost periodic.IR n is said to be stationary if the following limit exists.1) is completely deterministic. But we emphasize that the autocovariance defined in ( 1.6.9) shows that Re ( Su) is an even function of w.t. .9) give a technique for obtaining the spectral content of stationary deterministic signals. ( 1. ( 1. 0 From proposition 1.1.1 Ru is a positive semidefinite matrixvalued function.• +oo Su(dw) = f eiwr Ru(r)dr ( 1.3) Then. the autocovariance Ru(t.2) I Proof of Proposition 1. • Further.6. The concept of autocovariance is well known in the theory of stochastic systems. I. for all T > 0 I T The matrix measure Su(dw) is referred to as the spectral measure of the signal u.• tk E IR. Autocovariance A signal u: IR + .6.j = ( 1.8) oo The corresponding inverse transform is .6. w) exists.5) l.
6.14) exists absolutely.6.6.12). by considering if •(Jw) Su(dw) if T(jw) (1.13) Note that ( 1.6. with autocovariance 00 ::.lRm be stationary.6. to obtain Jf H(Tt)dTI [ ~ lor 1 +T L 1 u(u)uT(t +u+T 1 T 2 )du (HT(T2)dT2) l Note.6.18) 0 ~ to+ T to J y(T)yT(t + T)dT ~ 10 + T J (JH (TI) u (T .11) The expression in parenthesis converges to Ru(t + T1 .2 We first establish that y has an autocovariance.6.jw)Su(dw)H (jw) A Proof of Proposition 1. as Too .6 Generalized Harmonic Analysis 43 (1. using the fact that H (t) is real Sy(dw) AT H(. sup T to.6.T2)HT(T2)dTI dT2 00 (1. to ( 1. (j UT(t +TT 2)HT(T2)dT2)dT For all T.r 1 u(u)uT(t+u+TtT2)dul t0 + T showing spectral content at w 1 and w2• The autocovariance of the input and of the output signals of a stable linear timeinvariant system can be related as in the following proposition. we substitute (1. Definition Cross Correlation Let u : lR + IRn and y : lR + . the integral in (1.16) Hence.15) .7) in (1. If u is stationary. that if if(s) has a zero of transmission at w0 . and the intensity of the spectral line of y at w0 is given by ( 1.2 Linear Filter Lemma Let y = if(u). with real impulse response H(t).6.18) .TI) dT t) to Remark (1.14) It is easy to see that if u has a spectral line at frequency w0 .42 Preliminaries u(t) = sin(w 1 t) + sin(w2t) + f(t) Chapter 1 Section 1. with autocovariance Ru(t) Then y is stationary. stable m x n matrix transfer function. it is bounded as a function ofT. so that. since if is stable..6.oo.6. t 0 . We will also need to define the cross correlation and cross spectral density between two stationary signals. uniform in t 0 10 (1. since. we may change the order of integration. To complete the proof.6. and therefore H e L 1• Therefore.6.6. T 1 J I u(T)i 2dT to (1. . The signal u has autocovariance (1.12) (1.17) is the Fourier representation of Ry. so does y.6.6. by dominated convergence. Proposition 1. ( 1.6.12) to get Ry(t) = 21 71" 2 71" JJH(TI)dTI Jejw(l +rlrl) Su(dw)HT(T2)dT2 Jejwt ( J~"'rlH(T 1 ) dTI )Su(dw) ( Je. The cross correlation between u andy is defined to be the following limit. so that y has an autocovariance given by ( 1. however. T 1 and T 2 . then the amplitude of the spectral line at the output is zero. uniformly in t 0 . where if is a proper.12). Further.17) oooo and spectral measure (1.10) where f(t) is some continuous function that tends to zero as t .T 2) as Too.jwr Hh) dT2 )T 2 Ry(t) = J JH(TI )Ru(t + TJ.6.6.15) converges uniformly in t 0 . by the Schwarz inequality ~~ t 0 r 1 +T J t0 .
3 Linear Filter LemmaCross Correlation Let y = H(u).2. i. The cross correlation between the input and the output of a stable LTI system can be related in much the same way as in proposition 1.2) The steadystate response is then given by 1t) where = ~ sin(w 0 t + c/J) (2.mxn lim T.1) The parameters kp and ap are unknown and are to be determined by the identification scheme on the basis of measurements of the input and output of the plant. Frequency Domain Approach A standard approach to identification is the frequency response approach. To introduce the subject.6.6.3 The proof is analogous to the proof of proposition 1.6. linear time invariant (LTI) systems.e. Let the input r be a sinusoid r(t) = sin (w 0 t) (2.t) ( 1. aP > 0.3) 45 . The plant is assumed to be stable.0. we first informally discuss a simple example.20) CHAPTER 2 IDENTIFICATION and the Fourier transform of the cross correlation is referred to as the cross spectral measure of u and y.6.6.6. Proposition 1.21) 00 and the cross spectral measure is Syu(dw) H*(jw)Su(dw) ( 1.0. We consider the identification problem for a first order SISO LTI system described by a transfer function Yp(s) f(s) Proof of Proposition 1.0. If u is stationary Then the cross correlation between u andy is given by 00 Ryu(t) = I H(TJ )Ru(t + = TJ )dTJ (1.6.6. Pes) _!:g_ s + ap (2. with impulse response H(t).2 and is omitted here.19) It may be verified that the relationship between Ruy and Ryu is Ryu(t) R.oo ~ to+ T to I y(T) uT(t + T) dT = E (1. where fi is a proper stable m x n matrix transfer function.44 Ryu(t) := Preliminaries Chapter 1 IR.22) 2. we review some identification methods for singleinput singleoutput (SISO).{y(.0 INTRODUCTION In this chapter.
20) CHAPTER 2 IDENTIFICATION and the Fourier transform of the cross correlation is referred to as the cross spectral measure of u and y.6. Let the input r be a sinusoid r(t) = sin (w 0 t) (2.2 and is omitted here. we first informally discuss a simple example.6. The cross correlation between the input and the output of a stable LTI system can be related in much the same way as in proposition 1.mxn lim T.{y(.0.t) ( 1.oo ~ to+ T to I y(T) uT(t + T) dT = E (1.2) The steadystate response is then given by 1t) where = ~ sin(w 0 t + c/J) (2.3 The proof is analogous to the proof of proposition 1. i.6.6.1) The parameters kp and ap are unknown and are to be determined by the identification scheme on the basis of measurements of the input and output of the plant.6. linear time invariant (LTI) systems.22) 2.0 INTRODUCTION In this chapter. Proposition 1. where fi is a proper stable m x n matrix transfer function. Pes) _!:g_ s + ap (2. We consider the identification problem for a first order SISO LTI system described by a transfer function Yp(s) f(s) Proof of Proposition 1. we review some identification methods for singleinput singleoutput (SISO).6.0.6.19) It may be verified that the relationship between Ruy and Ryu is Ryu(t) R.6. aP > 0.e.3) 45 . with impulse response H(t). The plant is assumed to be stable. If u is stationary Then the cross correlation between u andy is given by 00 Ryu(t) = I H(TJ )Ru(t + = TJ )dTJ (1. Frequency Domain Approach A standard approach to identification is the frequency response approach.0.2. To introduce the subject.44 Ryu(t) := Preliminaries Chapter 1 IR.21) 00 and the cross spectral measure is Syu(dw) H*(jw)Su(dw) ( 1.3 Linear Filter LemmaCross Correlation Let y = H(u).
0. because our goal is adaptive control. However. in the time domain r 1 S+A A (2.11) into the standard framework used later in this chapter.0.0. we transform (2. However.0.4) or.46 m Identification Chapter 2 Section 2. Letting w(t) := w<1l(t)] [ w<2l(t) I (2. Nonzero initial conditions would only contribute exponential decaying terms with rate A (arbitrary). But first. in the time domain. We will therefore look for algorithms that use the complete information and preferably update esti·mates only on the basis 'o~ new data.1) and divide both sides by s +A for some A> 0 (2. that is when the input signal is constant. giving the DC gain. O(t) is the estimate of o• based on inputoutput data up to time t.. and kp· As few as two time instants may be sufficient to determine the unknown parameters from w<1l = . w< 2l may be obtained by stable filtering of the input and of the output of the plant.0 Introduction 47 kP A Y wfi+ af r/> = arg P Uwo) = . At w0 = o. Again.0. Conversely. called the adaptive identifier parameter. where parameter estlmat~s are ~pdated m realtime.11) may be written Yp(t) = o• T w(t) = wT(t) o• (2.ap)w(2)(t) (2.0. we will assume that measurements of r and Yp are made continuously between 0 and t.Is lost. If several frequencies are used. We have assumed zero initial conditions on w(ll and w< 2l.. Then. without storing the entire signals. measurements at one time instant give us one equation with two unknowns.13) equation (2.ap [ kp ) = Yp(tl) [ Yp(t2) r(t1)] I [yP(t 1)) r(t2) yp(t2) (2. We may refine this approach to avoid the measurement of Yp(t ). ~nly the ratio of the parameters k and ap is determined. y and r at one time instant t give us one equation with two unknow. we do not require differentiation.1 ).0.11) a: The signals w(ll. In the sequel.8) reads yp(t) = kpw< 1l(t) +(A. Equation (2. but instead stable filtering of available signals. :ach contributes two equations and the parameters are overdetermined.8) wo .0.0.5).1 0) Then.0. phase mformat10n .0. Time Domain Approach We now discuss schemes based on a timedomain expression of the plant (2. but are not considered in this simplified derivation.0.AW(I) +r (2.9) Measurements of the gain m and phase r/> at a single frequenc. (2. Frequency response methods will not be furt~er discussed in this book. we will still analyze these algonthms m the ~re quency domain and obtain similar results as above for the recursive schemes. that is (2.12) .0..0.0. Define the vector of nominal identifier parameters o• := [A :PaJ (2. a constant input r(t 1 ) = r(t 2 ) with constant output Yp(t 1 ) = Yp(t 2 ) will prevent us from uniquely determining ap and kp.11) is to be compared with (2.0.ap A (2.r wo ~ 0 uniquely determine kp and ap by inversion of the above. relatwn~hip~. We ~Ill therefore conce~ trate on recursive approaches. w(t) may be . Only one equation is left.14) This is equivalent to Based on measurements of r(t) and yp(t) up to time t. as in the frequencydomain approach. define O(t) to be a vector of identical dimension.5) Measurements of y .0. 1 (2.arc tan A ~ y + r S+A p S+A Define the signals ( A. Consider (2.6) assuming that the inverse exists.7) Knowledge of o• is clearly equivalent to the knowledge of the unknown parameters kp and aP' Similarly. Note that.0.
T (. and an estimate IJ(t) derived.20) exists.) _ ) {"t.24).Yp(t)] OLS(t)  [ [ w(r) wT(r) dr] [ l w(r) Yk) = dr] .0.0. on the basis of measurements of e 1(t) and w(t) up to time t.20) may be written IJLs(t) = P(t) (2. Such an expression may be obtained for the leastsquares algorithms by defining P(t) " [ [ w(r) wT(r) dr ]' (2.20) shows that IJLsCt) inverse in (2. o·. Since each time instant gives us one equation with two unknowns.0. Another approach is the leastsquares algorithm which minimizes ~ the integralsquarederror (ISE) ( T ( & \.0.0.) ) 6 df . using (2.17).1(t)] (2.P(t) [ p.0.1(t) + P(t) ![ p.P(t) w(t) [ wT(t) IJLs(t) .0.26) (2.22) Since Gradient and LeastSquares Algorithms The gradient algorithm is a steepest descent approach to minimize e[ (t ).0.1(t)) P(t) where g is an arbitrary gain.19) ~ [ IJLs(t)] =  P(t) w(t) wT(t) IJLs(t) + P(t) w(t) yp(t) so that the leastsquares estimate is given by 0 .25) Owing to the linearity obtained directly from 1 a ao [ [ ~dT 2 I w(T) [ wT(T) IJ(T).0. where parameters are updated continuously on the basis of inputoutput data.g 0 ! (I) = ![ ~ P(t) p..15) Note that the identification error is linear in the parameter error IJ . (2.0.24) ISE j ~dr. (2.P(t) w(t) e 1(t) (2.0. The purpose of the identification scheme will be to calculate IJ(t ). We will therefore call (2.26) should be started with the correct initial conditions at some t 0 > 0 such that .0.0.0. For adaptive control applications.18) may be so that.0. 0 On the other hand.o•.17) .Yp(T)] 0 t dT 0 (2.0 Introduction 49 calculated.14) into (2.0.23) e1 w g>O (2.48 Identification Chapter 2 Section 2.1(t)) ! [ it follows that P(t)) p.0.0.20) Plugging (2..P(t) w(t) wT(t) P(t) (2.16) we let the parameter update law IJ = . it makes sense to consider the estimate that minimizes the identification error (2.21) so that (2.0.0. assuming that the Note that the recursive algorithm (2. Since (2.24) I w(T) Yp(T) dT 0 t (2. called the adaptation gain.15) a linear error equation. we are interested in recursive formulations such as (2.0.
. the parameter error is given by ou) . P{t) P(t) w(t) wT(t) P(t) 1t may be verified that the solution of c2.0. The algorithms have similarities to the previous ones.!O + 1 Yp(r) W(T) dr l (2.0.and are closely related to the above condition. kp.0.o· _. based on the new error equation? The answer is yes. denoted a~.15). (2. Specifically.0.0. The motivation is that there exist nominal values of the parameters. but it involves the dynamics of the transfer function M.31) where am.0. We first define a reference model transfer function Ym A O(t) ao(t)l = [ bo(t) w(t) r(t) [ Ym(t) l (2. b~ is equivalent to knowledge of ap.39) which may .28) l =PCo) r P.0. This will be a goal of this chapter.29) so that the nominal values of a0 .28) is 1 8(t) = [ P. In practice.0 Introduction Ym(t) = am YmU) + km u(t) 51 (2. Since Yp = o*r w.32) P(t ) 0= [ l w(r) wT {T) dr ]' (2. knowledge of a~. These are called persistency of excitation conditions .0.36) and the identification error so that et(t) (am/kmbo(t))ym(t) + kmao(t)r(t) + apyp(t)kpr(t) amet(t) + km((ao(t)a~)r(t) +~o(t)b. + 1W~> = (r) dr r[ F(O) =Po> 0 (2. It is not obvious at this point how to relate the time domain condition on the vector w to frequency domain conditions on the input.0. I . In this chapter.32) and (2.]:>e compared to the linear error equation (2.30) It follows that O(t) converges asymptotically to o* if [ w(T) wT(T)dT is unbounded as t . b 0 are (2.P(t) w(t) =  ao(t)r(t) + bo(t)ym(t) (2.37) ~ (2.0.0. Model Reference Identification We now discuss another family of identification algorithms.35) Clearly. We define the following vectors instead of (2. This equation is still linear.33) [ oT(t) w(t).33) give the closedloop system Ym(t) = (amkmbo(t))ym(t) + kmao(t)r(t) b.20). and are useful to introduce adaptive control techniques. the recursive /eastsquares algorithm is started with arbitrary initial conditions at t 0 = 0 so that O(t) . ]I Po+~ w{T) wTtr P~~Oo[ I 8") (2.)yp(t) amet(t) + km(OT(t)O*T)w(t)  et(t) (2. such that the closedloop transfer function matches any first order transfer function. b 0(t) are adaptive parameters and r is the input to the plant.0.50 Identification Chapter 2 Section 2.0. km > 0.o. we will study conditions that guarantee exponential convergence of the parameter estimates to the nominal parameter.0.Yp(t)) 0(0) = 00 where a 0(t ).0. but the approach is now based on the Lyapunov function u = M(s) = A km s +am (2. based on the socalled model reference approach. Can we still find update algorithms for O(t ). In the time domain .0.0.oo.27) Let the input u to the model be given by u(t) = exists.38) In short (2.34) (2.
(f3n snI + . where the 2 n coefficients a 1.42) Yp(s) f(s) (2.. A rational transfer function is called proper if its relative degree is at least 0 and strictly proper if its relative degree is at least 1. These are very important conditions in adaptive control. + {3.2. described by a transfer function so that (2.0.. we consider the identification problem of SISO LTI systems.. 2.0.) is piecewise continuous and bounded on IR+. The relative degree of a transfer function is by definition the difference between the degrees of the denominator and numerator polynomials. Reference Input Assumptions The input r(. (2.42) is identical to the gradient algorithm (2. a few definitions. and {3. m is unknown.39). + (3t) Yp(S) (2. Note that the derivation requires am> 0 (in (2.1) (2.. To avoid this problem. and fip(s) and dp(s) are monic. The transfer function P(s) can be explicitly written as _P_ • M is stable • Re(MUw))>O for all w~O. an and {3. the expression sn Yp(S) = (an sn. In this case however. we . due to the presence of the transfer function M.1) 2. Assumptions (Al) Plant Assumptions The plant is a SISO LTI system.1 + · · · + at) f(s) .. sn + f3nsnl + . Rational transfer functions are called stable if their denominator polynomial is Hurwitz and minimum phase if their numerator polynomial is Hurwitz.oo and that e.0..med. They will be discussed in greater detail later in this chapter. it is not the gradient algorithm for (2. f3n are unknown. it would require explicit differentiations to be implemented.43) Therefore. ~or identification purposes.e• under further conditions on the reference input.2. It may also be shown that e 1 .. respectively. However.1 Identification Problem 53 In this chapter. But first. Note that we do not assume that P is stable.. Ljung & Soderstrom [ 1983]).0.52 Identification v(e 1 .2 IDENTIFIER STRUCTURE The identifier structure presented in this section is generally known as an equation error identifier. (cf. given the following assumptions.0. The motivation for the algorithm lies only in the Lyapunov stability proof.43)) and km > 0 (in (2.1).40) Section 2.1. but the plant is strictly proper ( m :::.40)). In general.17) obtained for the linear error equation (2. y (s) f(s) = P(s) = A an sn I + .2) is linear in the parameters a.: n . coprime polynomials of degrees m and n respectively...0) Chapter 2 (2..41) By Jetting e it follows that (2. . The resulting update law (2. e 1 and e are bounded. This expression is a parameterization of the unknown plant.0. + a. For example.0.0. defining strictly positive real transfer functions.0 as t .0. the conditions to be satisfied by M are that (A2) where f(s) and y (s) are the Laplace transforms of the input p and output of the plant.1 IDENTIFICATION PROBLEM We now consider the general identification problem for singleinput singleoutput (SISO) linear time invariant (LTI) systems. that i~ a model in which only a finite number of parameters are to be determ. A The objective of the identifier is to obtain estimates of kp and of the coefficients of the polynomials fip(s) and dp(s) from measurements of Athe input r(t) and output Yp(t) only. it is convenient to find an expressiOn wh1ch depends linearly on the unknown parameters. A polynomial in s is called monic if the coefficient of the highest power in s is 1 and Hurwitz if its roots lie in the open lefthalf plane.15).0.
StateSpace Realization a (sf.1) 1 I 0 0 0 0 0 bA X )J!p(S) (s or.2. SUppose that there exist a*(s) + oa(s) b *(s) + ob (~ ). since the plant parameters a • .14) where JP is n. with (2..dp(S)) Yp(S) (2. this choice is unique when fip(s) and dp(S) are COprime: indeed.5) (2.2 0 0 A.1.)twbtl(O) (sf A. the description of the plant (2.1) Yp(S) = k{J fip(S) r(s) + (X (s). w(2) p + bAy p (2. r s ) + . fdent~fier Structure 0 55 nth order polynomial denoted A(s) = sn +An sn.2. using (2.1wb 2l(O) w~2 l(s) (2.2.2.w(ll + bAr p A.2.2. Further.)t bAr(s) +(sf.: equation (2. bA E IR n in controllable canonical form.2.+ · · · + At. such that (2. This polynomial is assumed to be Hurwitz but is otherwise arbitrary.).2.}p (s ) A (s) A (s) The transfer function from r.1 0) so that the new representation of the plant can be written A Yp (s )  a*(s) A A( b*(s) ".4) 1 A(S) s (2.t + ··· + 1 at = kp fip(S) A A In analogy with (2.6) . E IR n x n.2.).13) and..12) With this notation.2.6) becomes Yp(s) = a* w~tl(s) + b* w~2 l(s) 7 7 (2. Let a*(s) b *(s) an Sn.A.t bAyp(s) + (sf A. such .2.2.A. In Laplace transforms w~t)(s) and it is easy to verify that this transfer function is fi(s) when *(s) and ~ *(s) are given by (2. wFl(O).f3t) (2.3) a sn= I + .2.9) This expression is a new parameterization of the plant.1.8) However.2.2. A statespace realization of the foregoing representation can be found by choosing A. the same expression is valid in the time domain (2.7) with initial conditions wbtl(O). define A(s)dp(s) (An. + A(S) 0 0 At al 0 An n r(s) + (An/3n)Snt +···+{At/3d A (s) A Yp(s) (2. Then. b • are constant.that the transfer function was still kP fip(s )/ dp(s ): The followmg equatiOn would then have to be satisfied (2.5).15) . while the degree of ob is at most n .2. dp are coprime. (t) Wp A.54 introduce a monic fdentUication Chapter 2 Section 2.5).Yp is given by a*(s) X(s) (2..8) has no solution since the degree of and fip.2./3n)sn + · · · + (At.2.11) b *(s) (2.
the plant output depends linearly on the unknown parameters. In analogy with the expression of the plant output. however.11) and (2.2.16) are arbitrary. without knowledge of the plant parameters. even when the plant is unstable.56 Identification Chapter 2 Section 2. {3.2.15). Therefore. The plant output can be written Figure 2.14) define a realization of the new parameterization.(t) . Since r and Yp are available. given here by (2.1.2. We will first neglect the presence of the ~<t) term but later show that it does not affect the properties of the identifier. Ljung & Soderstrom [ 1983]) rely on a linear expression of the form obtained above. We note that the generalized state of the plant wp(t) is such that it can be reconstructed from available signals.20) We also define the parameter error cp(t) := O(t)8* E (2. we define the observer A wO> + b.2. bT(t)) E IR 2n (2.Yp(t) = cpT(t) w(t) + ~(t) (2.1 0)(2.2.3 LINE.2.4R ERROR EQUATION AND IDENTIFICATION ALGORITHMS Many identification algorithms (9f. . the output of the identifier is defined to be Y.2.2: Identifier Structure 2.21) and the identifier error e 1(t) := Y.2.2. the observer error w (t) .22) Figure 2.2.16). The vector wP is the generalized state of the plant and has dimension 2n. This plant parameterization is represented in Figure 2.J w< 2> A w< 2> + b>.2. but the (s) unobservable modes are those of X and are all stable.2 Identifier Structure Yp(t) =.2. where the signal ~(t) is to remind one of the presence of an additive exponentially decaying term.2. The initial conditions in (2. by (2. Knowledge of a set of parameters is equivalent to the knowledge of the other and each corresponds to one of the (equivalent) parameterizations. Identifier Structure The purpose of the identifier is to produce a recursive estimate 8(t) of the nominal parameter o•.Yp (2. the realization of P(s) is not minimal.(t) = oT(t) w(t) e: lR IR2n (2.wp(t) decays exponentially to zero. so that standard identification algorithms can be used.16) to reconstruct the states of the plant.19) This term is due to the initial conditions in the observer.10)(2. and are represented in Figure 2.1: Plant Parameterization These signals will be used by the identification algorithm. In the last form.17) By (2. that is .18) Equations (2. The vector o• is a vector of unknown parameters related linearly to the original plant parameters a. Eykhoff [1974]. o•T w(t) + ~(t) 57 (2.2..2. We also define the identifier signals oT(t) := (aT(t).
3. called the update law. that is and Oi < 0 and iJi > 0 (2. strictly positive gain called the adaptation gain.9) a 1 ao ( e 2 ( O) ) = 2 e1 w (2.6) is globally Lipschitz in f/J (using Y lR . convex and delimited by a smooth boundary).6) n Input r(t)..3. Pr (z) denotes the projection of the vector z onto the hyperplane tangent to a0 at 0 and Operp denotes the unit vector perpendicular to the hyperplane.3.3. The initial condition 0(0) is arbitrary. viewed as a function of 0. but it can be chosen to take any a priori knowledge of the plant parameters into account. An alternative to this algorithm is the normalized gradient algo The gradient algorithms can be used to identify the plant parameters with the identifier structure described in Section 2. even when w is unbounded. The new regressor is thus a normalized form of w.. The identification algorithm is defined by a differential equation.2) We arbitrarily separated the identifier into an identifier structure and an identification algorithm. Using the normalized gradient algorithm. The righthand side is proportional to the gradient of the output error squared. pointing outward. The righthand side of the differential equation (2. The vector w(t) is usually called the regressor vector.4) Oj E [Pi.3.3. With the expression of yp(t) is associated the standard linear error equation (2. The identifier structure constructs the regressor w and other signals. of the form (2. that is g e1 w g> 0 (2. The parameter g is a fixed. w(t) are known signals and o• is unknown.5) This update law can thus be seen as a steepest descent method. 2.3.3. pt) The update law is then modified to PiPt (2.g wwT r/J.3) where F is a causal operator explicitly independent of the evolution of the identifier parameter 8. it is useful to modify the update law to take this information into account.3. the normalized gradient algorithm with projection is defined by e1w 0 = g=1+')'WTW 0 E int(0) a0 and e1wTOperp < 0 Pr [. for example. Identifier with Normalized Gradient AlgorithmImplementation Assumptions (A1)(A2) rithm Data g. with w replaced by w I 1 +I' wT w in iJ = .2.7) o•. When the nominal parameter o• is known a priori to lie in a set 0 E IR 2n (which we will assume to be closed.8) defines the standard gradient algorithm. A frequent example of projection occurs when a priori bounds Pi. For example.2)). which defines where int 0 and a0 denote the interior and boundary of 0. This update law is equivalent to the previous update law.3 Identification Algorithms 59 where yp(t ).3.3.>O (2. the implementation is as follows. related by the identifier error equation.g 1 + :T w ] if 0 ~ E (2.58 Identification Chapter 2 (2.1) Section 2.3.1 Gradient Algorithms The update law 0 =  (2. Pt are known. and it allows us to vary the rate of adaptation of the parameters. Yp(t) Output E where g and I' are constants.
include (2. Indeed.. respectively. = (sf .1 ' . e1{t) E IR Initial conditions are arbitrary.2 LeastSquares Algorithms Leastsquares (LS) algorithms can be derived by several methods. Y.10)(2..3.g .1b/. all that is required for identification is that there exist unique a•.3..60 Identification Chapter 2 Section 2.{t). b(t) e IR.(s) is Hurwitz .!3t •. .A). An.1 0) Comment We leave it to the reader to check that other choices of (A. .. wT = (w<l>T' w<2f) Y.1b/. with minor adjustments. Another interesting approach is to connect the parameter identification problem to the stochastic state estimation problem of a linear time varying system.Yp + oT = (aT.) are possible.11) are perturbed by zero mean white gaussian noises of spectral intensities Q e IR 2n x 2n and 11g e IR. b T) estimates of {a 1 .. with the presence of the socalled correlation term w e 1• The matrix P is called the .. r w< 2>= A w< 2> b/. an. The update law for 0 is very similar to the gradient update law.A) = i..D e w 1+yWTW o"(t) with output = o (2.3..12) Q and g are fixed design parameters of the algorithm..3.3. A1 .. b • in the corresponding parameterization.3. (t) E IR Internal Signals w(t) E IR 2n (wO>(t).n) Y. b/..3 Identification Algorithms 61 O{t). The parameter o• can be considered to be the unknown state of the system oT w Y. • g' 'Y > 0. Identifier Structure 0 w l = A wO> + b/.i3n) 2. Alternate choices of (sf . the leastsquares estimator is the socalled Kalman filter (Kalman & Bucy [ 1961]) 0 p gPwe 1 QgPwwTP Q .. E IR n in controllable canonical form such that det (sf .g > 0 (2. One approach was presented in the introduction. b).A). Design Parameters Choose O(t) E a>O and • A E IR n x n. w< 2>(t) e IR n) s+a IR 2n(a(t).Yp Normalized Gradient Algorithm e1 = 6 = .11) Assuming that the righthand sides of (2.. .
The theorems are also valid for vectors w of any dimension.1 Gradient Algorithms In this section.4."(>0 assuming the linear error equation et = <PT w (2. directional adaptation gain. withe 1 defined by (2."( > 0 (2. where P is reset to a predetermined positive definite value. 'Y are fixed parameters and P(O) > 0. but only require that w be a piecewise continuous function of time.4. since the parameter o• is assumed to be constant.12) is called the covariance propagation equation. Goodwin & Mayne [1987]) by 0 2.4 PROPERTIES OF THE IDENTIFICATION ALGORITHMSIDENTIFIER STABILITY A.I dt ~ 0. g > 0 g ( . defined by 1 1 1 • ko > k 1 >0 Normalized LS Algorithm with Covariance Resetting 0 . The covariance propagation equation is then replaced by Identifier with Normalized LS Algorithm and Covariance ResettingImplementation The algorithm is the same as for the normalized gradient algorithm. e1 w 1 +"(W W T g.4 establish general properties of the gradient algorithms and concern solutions of the differential equations (2..g _.2) g 1 +"(WT p W PwwTP . The leastsquares algorithms are somewhat more complicated to implement but are found in practice to have faster convergence Theorems 2. whenever Amin(P) falls under some threshold. =  Pwe1 g 1 + 'Y wT P w p = g PwwTP 1 + 'Y wT P w P(O) = P(t.2..3 properties.g( AP + PwwTP) or 2. Then P will become arbitrarily small in some directions and the adaptation of the parameters in those directions becomes very slow.3.1 )(2.1 + ww T ) (2.. The normalized leastsquares algorithm is defined (cf.kl} 0 .1) and the normalized gradient algorithm ¢ = 0 = g . This socalied covariance windup problem can be prevented using the leastsquares with forgetting factor algorithm. not necessarily even.62 Identification Chapter 2 Section 2. we establish properties of the gradient algorithm ¢ = 0 = ge 1 w g>O (2..4.+) = koi>Owheret.3) or wwT = g~~1+"(WT{PI)Iw (2. except that P(O) > 0.. P(O) is usually chosen to reflect the confidence in the initial estimate 0(0).4.AP .4. the leastsquares algorithm referred to is usually the algorithm with Q = 0.12. = {tiAmin(P{t)):. Identification Algorithms 63 covariance matrix and acts in the 0 update law as a timevarying. in addition where g is a constant._==! + "(WT p W g.gPwwTP or gwwT g > 0 (2.3. . In the identification literature.4.3. The same modifications can also be made to avoid covariance windup.4.3). . The properties do not require that the vector w originates from the identifier described in Section 2.3.14) Another possible remedy is the covariance resetting.shows that dP .15) Again g. The initial conditions are arbitrary.4. The covariance update law in (2.4.2).13) Design Parameters Choose. to guarantee the existence of the solutions. The new expression for P. so that Pmay grow without bound. except Internal Signals In addition P(t) E IR2nx2n dP dt = .
It may also be useful.. and~ e L 00 • Since vis a positive.4. v(O) for all t ~ 0.. 0 ::.4.4.4) and/or the gradient algorithms are replaced by the gradient algorithms with projection.4. 0 00 • Hence.Lr:. where E(t) is an exponentially decaying term due to the initial conditions in the observer.4. 0::. so that /3 e L2.4.(gh)l¢1. c/JT ¢.1 ). v(t)::. such that (2.~ e L 2 nL 00 ¢Tw e. ¢. or necessary. D positive. Using the fact that xI I+ x ::. Note that Yl+ywTw 1 +II Let the Lyapunov function v jectories is given by = c/JT ¢ so that the derivative along the tra=  wtll oo where the first term is in L 2 and the second in L 00 .3 Effect of Initial Conditions and Projection If Proof of Theorem 2.64 Identification Chapter 2 Section 2.4..1 The differential equation describing ¢ is ~ = gwwT ¢ implies that edY 1 + 'Y wT w e L2.4.4.4.4.3) is replaced by e 1(t) Theorem 2. .(Jb vdt =fer dt < oo 0 Effect of Initial Conditions and Projection In the derivation of the linear error equation in Section 2. together with the gradient algorithm (2. monotonically decreasing function. Theorem 2.3).2 are valid. e 1 /YI +"fWTw.1 and 2.lR2n be piecewise continuous.4.4. we get that 1~1 ::.4. v (0) for all t 0. ¢ e L Since v is a it follows that ¢ e L2.3) together with the normalized gradient algorithm (2.3 (a) Effect of initial conditions Modify the Lyapunov function to Hence. The following theorem asserts that these modifications do not affect the previous results.2. the limit v(oo) is well defined and Then (a) (b) e1 e L 2 ¢ e L 00 . monctonically 00 decreasing function. v (t) ::.2 Let v = Then the linear error equation (2.Jvdt 0 < oo Proof of Theorem 2. we found exponentially decaying terms.4. = c/JT(t) w(t) + E(t) (2.. Let w: lR + .lR 2n be piecewise continuous.2 Linear Error Equation with Normalized Gradient Algorithm Consider the linear error equation (2. the conclusions of theorems 2.. to replace the gradient algorithms by the algorithms with projection.. Since v =  2g(¢T w) 2 ~ 2ge[ ::.2). so that v.4) Then (a) (b) (c) e L2nL 00 Yl+"(WTW ¢ e L 00 . I for all x ~ 0.4 {3 e L 00 • Properties of the Identification Algorithms 65 Theorem 2. the limit v(oo) is welldefined and .3) is replaced by (2.1 Linear Error Equation with Gradient Algorithm Consider the linear error equation (2.4. so that v. Let w : lR + . so that 2ge[ Proof of Theorem 2.y \.
4.4. Note that the interval between resettings is bounded below. = {tiXmin(P(t)) s k. Thus. We note that d I dt P.4.P.4. Therefore.T z. defined by the following update law ¢ (b) q. v = 2q.4.4. due to the resetting.2q. since .4 The covariance matrix P is a discontinuous function of time. as stated in the following theorem.4. Theorem 2.8* ) .1 ¢.7) This update law has similar properties as the normalized gradient update .6) s K. so that 7 z = Pr(z) + Zperp.66 Ident((lcation Chapter 2 Section 2.<ko}.4.4. we find that. } is a set of measure zero. It also follows that { t. Then (a) e.5)  f Note that the additional term is bounded and tends to zero as t tends to infinity.2 LeastSquares Algorithms We now turn to the normalized LS algorithm with covariance resetting. so that where t.4.1 (!/) = k 0. Between discontinuities.T ¢.1 ~ 0. so that P t(t) ~ P . so that P. Using the Lyapunov function v = q.4. Consider first the gradient algorithm (2.4 Linear Error Equation with Normalized LS Algorithm and Covariance Resetting Consider the linear error equation (2. (b) Effect of projection Denote by z the righthand side of the update law (2. z is replaced by Pr (z) in the update law.4. J 2(T)dT ao I law. (8T. 9) dP P W WT P = g dt 1+')'WTPw or d(P I) W WT = gdt 1+')'WT(pt)lw P(O) = P(t/) = k 0 I > 0 where we used the fact that x /1 + x s I for all x ~ 0.6)(2.T · Zperp so that Vpr s v. denote by z perp the component of z perpendicular to the tangent plane at 8.1I. Vl+')'WTPw e L2 nL ao The proof can be completed as in theorem 2.1(! 0) = k 0..1.1 is globally Lipschitz. (2.3). and similarly for theorem 2. e Lao. for all t ~ 0. Pwe 1 + ')' WT p W and a discontinuous covariance propagation g.3. so that v = IR ln be piecewise continuous.Tw wtll ao E Proof of Theorem 2. When 8 E ae and z is directed outside 0. the differential equation governing P .1 ).4. together with the normalized LS algorithm with covariance resetting (2. P(t) ~ k 1I for all t ~ 0. ¢ e L 2 n Lao L2 n Lao (c) f3 = 1 +II q.4. Zperp = q. Note that it is sufficient to prove that the derivative of the Lyapunov function on the boundary is less than or equal to its value with the original differential equation. On the other hand. 1 = 8 = g _ ____:__ . II 'Y P 'II (2.. Let now v = q.1(! 2) ~ 0 for all t 1 ~ t2 ~ 0 between covariance resettings. 0 2.2.4 Properties of the Identification Algorithms 67 V = q.4. so that (2. 'Y > 0 (2.4. 7). ~at is *ht projectigg can on'y jmprgye the convergepce gf tbk alagritRm_ The proof can again be completed as before. the evolution is described by the differential equation in (2. At the resettings. +. 7).8) where we used results of Section 1.1) or (2.T P. noting that f E L 2 n Lao.Tq.T Zperp ~ 0. for the original differential equation.4. P. Let w : IR + (2.1(! 1) .1I. Since 8* E 0 and 0 is convex. Vpr = 2¢TPr(z) = v. For the differential equation with projection.2).
and the stability of A.12.1 0)(2. Similar conclusions follow directly for ~.68 Identification Chapter 2 Section 2.4.3 Stability of the Identifier We are not guaranteed the convergence of the parameter error cf> to zero. we deduce that cf> .12. Proof of Theorem 2.4. Also e 1 e L 2 .oo and the parameter error c/>.oo in an L 2 sense. The conclusions follow from this observation. Since only one output Yp is measured to determine a vector of unknown parameters.0 as t .4.) = cf>T (P. and. 0 Regular Signals Theorem 2. c/> e Lao . no assumption is made on the boundedness or on the differentiability of w.1 ).11) where the first terms in the righthand sides of (2.+).5 Stability of the Identifier Consider the identification problem.12.4. (2.2. we are not even guaranteed the convergence of the identifier error e 1 to zero.17).2) or the normalized LS algorithm with covariance resetting (2.4 state general properties of differential equations arising from the identification algorithms described in Section 2.4. In particular.4. The theorems can be directly applied to the identifier with the structure described in Section 2.5 Since r and Yp are bounded. because it is automatically (2. 0 Comments a) Theorems 2. since none of the errors involved actually converge to zero. and by corollary 1. The derivative of the parameter error ~ e L 2 n Lao and ~oast oo. c) The conclusions of theorem 2.4. .4.13).11) are in L2 and the last terms are bounded.4. (2.4 Properties of the Identification Algorithms 69 between resettings. e 1 . fundamental in excluding pathological signals in the course of the proofs presented in this book.4. {3 e L ao . The reader should note however that the conclusions are valid under very general conditions regarding the input signal w. At the points of discontinuity of P.P.1(t/) .4. It is of interest to relax this condition and to replace it by a weaker condition. this gain is guaranteed to become small as t .4.0 as e e t . with (A 1)(A3).12).oo .1 0) Pw v'1+ywTPw (2. In (2.4.5 relies on the boundedness of w.4. Also as a normalized error.))cf> ~ 0 It follows that 0 ~ v (t) ~ v(O). This can be obtained under the following additional assumption (A3) Bounded Output Assumption Assume that the plant is either stable or located in a control loop such that r and Yp are bounded.2.4.2. some additional condition on the signal w (see Section 2.2. In discrete time. guaranteed by (A3).4. This condition guarantees a certain degree of smoothness of the signal w and seems to be.6).4. {3 can be interpreted . e 1 . for all t ~ 0.2.4.4.4.1 e L 2 n Lao. 2.5) must be satisfied in order to guarantee parameter convergence.1(t. Theorem 2.4. cf> and ~ are bounded so that e 1 and 1 are bounded./!3( can be interpreted as the gain from w to cf>T w.13) The purpose of the identification algorithms is to reduce the parameter error c/> to zero or at least the error e 1• In (2.16). it follows from (2.2 and the gradient algorithms (2. ~. (2.=.==== v'1+ywTPw I+ II et Wt II ao (2.4. the identifier structure of Section 2. v(t. By theorems 2. w.4.Jvdt < oo 0 ao implies Note that V1+ywTPw 1 +II Wt II ao g. In fact.3. such a condition is not necessary. that is e 1 normalized by II w1 II ao. 1 e L ao and e 1 e L 2 implies that e 1 . that w and w are bounded. The function {3(t) is defined by Then the output error e. from the bounds on P.2.4 may appear somewhat weak.12) so that will I et{t)l = I c/>T(t)w(t)l = f!3<t)l1 wt II ao +tJU>I (2.v(t. We present such a result using a regularity condition on the regressor w.2 can be interpreted in the following way.4. From theorem 2.4.7). and the results interpreted in terms of the parameter error c/> and the identifier error e 1• b) The conclusions of theorems 2.
4. ular.17) 1 +II w1ll oo (d I dt 1 +II II WI II 00) I Since r is bounded by (A2). we derived results on the stability of the 1 identifiers and on the convergence of the output error e 1 = oTw. Therefore.4. rp E L 00 Proof of Proposition 2.4.15) is bounded. {3 E L 00 .rJ>T w tends to zero as too.4.4. Note that it will be sufficient for (2. Then The normalized error {3 = + II 1 rj>Tw w II 00 E 1 L 2 n L 00 .4. (2. The conclusions of proposition 2.4. or the normalized LS with covariance resetting (2.6 ). but also unbounded signals (e.4. we are left to show that~ We have that Proof of Theorem 2.~ E L 00 .6 Clearly.2.4.4. (2. we are now allowed to consider the case of an unstable plant with bounded input.0 as E t  oo (corollary 1. extending the properties derived in theorems 2.4.4. and hence {3 E L 00 . but nevertheless satisfies the regularity condition.6 are useful in proving stability in adaptive control. The definition presented here corresponds to a definition in Narendra.22. piecewise differentiable signals can also be considered.14) The class of regular signals includes bounded signals with bounded derivatives. We will also derive some properties of regular signals in Chapter 3.w 2 E e1I 1 +II or gain from w.4.4.2 or 2. the identifier structure of Section 2.15) The first and second terms are bounded. it follows that L 00 + rJ> I 0 T W A+ bhb* .6.16) The regularity assumption then implies that the last term in (2.g.70 Identification Chapter 2 Section 2. On the other hand will 00 since ¢. 7 It suffices to show that w is regular.4. to apply theorem 2. Proposition 2.6 shows that when w is possibly unbounded.2).4 Properties of the Identification Algorithms 71 verified. that is.¢ E L 00 If (a) w is regular T Theorem 2.4. and {3  0 as t  t  oo and ¢. {3 E L 00 and since {3. 0 Stability of the Identifier with Unstable Plant Proposition 2.2 and the normalized gradient algorithm (2.14) to hold everywhere except on a set of measure zero.6 Let¢. to relax assumption (A3).2). (2.5 PERSISTENT EXCITATION AND EXPONENTIAL PARAMETER CONVERGENCE In the previous section. the relative error zE L ooe :::::: z is called regular if. 00 w1 II L 00 eand¢. for some k 1. where the boundedness of the regressor w is not guaranteed a priori. k 2 0 for all t :::::: 0 (2. In the identification problem. (2.4 to the case when w is regular. ¢ E L 00 and w is reg 2. e 1 ). Definition Let z : IR + Regular Signals IR n.1 7) shows that w is regular. We are now concerned with the convergence of the . (b) {3 = rJ> w E t +II w1ll oo L2 oo. 7). & Valavani [ 1980]. Combining (2.w:IR+IR nbesuchthatw.4.4.7 Stability of the IdentifierUnstable Plant Consider the identification problem with (A 1) and (A2). Lin.2.2.4 followed by proposition 2.4.16)(2. 1 ·l w(t) + [ bh 0 l r(t) 0 (2.18). such that z.0* w = ]' c/> w to zero.4.4.. {3  0 as Then {3. This definition allows us to state the following theorem. It typically excludes signals with "increasing frequency" such as sin (e 1 ). {3 E L 2 implies that {3.
3. is equivalent to a uniform complete observability condition on system (2.:. we can expect </> to still converge to 0 if w completely spans IR 2n as t varies. thereby leading to estimates of the convergence rates in the proof of theorem 2. For consistency.5.5. . the PE condition requires that w rotates sufficiently in space that the integral of the matrix w(T)wT(T) is uniformly positive definite over any interval of some length o. such is never the case.5.5. Lemma 2.5). It is an alternate proof to the original proof by Anderson & Moore and relates the eigenvalues of the associated observability grammians. However.1). (2.11 )). (2.1 0). A + KC].5).5) A vector w: IR+. We recall that the identification problem is equivalent to the state estimation problem for the system described by (2.72 Identification Chapter 2 Section 2.5. lxl (2. Using the Lyapunov function v = <f>T <f> v=  <I>T (A +AT)</> When A (l) is uniformly positive definite. with Amin (A +AT) ~ 2 a. the dimension of w is assumed to be 2n..5.3. for all o> 0.5. The proof by Anderson has the advantage of leading to interesting interpretations. we establish the following convergence theorem. This leads naturally to the following definition.6) .5.1 PE and Exponential Stability Let w : IR + . The condition has another interpretation.5.2) Comments The proof of theorem 2.5. which turns out to be an identifiability condition.5. Kreisselmeier [ 1977]). Anderson [ 1977].5.3) Although the matrix w(T) wT(T) is singular for all T. that is the convergence of the parameter error </> to zero. then v :. We now find that the persistency of excitation condition. The convergence of the identification algorithms is related to the asymptotic stability of the differential equation tP(l ) = .5. but it is in fact arbitrary.2 a v. Theorem 2. Definition Persistency of Excitation (PE) o·ct> y(t) o (2.5.1 given afterward.5.5 Persistent Excitation & Exponential Convergence which appears as a condition on the energy of w in all directions.IR 2n is persistently exciting (PE) if there exist a" a 2.1 uses the following lemma by Anderson & Moore [ 1969]. by reexpressing the PE condition in scalar form to+ o a2 which is the system described earlier in the context of the leastsquares identification algorithms (cf.4) J II K(T)II to 2 dT :. The proof of theorem 2.IR2n be piecewise continuous.g W (l ) W T(l ) <f>(l ) g>O (2. while those by Sondhi & Mitra and Kreisselmeier give estimates of the convergence rates. which states that the UCO of the system [C . however. Morgan & Narendra [1977a&b]. there exists k 6 ~ 0 such that.1 can be found in various places in the literature (Sondhi & Mitra [1976].1) is globally exponentially stable 73 parameter () to its nominal value 8*.2) is exponentially stable with rate a. since at any instant the matrix w(t)wT(t) is of rank 1. The proof of the lemma is given in the Appendix. k 6 (2. for allt 0 ~ 0 lo+o ~ J (WT(T)X) 2dT to ~ a1 for all to~ 0. The idea of the proof of exponential stability by Anderson [ 1977] is to note that the PE condition is a UCO condition on the system where A (l) E IR 2n x 2n is a positive semidefinite matrix for all t. A] is equivalent to the UCO of the system with output injection [C. since w varies with time.Z Uniform Complete Observability Under Output Injection Assume that. For the original differential equation (2.:.1) which is of the form rP(t) = A(l)<f>(t) (2. o> 0 such that 10 + a 2/ ~ J W(T)WT(T)dT to o ~ a 1f for all t 0 ~ 0 (2. With this. If w is PE Then (2. any vector <1> perpendicular to w lies in the null space of w wT and results in ¢ = 0. In fact.5. which implies that system (2.
g a t1 o. then al> a 2 are multiplied by 4.3 Exponential Convergence of the Identifier Consider the identification problem with assumptions (A 1)(A3).5. Proof of Theorem 2. Theorem 2.6).2 and the gradient algorithms (2.5 Persistent Excitation & Exponential Convergence 75 Then the system [C . if the observability grammian of the system ti\ Cl satisfies . (2.5.62 I ~ N (to.4. the rate of con vergence a .3 This theorem follows..5.5. For the standard gradient algorithm (2.5. the convergence rate is proportional to the adaptation gain g and to the lower bound in the PE condition.9) By the PE assumption. for example. In the limiting case when the adaptation gain g or the reference input r is made small. The constants ai> a 2 and o depend in a complex manner on the input signal r and on the plant being identified. so that the exponential convergence is guaranteed for both gradient update laws. Therefore. Through the PE condition. so that (2. to + o) ~ . A] is uniformly completely observable if and only if the system [C.2).11) 0 Exponential convergence then follows from theorem 1.f .5. In this case.6 2 :! 1 )2 l¢{to)l 2 (2.1). However. 7).2 in Appendix.2g (wT ¢) 2 ~ 0 t 0 +6 along the trajectories of bounded. Note that when w is . with t0 + 6 1 U In (2.4. A + KC] is uniformly completely observable. Moreover.4. or the normalized LS algorithm with covariance resetting (2.8) . .'\J Exponential Convergence of the Identifier Theorem 2..3) and n is the order of the plant.5.gw(t).5.1.62) Proof of Lemma 2.Bt/(1 +Jk. directly from theorem 2. if r is multiplied by 2. and o come from the PE definition (2. the system fo'lwT(t)l is UCO. This result will be found again in Chapter 4.4 allow us to extend the proof of exponential convergence to the LS algorithm.1 can be applied to the identification problem as follows.1). the. for all t 0 ~ 0 to+6 to J vdr ~ (l + 2ng.5.. the system with output injection is UCO.5.ri.5.62 exp ( k6 .1. The bounds on P obtained in the proof of theorem 2.4. By lemma 2.5.10) where 2n is the dimension of w.4. injection with K(t) = . using averaging techniques. where g is the adaptation gain.1 Let v = cpT ¢..6t I then the observability grammian of the system satisfies these inequalities with identical oand [~ (2.2. the identifier structure of Section 2.s~ becomes wT(t)]. Forallt 0 ~0 t 0 +6 v = .12) k6 = J to !gw(T)! 2dr • g tr [ '"{ w(T) wT(T) 2 dT] (2.2. 0 Exponential Convergence Rates Estimates of the convergence rates can be found from the results in the proof of theorem 2. If w is PE Then the identifier parameter 8 converges to the nominal parameter o· exponentially fast. The influence of some design parameters can be studied with this relationship.5. it is also proportional to the square of the amplitude of the reference input r. the convergence rate is given by a = J VdT to 2g J (wT(r)cp(r)) 2 dr to (2. al> a 2 .4.1) or (2. w PE is equivalent to w/V 1 + 'Y wT w PE.5.5.74 Identification Chapter 2 Section 2.7) (2. Proof of Theorem 2.5.5.
Proposition 2. dm is Hurwitz.1. In that case.1. so as to match the plant transfer function. i.12) shows that above some level. a•.22. The polynomial Xis a monic.1 There exist unique Figure 2. although its stability may only be guaranteed from prior knowledge about the plant. identifier structure which led to a linear error equation. Sondhi & Mitra [ 1976]). Kreisselmeier [ 1977]). respectively. The scheme discussed in Sections 2. b*(s) in Figure 2. We will have to require that the zeros of X contain those of fim(s) and therefore we write (2. we may already interpret it as being obtained by modifying the reference model through feedback and feedforward action.2. the convergence rate estimate saturates and even decreases (cf. It is also possible to show that the presence of the exponentially decaying terms due to initial conditions in the observer do not affect the exponential stability of the system.2. that is 1 :::. b• such that the transfer function from r  Yp in Proof of Proposition 2.2. We start from an arbitrary reference model. and although the degree is now n.1) that the plant can indeed be represented this way. The rate of convergence will. This structure was based on a convenient reparametrization of the plant. dm(s) are monic. this approximation is no longer valid and (2. A simple choice of M is 1Is+ l.6. n. Another approach. It serves a similar purpose in Section 2.Yp'in Figure 2.1) where fim(s). Figure 2.3 have degrees at most n . As previously. k I :::. It is worth pointing out that there exist several ways to reparametrize the plant and many error equations that may be used for identification.1 Model Reference Identifiers In Section 2.1. instead of n in Section 2. the model transfer function was simply the identity.m. coprime polynomials of degrees I and k :::. Assume that the reference model is strictly proper. we might interpret this structure as a more general parametrization of the plant than the one in Figure 2. described by a transfer function Although we will have to show (see proposition 2. The polynomials a*(s).6.2) where X0 (s) is a monic.6. Landau [ 1979] discussed an interesting alternative called the output error approach. but that its relative degree is no greater than the relative degree of the plant. we start with the following proposition. The resulting scheme has significant advantages in terms of noise bias.76 Identification Chapter 2 Section 2. the plant is assumed to satisfy assumptions (A 1) and (A2) (cf. we presented ar.3: Model Reference Reparameterization 2. Hurwitz polynomial of degree n I  1. The reference model is stable. n.6. Section 2. with transfer function M satisfying the following conditions (A3) Reference Model Assumptions The reference model is a SISO LTI system.3 is given by . Alternatively.6. Luders & Narendra [1973]) is discussed now. however.1 and serve similar purposes as before.5. the following derivations can also be carried out with a degree equal to n.e.1 ). Therefore. which we will call the model reference approach (cf. Yp 77 When the adaptation gain and reference input get sufficiently large. Hurwitz polynomial of degree n . be as found previously only if the rate of decay of the transients is faster than the rate of convergence of the algorithm (cf.3 is P (s ). M(s) = km ~m(s) dm(s) (2. 3.3 is generally known as an equation error identifier.1 Existence The transfer function from r.6.6 Model Reference IdentifiersSPR Error Equation Consider now the representation of the plant of Figure 2. with only minor adjustments (the reference model may then be assumed to be proper and not strictly (s) proper).6 MODEL REFERENCE IDENTIFIERSSPR ERROR EQUATION 2.
w< 2> e IR n.10) b • appropriately. we find that any plant can be represented uniquely as oa ob P (2. b• + ob satisfying = and the nominal parameter is (2. let (2.6. 7).A). b• of degrees Given any plant and model transfer functions satisfying (A1) and (A3). 0 Identifier Structure The identifier structure is obtained in a similar way as in Section 2.5) This defines by ao. ·cs ). bo E IR and a•. E IR.3) This transfer function will be equal to P(s) if and only if Xodm.6. a•.1 and the initial conditions are arbitrary.1.At bx (2.r (sf .6.l.6.6.6 Model Reference IdentifiersSPR Error Equation 79 ~· a Recall that fip.4) from (2.1 in controllable canonical form such that det (sf .14) . Therefore.1.2n (2.7) (j oT · Xodm.6.1. a Equality (2.km(b• + ob) %m P JP (a+~oa) np .1 x n.6.1 by assumption (A3). we find that Using proposltwn 2.1. equation (2. respectively.I .6.6. The regressor vector is now wT(t) := [r(t).6.6.A). Divide Xodm by dp: denote by q the quotient of degree k . In other words. b. b o' b .6) (2. wOl(t). n (2.2.6.. A solution can be found by inspection. bx e IR n.4) is satisfied if a• is defined We define the observer through A w<l) + bx r ~· kp ~ ~ km q np (2.8) cannot have any solution.9) X (s) The problem is therefore to find polynomials n . a •T . Uniqueness Assume that there exist w< 2>\t)] e 2 IR.6.6.km b • ~ ~ ~ = k dp ~ P km ~ a: np (sf.11) where w< 1>. while the degree of dp and ob are n and at most n .6. [ • • ao. The degree of the polynomial in the righthand side is m + k .  78 Identification Chapter 2 Section 2.4) at most 1 1 s (2.4). so that the degree requirements are satisfied.A) = X and (s) (2. we showed that there exist unique polynomials b We now let a·cs ).13) Subtracting (2.1 and let km b • be the remainder of degree n .6. yp(t).r (sf ..1 bx (2.6.I .6.1 bx b0 + b .12) a• + oa. We let A e IR n. dp are assumed to be coprime. which is at most n .1 that satisfy (2..8) (2.n I such that a·cs) X(s) G ·cs) X(s) w< 1> = ao + a .r ] E IR.
A) is Hurwitz. We will define and discuss this assumption in the next section.g. but it now involves the dynamics of the transfer function M.1 in controllable canonical form such that det (sf .80 Identification Chapter 2 Section 2. • g> 0 Identifier Structure w(l) = A w<l) +bAr w< 2> = A w< 2>+ b>.16) is the vector of adaptive parameters. The overall identifier structure is represented in Figure 2.14). provided that the following additional assumption is satisfied. for the purpose of the analysis. We define the parameter error Figure 2. aT.Yp .1 x n . one should add an exponentially decaying term due to the initial conditions in the observer and in the reference model transfer function. by r(l). but we will neglect this term for simplicity. the identifier output is defined to be Yi(t) := M ( ao(t)r(t) + aT(t)w(ll(t) + bo(t)yp(t) + b T(t) w(2l(t)) := M[OT(t)w(t)) where (2.1.yp(t)elR Output O(l). Yi(l) E IR Internal Signals w(t) E IR 2n [ w(ll(t). upper bound on n. In similarity with (2.o· so that the identifier error (2. n) Input ()T = [ao. w< 2>(t) O(l)eiR 2n[a 0(t)eiR.Yp Assumptions (Al)(A4) Data n. and contains the zeros of fim(s ).b(t) E IRnl] The error equation is linear in the paramder error.6.a(t) E IRn I] E e1 (t) = M [ cJ>T(l) w(t)) (2.1 .6.19) IRn. We will show in the next section that the same gradient algorithm as used previously can still be used for identification. Model Reference IdentifierImplementation Yi(t).6 Model Reference IdentifiersSPR Error Equation 81 Rigorously.Yp(t) (2. e1(t) E IR Initial conditions are arbitrary.18) is also given.m (e.4 and we summarize the implementation of the algorithm hereafter. e IRn.6. Design Parameters • M satisfying (A3)(A4) • A e IRn.6.17) e1(t) = Yi(t). b>.6. (A4) Positive Real Model M is strictly positive real.6. bo.15) (2.b 0(t) E IR.. bT] wT = [ r w<l)r y w<2Jr] ' ' p' Yi e1 = = M(OTw) Yi.4: Model Reference Identifier Structure cJ>(t) := o(t).
It may result in savings in computations. the state variable filters are of order n . for all u e IR. the algorithm is not the true gradient anymore. stable. For uniformity with previous discussions. We will also show in Chapter 4 how the transfer function M influences the convergence properties of the gradient algorithm.2.6. This error equation is still linear. !he de~~itions of positive real (PR) and strictly positive real (SPR) functwns ongmate from network theory.19). we encountered a more general error equation. Anderson & Vongpanitlerd [1973].s section. although we will keep using this terminology for the similarity.2 Minimal Realization of an SPR Transfer Function Let [A. The following definitions are deduced from these properties. precludes the use of the leastsquares algorithms.1. and Re(M(cr+ jw)) ~ 0 for all ~ > 0. Q.20) Proof of Lemma 2. rational transfer function M(s). SPR transfer functions form a rather restricted class.4. SPR Error Equation with Gradient Algorithm A remarkable fact about SPR transfer functions is that the gradient update law c/>(t) O(t) = ge 1 (t)w(t) g>0 (2.6 Model Reference IdentifiersSPR Error Equation 83 Gradient Algorithm 0 = ge 1w 0 • M(s) is stable • Re (MUw ))> 0. 2. we will establish general properties involving this error equatiOn. it is the driving point impedance of a dissipative network if and only if it is SPR.6. however. the realization requires one less state.21) A rational function M(s) of the complex variable s = u + jw is positive real (PR).IR 2n.6. It may be shown (cf.6. and when e 1 is defined by the linear error equation (2. a statespace realization of M(s) with state em can be obtained so that . It is strictly positive real (SPR) if.6. In particular. instead of n previously. an SPR transfer function must be minimum phase and its phase may never exceed 90•. Lemma 2.1.f) lS PR. a + b > c > 0. M(s. but it now involves additional dynamics contained in M. In thi. the transfer function S+C M (s) = (s + a )(s + b ) is SPR if and only if a > 0.6.3). for some f > 0. cT] be a minimal realization of a strictly proper. Indeed. If the transfer function M is of order 1.2 cf. Using lemma 2. Definition tions Positive Real (PR) and Strictly Positive Real (SPR) Func M(s)isSPR There exist symmetric positive definite matrices P. b > 0. w ~ 0.82 Identification Chapter 2 Section 2. Then. Similarly. we assume that w : IR+ . the following statements are equivalent (a) e1(t) = M ( c/>T(t)w(t)) where M is a strictly positive real transfer function.ction 2. which are usually faster. The presence of M. An important lemma concerning SPR transfer functions is the KalmanYacubovitchPopov lemma given next. but the dimension of w is in fact completely arbitrary. if M(u) e IR.6.2 Strictly Positive Real Error Equation and Identification Algorithms In Se. such that (b) PA +ATP Pb = c Q (2. Note that in the case of the SPR error equation. A rational transfer function is the driving point impedance of a passive network if and only if it is PR.6. for all • lim woo w~ 0 w2 Re (MUw ))> 0 A Comments The model reference identifier is an alternative to the equation error scheme discussed previously. which we w11l call the strictly positive real (SPR) error equation For example. Ioannou & Tao [ 1987]) that a strictly proper transfer function M(s) is SPR if and only if has similar properties when e 1 is defined by the SPR error equation (2. b.
22) where [A .6 Model Reference IdentifiersSPR Error Equation Theorem 2.22) and (2. so that rJ> e £2.6.6.6. cf> are bounded. Along the trajectories of (2. Theorem 2.6. cT] satisfy the conditions of lemma 2.2.7 hereafter). Pem + r/JT r/J.6.6.5 Exponential Convergence of the Gradient Algorithms with SPR Error Equations Let w: IR.IR. 0 2.4 85 Modified SPR Error Equation with Gradient Algorithm Theorem 2. it follows that em.6. and again.25) (2.3 Let P. Q be as in lemma 2. The same gradient algorithm may be applied with this error equation.6. Let [A. e 1 . If Then w is PE and w.6.21 ).4 Let P.20).23) where we used (2.25). Along the tra Again.6. where [A . The additional boundedness requirement on w guarantees that PE is not lost through the transfer function M (cf.2. Consider the SPR error equation (2. we (2.PAem + ge!. Then (a) (b) Let w : IR. lemma 2. Q be as in lemma 2.3 Exponential Convergence of the Gradient Algorithms with SPR Error Equations As stated in the following theorem.6.6. under the PE condition.6.2. cr] satisfy the conditions of lemma 2.6. as required previously. Equivalently.2 and v jectories of (2. the main idea is to interpret the condition as a UCO condition.'Y wT(t) w(t)cT em(t)) cTem(t) g. so that piecewise differentiable signals may be considered.6.6.25) . 2n be piecewise continuous.6. v = ge!. b .19) with M(s) SPR. e1. rjJ em.6. together with the gradient update law (2.25) are globally exponentially stable. + . consider the statespace realization (2. Equivalently. it also follows now that e 1w e £ 2. The proof given hereafter is similar to the proof by Anderson [ 1977] (with some significant differences however). cT] satisfy the conditions oflemma 2.+ . The main condition for exponential convergence is the PE condition.6.6. Qem :::.21 )."(>0 L 00 (2.AT Pem + grjJTwbTPem2gcTemcf>Tw .1. Instead.26) Proof of Theorem 2. together with the gradient update law (2.2 and v = ge!.6.6. b. Pem + cf>T cf>. Then (a) (b) em.6.3 gcTem(t)w(t) g>O A l~) SPR Error Equation with Gradient Algorithm ~(t) (2. 0 (2. and em. 0 Modified SPR Error Equation The normalized gradient update law presented for the linear error equation is not usually applied to the SPR error equation. b.84 Identification Chapter 2 Section 2. The conclusions follow as in theorem 2.24) where 'Y is a constant. the gradient algorithm is also exponentially convergent with the SPR error equations.22) Let w : IR. so that in statespace form A em(t) + b [ rjJT(t)w(t).IR 2n.6. e 1 .4.6.6. a modified SPR error equation is considered "(>0 (2.ge!.6. e 1 e £ 2 • Moreover. consider the statespace realization (2.6.22) = ge!.IR. 2n be piecewise continuous. rjJ E E L2 L 00 Proof of Theorem 2.24) with M(s) SPR. It is sufficient that the boundedness conditions hold almost everywhere.+. since P and Q are positive definite.PbrJ>Tw + ge!. Consider the modified SPR error equation (2.
6.6.2 about UCO under output injection. We are thus left to show that system (2.5.2.+ . We now prove theorem 2. e : IR. Again. w e L 00 H is a stable.29) An additional term . For this.5.86 Identification Chapter 2 Section 2.6 in the Appendix. 7 shows that PE is not lost if the signal is filtered by a stable.e.30) is UCO. we find that inequality (2.6 Model Reference IdentifiersSPR Error Equation 10 + 6 87 Auxiliary Lemmas on PE Signals The following auxiliary lemmas will be useful in proving the theorem.6.6. exponential convergence will be guaranteed if. rJ>(to) · Derivation of 2.7 PE Through LTI Systems Let w: IR+. this implies that terms due to initial conditions do not affect PE. fJ2 (I £m{to)l 2 {) 2 . provided that the signal is sufficiently smooth.5 As previously.6. for some {J 1 . If w is PE e e £2 Then w + e is PE.6. em (to). but the dimension is in fact arbitrary. Using lemma 2. (2.28) will be satisfied if the following system em rJ> el Aem + brj>Tw 0 (2.6.27) to J e[.5. Pem + rJ>T rJ>. Qem dr •0 o> 0 lo + 6 A :S: g ~i~~Q) J e[ dr C lo+6 + I rJ>{to)l 2 ) :2: :S: 0 (2. Proof of Lemma 2. Lemma 2.6. an £ 2 signal is necessarily not PE. Note that the sum of two PE signals is not necessarily PE. Lemma 2. that I e 1(t) = cTeAUto)em(to) +JeT eA<tT)b wT(r)dr rJ>(to) to := Xt(t) + x2(t) satisfies.30) is UCO.6. respectively. we Jet or (2. 7 in the Appendix.6.IR. for some a3 > 0 . rational transfer function H(w) is PE. (2. we assume the dimension of the vectors to be 2n. minimum phase transfer function.3 is satisfied.28: This condition can be interpreted as a UCO condition on the system (2. A Proof of Lemma 2.31) If Then w is PE and w .6.6 PE and £ 2 Signals Let w. since w is bounded.(r )dr to By theorem 1. minimum phase.IR 2n. so that for both SPR error equations to+ d 10 + d for the basic SPR and modified SPR error equations.6.6. for uniformity.6.6 asserts that PE is not altered by the addition of a signal belonging to L 2• In particular.}yy wT weT em is added in the differential equation governing em in the case of the modified SPR error equation.28) J ef(r)dr :2: aJ(Iem{to)l 2 + lr/>(to)l 2 ) to for all to. Lemma 2.5. The condition on K in lemma 2.32) J ~ vdr :s: . i.g r e!:.6. Proof of Theorem 2. On the other hand.6. 2n be piecewise continuous.6. let v = ge!:. Lemma 2.
33: By assumption. em(to) and m > 0. b) The steps of the proof can be followed to obtain guaranteed rates of exponential convergence.36) I lo x[(r)dr ~ 'YJ(ma)lem(to)l 2 (2.38).6.(m+n)a 2) (2.37) for all t 0 .30)). ~) lo+h t0 +mu t0 +mu I ef(r)dr ~.6. we have that.42) t0 +mu for all t 0 .~ ~4f..b'l = (m + n) a. )' 2 > 0. such that 00 The lower inequality in (2. They do not rely on the structure of the identifier.6. there exists )' 3(m a)> 0.ma2l </>(to )1 2 Comments a) Although the proof of theorem 2. Although such rates would be useful to the designer. </>(to).6. 2m ~ Let m be large enough to get . for all t 0 ~ 0.6. Since [A1cT] is observable.39) w1(t) = IcTeA<trlbw(r)dr to (2..41) for all t 1 ~ t 0 ~ 0 and </>(to). Let n > 0 be another integer to be defined and let Using the tMR@l& iR&'f'la~ c. which is a filtered version of w. theorem 2.34) and n sufficiently large to obtain J Further. define {31 = na1 . o 1. c) The results presented in this section are very general. (2. The UCO condition leads then to a PE condition o~ a vector w1 .6.5. using averaging techniques. (2.6. while the upper inequality is easily found to be valid with {32 0 = max('Y 1.6. but only on the SPR error equation (2. w E L 00 fore. the signal I There I _. (2. lo lo r:.6.38) for all t 0 . 'YJ(m a) .35) min(aJ. the expression one obtains is quite complex and examination of the proof leaves little hope that the estimate would be tight. using lemma 2.{~. w is PE and w.I x[(r)dr loto+mu lo+mu ~l'YJ(m a) I em(to)l 2 ..2m 11 ~ "YJ(m u} / r'11'f .33) follows from (2. First. 10 +h 10 +h +.6.:. We leave it to the reader to specialize these results and obtain stability and convergence properties of the model reference identifier of Section 2.2 concerning UCO under output injection to transform the UCO condition to a UCO condition on a similar system.6.1. Derivation of 2.ma2 ~ a1 is PE..6. . em(to) and an arbitrary integer m > 0 to be defined later.)'J(ma).6. with 'YJ(m a) increasingVtth m a such that t 0 + mu I (2. This means that. through the LTI system M(s).1 ).I x!(r)dr to ').6. em (to).40) (2.6 Model Reference IdentifiersSPR Error Equation 89 +~na1l ¢(to)l 2 . with {3 1 as defined. It interprets the condition for exponential convergence as a uniform complete observability condition. a 2 . Then.88 Identification Chapter 2 (2.'YI e 11 1em(to)l 2 (2.6. it uses lemma 2./) ."1'1.33) Section 2.6. since A is stable. a > 0 11 +u 2 a2 I ¢(to)l ~ I xhr )dr ~ a1 I ¢(to)l 2 II ~ (2. . but where the vector </> is constant (cf.! I x!(r)dr. On the other hand.6. A more successful approach is found in Chapter 4.19). it relies on the same basic idea as the proof of exponential convergence for the linear error equation (cf.5 is somewhat long and tedious._ ..5.6. for some a 1.1 I x[(r)dr. there exist 1' 1. it has some interesting features.7.1'1 e _.
Since Yi .1 Let w (t) E Frequency Domain Conditions PE and Autocovariance 91 2. for all T ~ J to (wT(r)c)zdr (2.3 and 2. .2. In this section. namely that w be PE. Roughly speaking.1 Parameter Convergence From the definition of the regressor w in (2.5.. 2. we see that w is the output of a LTI system with input r and transfer function ~ to+ T to J (wT(r)cfdr ~ ~ t 1 (2.1 (a) implies (b) From the PE condition on w. using the linear .7.7. (a) w is PE if and only if (b) Rw (0) > 0.1) We will assume that the input r is stationary and that P is stable.3) and (2.7.4) yields that Rw(O) (b) implies (a) I> 0.Yp as t . it follows that w is also stationary. where 2n is the number of unknown parameters in the adaptive scheme. Thus." that is. the asymptotic identifier transfer function must match the plant transfer function at s = jw 1. as stated in the following proposition. 7 Proposition 2. . wk.7. that the parameter error converges to zero. Note that Rw (0) is a symmetric positive semidefinite matrix. to guarantee exponential convergence of the parameter error. we will show IR Zn be stationary.6.7.4) Combining (2.5) Therefore (2. the plant loop is "asymptotically timeinvariant. In turn. that is has an autocovariance. From the definition of autocovariance.7. we will show that the reference signal must be "rich enough.7. will too. we give frequency domain conditions on the input r to guarantee that w is PE.2) o (2. We will use the results of generalized harmonic analysis developed in Section 1. D We may relate the PE condition to the frequency content of w through the formula (2.90 Identification Chapter 2 Section 2. tJ to t0 + 6 (wT(r)c) 2 dr ~ Amin(Rw(O))icl 2 t (2. we expect that Yp and Y.7.oo. For stationary signals. Then.7.7.6. this will imply that the asymptotic identifier transfer function is precisely the plant transfer function and. "contains enough frequencies.. If k is large enough.7) where Sw(dw) is the spectral measure of w.6) which implies PE for t sufficiently small. therefore. The difficulty with the PE condition is that it is not an explicit condition on the reference input r. by the linear filter lemma (proposition 1.7.7. we have that for all c E IR 2n and for any positive integer k t0 + k 6 _k1 0 Hence." for the parameter error to converge to zero. persistency of excitation is directly related to the positive definiteness of the autocovariance. for all that > 0. We will also discuss partial parameter convergence when the input is not sufficiently rich.7 Frequency Domain Conditions for Parameter Convergence Theorems 2.3) Since w has an autocovariance lim T+ oo A reference input r results in parameter error convergence to zero unless its spectrum is concentrated on k < 2n points. ••• .2). Proof of Proposition 2. there exists o such (2.5 give a condition on the regressor vector w. Then. }wk.16).6 and restrict our attention to stationary reference signals. The result of this section will make precise the following intuitive argument: assume that the parameter vector does converge (but not necessarily to the nominal value)." If the reference input contains frequencies w 1 .
2 2 JIHw.3 and 2.4. (b) implies (a) We also prove this by contradiction.{jw)cl Hw.8). 2n (2.13) PE and Sufficient Richness Let w (t) e IR n be the output of a stable LTI system with transfer function ifw.(dw) = 0 (2.8) allows us to relate the frequency content of r to the persistency of excitation of w. Assume that r is not sufficiently rich of order 2n. .w0 • On the other hand..1)) are linearly independent in <r 2n. positive measure. have that Hwr(Jw.7. w2n e IR. .(s) given in (2.}) ~ • ~ T (2.2) and (2. 7.92 Identification Chapter 2 Section 2. (dw) •.7. 7.7. . (Jw. b>. w2n such that Rw(O) = ~ Hw. .11) Since S.• .8). Assume that ifw.)S.(jw)HwrUw)S. Proof of Theorem 2. . we are left to show that. • • • .1) or (2. . .( { W.3 Using theorems 2. .5.10) implies that The expression (2.{jw)c ~ ~ T · 2 S. namely.2 and the gradient algorithms (2. the vectors fi (Jw. if the support of the spectral density of r. . (dw) is a continuous.. (dw) has at least 2n points. 2n. . Theorem 2.7.14) for some c e <r 2n and for all i ~ 1.10) filter lemma (proposition 1. with assumptions (A1) and (A2) and P stable. are linearly independent on <r 2n for all w 1 . Theorem 2. This leads us to the following definition.7.12) w2n.1). 's.(Jw. 2n. If r is stationary and sufficiently rich of order 2n the identifier parameter 8 converges to the nominal parameter Then e• exponentially fast.7. (jw 1).9) is the sum of k dyads. •• . Using the fact that (A.7.)Hwr(Jw. w2 . It remains to be verified that this property holds for the ifw. 7). We show the result by contradiction.7.) are in controllable canonical form and that Pes) = kpfip(s )/ dp(s ).1. 7. W2n E IR.(dw) is w1 . . .)C = 0 ~T . (Jw. we see that Rw(O) = • JHw..7.4.7 .7.2.) for i = 1.8) Using (2.. say w 1 .2 By proposition 2. . Definition Sufficient Richness of Order k A stationary signal r : IR + . 7). for every w 1 . (a) implies (b) We prove this by contradiction. . the identifier structure of Section 2.7. .6. (a) is equivalent to Rw(O) > 0.2). The right hand side of (2.. 0 We saw that sufficient richness of order 2n in the reference input translates to w PE.) are independent for every set of 2n w. (a) w is PE if and only if (b) r is sufficiently rich of order 2n. (dw) contains at least k points. we see that (2. a DC signal contributes only one point in the support of the spectral density.6)(2. (s) and stationary input r(t ).7. . HwrUw2n). provided that the ifw. it follows that k i =I contradicting the hypothesis of linear independence of ifw. Comment Note that a single sinusoid in the input contributes 2 points to the spectrum: at + w0 and at . (2. S. from (2. • • • . Proof of Theorem 2. 7. Then. = 0 for all w e support S.4. 7 Frequency Domain Conditions cT Rw(O)c = 0 93 (2.IR is called sufficiently rich of order k.(dw) ~ ~ T (2. 7. that is the support of S. we may conclude that T .7.7. wk with k < 2n.9) w. contradicting the PE of w. defined in (2. there exists c e IR 2n such that (2. or the normalized LS algorithm with covariance resetting (2. If w is not PE.1. . i = 1.4.) (with fi w.7.3 Exponential Parameter Convergence and Sufficient Richness Consider the identification problem of Section 2. we Since the support of S. Assume that there existed w 1 . so that the rank of Rw (0) is at most k < 2n.
kpfip(s).I.7. snIJp(s). 0 has dimension 2n . . 7..) A Identification Chapter 2 Section 2.7 Frequency Domain Conditions 95 converge to the set of fJ's for which the identifier transfer function matches the plant transfer function at the frequencies s = jw 1.(r) (2.7.7. sdp(s). pactly as jw 2n. ..)T are linearly independent.7. we see that "the plant output is given by (2.20) and the identifier output by Y. 7. S. dp(s) is of degree n and fip(s). .. if the identifier output matches the plant output at = jw 1. 7. Consequently.. we have kp fip(S) dp(s) _ C(l )(S) Let us then call 0. }wk> the asymptotic value of fJ must satisfy c0 l(s) and = (} (}* (2. . (jwi)nl . let us discuss the idea informally. .7.21) C(I)(S) Jp(S) + C(2)(S) kp fip(S) 0 (2. .16) for s = Jw1. In terms of the parameter error vector ¢ description 0E 0 ~ = fJ. is concentrated on k < 2n frequencies w1. wk. Before stating the theorem. Therefore...22).7. If w is not PE.15) for i I . recalling that .) ( }WI p dp(}w. Equation (2.7.. dp(s) are coprime. it should (2. 8* e 0 and for all s (2. . (2. 2n. 7.k.17) for s = jw 1. .94 (2.14) becomes [I. . .(Jw.22) (2. it must be identically zero.I 7) is of degree at most 2n .. . = (2. From the structure of the identifier.19) cannot be satisfied since the degree of c(2) is at most n . This establishes the contradiction..19) 0 8* + null space HwrUwk) A c(2)(s) By assumption.18) The polynomial on the lefthand side of (2.23) T The k row vectors Hw.7. D Comments It would appear that the developments of this section would enable us to give an explicit expression for the rate of exponential convergence: it is in fact possible to carry out this program rigorously when the rate of adaptation is slow (see Chapter 4 for a further and more detailed description of this point). . 0 has the simple (2. . We now consider the case when the spectrum S.7.16) may be written more comfJT(t)Hw. . . (dw) is concentrated on k < 2n points.24) Rw (0) ¢ = 0 We leave the verification of this fact to the reader. This is indeed the case. In this case. . 2..7. then the parameter error need not converge to zero.7. that is [dp(s). which satisfy (2. Consequently.8*.Jwi. ·)k fip(Jw.I.7.2 Partial Parameter Convergence 0 (2.. . }wk. Jw2n' where s cl + c2s + Consequently. Since it vanishes at 2n points. the set of fJ's. Clearly. Intuition suggests that although fJ need not converge to 8*..
theorem 2.0 as t . 7.7. with assumptions (Al)(A2) and P stable. and so does the derivative of the parameter error ¢. While the 2n .4 is an alternate proof of the parameter convergence results of Section 2. We will show that.3.5. and the updating slows down. it also implies that I Rw (0) ¢(t )I tends to zero as t .4 does n~t rely on theorem 2. These are very common properties of identification schemes.96 k Identification Rw(O) Chapter 2 (2.7. HwrUw.28).5) and rPT w . or the normalized LS algorithm with covariance resetting (2. S. we find that there exists t 1 such that.)S.4 Partial Convergence Theorem Consider the identification problem of Section 2.4.29) = 2. 7.4. the rate of convergence of 0 to e depends on both.4.}) i =I I ~(t)l ~ 6kE3 T Proceeding more rigorously. (2.4.7. that the output error e 1 = ¢T w .25) Section 2. 7 Frequency Domain Conditions 97 (2. for all E > 0. the identifier structure of Section 2.7.7.7.)H. .oo.31). we have that.¢(t)T ~ I lo w(r)w(rf dr ¢(t) (2. this implies that.4.4 The proof relies on the fact. ( { wk } ) contained in the reference signal at those frequencies. Theorem 2.7.32) D IRw(O)for all t 0 ~ ~ lo+ T I lo w(r)wT(r)dr I~ I (2. proved in theorem 2. with the additional assumption of stationarity of the reference input r (t ). t1 ¢T(t)Rw(O)¢(t) ~ which completes the proof of the theorem.1) or (2. for all t ~ t1.0 as t .7. I w(t)l ~ k.(( w.31) fort~ Now. 7). ••• .oo.oo (cf. there exists t 1 ~ 0 such that.1.28) and Comments The proof relies on the fact that for both the identification schemes discussed the parameter error eventually becomes orthogonal to the regressor w.0 as t .7. let k be such that I ¢(t}l. fort 1 T J 1 I+ T ¢T(r)w(r)wT(r)¢(r)dr ~ . for all t ~ t 1 (2. fort ~ t 1 Proof of Theorem 2. for T large enough  < ~ 3 ~ (2.(Jw. for all t ~ 0. (2.6)(2.3 relating PE and exponential convergence. we have the following theorem. ( { w1 } ) . As opposed to the proof of theorem 2. we have that lrP(r)¢(t)l ~ E(rt)/6k 3 T for all r ~ t ~ t 1• Together with the boundedness of¢ and w. If r is stationary Then lim Rw (0) ¢(t) = 0 I+ 00 From (2.k dimensional set 8 to which O(t) converges depends only on the frequencies w1 .oo and since Rw (0) is symmetric positive definite.7. Since w has an autocovariance.2). 7.29).5.30) and (2..4. This means that ¢T(t)Rw(O)¢(t) converges to zero as t .5. the proof of theorem 2. wk and not on the average powers S.30) t1 I Using (2. 7.2 and the gradient algorithms (2.7. combining (2. ¢T(t)Rw(O)¢(t) ~E. E (2. we have that.7.27) From the fact that ¢ . the proof of theorem 2.27). If w is PE and Rw (0) > 0.7. Since ¢ and w are bounded.oo. ••• .26) 0 and therefore lo+T l¢(tf Rw(O)¢(t). 7.
Finally. Several gradient and leastsquares algorithms were reviewed and common properties were established. the objective is to design an adaptive controller such that the behavior of the controlled plant remains close to the behavior of a desirable model. the SPR error equation was found to have similar stability and convergence properties. If the input was a sum of sinusoids. CHAPTER3 ADAPTIVE CONTROL 3. Then. Guaranteed rates of exponential convergence were also obtained and showed the influence of various design parameters. the regressor and the parameter error. so that a frequencydomain analysis could be carried out. for example. When the regressor was not bounded. that are valid under general conditions. The scheme involved a generic linear error equation. the PE conditions were transformed into conditions on the input. In particular. The unknown plant P has input u (t) and output Yp(t ).0 INTRODUCTION In this chapter. The exponential convergence of the parameter error to its nominal value followed from a persistency of excitation condition on the regressor. It was shown that parameter convergence was guaranteed. if the input contained the same number of spectral components as there were unknown parameters.1) where kp and ap are unknown. It was shown that for any of these algorithms and provided that the regressor was a bounded function of time. with all generated signals remaining bounded. The reference model is a stable SISO LTI system of identical order 99 . relating the iden~ifier error.8 CONCLUSIONS In this chapter. Our attention is focused on model reference adaptive control. s + ap A (3.0. the identifier error converged to zero as t approached infinity. we start with an informal discussion for a first order system with two unknown parameters. with input r(t) and output Ym(t ). The stability and convergence properties were further extended to strictly positive real error equations. PE appeared as a fundamental condition to guarantee exponential parameter convergence. but satisfied a regularity condition. despite uncertainties or variations in the plant parameters. In this section however. More formally. This will allow us to introduce the algorithms and the stability results in a simpler context. their number should be greater than or equal to the order of the plant. We will consider linear time invariant systems of arbitrary order.98 Identification Chapter 2 2. In particular. The parameter error was also guaranteed to remain bounded. Although more complex to analyze. we derived a simple identification scheme for SISO LTI plants. and establish the stability and convergence properties of the adaptive algorithms. The control objective is to design u (t) such that Yp(t) asymptotically tracks YmU ). Consider a first order singleinput singleoutput (SISO) linear time invariant (LTI) plant with transfer function kp p = . a reference model M is given. the reference input was found to be a dominant factor influencing the parameter convergence. We assumed stationarity of the input. then it was shown that a normalized error still converged to zero. we derive and analyze algorithms for adaptive control.
1) where kp and ap are unknown. It was shown that for any of these algorithms and provided that the regressor was a bounded function of time. with all generated signals remaining bounded. In particular.8 CONCLUSIONS In this chapter. Guaranteed rates of exponential convergence were also obtained and showed the influence of various design parameters. PE appeared as a fundamental condition to guarantee exponential parameter convergence. Our attention is focused on model reference adaptive control. Although more complex to analyze. we derive and analyze algorithms for adaptive control. the SPR error equation was found to have similar stability and convergence properties. with input r(t) and output Ym(t ). We assumed stationarity of the input. so that a frequencydomain analysis could be carried out. When the regressor was not bounded. the objective is to design an adaptive controller such that the behavior of the controlled plant remains close to the behavior of a desirable model.98 Identification Chapter 2 2. Several gradient and leastsquares algorithms were reviewed and common properties were established. we start with an informal discussion for a first order system with two unknown parameters. In particular. It was shown that parameter convergence was guaranteed. for example. the regressor and the parameter error. We will consider linear time invariant systems of arbitrary order. that are valid under general conditions. a reference model M is given. The control objective is to design u (t) such that Yp(t) asymptotically tracks YmU ). but satisfied a regularity condition. if the input contained the same number of spectral components as there were unknown parameters. relating the iden~ifier error. The stability and convergence properties were further extended to strictly positive real error equations. The exponential convergence of the parameter error to its nominal value followed from a persistency of excitation condition on the regressor. In this section however. then it was shown that a normalized error still converged to zero. The scheme involved a generic linear error equation. The parameter error was also guaranteed to remain bounded. CHAPTER3 ADAPTIVE CONTROL 3.0. The unknown plant P has input u (t) and output Yp(t ). Consider a first order singleinput singleoutput (SISO) linear time invariant (LTI) plant with transfer function kp p = . If the input was a sum of sinusoids. the PE conditions were transformed into conditions on the input. s + ap A (3. we derived a simple identification scheme for SISO LTI plants. The reference model is a stable SISO LTI system of identical order 99 . Finally. More formally. Then. the identifier error converged to zero as t approached infinity. and establish the stability and convergence properties of the adaptive algorithms. their number should be greater than or equal to the order of the plant. the reference input was found to be a dominant factor influencing the parameter convergence.0 INTRODUCTION In this chapter. This will allow us to introduce the algorithms and the stability results in a simpler context. despite uncertainties or variations in the plant parameters.
r+¢yYp) t + km r(t) (3.11) We may represent (3. r + ¢y Yp) and filtering by the LTI operator M. 7) which becomes Yp(t) = am Yp(t) + km r(t) (3. Therefore.0. ¢y) am (yp . the error formulation is eo . it is convenient to introduce an error formulation.6) such that the closedloop transfer function matches the reference model transfer function.g e0 r .4) k kp m M (¢. so that Ym is also bounded.am Ym(t) s +am kp r ·A/(¢.am eo where km and am> 0 are arbitrarily chosen by the designer.0.ap Yp(t) + kp(¢.3)(3.12) is not exogeneous.5) the motivation being that there exist nominal parameter values (3.100 Adaptive Control Chapter 3 Section 3. Note that there is a significant difference with the model reference identification case. but is itself a function of e0 • However. Define the output error Yp. namely that the signal Yp which appears in (3.5) yield .12) is of the form of the SPR error equation of Chapter 2..g eo Ym Subtracting (3.r + ¢yYp) A (3. Specifically.r + ¢y Yp) A 1 =  The next steps are similar to those followed for model reference identification in Section 2.0. the righthand sides only contain states (e 0 .0.0.0.5) and (3.12) (3.ap c 0(t) c~ ] ¢.0..0. we tentatively choose the update laws c0 .0. For the analysis.kp do(t)) Yp(t) Although this notation will be very convenient in this book. (3.0.km r v(e 0 . ¢.1 0) .d~ (3.0.14) [ do(t) .0.8) when co(t) = c~. Consider then the Lyapunov function kp Cor .0.3) and the reference model by Ym(t) = . ¢. while the second condition restricts the class of models which may be chosen.g e0 r (3. In the time domain.13) + kp co(t) r(t) (3.( ap . the stability proof proceeds along exactly the same lines. Ym).3) and (3.0.Ym) + + kp do) Yp + (3.15) .Ym assuming that kp I km > 0 and that M is SPR.r + ¢y eo + ¢y Ym) ¢ = [ ::~:~] (am . d 0(t) = d~. we caution the reader that it rnixes time domain operations (¢.0.0 Introduction 101 M (3.0.0.13). First assume that r is bounded.0. . ¢y) and exogeneous signals (r. Equation (3.g eo Yp g>O (3.ap Yp(t) + kp ( co(t) r(t) + do(t) yp(t)) .0. The first condition requires prior knowledge about the plant (sign of the high frequency gain kp).0.2) .9) and the parameter error + kp(¢. Alternatively.4) from (3.r + ¢y Yp) (3.am eo (3.0. Let the control input be given by u(t) = co(t) r(t) + do(t) Yp(t) c~ M (¢.11) in compact form as eo= + kp u(t) (3.0.7) eo =  In this representation.0. the plant is described by Yp(t) = .g e6 . The adaptive system is described by (3.0.0.
18) to the signal u (3. and </>y are bounded. Yp  d.. which will be discussed later. or to the other schemes. 0 is also bounded.0. singleoutput (SISO).0 as t . but has significant advantages.9) in the output error scheme). it is not straightforward to extend it to the case when the relative degree of the plant is greater than 1.kp <!>y e6 . the equations in (3. and relatively simple.kp Q>.0.17) determines how e2 is calculated in the implementation (as (3. we relate e2 to the parameter error (as (3. Therefore. along the trajections of (3.6) would determine the nominal controller parameters c~ and In an indirect approach. thereby defining c0 and do.0. eo r .kp <!>y eo Ym ~0 + <!>y M(yp) (3. However. one replaces kp and ap in (3. No condition is posed on the reference model by the identification/adaptation algorithm.19) Assumptions (AI) Plant Assumptions The plant is a singleinput.14). applying (3.0.18) where we used (3.0. Instead of using e 0 in the rdaptation procedure.0. From (3.0.0.0. lim v{t) as too exists. Using any of the procedures discussed in Chapter 2. we may now use any of the identification algorithms. for which update laws are designed to directly update the controller parameters c0 and d 0 . This is a very intuitive approach to adaptive control which we will also discuss in this chapter.16) It follows that the adaptive system is stable (in the sense of Lyapunov) and that.0. Conditions for parameter convergence will also follow. For the derivation of the adaptation scheme and for the analysis. we may design a recursive identifier to provide estimates of the plant parameters kp and ap· If kp and ap were known. so that e0 E L 2 • Since e0 e L 2 n L 00 . we will use tools from functional analysis. the simple Lyapunov stability proof presented for the output error scheme does not extend to the general case. linear timeinvariant (LTI) system.0 and with (3.am M p kp (3.17) Introduction 103 so that. The above approach is elegant.0. Unfortunately. eo r + kp <!>y e6 + kp </>y eo Ym (co . and is not addressed in this introduction. we will study the stability and convergence properties of the adaptive systems. The global stability of the adaptive control schemes will be established. described by a transfer function . e0 . After extending the above algorithms to more general schemes for plants of arbitrary order.17) The motivation behind this choice will be made clear from the analysis.am kp kp s + am s + ap + p+ ap . Further. together with a set of standard lemmas which we will first derive.6) by their estimates.0.12) in the output error scheme).1 MODEL REFERENCE ADAPTIVE CONTROL PROBLEM We now turn to the general model reference adaptive control problem considered in this chapter. and e0 e L 00 . Q>. it follows that e0 .c~ )Yp + (d 0 </>.0. The resulting scheme is slightly more complex for this first order example. we use the error e Equation (3.0.oo. (3. including the leastsquares algorithms. Equation (3. Since v is monotonically decreasing and bounded below. An alternate approach is the indirect approach. The following assumptions will be in effect. We now discuss an alternate approach which we will call the input error approach.0.6).) M(yp) (3.0.20) . Instead. 3. Therefore. First note that M = km s + am km kp kp s + ap km kp km s + ap ~ kp s + am s + ap km ap . Proving stability is however more difficult. it places restrictions on the reference model which are in fact not necessary. together with input signal conditions similar to those encountered in Chapter 2 for identification.20) is of the form of the linear error equation studied in Chapter 2. d.102 Adaptive Control Chapter 3 Section 3.14) v =  am e6 + kp </>. for all initial conditions. The algorithms described above are both direct algorithms.
(s) and X are polynomials of degrees (s) n .(u) dp (3.1. am(s) are monic. ap(s) are monic. this would correspond to the exact cancel(s lation of modes of c(s )I X and a (s )I X ).1) where Co is a scalar. Reference Input Assumptions The reference input r(. we see that u = cor+ where X (s) is an arbitrary Hurwitz polynomial of degree n .1: Controller Structure By inspection of the figure.2. coprime polynomials of degree m and n.e to 1achieve the control objective. it is clear from (3. a (A2) where fip(s). so that we (s) write (3.1. respectively.5) Figure 3.2: Since Controller StructureEquivalent Form np Yp = kp::. The reference model is stable. respectively.2. n . Physically.1 ).2.2. without loss of generality.2 CONTROLLER STRUCTURE To achieve the control objective. we will assume kp > 0. For this.2.2. The sign of the socalled highfrequency gain kp is known and.m .2 Controller Structure 105 ~p(s) dp(S) (3. that is. minimum phase and km > 0. Reference Model Assumptions The reference model i~ described by Ym(s) f~) = M(s) = u = A x (cor+:::.2) (A3) where fim(s ). that it is possible to make the transfer function from r to Yp equal to M(s). c(s ). we consider the controller structure shown in Figure 3. but is not assumed to be stable. Yp the transfer function from r to Yp is Yp c0 kp Xfip (xc)apkpnpa (3.104 yp(s) U(S) Adaptive Control = P(s) = kp Chapter 3 Section 3.2. o ~(s) A (s) (u) + ~(s) A (s) (yp) (3. the same degrees as the corresponding plant polynomials).1. For numerical considera(s) (s tions.2.4) relies on the cancellation of polynomials X ). The plant is strictly proper and minimum phase. coprime polynomials of degree m and n respectively (that is. 3. Figure 3. (s) The following proposition indicates that the controller structure is adequat. we will therefore require that X is a Hurwitz polynomial. km ~m(s) dm~) (3.4) f Note that the derivation of (3.2) Ac which is shown in Figure 3.(yp)) A a (3.I.2.4) that X must contain the zeros of fim(s ).I and n .) is piecewise continuous and bounded on IR+. From (3.1) .2.1.3) Note that P(s) is assumed to be minimum phase.
Co be given by (3.11) characterizes the set of solutions in this case.2. the controller structure can be expressed as in Figure 3. aP are coprime. Xo and fip. M if and only if the following Comments oc od ~ =  k fip p dp ~ P (3.1). so that are also unique.1 and the lefthand side is a polynomial of degree at most 2n .13) co km kp A (3. 7).2 extra modes are those of X. b) Using (3. 7)(3.1 1. p. r .2. Thus the forward block actually cancels the zeros of P and replaces them by the zeros of M.2. Consequently.10) becomes Let. c) The transfer function from r to Yp is of order n.8) show that the compensator becomes c a (3. fip. so that (3.2. (3. a . c.kp n.1 0) Recall that ap.9). then. If this assumption is not satisfied.2 states. and oc and oa have degrees at most n. Xo and am have degrees n. 0.2. When matching with the model occurs.2. It can be checked (see Section 3.8) (3. In particular.2.2. Divide Xoam by aP.2.6) The solution can be found by inspection. The modes corresponding to X. 0 c• and The transfer function from r to Yp is matching equality is satisfied (3.2. with a forward block X/X. for example.2 Controller Structure 107 There exist unique c0.c· and A (3.14) The motivation in using the previous controller structure is to obtain an expression that is linear in the unknown parameters. oc 0 = No solution exists unless oc0 = 0.1) and .2. so that c0 is unique.2 and n . and represented in Figure 3. (3. while the plant and controller have 3n .6).2. The polynomials found in this case are related to the previous ones through (3. The coprimeness of fzp. m.2. Thus is given by a• a. 7)(3.2.2.2.2. 2. the righthand side is a polynomial of degree 2n. n.and a feedback block /X.2.9) define a solution to (3.2. Proof of Proposition 3. Uniqueness Assume that there exist Co = Co+ oCo. a = a • + oa satisfying (3.5) that the 2n .9) kp Xofim Equations (3. Equation (3. Existence a•(s) such that the transfer function from a.2. in Callier & Desoer [ 1982]. respectively.1 and n.2. as can easily be seen by substituting c0.m.11) This equation has no solution since fip.1) shows that the control signal is the sum of the parameters multiplied by known or reconstructible signals. c• and d• in (3.1 Adaptive Control Matching Equality Chapter 3 Section 3.6).2). a solution can still be found using (3.7) Let x.2. let q be the remainder (of degree the quotient (of degree n . The expression in (3. it is equivalent to the familiar structure found.m .2. C = c• + oC.2.106 Proposition 3.1. aP is only necessary to guarantee a unique a) solution.2.6).l.Yp is M(s). c•(s ).12) c• (of degree n  2). These parameters are the coefficients of the polynomials and the gain Co. Xo are stable by choice and those of fip are stable by assumption (A 1). d) The structure of the controller is not unique.3. with m ~ n. 164.2. The following equality must then be satisfied ocdp+kpfipod ~ ~ = kp ~ ~ ~ oCokAonpdm m (3.
2n (3.24) .23) and w(2) A w< 2> + b.18) (3.2. We also define the parameter errors ' r/J := () . we let d(s) A(S) = do+ dT(sl A) 1 b. we consider a statespace representation of the controller. o•r := (c 0.4: U = Controller StructureAdaptive Form Cor+ CT w<I) + d 0yp + JT w<Zl (3. The controller can be represented as in Figure 3..CT. Similarly.. E lR 2n (3.. w<2>r) E lR.n.yP be the vector of nominal controller parameters that achieves a matching of the transfer function r.1.1 and b. this transfer function can be realized by ()T := (co.19) where the state w< 2> e lR n.2..2.dT) is the vector of controller parameters and wT := ( r.2n I (3.d0.2..2. b. It follows that 1 A (s) s (3. there exist d0 e lR and d e lR.. In analogy to the previous definitions.108 Adaptive Control Chapter 3 Section 3.. e lR n.2 Controller Structure 109 (3.A).(u) A A wO> + b.3: Alternate Controller Structure Yp StateSpace Representation To make this more precise.A) = ~ (s ).2.2..17) where the state wO> e lR. Choose A e lR n..yP.()Tw (3.1 x n..do..0. with r u Figure 3.d..20) Let c that E lR n I be the vector of coefficients of the polynomial c(s ).2.r) := (c 0.r) e lR 2n (3. .16) A(s) Consequently. so where c(s) I A=c T ( s.I b.1 and the initial condition w< 2>(0) is arbitrary.c•r. such that (A.) is in controllable canonical form as in (2..n.2.22) c (3.u wT) := ( r. such that is a vector of signals that can be obtained without knowledge of the plant parameters.21) w< 1> .2..4.Yp to the model transfer function M.2. w< 1>r. Note the definitions of 0 and w which correspond to the vectors () and w with their first components removed.1 and the initial condition wO>(O) is arbitrary..1..15) Figure 3..0• E lR 2n i := 0 o• E ]R.OT) := (co.9) and det (sf ...
dp(s )) are identified using a recursive identification scheme. the plant parameters (i.3) With the definitions of P and M.. or a modification of it.. By definition.3. Similarly. an input error adaptive control scheme is a scheme based on this error.3).3. The identification algorithm defines the evolution of the identifier parameters. In a direct adaptive control scheme. for efficiency.4) "X "X We may interpret this last equality as an equality of two polynomial ratios.5) aP.5) Now. An output error direct adaptive control scheme and an indirect adaptive control scheme will be described in Sections 3.2.3. Narendra... with (3.Ym) = M.1(yp) A).3. (3.2. & Valavani [ 1980].3.. an identification scheme is designed that directly identifies the controller parameters c 0 . within the identifier. The identifier obtains estimates of these parametersdirectly or indirectly. Consider the matching equality (3. .3.1 Input Error Direct Adaptive Control Define rp = a M.. and Morse [ 1980].2.Ym is the output error.19) give a meaning to (3. c.2.1P(u) (3.17) and (3. Given an identifier structure with linear error equation for example. usmg some controller parameters as in a nonadaptive context.3.6).3. In an indirect adaptive control sc~eme.3. Applying both transfer functions to the signal u . 7)(3.r I I M ~ (yp.e.5) ..1bx(yp) (3.YmCt) is the basis for output error adaptive control schemes such as those of Narendra & Valavani [1978]. 3. Equations (3.2. Although we make the distinction between controller and identifier we will see that.1(yp) M. The identifier structure constructs signals which are related by some error equation and are to be used by the identification algorithm. We now consider the case when the plant is unknown and the control parameters are updated recursively using an identifier.1 (this scheme is discussed in Bodson [ 1986] and Bodson & Sastry [ 1987]).20).. As in Chapter 2. The output error eo(t) = yp(t).2) where eo= Yp.3. Input Error and Linear Error Equation The interest of the input error in (3.. we showed how a controller can be designed to achieve tracking of the reference output Ym by the plant output Yp. we also make the distinction. The estimates are then used to compute the control parameters through (3. Section 2. such as those described in Chapter 2. when the plant transfer function is known.6) so that.(eo) ~ (3. we will consider adaptive control algorithms and the parameter 0 will be a function of time. The controller is by. but also as an equality of two LTI system transfer functions..3 in the context of identification. even in the case of direct adaptive control.kp::~A dp A c* c0 ~dp • kp fip am kmfim (3. (s) will be polynomials in s whose coefficients vary with time. and let the input error ei be defined by ei := rp.2) is to lead to a linear error equation such as studied in Section 2. both systems.P = c0M.1 b~.3.3 Adaptive Control Schemes 111 The linear dependence of u on the parameters is clear in (3. some internal signals will be shared b. c* fip a..3. Lin.110 Adaptive Control Chapter 3 Section 3. between the identifier structure and the identification algorithm. d 0 and d. but we will first turn to an input error direct adaptive control scheme in Section 3. from which the controller parameters depend.(u) .3 ADAPTIVE CONTROL SCHEMES In Section 3.2. and divide both sides by "X Using (3. we have u . definition the system that determines the value of the control input. c(s ). Note that we made the distinction between controller and identifier.3.1 P (3.~ (yp) a· ~ c* a· A A = coM.1) in that case.A). In the sequel.1) 3. we recall that we IR 2n.3) becomes . We first present an intuitive derivation of this linear error equation. kp and the coefficients of fip(s).2 and 3.(u)l (3. Several approaches are possible.3.9).2.1 is given by _ w = [ (sfYp (sf.2. several identification algorithms exist from which we can choose (cf. A typical procedure is to derive an identifier error signal which depends linearly on the parameter error ¢..2..
These will be unobservable or uncontrollable. =  I~· + ca<. as will be seen later).3. so that rp and e.9) becomes e. the gain of the operator M. the transfer function Mi has a proper and stable inverse.1 0) so that (3.2.8) so that o·T w cor + oT w.7) We now derive a modified input error. For example.3.1 of relative degree n .T(} w = c· A A (u) + A a· (yp) A (3. Although this is not a crucial problem. Therefore. using (3. + OTw + (coco)M. Consider (3.) is well defined.3. As presented.6) by X i dp (3. u . stable transfer function.3.4) to (3.3.112 Adaptive Control Chapter 3 Section 3.8) being satisfied at all times.I (yp) Fundamental Identity Since M is minimum phase with relative degree n.3. however.1 (yp)) + oTw + (coCo)M. If the lefthand side is realized by P followed by L 1 the unstable modes of P able. due to the presence of measurement noise. depending on the realization of the transfer function. These are decaying because the transfer functions are stable and the unstable modes are unobservable.11) This equation is of the form of the linear error equation studied in Section 2.M.wT) e IR 2n Divide both sides of (3. for any stable.3.20) (3.CoM I (yp) Co(rM.. The signal i I (rp) is available since (3.3.3. (i so that it becomes. since P may be unstable.1 (yp) = A L.1] fi = i I  i I ~· (3. while the lefthand side is possibly unstable (since P is not assumed to be stable).3.2.5). we can let i be a Hurwitz polynomial of degree n .m.wT) .1 is a proper.12) where (Mi). unobserv JA• £I aA·(ypt+co(Mi).3.) in the analysis.13) as an equality of two transfer functions.1 is arbitrarily large at high frequencies.1 (u)iIC: A (u)+t{t) (3. are not available.3.2.1(.13) can then be transformed to a signal equality by applying both operators to u . provided that the argument is sufficiently smooth. The righthand side is a stable transfer function.13) into an equality in the time domain.3. The operator equality (3.m. so that + Co (Mi). c) We were somewhat careless with initial conditions.1 is not desirable in practice.14) where t(t) reminds us of the presence of exponentially decaying terms due to initial conditions. care must be taken of the effect of the initial conditions related to the unstable modes of P. we will discuss the advantages of avoiding it.3. The control input is given by (3.13) I co </> T z (3. It could thus be used to derive an identification procedure to identify the parameter (} directly. To transform (3.3 Adaptive Control Schemes 113 .(M. we will consider it not implementable. b) The derivation of the error equation (3. its inverse is not proper.1 A will be controllable and.co . Although M.1 (yp) coe.11) relies on (3. the scheme and its derivation show several problems: a) Since the relative degree of M is at least 1 .1 (yp) (3. .1 (yP). minimum phase transfer function i.5) and the definitions of P and M.3. Therefore.1(. and resolve the technical question in c). Although we will use M.3. or z. leading to a scheme that does not have the above disadvantages a) and b). the use of M. therefore.3 (the gain I I co being known may be merged either with e.9) We now define the signal zT := (rp. going from (3.Mi).m.
Assumptions The algorithm relies on assumptions (A I )(A3) and the following additional assumption. d(t) E IRnl] v(t) E IR 2n.5.3.9).17) where 8* is defined in (3. 7)(3.1(wr)] EIR 2n (3. for all piecewise continuous u E L ooe . Therefore.1(yp).16) so that (3.16) and (3. with the required assumptions. Proposition 3.3.3.1(rp).1 (i*\v) A.2.3. in connection with alternate schemes.23). because it would require the realization of a nonproper transfer function. using (3.1(wT)] = [(Mf).3. The need for assumption (A4 }.1 (w) is given by ii*T i. not to be available. Both the signal i .f.3.17) is essential to subsequent derivations. while the complete algorithm is summarized hereafter.2. with (3. Yp(t) Output u(t) E IR Internal Signals w ( t) E IR zn ( w< 1>(t) . 7. Define now vT which is of the form of the linear error equation studied in Section 2.14) is valid for arbitrary initial conditions in the realizations of I I L.3 Adaptive Control Schemes 115 (3. We also observed in Chapter 2 that standard properties of the identification algorithms are not affected by the f(t) term.3.3. The structure of the controller and identifier is shown in Figure 3. e2(t) E IR so that.3.23).be any stable.17) is of the form studied in Section 2.\.2. We now consider the practical implementation of the algorithm. with (3.2. c(t}.14). that kp ~ k max for some k max.3 for recursive identification.3. so that we summarize the result in the following proposition. (A4) Bound on the HighFrequency Gain Assume that an upper bound on kp is known. A A A A Since ii* is constant.17) is satisfied.15) can be written (3.f.7)(3..m. a· i.3.18) 8(t) E IR 2n [co(t). Equation (3.2. It will be more obvious from the proof of stability of the algorithm in Section 3.114 Adaptive Control Chapter 3 Section 3.3. Let L .m. Let v and w be as defined by (3. For simplicity. minimum phase transfer function of relative degree n . given these considerations. d 0 (t) E IR.19) Initial conditions are arbitrary. that is.I [ ex (u) + T (yp) ] (3.and (ML). w(2>(t) E IR n 1] E IR i l(u) (3.3.9).6). the approximate input error e 2 ..1(u) and v are available from measurements. 7 ii* i. we define the modified input error to be e2 8T V  I n. and for the projection of c 0 . and the expression is linear in the unknown parameter 8 •. Then Input Error Identifier Structure Equation (3.3.2.3. will be discussed later.kmax Input r(t}. Input Error Direct Adaptive Control AlgorithmImplementation Assumptions (AI)(A4) Data := [f. we will omit this term in subsequent derivations.15) where we used (3.3. except co(O) ~ c min = kml k max > 0. and the signal v are available. Let 8 • be defined by (3. (3. Although we considered the input error e.1 Fundamental Identity A A A Let P and M satisfy assumptions (A I) and (A2).17) (3.1(w) = i. .. with arbitrary initial conditions in the realizations of the transfer functions.
while those used to compute e 2 are associated with the identifier. This is also useful.3. The signal generators (sometimes called statevariable filters) for w< 1> and w< 2> form a generalized observer. The states are used for the state feedback of the controller to the input in a certainty equivalence manner. in case of input saturation (cf. = g = e2 v 7 i+')'V V = . c. The identifier with the generalized observer is sometimes called an adaptive observer since it provides at the same time estimates of the states and of the parameters.1 stable.3 Adaptive Control Schemes 117 u = or w Structure ((M i).17) and (3. i l(w< 2lr)] Normalized Gradient Algorithm with Projection 0 .1(ll) Ident~[ier v1 = e2 = i l(w(ll). • A E km. The gains c0 . d 0 and d serving to generate u are associated with the controller. 0Tv_i. the identifier can be used offline.A) is Hurwitz and contains the zeros of fim(s ).116 Adaptive Control Chapter 3 Section 3. Goodwin & Mayne Figure 3.3. U A w< 2l + b>. for example. the identifier and the controller can be distinguished.5: Design Parameters Choose Controller and Input Error Identifier Structures • M (i.19) was derived using the fundamental identity (3.19). which is valid no matter how u is actually computed. E IR n  1 in controllable canonical form such that det (sf. b>.1(yp). meaning that the parameters used for feedback are the current estimates multiplying the states as if they were the true parameters. • g' 'Y > 0. if Co 0 c min and co< 0 . without actually updating the controller parameters if necessary. This parameterization has the characteristic of allowing the reconstruction of the states without knowledge of the parameters. i l(yp). 1 x IR n  n 1 .I instead of n previously). Separation of Identification and Control Although we have derived a direct adaptive control scheme. it is not necessary that these be identical for the identifier error to be as defined in (3.e. ' • i .3. Comments Adaptive Observer The signal generators for w<ll and w< 2 > ((3. In other words. They are shared by the controller and the identifier. minimum phase transfer function of relative degree n m.19)) are almost identical to those used in Chapter 2 for identification of the plant parameters (their dimensions are now n . fim.17). reconstructing the states of the plant in a specific parameterization. Jm) satisfying (A2). Controller Structure w<l) w(2) = A w<ll + b>.2. then set Co 0.2. This is because (3.Yp . In fact.
oTi l)(w) + yvTve1) (3.6 can there~ore be used. Ya =  eo = Yp. such that Mi is SPR. in contrast to the original output error eo = Yp. Practically.. M(c/>T w) e1 1 = .22) 1s not implemented as such. The practical implementation of the algorithm is summarized hereafter.!.3. since this requires M to have relative degree at most l . The controller structure of the output error scheme is identical to the contr?lle~ structure of the input error scheme. for example).2 Output Error Direct Adautive Control An output error s~h:_me is based on the output error e0 = Yp. this scheme does not work for plants with relative degree greater than 1.M(r). As previous!y.22) Assumptions The algorithm relies on assumptions (A1)(A3) and the following assumption. M(OTw) co +..17). is denoted (3. the control input u is set equal to u = () T w but now this equality is used to derive the identifier error equation ' ' Yp.Ym· Note that by applying M L to both sides of (3.Ym· The error (3.!.3. & Valavani [1980]. We now review their scheme for the case when the highfrequency gain kp is known. If this condition is satisfied..M (r) M ( (Co..(w)yv ve 1) co co A A l 1 (3.!. However.3. (3. 0 Again. M i co l + Co(Mi).3.20) e1 = Yp.!.1(w).1 ) (w ) + 'Y v T v e 1 )  ..Ya is called the augmented error.3. The ap~roach can ho~ever be saved by modifying the scheme.3.T ()TLA . v is identi~al to v. provided that M is SPR. 7. (AS) .3 e1 Adaptive Control Schemes (OTv. It relies on the identifier error 1 MALA ( (LA . which appeared starting with the work of Monopoli [1974].!.3.17) to obtain Assumptions (A1)(A3). T pute d. we use (3.() * ) w)  T co (3. we find .Co) r + ( () A A  T A co T .3.25) co and the resulting error e 1 = Yp. Instead. but with the first component removed..23) As before. the Identifier wlll still have consistent input signals. co ML(cJ> vyvTve 1) A A T (3.. Output Error '})irect Adaptive Control AlgorithmImplementation which is now of the form of the modified SPR error equation of Chapter 2. mpu~ u := () w (due to actuator saturation.l.. Mi(oT i co . Mi ((i loT. as for example m Narendra.Ym... the identifier error involves the output error eo = Yp.Ym =  l 1 M ( u. a stable.Ym.3.22) is of the form of the modified SPR error equation of Chapter 2 provided that Mi is a strictly positive real transfer function. provided that the signal u entering the identifier is the actual input entering the LTI plant.3.1 ().Ym· The additional term. the control signal is set equal to u = () Tw = Cor + oT w ' and the equality is used to derive the error equation for the identifier M(u) A = CoYp + M(O* w) A  T (3. the properties of the identifier will follow and are the basis of the stability proof of Section 3.yvTvet) As before. Lm.118 Adaptive Control Chapter 3 Section 3. and we let c0 = c0.() • w) ..m . The gradient identification algorithms of Section 2.1(yp)T A yvTvet) T yPM(u)+ML(O L.3.i l(u) A 119 [ 1987]).21) co ~hich has the form of the basic strictly positive real (SPR) error equation of Chapter 2. while the identifier structure 1s d1fferent. 3. (AS) HighFrequency Gain and SPR Assumptions 1 Assume that kp is known and that there exists i .24) c* . If the actual input to the LTI plant is different from the com. minimum phase transfer function of relative degree n .
e. & Valavani [ 1980]) requires an overparameterization of the identifier which excludes the possibility of asymptotic stability even when persistency of excitation (PE) conditions are satisfied (cf.A) is Hurwitz.24) from (3. Further. When the input error scheme is used. the solution proposed by Morse [1980] (and also Narendra. Ya(t). are available.19) is based on (3.1 stable.1 . These algorithms are an advantageous alternative to the gradient algorithm.3. in controllable canonical form.120 Data Adaptive Control Chapter 3 Section 3. and contains the zeros of fim(s ). such that det (sf .1 (w) Ym Ya = M(r) = _!__ cci M Lei lcoTw) oT i l(w).Ym· Using the error between the plant and the reference model to update controller parameters is intuitive. it was shown wT = ( w<llr' Yp' w<2l) c(. Anderson.3. and the identifier can be used "offline. do(t) e IR] v (t) E lR 2n ."fVTVel) . • g' 'Y > 0.m .Ym. 1 . = kml kp>O U =Cor+ OT W Identifier Structure vr=i. 1 fim. However. & Tsoi [ 1985]). stability proofs suggest that SPR conditions must be satisfied by the model and that an augmented error should be used when the relative degree of the plant is greater than 1. and the connections between exponential convergence and robustness (see Chapter 5). In view of the recent examples due to Rohrs. the parameters used for identification and control can also be separated. Controller Structure w(l> = A w< 1> + b11 u w(2) = A w< 2> + b11 yP OT = (CT' do. The derivation of the input error scheme shows that model reference adaptive control can in fact be achieved without formally involving the output error and without SPR conditions on the reference model. m. this appears to be a major drawback of the algorithm. Boyd & Sastry [ 1986]. provided that the actual input entering the LTI plant is available and used in the identifier.3." A second difference appears between the input and output error schemes when the highfrequency gain kp is unknown. Another advantage of the input error scheme is to lead to a linear error equation for which other identification algorithms. The first is that the derivation of the equation error (3. dm) satisfying (A2) and (AS). kp Input r(t). If the input saturates.3. dT) Traditionally. and the relative degree of the plant is greater than l. • A e IR n.yp(t) e IR Output u(t) e IR Internal Signals w(t) E IR2n I [ w(ll(t). w< 2>(t) e IRn I] O(t) E IR 2nl [(c(t). • i. The error e 1 derived in (3. such as the leastsquares algorithm.1 . This is because (3. Lin. such that Mi is SPR. updates of the identifier will be erroneous.22) is not implementable if c0 is unknown.Ya Gradient Algorithm Adaptive Control Schemes 121 n. d(t) E IRnl.I el(t). this problem can be avoided. Important differences should be noted between the input and output error schemes. Ym(t) E IR o= 0  ge 1v Differences Between Input and Output Error Initial conditions are arbitrary. km. Design Parameters Choose • M (i. at all times. If needed.xn  b11 e IR n. Although an SPR error equation can still be obtained in the unknown highfrequency gain case. Dasgupta. minimum phase transfer function of relative degree n .3 e1 = Yp. the starting point in the derivation of model reference adaptive control schemes has been the output error e0 = Yp.22) relies on the input signal u being equal to the computed value u = () T w .3.17) and does not assume any particular value of u.
d(t) e IRnl] Ym(t). C and d are then computed using the relationships resulting from the matching equality (3. b>. • A e JR n ..1 in controllable canonical form and such that det (sf. The highfrequency gain kp (and consequently co) can be assumed to be unknown. and dp are obtained using the standard equation error identifier of Chapter 2.2. Relative Degree 1lmplementation Assumptions (A1)(A3). do. i .3 Indirect Adaptive Control In the indirect adaptive control scheme presented in this section. Controller Structure w< 1l =A w< 1l + b>. as discussed now. instead of n .6). Yp.u w< 2l = A w< 2l + b>. w< 2l are on order n.Yp 8T =(Co. 3. however. The relative degree of i . We assume therefore the following: (A6) Relative Degree 1 and SPR Assumptions n .1. the input error scheme requires more computations. The controller parameters Co.I x n .. The simplicity of this scheme makes it attractive in that case. it is sufficient that this dimension be n.1 . dm) satisfying (A2) and (A6).1 is unnecessary along with the additional signal Ya· The output error direct adaptive control scheme then has a much simpler form. Note that the dimension of the signals wOl. np. the filter i I is one order higher. Bodson [ 1988]). but the sign of kp must still be known to ensure that c0> o.A) = nm(s ). but the degrees of the polynomials 8(t) e IR 2n [co(t).. w< 2l used for identification in Chapter 2 is n.1 for the output error scheme. Output Error Direct Adaptive ControlThe Relative Degree 1 Case The condition that Mi be SPR is considerably stronger than the condition that Mi simply be invertible (as required by the input error scheme and guaranteed by (A2)). W(I)T. In the case when the relative degree n .3.ge 0 w 0 Comment The identifier error equation is (3.24) and is the basic SPR error equation of Chapter 2. This is because the observers for w< 1l. Output Error Direct Adaptive Control Algorithm. e IR n .e.3 Adaptive Control Schemes 123 recently that there are advantages of the input error scheme in terms of robustness to unmodeled dynamics (cf.m . kp Input r(t). dT) WT = ( r. in which the error equation used for identification involves the output error eo = Yp. km. is only required to be n . CT.2. the order of the plant. M is SPR.m = I .1.Ym Gradient Algorithm 8 = . estimates of the plant parameters kp. do(t) e IR. c{t). . nm. For control. as compared to n . we will let their dimension be n.m for proper invertibility.. w<2l) u = oT w Identifier Structure Ym = M(r) eo= Yp. (A6) Data n. When the relative degree is 1. w< 2l(t) e IRn I] Design Parameters Choose • M (i.122 Adaptive Control Chapter 3 Section 3. Proposition 3. in order to share the observers for identification and cOI{trol. eo(t) E IR Initial conditions are arbitrary.I . Also. • g>O. In some cases. significant simplifications arise in the output error scheme. so that ( 1I c0)M is SPR.1 is still true then. However.yp(t) e IR Output u(t) e IR Internal Signals w(t) e IR 2n [wOl(t).m of the model and of the plant is I .Ym only.3.
e. so that W u(t) e JR.I.2.A).1b.2.(a I. kmin Input (3.9).I.n] E JR.m. as in (3. we will keep our previous notation. n in controllable canonical form. w e3(t) e JR.2.u w< 2> = A w< 2>+ b>Yp fJT=(Co. 2n+ 1 [a(t).I. 2n W(t) Y. The practical implementation of the indirect adaptive control algorithm is summarized hereafter. km. dm) satisfying (A2). o o o . there is no need to update the parameters am+ 2so that we let these parameters be zero in (303. without the direct gain d 0 from yP' This a (minor) technical difference. • A e JR. am+ 1.T . that is a*(s) b*(s) bj + bis + 0 Design Parameters Choose • M (i.(aT . Indirect Adaptive Control AlgorithmImplementation The controller structure is otherwise completely identical to the controller structure described previously. Initial conditions are arbitrary. ax= n. o. W (2)T) E IR.m.bT) ·.CI••:o.m..(t). n 2 Assumptions (Al)(A3). 7) is that the relative degree of the transfer function from u. we and make the following assumption. (A 7) Bound on the HighFrequency Gain Assume kp .CT.28).17) 1rT . dn) The transformation 1r.3.3. 8Xo = n. such that o • + b~snl = X(s)dp(s) JR. := 7r _ 1r* E (3. fim.. o. bn) E JR. o o o. E IR2n A consequence (that will be used in the stability proof in Section 3. Controller Structure = X (s) is Hurwitz and X (s) = X o(s) fim (s ).A) • g.5).>O. The identifier parameter error is now denoted y. we will let Co = km I am+ I· The control input u will be unbounded if am+ 1 goes to zero. 2n +I [ w(ll(t)' w<2>(t) IR. aq = n. o_ ( W \1) . n x n. we define am+ 1• Since c(j = km I kp. · · · .124 Adaptive Control Chapter 3 Section 303 Adaptive Control Schemes 125 become. (I) 0. and ac = n . c(t). and for simplicity.is at least n. We let w be equal to w except for those components which are not used and are thus set to zero.2n (3o3o30) det (sf . The identifier parameter is now different from the controller parameter fJ. and to avoid this problem. . b(t) E IR. Since ad = n . o. Thus. n] fJ(t) e [c 0 (t) E JR. The corresponding components of w are thus not used for identification.n] . Note that our estimate of the highfrequency gain kp is . Internal Signals w(t) E JR. Output Since the relative degree is assumed to be known.::: k min> 0.29) 1r(t) E IR. it can be realized as d T (sf . Wm + 1.. b 1.2n (3o3. b'/o. (A7) n. respectively. . We will denote.7)(3. ad = n. except am+ 1CO) > k min· The nominal value of the identifier parameter 1r• can be found from the results of Chapter 2 through the polynomial equalities in (202. .dl. e JR. d(t) E IR. . 0.. in analogy with (2. Yp(t) e JR.31) w(l> = A ~<o + b>. Cn.dT)= (Co..28) r(t).. m.o•.fJ is chosen following a certainty equivalence principle to be the same as the transformation 1r*.
1 Divide ~oam by (~b). Its coefficients can be expressed as the sum of products of coefficients of ~oam and ~G.3. i are used for similar purposes as in the input error scheme.Yp is made to match i.G is a monic polynomial. ' c.126 Adaptive Control wT = Chapter 3 Section 3. ~(s) (r) (3.3. The analysis is also considerably simplified. and if is guaranteed by the projection). adchtions and a division.1. b r) = (a l• W  · · · • am + 1' 0 • · (2)T) · · • b l• · · ·.3. • .Yp  Normalized Gradient Algorithm with Projection .32) A(s)L(s) where fz.33) A(s) is used as input to the actual controller. In our notation. The same is true for a' and Co with an additional division by am+ 1 • Therefore.4 Alternate Model Reference Schemes The input error scheme is closely related to the schemes presented in discrete time by Goodwin & Sin [1984]. c.1 can be used to achieve this new objective and is represented in Figure 3.m ). Since the new model is i . Therefore. Their identifier structure is identical to the structure used here. Thus.3. However.ii G. n. Goodwin & Mayne's algorithms essentially control the plant by reducing the transfer function to an allpole transfer function of relative degree n .~oam) Transformation Identifier Parameter . but their controller structure is somewhat different. Note that ~. then let am+ 1 = 0.Controller Parameter Let the polynomials with timevarying coefficients C(s) = C1 + · · · + CnSnl a (s) = d 1 + . of and G are bounded. by prefiltering the input.3 Adaptive Control Schemes 127 (r. Co are bounded. and let q be the quotient. a. so that the transfer function from r . bn ) . Transformation Identifier Parameter. and n . . .6. the transfer function r.1. except for possible polezero cancellations. the input error scheme presented in Section 3. . in this instance. The filtered reference input = A A M(s ) km a(s) =a!+··· + am+!Sm b(s) = b1 + · · · + bnsnl l A A a =  1  <ii ~. Thus. W Yi e3 = = w e3 w 7r T w. n. ~ i now also defines the model poles in (3. the differentiable and has bounded a . This scheme is the one obtained by Goodwin & Mayne (up to a small remaining difference described hereafter). w<2l) U = ()Tw Identifier Structure 1rr = (aT. Wm + 1 .. given n and m the transformation consists of a fixed number of multiplications. Goodwin & Mayne choose n(s) (3. .Yp is M.Controller Parameter We assumed that the transformation from the identifier parameter 1r to the controller parameter 0 is performed instantaneously.m. 0 is given by the coefficients of the polynomials c=l\qa am+ 1 am+! and by km co=am+ 1 0 A \A g 3. Then. ( W( . the input and output errors are identical and the input and output error schemes are very similar.3. the new transfer function Mi is equal to 1. 7r =  i+ywrw If am+ 1 = k min and am+ 1 < 0. respectively.32). allowing for the placement of all the closed r = km . wO>r. Note also that if the coefficients am+ 1 is bounded away from zero (as then the coefficients of q. and in continuous time by Goodwin & Mayne [ 1987]. .3.1• The input error adaptive control scheme of Section 3. (1) (!) 1rT 0.. the control problem of matching a transfer function M is altered to the problem of matching the arbitrary allpole transfer function i . and transformation is also continuously derivatives.1 is a more general scheme. The polynomials ~. so that q is also a monic polynomial (of degree n . The additional dynaq1ics are provided by prefiltering the reference input. ~ and i are polynomials of degree : : . Note that.m. + dn sn .
x.34). If c0 = 0 and 0 = 0.36) A necessary condition for a solution y to exist is that any common zero of b is also a zero of A s~fficient condition is simply that a. of kp). we may identify 1/co and 0I c0 rather than c0 and 0 . of kp) and a lower bound on c 0 . . the appropriate linear error equation may be found and used for identification. and consider the same controller as previously. since we basically lose information in the regression vector.34) must satisfy loop poles directly at the desired locations without prefiltering.35) This equation is a Diophantine equation.2.2. If we identify c0 directly.4) is valid. If we identify 11c 0 .3. and of a lower bound on 1I c0 . c.e.3.e. we require the knowledge of the sign of c0 (i. in this case. we require knowledge of the sign of 1lc 0 (that is. SUCh that the closedloop transfer function is equal to the M(s) defined in (3. that is a polynomial equation of the form = (3. so that (3. a• <~ c·)apkpfipa· ax+ Gp = [co::] ~am c (3. By dividing the identifier error e2 by co.3 in the a. a different problem appears. We now discuss some of the differences between this and the model reference adaptive control algorithms presented above.6: Alternate Input Error Scheme that is. b be coprime. First note that in (3. with fip replaced by its estimated value. Figure 3. One might consider the approach of letting the model be M(s) = Yp km ~p(S) dm(S) i. fip. This results from the necessity to cancel the plant zeros in order to replace them by the model zeros.128 Adaptive Control Chapter 3 Section 3. (3. that is a lower bound on kp to be used with the Co. Now. let ~ be an arbitrary Hurwitz polynomial. The nominal values c*. 3.3. To avoid it.3.34) Figure 3. as we did in the indirect scheme.Ym does not tend necessarily to zero and may even be unbounded. therefore an upper bound on kp. Note that since identification and control can be separated in the ~nput error scheme. No adaptation will occur (~ = 0). although Yp. This is an identification problem. It is curious to note that the problems encountered are different depending whether we identify c 0 or 1I c 0 . This is also shown in Figure 3.3 Adaptive Control Schemes 129 t1 projection algorithm. then u = 0 and e2 = 0 (cf.6.. the control input U =Cor+ 0TW will be unbounded if the estimate of 1I c0 goes to zero. to be used by the projection algorithm. it is impossible to choose ~ = ~ 0 fim as before. the reference model itself becomes adaptive. Since fip is unknown a priori and it is not Hurwitz. An adaptive control based on this idea is called an adaptive pole placement control algorithm.3.5 Adaptive Pole Placement Control The model reference adaptive control approach requires a minimum phase assumption onthe plant fi(s).3.5). aP coprime (see lemma A6. as we did in the input error scheme. To avoid the zero crossing. require that only the closedloop poles be assigned.
A natural stability concept is then the boundedinput boundedstate stability (BIBS): for any r(. The identifier error may be the output error eo = Yp. dp was automatically a zero of the right hand side..4.7: Generic Model Reference Adaptive Control System The signals and systems defined previously can be recognized. we do not have any control on how close initial conditions are to equilibrium values. and the identifier. In t~e c~se of direct control. these algorithms have the significant advantage of not requiring minimum phase assumptions.35) was replaced by (3. Nevertheless. There. stability in the sense of Lyapunov is not sufficient for adaptive systems. or any other error used for . The stability of the overall adaptive system. guaranteeing that the trajectories will remain arbitrarily close to the equilibrium. In adaptive control.x. although not unique when fip. In adaptive systems. An indirect adaptive pole placement control algorithm may be obtained in a similar way as the indirect mudel reference adaptive control algorithm of Section 3.Ym• the input error e.) bounded. The adaptive systems described so far are of the special form x = f(t. This is the concept of stability that will be used in this chapter. (3. this definition is a local property. A further discussion is presented in Section 6. the stability of the plant was assumed exphcttly. This was not necessary in the model reference case where the solution always existed.sumed that the input is sufficiently rich and that some procedure guarantees coprimeness of the estimates. Then c.130 Adaptive Control Chapter 3 Section 3. To understand the nature of the problem.x) (3. and 1r is the identifier pa:am~ter .5.4.r. since the solution of (3. dp replaced by their estimates. = rp. the stability of the controll~d plant must be guaranteed by the identifier. For practical reasons. 0 is the controller parameter.2.35) with kp. MODEL REFERENCE ~. the controller.3. A difficulty arises to guarantee that the estimates of fzpdp are coprime.35) will usually not exist otherwise. As we recall. IDENTIFIER ERROR Figure 3. the solution always existed. so that any common zero of fzp. with fzp appearing on the right hand side. Previously. d are obtained by solving (3.1) is the stability in the sense of Lyapunov defined in Chapter 1. we will take a general approach in this section and consider the generic model reference adaptive control system shown in Figure 3. the solution x(. which seriously ~o~phcates the problem. + c + rp y INVERSE MODEL REFERENCE INPUT 111 ERROR M OUTPUT ERROR + Yp eo + x = f(t. A classical definition for systems of the form The Problem of Proving Stability in Adaptive Control The stability of the identifiers presented in Chapter 2 was assess~~ in theorem 2.3.2) where r is the input to the system and x is the overall state of the system.6). when started sufficiently close.3. Therefore. that is.4. wht~h mcludes the plant.3. Proving stability for an adaptive pole placement algorithm is somewhat complicated. including the plant. 3. dp were not coprime. and the identifier.. It is often a<.r(t)) (3. the parameter bemg tdenttfied ts dtrectly the controller parameter.3.2. must then be constdered.) remains bounded.4 The Stability Problem in Adaptive Control 131 Appendix for the general solution in that case). () = 1r. 7.. and x 0 e JRn.4 THE STABILITY PROBLEM IN ADAPTIVE CONTROL Stability Definitions Various definitions and concepts of stability have been proposed.2. fip. the controller.
Tc• rp + () w 0 (3. and the proof is consequently simplified.5.5)).5.1 0). that is.3. We note that the input error ei = rp. Such properties were already derived in Chapter 2.7) and the vector w is similarly expressed as . as we saw in Chapter 2. Recall that we defined rp in (3. since u = o• is constant.5. the convergence of the identifier is dependent on the stability and persistent excitation of signals originating from the control loop. Most results also rely on the control input being defined by U = fJTW (r ' wOl y ' p' w<2l) (3.2) while Ym Applying = M(r) (3. Desoer & Vidyasagar [ 197 5]).1) as (3. the controller parameter fJ will be arbitrarily close to the nominal value that stabilizes the plantcontrol loop. Intuitively. while the output error e 0 = Yp.1M(rp) = fo. Recall that the identifier parameter error 1r. First. that is. the identifier error to the control error. rP = 0.5 Analysis of the MRAC System (c 0 . the proof of stability will basically be a small gain theorem type of proof. the plant with the control loor will be stable if fJ is sufficiently close to the true value o•. we cannot argue that for t sufficiently large. the control input can also be expressed in terms of the control error as 1j = p. This step is similar to transferring the equilibrium point of a differential equation as (3. Since Yp = P(u) = M(rp). Provided that the resulting control error is a "small" gain from plant signals.1 ).1 M(r+~rPTW) co (3.r is directly proportional to the control error rP Tw (cf. we can express the control signal as a nominal control signalthat makes the controlled plant match the reference modelplus a control error. the transformation is the identity. Many identities involve signals which are not available in practice (since P is unknown) but are well defined for the analysis. (3. and were expressed in terms of the identifier error. A major step will be to transfer properties of the identifier involving v to properties of the controller involving w.3. do. The problem of stability can be understood as follows.4) and.5) co 1 Further.5.5. since u is given by (3. and prove stability.1) Error Formulation It will be useful to represent the adaptive system in terms of its deviation with respect to the ideal situation when fJ = 0*. that . Another difficulty arises however from the different signals v and w used for identification and control. The problem then is to transfer the properties of the identifier to the control loop.1) to x = 0 by a change of coordinates. 3. we must first express properties of the identifier that are independent of the stability and persistency of excitation of these signals.5) Yp = Ym +  co M(IP w) A T (3. Instead of relying on the convergence of fJ to o• to prove stability.132 Adaptive Control Chapter 3 Section 3.5. Several difficulties are encountered here.1r* does not converge to zero but that only the identifier error converges to zero in some sense. To break this circular argument.3) i to (3. In direct adaptive control. dT) 133 identification. rp =r+~Pw 1 • T (3.Ym is related to the control error through the model transfer function M (cf.5.6) The signal rP Tw will be called the control error. a generic proof to assess the stability of nonlinear time varying systems (cf.5.5.3.6)). the transformation fJ(1r) is usually nonlinear.4. (3.5.5. The results derived in this section are the basis for analyses presented in this· and following chapteis. cT. However. Thus. applying M to both sides of (3.5 ANALYSIS OF THE MODEL REFERENCE ADAPTIVE CONTROL SYSTEM We now return to the model reference adaptive control system presented in Sections 3. it follows.13.
19).b. StateSpace Description We now show how a statespace description of the overall adaptive system can be obtained. (3.· w(ll m Ym ·.5.rjJTw) co (3.3.5.M M I ~ (sf. respectively (cf. The plant has a minimal statespace representation [Ap. the model signals corresponding to Yp and rp are Ym and r.2.134 Adaptive Control Chapter 3 Section 3. In particular.5)). (3.A)P.1). Similarly. as u = 8Tw 0*T W + r/JT W (3.11) The control input u can be expressed in terms of its desired value.15 ) Wm . 1 l With the definitions of w(ll. Consequently.and Hw m'(r) (3. we define fip(S) T I (3._cJ bp c•T bp d*T b._ cJ Xp (3. the plant with observer is described by Xp ApXp + bpu (r) w(l> A w<l) + bx u A w< 2>+ hxYp = A w< 2l + b._d•T Xp A+ bxc•T By defining 0 A w(ll w<2l .bxM  so that (r) (3.1 0) are all stable (using assumptions (AI)(A2) and the definitions of A and i .13) while v (cf. rjJ = 0. · (3.b1M .5.17) I (sf .3. that is. w<2> Ap + bpdo cJ bxdocJ b.5. (3. plus the control error rjJ Tw.5.6) and (3.16) ·.5.14) v i I (3.5.5.A).10.10) Note that the transfer functions appearing in (3.12) [ w<l> x.5.bp dp(s) A M (sf. w<2> m [ (s/ .5 Analysis of the MRAC System 135 I (r+.5.= cp (sf Ap).5.A).5. if the differences between plant and model signals are bounded. w< 2> in (3. bp.16)) is given by we note the remarkable fact that (3.1 0)) I (sf.5. we also define (cf.17)(3.5.A).M bJr 1M (3. it follows that all model signals are bounded functions of time.9) For the purpose of the analysis. and since r is bounded (assumption (A3)). Model Signals The model signals are defined as the signals corresponding to the plant signals when 8 = 8*.6)(3. the plant signals will be bounded.2.~bJ) 1 M M I ~ (sf. cJ ] such that P(s) = kp A.1 Since the transfer functions relating r to the model signals are all stable..8) Wm . As expected.A). we will check that no cancellation of possibly unstable modes occurs when 8 = 8*.A).
i.18) Defining Xpw e IR 3n ..C")dp. Using standard transfer function manipulations.8. and by assumption (A 1). Therefore. but avoiding cancellations. since the order of M is n. Since the transfer function from r. while the dimension of Am is 3n . i.kp fzpd") and.20) Note that although the transfer function M is stable.5.2 are defined through (3.18).5.c")d0 kp fzp i. the state vector trajectory Xm is bounded. In other words.2 and Cm e IR 3n. this equation is rewritten as (3.ofzp =  1 M ~ (3.. its representation is nonminimal. bm e IR 3n.25) where . letting the state error e = Xpw. and fzp.5.Xm E JR 3n . bm. We can represent the plant states as their differences from the model states. that is.Yp is M when </> = Figure 3.5. the additional modes are those of i. Note that _(3. which are all stable by choice of i. c~] is a representation of the model transfer function M. i.5. ((i. we get Thus. Am is a stable matrix.21) eo = Yp.5. <i. SO that e and Arne+ bm</>TW (3.2.23) is not a linear differential equation representing the plant with controller.2 x Jn. that T . <i.2 .Ym = __!__ M(</> Tw) = M( __!__ </> Tw) co (3. We can find where the additional modes are located by noting that the representation of the model is that of Figure 3.1b Cm (sf Am) m =  [Am. because w depends on e. Since r is assumed to be bounded and Am is stable.136 Adaptive Control Chapter 3 Section 3.c")dp (3. divided by co.2. This can be resolved by expressing the dependence of w on e as w = w~ + Qe (3.24) 1 a· co i.5.6). 0 .5. we can also represent the model and its output by co fzm i_ i_ 0 fzp m ~dm i.23) kp fzp i.6) which is equation (3.5.5..8: Representation of the Reference Model 1 k 0.5. using the matching equality (3.22) co Ym (3. 0 .19) where Am e IR 3n.1 bm = (I 1 c 0) M(s ). we must have that c~ (sf Am). (derived above) through a somewhat shorter path.2.5 Analysis of the MRAC System 137 (3.2 to be the total state of the plant and observer.
1 is a standard result in linear system theory and relates the Lp norm of the output to the Lp norm of the input.6.2nx3n2  IR.2. rat~onal transfer function.6 USEFUL LEMMAS The following lemmas are useful to prove the stability of adaptive control schemes. where if is a proper.1 Input/Output Lp Stability Let y = if(u).6.6. with u and y interchanged in (3.6. where decaying term due to the initial conditions.4).:::: 0.5. The systems considered in this section are of the general form y = for all t .) and 0(. with a completely different proof (see the Appendix). rational transfer function.3) c~e (3.Lpe is a SISO causal operator. the overall adaptive system (including the plant.26) Lemma 3. It is useful. IR. oo].1 cf. then it has a proper and stable inverse. for p = oo .6.1. Such a lemma can be found in Narendra.3)(3.6. Annaswamy. For example. 241. Desoer & Vidyasagar [ 197 5]. the adaptive control scheme is described by a nonlinear. where if is a proper. Complete DescriptionOutput Error. IR. although not standard. H (u) (3.6. Lemma 3. 3.6.2 0utput/lnput Lp Stability Let y = II hll1ll Ur 11 00 +I E{t)l E(t) is an exponentially (3. bounded input. to obtain a result that is the converse of lemma 3. & Singh [1985]. one must simply add to this set of differential equations the set corresponding to the identifier.28) will be used in subsequent chapters as a convenient example. and the converse result is true by lemma 3. Let p e [1. Note that if if is minimum phase and has relative degree zero.n1xn1 JR1xn1 IR. that is.5 further restrict ~ the attention to LTI systems with proper transfer functions H (s ).13. time varying. This specific case (3.6. Lin.)).5. oo ]. & Singh [ 198 5].1.6. oo] and for all u E =Arne+ brnc/>TWrn + brnci>TQe = Lp (3.n1xn E JR1 xn (3.n1xn1 JR1xn1 IR.27) for all u E II YIIP L 00 e ~ II h lit II ullp + II EIIP where wm is an exogeneous. Sastry [1984].n1xn1 for all u e L and for all t 2:: 0. we do not use any ordering of signals (order relations o(. but has relative degree greater than zero. p. Lemmas 3. but keep relationships between signals in terms of norm inequalities. If if is minimum phase.6. Most lemmas are inspired from lemmas that are present in one form or another in existing stability proofs.6. Lemma 3.n1xn A differential equation representing the plant with controller is then 'l• e If Then if is stable for all p e [I . Relative Degree 1 Case To describe the adaptive system completely. ordinary differential equation.4) if(u).6.1) . In contrast with Sastry [1984] and Narendra.5. Narendra.28) As for most adaptive control schemes presented in this book. such that Yt = (H (u 1)) 1 (3. controller and identifier) is described by IY(t)l ~ e Arne + brn cP TWrn + bm cP TQe (3.6.5. that is.2 is a version that is valid for p e [ 1. in the case of the output error adaptive control scheme for relative degree I plants. Lemma 3.6.n1xn1 JR1xn1 IR. Proof of Lemma 3. Narendra [1984]. & Valavani [1980]. Let h be the impulse response corresponding to H.6 Useful Lemmas 139 where H : Lpe.2) 0 0 0 0 I JR1 xn JR1xn1 IR. Annaswamy. This is the result of lemma 3.138 0 0 0 0 I 0 Q CT p Adaptive Control Chapter 3 Section 3. then the converse result will be true provided that additional conditions are placed on the input signal u.6.
If if is stable..6. through Y.6.9. (3.6.).9./3 2 e L 00 and J3 1(t). + . for all u.y2(t)Oast oo Proof of Lemma 3.2 is that u is regular (cf. r 2 E L2 such that.6. definition in (2. Proof of Corollary 3.6. related to L oo norms.5 Swapping Lemma Let rf>.8) II Ut liP ~ a ill Yt liP + a2 Then for all t .6. with a minimal realization Let H be a proper.::: 0 I y(t)l for all t . The following lemma is the socalled swapping.::: 0./32 E L2 there exist 'YI.2 therefore leads to the following corollary.5) of lemma 3..6. for p = oo.6. and is essential to the stability proofs presented m Sectwn 3. (a) (b) If ~ 'YI(t)ll Xt 11 00 + 'Y2(t) (3. Corollary 3.::: + If if is stable.IR. w: IR.6.6. If u is bounded and if is strictly proper Then y is regular. with inputs u 1. for all t . either if is strictly proper.3 Properties of Regular Signals Let y = H(u ).6.6.6) 11 00 + i32(t) (3.0 as t y 1. u E Lpe and for all t Utllp .7) The same is true if the righthand side of the inequalities is replaced by any positive. Lemma 3.6.6.9) If in addition.6.4 Let y = if (u ). Roughly speaking.11) where (3. if be (3.6.6. lemm~ (Morse [ 1980]).6 is the socalled small gain theorem (Desoer & Vidyasagar [1975]) and concerns general nonlinear timevarying systems connected as shown in Figure 3. of y is related to the corresponding u. = H(u.6 Lemma 3. b E L ooe .14)). Therefore.1 0) u is regular Then lu(t)l ~ad1Ytll 00 +a2 forallt . a key property of the identification algorithms was obtained in terms of a gain b( longing to L 2• Lemma 3. In Chapter 2.6. Proof of Lemma 3.6. 7. For all a. u2 and outputs y 1. u L 00 e k2 0 such that I u(t)l (3. Y2· is BIBO stable.5) if E is a proper.4.n and ¢ be differentiable.3 in Appendix.:::o.140 If Adaptive Control Chapter 3 SectiC'n 3.::: 0 and for some !31 .5 in Appendix.::: 0 (3. Let stable and minimum phase andy E L 00 e. .i3 2(t). The properties are also valid if u and y are vectors such that each component Y. Lemma 3.6.12) Proof of Lemma 3.2 in Appendix. Lemma 3. L 00 e' and for some x ~ i31(t)ll Xt E II Then there exist a 1. or J3 1. Then (3. It is also interesting to note the following equivalence.. then u is bounded. provided that H 1 and H 2 are BIBO stable and provided that the product of the gains of H 1 and H 2 is less than 1. rational transfer function.4 is useful for such gains appearing in connection with systems with rational transfer function H. rational transfer function.6. where Useful Lemmas 141 if is minimum phase For some kb k 2 :<! 0.4 in Appendix. In particular.. rational transfer function.. lemma 3.. the small gain theorem states that the system of Figure 3.2 shows that if u is regular andy is bounded. (c) If u is regular Then y is regular. the assumption (3.y 2 E oo Then L 00 andyl(t). where if is a proper. a 2 Utllp ~kill . monotonically increasing function of time.
7.12.11 + !32 + 'Y2f3t) II ed ~ If {l'Yt'Y2). 3. p.13) Suppose that there exist constants {3 1 . with initial conditions in an arbitrary Bh. the error is bounded by an L2 function.ll + /3.9: Feedback System for the SmallGain Theorem Lemma 3.1 is also valid if the normalized gradient algorithm is replaced by the normalized leastsquares (LS) algorithm with covariance resetting. ')' 2 ~ 0. Yt = H 1(e1). 41.6. too .6.6 cf.4.6. ed + !32 for all t ~ 0 (3.4.11 + 'Ytll Ut. the regressor error v .)t II ~ 'Ytll II H 2 (e2)tll ~ 'Y2II Then (a) all states of the adaptive system are bounded functions of time.oo.4.Vm e L 2 and tends to zero as t . as Derive properties of the identifier that are independent of the boundedness of the regressorExistence of the solutions. Let e 1.(II Ut.oo and is in L 2• This result will be useful to prove exponential convergence in Section 3. 7. Therefore.3.4. It shows that. 1 H2: Lpe. the output error e0 = Yp .142 Adaptive Control Chapter 3 Section 3.4.ll ~ (l'Yt'Y2).6. Although the theorem concerns the adaptive scheme with the gradient algorithm. Y1 = H 2(e 2 ) all subscripts t dropped.Lpe be causal operators. in addition. theorem 3.6 Small Gain Theorem Consider the system shown in Figure 3.1 {a) Then e" e2. 7 Stability Proofs 143 Further.6.ll + 'Y2II 1 Then u2. (e. (b) ed + /3.1. Desoer & Vidyasagar [ 197 5].4.9. examination of the proof shows that it only requires the standard identifier properties resulting from theorems 2.8.. Properties obtained in theorems 2.15) is valid with Proof of Lemma 3.7.14) If 'Yt ''Y2< 1 II e. 7 STABILITY PROOFS 3. u" u2 e Lp E Comments The proof of the theorem is organized to highlight the main steps that we described in Section 3.4 led to l4>T(t)v(t)j =(!3U~l Vtll 00 +(!3(t)\ . + 'Yi/32) (3.1 Consider the input error direct adaptive control scheme described in Section 3.12. For example. oo ]. Theorem 3. Proof of Theorem 3.Ym e L 2 and tends to zero as t . such that II H. e2 E Lpe and define u" u2 by u1 e 1 + H2(e2) (3. Lp and (3. because although the properties are valid for any initial conditions.oo . We also obtain that the difference between the regressor vector v and the corresponding model vector Vm tends to zero as t . We insist that initial conditions must be in some small Bh. the convergence of the error to zero and the L 2 bounds are not uniform globally. the states of the adaptive system remain bounded (BIBS stability) and the output error tends to zero.7.1 StabilityInput Error Direct Adaptive Control The following theorem is the main stability theorem for the input error direct adaptive control scheme. {32 and 'Yt. Let p e [ l. Figure 3.15) for all t ~ 0. there does not exist a fixed L 2 function that bounds the output error no matter how large the initial conditions are.<II 1 u2.6. given any initial condition and any bounded input r(t). Let H .
(z) This relationship is obtained through the swapping lemma (lemma 3.v .7. rp. c0(t) ~ Cmin.. namely. The proof of (3. o· + I cp(O)I for all t ~ 0.3) and using the fact that the control input With (3. the controlled system is a linear time invariant system with a linear time varying controller and bounded feedback gains.7.144 Co(t) ~ Adaptive Control Cmin Chapter 3 ~ Section 3.5.7) r + . (d) Establish the regularity of the signals.2). so that any of its component is in L 2• In particular co e L2.rP c/J w 1 r .7.7) with (3. The need to establish the regularity of the signals can be understood from the following.1(z). The question of the existence of the solutions of the differential equations describing the adaptive system may be answered as follows. The gain c/JT operating on v is equal to the gain (3 operating on II v1 II 00 . The two can be related using (3. 7. this equation leads to Figure 3.7.5 and led to the control error c/JT w. From proposition 1.6) __!_ c/JT W co = __!_ rJ>T Z co A (3.I ( __!_ c/JT z) = __!_ c/JT v + c0 c0 and.7. (cl) Relate c/JT w to c/JT z Only the first component of w.1.2) where the transfer functions M and k.6. are stable and strictly proper.7.ru = corp+ 0 w (3. it follows that all signals in the feedback loop.1) The inequality for c0(t) follows from the use of the projection in the update law.4. in Figure 3. and this gain belongs to L 2• On the other hand.9).5). with rp and (3. We have. It represents the plant as the model transfer function with the control error c/JT w in feedback.I (zT)( _p_)) co (3. Also. I c/J (t )I :$. ~ e L2. that is * .4) > 0 for all t 0 (3. so that 1/co e L . 7 u to obtain Stability Proofs c0 r + 8Tw 1 T (cor+ c/J w) co 1 r + . is different from the first component of z.7.7. I c/J(O)I . The control error has now been expressed as a function of the identifier error c/JT v using (3.M(cp w) I A T co = Wm+H. We are not only concerned with the boundedness of . namely r. Therefore. The properties of the identifier are stated in terms of the error c/JT v = c/JT L .(cp w) m A 1 * co T (3. 7.c/JTW co 1 T I (c2) Relate c/JT z to c/JT v = rJ> L..1 0. the controlled plant appears as a stable transfer function M with an L 2 feedback gain.4)..1 0.c/JTW co 145 (3. so that I O(t )I :$.5) r It follows that 1 • T ) co (c0 . The relationship between the two can be examined in two steps.8) _!_M(c/JTw) co Mi(i. Therefore.1) indicates that c/J e L 00 as long as the solutions exist. (c) Relate the identifier error to the control error. with notation borrowed from the lemma i. In fact.1(_!_ cjJTw)) co (3.7.9) u PI M(rp) M(rp) A Lc. using (3. while the control error is c/JT w. 7. belong to Looe' (b) Express the system states a1 :d inputs in term of the control error.8) = Ym + . This was done in Section 3.c/J T z co P (3. Thus.7. and therefore in the whole adaptive system.I (ib.v m.7. d/dt(c/J/co) e L2.
11) Recall from (3.r +  co co TrJ> w (3. 7 Stability Proofs 147 (cf.:.7.4.10) lwl : :. Since rJ> is bounded.:. k11Yp.ll 00 + k (3.15) .6.. To prevent proliferation of constants.7. using lemma 3.lloo + k hence.!]_ wl . kll (tPTw).2. with (3. It follows. rJ> Similarly. kll I. we will hereafter use the single symbol k. by the results of (a) and by assumption (A3). . w.:. . note that since rJ> e L 00 .7. k II Yp.10)) (3.13) iT Yp = M(rp) = Figure 3. ki1Yp.:.:. . w is related to rp through a strictly proper.2.7.10) and (3.:..7.16) show that the boundedness of Yp implies the boundedness of rp. the last inequality implies that Yp (iT w).146 Adaptive Control Chapter 3 Section 3.6. I<.. from proposition 1. (3.:.ll 00 + k that is.14) iT w is I i~ wl .7. since and~ are bounded and using (3.7. whenever such an inequality is valid for some positive constant. that all signals belong to Looe' where M(c 0 r) is bounded.12) w is regular. and th~refore of all the states of the adaptive system.10: Representation of the Plant for the Stability Analysis A 1 co M(cor) + . Therefore.5) that lrPI : :.:. with the fact that regular and then (3.ll 00 +k 1. kll Yp. kll(tPTw).ll 00 + k.7. <iTw)l : :. By ensuring the regularity of the signals in the loop.ll 00 + k (3. the controller parameter 8 is also bounded. that (3.6. we guarantee. u.ll kll <iT w).11) 1.:.:.7.7.11) +k 00 dt for some constant k ~ 0.16) Note that c0 and r are bounded.7.1 1wl .10) and lemma 3.ll 00 +k so that w is also regular.:.7.7.M (4> A 1 A T co w) (3. Using lemma 3. iT>wl +I iT<~ w)l : :. kll (M(tPT w))til 00 + k :::.1. Now. that boundedness of one signal implies boundedness of all the others.7.2)).14) the output Yp but also of all the other signals present in the adaptive system. stable transfer function Inequalities in (3.ll 00 + k co 1 rp = . wl .7.ll 00 + kll (M(cor)).:. kll w. (3. The output Yp is given by (using (3. with (3. 7.
are all stable and the the output error e0 the regressor error as too.12) and (3.2 The proof is very similar to the proof for the input error scheme. The gain .M(cpTw) 1 A co (3. 7.1. 7 Stability Proofs 149 It also follows that v is regular.5. all signals belong to L 00 • From (3.5. (3.Ym e L 2 and tends to zero as t . Then all states of the adaptive system are bounded functions of time.M I A (3.21) Note that these results are valid. (e) Stability proof.1(w) .Ym e L 2 . although the realization of minimal (but is stable). 7.oo. while the second term is the output of i.P I M £l_!_. for T sufficiently large.7. 7 . as too.vm also belongs to L 2 and tends to zero. and is just sketched here.7. But since Yp• Ym e Looe' it follows that Yp e L 00 • Consequently.3). 7. tb~eb" 2.7.b>.7. all signals in the adaptive system belong to L 00 e.9) (s/.1(wm) e L 2 and tends to zero 1 last two are strictly proper. (a) We now have.2 Consider the output error direct adaptive control scheme described in Section 3. 7.oo . o f t he proJectiOn m the update law.i . using results obtained so far and lemmas 3.is bounded by (3. 3.9). as too .4.1 and 3. since it is the sum of two regular signals. following the steps of the proof of theorem 3.7.20) so that v. specifically i I [ £1~ l £1~ r Therefore. {j II :::.2 StabilityOutput Error Direct Adaptive Control Theorem 3. 7 . :::.. II 00 + {j II zl II 00 + {j PII r p II 00 + PII w1 II 00 + P VI 1 Pro<'f of Theorem 3. it also follows that e0 = Yp. because co . ~~ Since v is regular.7. corollary 3.3. .148 Adaptive Control Chapter 3 Section 3.6. ib.1 with a regular input (cf.17) V = Vm + [  w~e~e the first term is the output of i I (a stable and strictly proper. instead Pll Yp.7.b>.6 shows that {j. .1.18) We will now use the single symbol {j in inequalities satisfied for some function satisfying the same conditions asiPf that is {j e L 2 n L and 00 {j(t)0..ll 00 + P Pll (yp. From (3.19) (3.2) and (3.6) with (3. The transfer functions Mi.19) shows that Yp..Ym is bounded for t ~ T.. and tends to zero as t .1 and ic. Similarly.A)_M}. ' M is not Recall that since fJ e L 00 . On the other hand. Therefore. = i Yp. .'7.9) (sf.19).Ym)llloo + {j (3. using (3.4 IYp.6.2).A).6. application of the small gain theorem (lemma 3. :::. {j (t ~ T) < 1..2.0 as too.rw + co 0 co £lw (3. mmtmum phase LTI system) with bounded input.Yml :::. with initial conditions in an arbitrary Bh. 7.6. (a) (b) Ym + .
3.ll 00 +k (d) As in theorem 3. that __!_ co Mi <iTV) e1+.7.6.16}. with initial conditions in an arbitrary Bh . A w< 1l + b>..16) still holds.31) :. 7.l..1(ib.26) Recall that e 1 is bounded (part (a)) and that Mi is strictly proper (in the output error scheme).32) so that.25) that P{w(ll) A w<I) + b>.Ym e L 2. k IIYp.7..7. and tends to zero as t .. (c) Since c0 = and.7.16) (cf. a1 IIYriiP + a2 + aJIIzriiP (3..25) At 1 b>. 7.1(ib. Proof of Theorem 3.6.ll w< 2l where = To apply the modified lemma 3.7. (e) Recall.)) for some a3 ~ 0._J__M i (.ll 00 + llwi 2lll 00 (3.1 co co (3.1.7.7 Stability Proofs 151 (b) As in theorem 3. (b) .2 may be modified as follows. and tends to zero as t . (3.7..7.7.16) and the definition of the gradient update law. (3.3 StabilityIndirect Adaptive Control so that llwf2lll :.28) It follows that the output error eo = Yp . we note from (3.7. 7.ll 00 +k (3. k 3 ~ 0 such that (3.7.150 Adaptive Control Chapter 3 Section 3.7. V) + _!_if i (ic...7..7. k l wt(2)11 00 + +k IIYp. To prove (3.24) _!_Mi(iTv)+_!_Mi(ic.. 7.1(wT))...2. first note that lemma 3.6. Unfortunately.Wm e L 2. 7.33) (3.27) Theorem 3.6..)} _!_ifi(iTv)+ _!_ifi(ic..1.1. 7.. The proof can then be completed as in theorem 3. D 3.) A~: 1 b>. Now.7. (3. w(ll = Further U = all states of the adaptive system are bounded functions of time. (3. = (wT}.1(w T)....6. We leave it to the reader to verify this new version of the lemma.26) :. (u) I gco co (3. from (3.12).3. 7.7. k IIYp.7.Mi(vTveJ) co e1ML(¢ 'Y A A 'T lluriiP :.. c~ r + b>.5) is replaced by k1 lluriiP + k2 + kJIIzrllp then lemma 3.7)).1(ib.23) gco v) (3.oo .3. it follows that w is regularcf.6. which is all we will need for the stability proof.3 Consider the indirect adaptive control scheme described in Section 3. If there exists z e Lpe.6) replaced by which is equivalent to (3.22) llwfllll 00 11Yp.9) becomes llwfllll 00 llwtll 00 :.7. iT w is not necessarily regular because .7. (yp) +k (3.25)(3. with (3. is not bounded. However.. we will show that (3.22) YpYm = llurllp :..7. recall that (sf [ (sf  co co e1. applying the modified lemma 3.6.30) __!_ M(iT w) co Putting (3. (3.30) with (3. OT w (3. Then (a) P is minimum phase.2 is valid with (3.3 in Appendix.oo the regressor error w.2 c0.
the proof of theorem 3. = b(s) A (3.0. let kp = am+ 1 = km = 1.7. 7. 7.40) Since the degree of q is at most equal to the relative degree of the plant.qnp " qa . but only requires more complex manipulations. For simplicity.q~ + Qap so that (a. np s· c ~ . c· It follows that ~ q•a· = ~ q*fzp s A*) AA A A*) .ap qa (3.a . To understand the approach of the proof.7. assume that the parameters 1r and () are fixed in time and that kp is known.36) Q~.a·)+ (q. If we apply the operator equality to the input u. the transfer function Mq is proper and stable. The first is related to the unknown highfrequency gain.43) ~(s) ~(s) where and b vary with time but ~ is a constant polynomial. 7 Stability Proofs 153 Comments Compared with previous proofs.~oam + qap) (a. 7.43) implies the following operator equality a a A A.7. Equality (3.~Oam . G. The real difficulty comes from the fact that the polynomials q. since we assumed that the .152 Adaptive Control Chapter 3 Section 3.7.7.41) (3.3) (3. A major step is to relate the identification error 1/. The exact formulation is left to the Appendix.3 presents additional complexities due to the transformation 1 r . We now discuss the basic ideas of the proof. but transforming it to an operator equality leading to (3. it leads to (with the definitions of Section 3.38) A a This equality of polynomial ratios can be interpreted as an operator equality in the Laplace transform domain. 7. To make sense of time varying polynomials as operators in the Laplace transform domain. c. The nominal values of the identifier parameters are then given by parameters were fixed in time. Two difficulties arise when using this approach to prove the stability of the indirect adaptive system.c + (~ . 7. Equation (3.38) is still valid as a polynomial equality.34) The controller parameters are given as a function of the identifier parameters through A a while the nominal values are given by a a.1 and the properties of the identifier would then lead to a stability proof. consequently (3.T w to the control error cJ>T w.a·)+ (a·.39) and. and vary as functions of time. The techniques used in the proof of theorem 3.7.40) requires some care.c . 7.q*)ap Therefore q [ A (3.7.37) ~~:~ a(s) = aT [ ~) (3.(c .42) Consider the following equality of polynomial ratios (3.qa• and (3.7. we define a A.
8.154 Adaptive Control Chapter 3 (3. we are concerned with the exponential stability of the overall adaptive system. by theorem 2. we are not only interested in the convergence of the parameter error to zero.5 and 2.7.2 Consider the eutput error direct adaptive control scheme of Section 3.8 Exponential Parameter Convergence 155 words. the parameter error cp converges exponentially to zero.vm e L 2 • Therefore.Vm is not bounded by a unique L 2 function for any initial condition. D Comments Although theorem 3.5. Recall that in Section 3.3. Therefore. It is. by first operating the matrix transfer function on the argument and then multiplying by a and b in the time domain.1.8.3.6~ Vm PE implies that v is PE. 7 . provided that the regressor v is persistently exciting. since the signal v is generated inside the adaptive system and is unknown a priori. Proof of Theorem 3. wm 1 Similarly.3.7.1). using averaging techniques.5 will actually show that the adaptive control system is not globally exponentially stable.3). it was established that the errors between the plant and the model signals are the outputs of stable transfer functions with input cpr w.1 establishes exponential stability in any closed ball.3.46) so that the operator corresponding to (3. A more successful approach is that of Chapter 4.7.3. J ar[~ ['{()bll but the two operators can be related using 3.8.5.e.8.2 (or the indirect scheme of Section 3. doubtful that these estimates would be of any practical use. Results in Section 4. Since w is bounded (by theorem 3. Note that in the case of adaptive control. Finally. Theorem 3. b(s) 1' C (3.1 shows that it is sufficient for the model signal Wm to be persistently exciting to guarantee exponential convergence. However.1 can be used to obtain estimates of the convergence rates of the parameter error.1. all errors between plant and model signals converge to zero exponentially fast. we found that v. We note that the product of polynomials can be expressed as £2(s)b(s) = aT(Snsi)b a(s) .1 Since Wm.3.7. since v is PE.7. If Wm is PE Then the adaptive system is exponentially stable in any closed ball. The various theorems and lemmas used to prove theorem 3.1 has direct parallels for the other adaptive control algorithms presented in Section 3.47) i.45) ~ (s) ~ (s) This can be interpreted as an operator by multiplying the coefficients of the polynomials to lead to a ratio of higher order polynomials and then interpreting it as previously. cpr w converges exponentially to zero. however.(wm) is PE. The result of theorem 3. Note that this operator is different from the operator are bounded.8. Consider now the input error direct adaptive control scheme of Section 3. In other .5. 7. Theorem 3. em converge exponentially to their nominal values. Theorem 3. ~he (3.6. it does not prove global exponential stability. consider the product (3.1(zm) = i .8. but also in the convergence of the errors between plant states and model states. it would be straightforward to show that the parameters of the adaptive sys·. This is because v. such result is useless. If Then Wm is PE (wm is PE) the adaptive system is exponentially stable in any closed ball.3.6~ implies that Vm = i . In theorem 3. lemma 2.44) Section 3. due to their complexity and to their conservatism. using lemma 2.5).7. Using theorem 2.48) swapping lemma (lemma 3.45) is (3.8.1 Consider the input error direct adaptive control scheme of Section 3.8 EXPONENTIAL PARAMETER CONVERGENCE Exponential convergence of the identification algorithms under persistency of excitation conditions was established in Sections 2.6.1.
The input error direct adaptive control scheme and the indirect scheme are attractive because they lead to linear error equations and do not involve SPR conditions. the normalized gradient identification algorithm can be replaced by the least squares algorithm with projection without altering the results. . We proved basic lemmas that are ft ndamental to the stability proofs and we emphasized a basic intuitive idea of the proof of stability.8. to the us~ . The various model reference adaptive control scheme~ also sho. Differences appear between the schemes however. Then. All had a similar controller structure but different identification structures.8. The final result however shows that stab1hty 1s not an argument to prefer one over the other.a.8. An interesting conclusion is that the stability and convergence properties are identical for all three adaptive control schemes. where the parameters updated by the identifier were the same as those of the basic identifier of Chapter 2. The stability of the model reference adaptive control schemes was proved. The third scheme was an indirect scheme. of the output error m the identification algorithm. we derived three model reference adaptive control schemes. This condition was to be satisfied by an exogeneous model signal.2 The proof of theorem 3. The first two schemes were direct adaptive control schemes. or to SPR cond1t10ns on the reference model. the indirect scheme had the same stability properties as the direct schemes. that was the existence of a small loop gain appearing in the adaptive system.9 CONCLUSIONS In this chapter. influenced by the designer and was basically a condition on the reference input.9 Conclusions 157 Proof of Theorem 3. the controller parameters were obtained from the identifier parameters through a nonlinear transformation resulting from the model refe:ence control objective. where the parameters updated by the identifier were the same as those used by the controller. in connection with the highfrequency gain and with other practical considerations. together with the result that the error between the plant and the reference model converged to zero as t approached infinity. Further. although more complex in implementation and especially as far . The difficulties related to the unknown highfrequency gain were also discussed. We used a unified framework and an identical stepbystep procedure for all three schemes.1 and is omitted here. The indirect scheme is quite more intuitive than the input error direct scheme.wed that the model reference approach is not bound to the cho1ce of a. Another advantage is that they allow for a decoupling of identification and control useful in practice. We investigated the connections between the adaptive control schemes and also with other known schemes.156 Adaptive Control Chapter 3 Section 3. with the additional assumption of the persistency of excitation of a model regressor vector. D 3. The exponential parameter convergence was established.2 is completely analogous to the proof of theorem 3. In particular.s t~e analysis is concerned. duect adaptive control scheme.
but in general as an approximation method.2) and corresponds to the identification of a single constant urements of (4. The method was proposed originally by Bogoliuboff & Mitropolskii [ 1961 ]. & Sastry [ 1986].1) and relates properties of the solutions of system (4. 4. together with their application to the adaptive systems described in previous chapters.0 INTRODUCTION Averaging is a method of analysis of differential equations of the form OneTime Scale Averaging Consider the linear nonautonomous differential equation x = t/(t. developed subsequently by Volosov [ 1962].0. and stated in a geometric form in Arnold [1982] and Guckenheimer & Holmes (1983]. the solution of(4. Balachandra & Sethna [1975] and Hale [1980]. In the following section.1. and Anderson et a/ [ 1986].0.0.4) is known analytically. These results are reviewed here.x) (4.0.1) to properties of the solutions of the socalled averaged system (4. it is not only useful in stability problems. and theorems were derived for onetime scale and twotime scale systems such as those arising in identification and control.3) Yp(t) o• from meas(4. Our recommendation to the reader not familiar with these results is to derive the simpler versions of the theorems for linear periodic systems. allowing one to replace a system of nonautonomous (time varying) differential equations by an autonomous (time invariant) system.0 Introduction 159 CHAPTER4 PARAMETER CONVERGENCE USING AVERAGING TECHNIQUES Soderstrom [ 1983] (the ODE approach). Bodson et a/ [ 1986]. This equation is a special case of the parameter error equation encountered in Chapter 2 Q>(O) = 1/>o (4.2) where fav(X) := to+ T where x is a scalar.x)dr to = o• sin (t) x(O) = x 0 I assuming that the limit exists and that the parameter E is sufficiently small.1) is simply . Mareels eta/ [1986]. Averaging methods were introduced for the stability analysis of deterministic adaptive systems in the work of Astrom ( 1983].3) lim Too ~ J f(r.1. and. This aspect was emphasized in Fu.1. Bodson.1.1) x =  E sin 2(t) x x(O) = x0 (4. in Ljung & 158 using a gradient update law. However. Averaging is very valuable to assess the stability of adaptive systems in the presence of unmodeled dynamics and to understand mechanisms of instability. Riedle & Kokotovic [1985] and [1986].1. The general solution of a first order linear differential equation of the form x = a(t)x (4. in a stochastic context. Sethna [ 1973].5) In particular. we present examples of averaging analysis which will help to understand the motivation of the methods discussed in this chapter.1. and is given by [ a(r) dr x(t) = e xo (4. Astrom [1984]. We also find early informal use of averaging in Astrom & Wittenmark [ 1973].Section 4.1 EXAMPLES OF AVERAGING ANALYSIS 4.
Also.00 I x(t) .17) 2 and the averaged system is I x(t) I ~ m ea(t. a product of two sinusoids at different frequencies has zero average. The method of proof is completely different. In the above example.6) and of the averaged system (4. and closeness of the convergence rates as E .1.8) OneTime S'ale AveragingMultiple Frequencies Consider the system x "' .15) Let us now compare the solutions of the original system (4.1.. and averaging becomes useful.1.e w2(t) x where x e IR. noting that The averaged system is exponentially stable as soon as w contains at least one sinusoid. t 0 ~ 0.1..3) is now given by Xav "' • .!.to) I x(to) I _ E[ (4.1. As is now shown. Then.6). at a fixed t To define the averaged system.0.1. X lim Too ~ to+ T I to sin 2(r)x dr "' ~x (4.2)(4.0.1.1. sin(2t) 4  ll ~ t0 + T I to w2(r)dr (4. we separated the integrand into its average and periodic part.'[ x(t) e e . c .1. the convergence rates are also identical.!.18) for all x(t 0 ).0. and w contains multiple frequencies n (4. A graphical representation may be obtained by plotting In( I x (t) I 2).!.18) is independent of the phases rPk· .16) (4.1.0.6) Therefore. sin (2t) 4 Xo (I. 2 E Xav Xav(O) "' Xo (4.!. the solutions are arbitrarily close as E . the averaged system defined by (4. the original and the averaged system have identical convergence rate a "' . using frequencydomain expressions. The difference between the solutions. the graph of In ( I x(t) I 2 ) is bounded by a straight line of slope . We devote the rest of this section to show how the averaged system may be calculated in more complex cases..9).13) cos(2r) )dr x0 (4.14) The solution of the averaged system is XavU) .10) (4.. so that we may approximate the original system by the averaged system. XavU)I "' e 2 I: le . Indeed.1. so that ~!:!:_ k=l n 2 (4.1. the analytic solution of the original system is not available.0 In other words.1. for all to. we need to calculate the average AVG(w 2(t)) :"' lim T.: e' 2 t Xo (4. . Note also that the expression (4. both systems are exponentially stable (and if we were to change the sign in the differential equation.1. However.7) Therefore.1 Examples of Averaging Analysis 161 . 2 x 0 "' e + . [ sin 2(r) dr . but the results are essentially the same: closeness of the solutions. Note that when we replaced sin 2(r) by ~  ~cos (2r) in (4.1.1. we will prove theorems stating similar results for more general systems.12) k ± aZ ~I 2 l Xav (4.I (4.160 Parameter Convergence Using Averaging I I Chapter 4 Section 4.2 a. Recall that the convergence rate of an exponentially stable system is the constant a such that the solutions satisfy Expanding w2 will give us a sum of product of sin's at the frequencies wk.1. both would be unstable).9) w(t) "' ~ k =I ak sin (wk t + rf>k) (4. In this chapter. ...' .11) sin(2t)l as e .
In the limit as E .1.0.1. so that M(wx) = The condition is the familiar SPR condition obtained for the stability of the original system in the context of model reference identifica~ion.19)( 4.20) is denoted y = _b_ (wx) = M(wx) s +a (4. The solution (4.X= EwM(wx) Equations (4.1.20) or (4.1.28) is positive.1. as do products of sin's with cos's of any frequency. so that the sum in (4. Now.26) so that the averaged system is given by (the gain g plays the role of E) 1>av =  where (4.1.1. Vector Case M(w)x (4.1.1.1. As before.19)( 4. Otherw1se.22) may indeed be approximated by Xav = EAVG(wM(w))Xav (4.1. as will be shown in Sectwn 4.30) w(t) = ~ ak sin (wk t) k =I (4.p l = [foYil r n (4.1.22) using (4. a > 0. The averaging analysis brings this condition ~n evid~~ce ~irectly m t~e frequency domain. The regressor vector is given by w Again.1. products of sinusoids at different frequencies have zero average.28) a y(t) + b w(t) x(t) where x.1.25) where P = kp 1(s + ap). 1t is sufficient that the wk 's be concentrated in frequencies where Re MUw) > 0.24) In identification. y(t) the fast state and the system (4. It is also evident that th1s cond1t1on 1s necessary.19) becomes (4.1.3. We lilustrate the procedure with the simple example of the identification of a first order system (cf. y e ffi.19) (4.0).1.28). but t~e frequency domain analysis does.24) is that ReM Uw) > 0 for all w > 0 (4. ~f o~e does not restrict the frequency content of the signal w(t ).5) does not extend to the vecto~ case.20) were encountered in model reference identification with x replaced by the parameter error ¢.1.29) . (4.1.1 Examples of Averaging Analysis 163 TwoTime Scale Averaging Averaging may also be applied to systems of the form x(t) y(t) .20) a twotime scale system.E sin (¢d. When E .20) w(t) y(t) AVG [ w M(w) ] = ~ k =I n  al 2 I MUwdl cos (¢k) (4.1.1. Let w contain multiple sinusoids n = [. Section 2. a sufficient condition for the stability of the averaged system (4. we let the input r be periodic r (t) = so that n ~ rk k =I sin (wk t) (4.162 Parameter Convergence Using Averaging Chapter 4 Section 4.1. sin(wkt + ¢k) = sin(wk t) cos(¢k) + cos(wk t) AVG(rAr) AVG(r P(r)) AVG(r fi(r)) AVG (fi(r) fi(r)) 1 1>av .0.1. x(t) varies slowly when compared to y(t ).1. Therefore (4. we encountered (4.21 ). E by the adaptation gain g.1.1. (4. x(t) is called the slow state. and the time scales of their variations become separated.27) g AVG (w wT) 1>av g ( The product w M (w) may be expanded as the sum of products of sinusoids.31) M(w) k ~ ak I MUwk) I sin (wk t + ¢k) ~I (4. and y by the identifier error ei. where 1> was a vector. Further. In short. a frequency domain expression brings more interesting insight..21) where M is a stable transfer function. x(t) may be considered frozen in (4.1.23) The result of the averaging theory is that (4.1.2).
. and even obtain frequency domain results. In that case eo Analyzing adaptive control schemes using averaging is trickier because the schemes are usually nonlinear. This is the motivation for the derivation of nonlinear averaging theorems in this chapter.42) . it is positive definite for all wk .1. Nonlinear Averaging .164 Parameter Convergence Using Averaging n Chapter 4 Section 4.1.(r) (4. The adaptive controller is defined by and it follows that AVG teo(eo + Ym)) n k ·1 rJ> fixed . Ym and e0 .cp is the closedloop polynomial. Assume again that r is of the form n r am Ym + r u r +do Yp ~ rk sin(wk t) k a (4. Indeed.37) a'j a'j The matrix above is symmetric and it may be checked to be positive semidefinite.cp (ym) c/J + Ym (4. k ·1 eo rl 2 (am . cp varies slowly compared to r.34) 1 where am > 0. we will show that it is possible to keep the nonlinearity of the adaptive systems. assuming that c/J is fixed.1.1 Examples ofAveraging Analysis eo Yp .cp s +am .1.g eo Yp (4.0.32) do .35) where g > 0 is the adaptation gain. Further. The analysis is therefore not restricted to a neighborhood of some trajectory or equilibrium. independently of the linearity or nonlinearity of the equations. Note that it is possible to linearize the system around some nominal trajectory. 0. we consider the output error model reference adaptive control scheme for a first order system (cf.1. The output error and the parameter error are given by ~ rl 2 (4.ap) (4.Ym do .d~ = 165 g ~k ·1 ri 2 c/Jav (4.cp kp = km = Co = 1) Note that s + am . Taking a Lyapunov function v = cJ>!:v c/Jav shows that the averaged system is exponentially stable as long as the input contains at least one sinusoid of frequency w "# 0. we directly recover a frequencydomain result obtained earlier for the original system through a much longer and laborious path. The averaged system is defined by calculating AVG (eo (eo + Ym) ). or around the equilibrium. However.cp s +am ..40) s +am (r) (ym) = s +am .36) The adaptive system is completely described by g j.t. Section 3. with s +am . As an example.1.g eo (eo + Ym) where Ym = (4.1.1.1.(am .33) . and ap is unknown.39) and eo+ Ym s +am.cp (ym) cp [ 1 s +am ( r ) ] cp (4.41) (4. Thus.38) f s +am When g is small (g takes the place of in the averaging analysis ).c/J) eo + Ymc/J ap kp w~ + k} w~ + (4. that is the polynomial giving the closedloop pole for cp fixed. averaging allows us to approximate a nonautonomous system by an autonomous system.1.1.
to assume that the function f(t. x e Bh. we do not make the distinction between the periodic and the almost periodic case. with rate of convergence a. 0) defined by the limit has zero mean value. p. the system is not only locally exponentially stable.3) is satisfied. n ¢av = . x. if the function d(t.X. This is equivalent to saying that there exists a convergence function 'Y (T) = a/ T (i. or almost periodic in t. 0) has mean value fav<x) if and only if the function d(t.e.x) = f(t. averaging did not alter the nonlinearity of the original system. but consider instead the assumpt~on of the existence of the mean value as the starting point of our analysts. such that y(T)0 as Too and t 0 +T ~av = .1. Note that the function f(t. x. It is common. Eo.44) where a(¢av) is a nonlinear function of ¢av· However.a(¢av) ¢av (4. only its time variation. In general.1. then the integral of the function d (t. x) is periodic in t and is bounded. Since h is arbitrary.4) X= Ef(t.2 AVERAGING THEORYONETIME SCALE In this section.oo.166 Parameter Convergence Using Averaging Chapter 4 Section 4. 346). Definition Mean Value of a Function.1. the variation of x with time is slow. t ~ 0.1. ~he existence of the mean value is guaranteed. x) is bounded. 344). The function 'Y (T) is called the convergence function. but we do distinguish the bounded integral case from the . Note that if the function d(t.45) 0). Again. 4. x) is also a bounded function of time. because a is not bounded below as h . The analysis of this section may be extended to the general identification and adaptive control schemes discussed in Chapter 2 and Chapter 3. 0 < E :::. E) is periodic in t. we consider differential equations of the form for all t 0 ~ 0. but also exponentially stable in any closedball. theorem 6. h = (4.43) The averaged system is a scalar nonlinear system. but the convergence fu~ction need not be bounded by a/ T as Too (it may be of order 1/Vr for example). centered at the origin. x. However. Indeed. Note that the averaged system is of the form assuming that the limit exists uniformly in t 0 and x. without further assumptiOn (Hale [ 1980].IR +. and is not required to be periodic but almost periodic. and f is piecewise continuous with respect to t. although the converse is true.g ~ k =I T r"f fav(X) = lim Too 1 am 2 ~ to+ T to J f(r.43) which may be used to predict parameter convergence quantitatively from frequency domain conditions. The function 'Y (T) is bounded (by the same bound as d(t.1) where x e IR n. even if its mean value is zero (Hale [ 1980].1. in the literature on averaging. Then. X. we recovered a result and a frequency domain analysis.2. For small E. An advantage of the averaging analysis is to give us an expression (4. x. for all h > 0. The method of averaging relies on the assumption of the existence of the mean value of f(t. O)dr. x)) and converges to zero as Too. Here.E) x(O) = Xo (4. p. of the order of 1/ T) such that (4. X. it is easy to see that (4.2) wt + (am  c/Jav) 2 Wk + am 2 ¢av (4. 0) is said to have mean value fav<x) if there exists a continuous function 'Y ( T): IR+ . then the integral of the function d(t. as compared to the rate of time variation of f. By tak I _!_ T J to f(r.x. obtained for the original system through a very different path. We will concentrate our attention on the behavior of the solutions in some closed ball Bh of radius h. On the other hand.3) as long as r contains at least one sinusoid (including at w ing a Lyapunov function v = ¢~v. This is formulated more precisely in the following definition.43) is exponentially stable in Bh.2.O)fav(X) (4. T ~ 0. x) need not be a bounded function of time.fav(x) I :.2. there exists a > 0 such that a (¢av) ~ a> 0 for all I ¢av I : :.2 Averaging TheoryOneTime Scale 167 so that the averaged system is given by . Convergence Function The function f (t.2. strictly decreasing. In this book. a zero mean function need not have a bounded integral. it is not globally exponentially stable. we do not make the ~ssumption of (almost) periodicity.2. We first present the averaging theory that supports the frequencydomain analysis. y(T) (4.. O)dr (4..
5) Note that the averaged system is autonomous and.2 in Appendix. for all t 2: 0.x) I :::.d(t.. lemma 4. A useful lemma is given by Hale ([ 1980]. Lemma 4. d(t. 349).7) = I If. .(t. x) may increase as E. x) 0.2. x) = 0. assuming the existence of the mean value for the original system.6).x)l:::. for all x e Bh. modulo a simple time scaling by E. Moreover.(t. Lemma 4.2. 450 and Hale [ 1980]. although the exact integral of d(t. (t. Then aw. x) : lR. Then the functron HE) can be chosen to be 2al. + x Bh IR n is piecewise continuous with respect to t.1. x) at . w.2. x).n is a bounded function. 346.2.9) :::. TIE]. (t. the averaged system is defined to be Xav(O) = Xo (4. the following lemma leads to stronger conclusions that will be useful in the sequel. that has some additional smoothness properties. for all x e Bh. If. x) has zero mean value. TIE] are identical. x ). x) has zero mean value.2. p. but the proof of (4. such that (4.2. To compare the solutions of the original and of the averaged system.(O.6) (4.2. x e Bh· Moreover. it increases slower than HE)/ E. System (4..8) IEw.x)l (4. HE)!xl Then there exists HE) e K and a function w. for T fixed and E varying. 0) = 0. We address the following two questions: (a) the closeness of the response of the original and averaged systems on intervals [ 0.2.2.1 in Appendix. x) in the proof is identical to that in Bogoliubotf & Mitropolskii [ 1961 ].2.2.n. there exists a bounded function w. 7) is different and leads to the relationship between the convergence function 'Y (T) and the function HE). y(T) =a/ T' for some a 2:0. 1] the function HE) can be chosen to be 2al.2..1 is that. and has zero mean value with convergence function 'Y (T) ad(t. moreover. x): IR+ x Bh IR.(t. + x Bh IRn such that l~w. x e Bh. It is necessary to obtain a function w. moreover. x). Proof of Lemma 4. p. x) : lR. We state a generalized version of this lemma.2. r e (0. HE) (4.0. ~(E) (4. x) may be an unbounded function of time. piecewise continuous with respect to t. with convergence function 'Y (T)I xi and Approximate Integral of a Zero Mean Function Then d(t. (t. for all t 2: 0. as in lemma 4. (0.2 Averaging TheoryOneTime Scale 169 general case and indicate the importance of the function 'Y (T) in the subsequent developments. (t.10) for all t 2: 0.(t. An important lemma that leads to this result is attributed to Bogoliubotf & Mitropolskii [ 1961 ]. Moreover. (4. has bounded and continuous first partial derivatives with respect to x and d(t. (t. 1]. Proof of Lemma 4.1 If Comments The construction of the function w. Although the bound on w. it is convenient to transform the original system in such a way that it becomes a perturbed version of the averaged system.2. as indicated by (4. r e (0.2 Smooth Approximate Integral of a Zero Mean Function If d (t. the solutions over intervals [ 0.168 Parameter Convergence Using Averaging Chapter 4 Section 4. At the price of additional assumptions on the function d(t. w.6).1) will be called the original system and.2. lemma 5.2. (b) the relationships between the stability properties of the two systems. 'Y (T) = a/ T' for some a 2: 0. with convergence function y(T) ax there exists HE) e Kanda function w. x ). whose first partial derivative with respect to t is arbitrarily close to d(t. x): IR+ x Bh IR. The main point of lemma 4. p.
x) has zero mean put unbounded integral. Eo. f(t. x E Bh.170 Parameter Convergence Using Averaging Chapter 4 Section 4.2. 192).2. (4. If (E) is of the order of E+ HE). E1).1I T).2.3 Perturbation Formulation of Averaging If the original system (4.2.16) and \f(E).16). z. x.Xav· The first component results from a pointwise (in time) transformation of variable.x is invertible and as E . This component is guaranteed to be small by inequality (4.fav(X) satisfies the conditions Lemma 4.2. Proof of Lemma 4.18) is a Lipschitz type of condition on p(t. Under the transformation.2.2.9).2. b) Lemma 4.2.. Inequality (4.1) and the averaged system (4.15) may be replaced by _x(t) • z(t) 0 (4.. It separates the error in the approximation of the original system by the averaged system (x. It may indeed be useful to the reader to check the lemma in the linear periodic case. 0) = 0 for all t .2 and Et > 0 such that the transformation X = + . that is.2.2.:::: 0. x ).2 Averaging TheoryOneTime Scale 171 Comments The difference between this lemma and lemma 4. !av(O) = 0 and !av(X) is Lipschitz in x. \f(E)Izl (4. (t.Aav has zero mean value. x. Comments a) A similar lemma can be found in Hale [ 1980] (lemma 3. If d(t.2.2. the perturbation terms go to zero as E .11) for some A(t): IR+IR. E). 0. [ (A l (T) . c) At this point.nxn. Xt.1 provides a direct relationship between the order of the convergence to the mean value and the order of the error terms. then both x .)dT] z(l) Z + E W. E) is Lipschitz in E. linearly in x..(t. x.18) guarantees that this perturbation is small as E .2.. Further. Inequality (4.2.17) where p(t.2.12) (A2) for all t. which is not found in Hale [ 1980] and results from the stronger conditions and conclusions of lemma 4. The following assumptions will now be in effect. Assumptions For some h > 0.15) is a homeomorphism in Bh for all E ::. E) satisfies ip(t. it tends to the identity transformation.0. that is. for some lav. Note that if the original system is linear.1) becomes z = E/av(z) + Ep(t. Ez ::.. E) I ::. x) with respect toxin (4. HE) as in lemma 4.2. the function d (t.. that is.3 in Appendix. For E sufficiently small (E ::.:::: 0 lfavCXt).2.2 is that there exists Aav such that A (t).3 relates the function HE) to the error due to the averaging.2.e.10) and the dependence on I xl in (4. z. 'Y (T). z. x 2 e Bh. the transformation (4. E) is Lipschitz in x. E) z(O) = Xo (4. The proof of lemma 4.2.2.8).A.:::: 0. Lemma 4. Et and (4. Z) (4.Xav in the approximation of the original system by the averaged system. that is. z..fav(Xz)i S: laviXtXzi (4.2.2. Xz E Bh.13) (A3) for all t :?:: 0.z and z . but possibly more slowly than linearly (as fi for example). Et. system (4. E). .2.0. If d (t. x) of lemma 4.2.2.e. (t. x) has a bounded integral (i. i.2. z.1 is in the condition on the partial derivative of w. 0) .2. E) are of the order of E with respect to the main term !av(z ). then the main assumption of lemma 4. the transformation z.18) x = EA(t)x x(O) = xo (4.1.3 is fundamental to the theory of averaging presented hereafter. p.1 ). Then.2. = f(t.0.2. E0• f(t. for some / 1 :?:: 0 (4. HE) are of the order of E. Eo> 0 (At) x = 0 is an equilibrium point of system (4. The relationship between the functions 'Y (T) and HE) was indicated in lemma 4.5) satisfy assumptions (A1)(A4) Then there exist functions w.2.:::: for some If (E) e K. we can relate the convergence of the function 'Y(T) to the order of the two components of the error x .14) (A4) for all Xt. E::.z and · p(t.2. z.Xav) into two components: x. The second component is due to the perturbation term p(t.2. f(t. for some / 2 .
. ET > 0.2..25) for all t :<! 0 and some b. using the Lipschitz assumption (A3) and proposition 1.2 Averaging TheoryOneTime Scale ~ 173 We now focus attention on the approximation of the original system by the averaged system. lxav(t)l : :. a3.4 We apply the transformation oflemma 4.(E)bT :::.2.4 Basic Averaging Theorem If the original system (4. 0 Comments Theorem 4.(E))Iz(t)Xav(t)l 'f(t){h + (1 + lJ. on this time interval. the function fav(x) has continuous and bounded first partial derivatives in x.26) . so that lxzl ~ HE)Izl ~ Y. ET.19) for some bT :<! 0. Consequently.24) is valid for all t e [ 0.1). Under the hypotheses.XavU)I ~ Ef(E) [I z(r)l e.2. TIE] are unbounded. a2l Xavl 2 (4. Theorem 4. This relationship is investigated in theorem 4.4 does not allow us to relate the stability of the original and of the averaged system.2..2.5) satisfy assumptions (A1)(A5). (4.(lr) dr ~ Combining (4.(E 1))aT) lJ. ·. TIE].Xav) and (x.2.Xav) can be made arbitrarily small by reducing E.1IT).4.2. If d(t. theorem 4.2. the errors (z.2.2.e. (4. z E Bh.23) f(E)h [ e..XavU)I ~ Elav [I z(r).t ·~~v 1] (4.3..(E)aT ~ V(Xav) ~ .Xav(r)l dr + Ef(E) [I z(r)l dr (4. Xav E Bh'• h'<h. It follows.1) and the averaged system (4. Integrating (4.2..3).2.2. and for as long as x.172 Parameter Convergence Using Averaging Chapter 4 Section 4.5 The proof relies on the converse theorem of Lyapunov for exponentially stable systems (theorem 1. The error is of the order of f(E).4.2. h'<h. TIE) (this is possible.2. from a simple contradiction argument. (4.2) I I z(t). We will now show that. E Ix(t).4. when E is sufficiently small. Theorem 4.5) satisfy assumptions (A1)(A5) Then there exists Y. and that the estimate in (4.20) On the other hand. whenever E :::.Xav(O) E [fav(z).fav(Xav)) + tp(t. given T :<! 0 (4. (ET )bT < h .n IR+ and strictly positive constants a 1.20)..2. a2.2.3 such that. there exists a function v(Xav): IR.. a I IXavl 2 ~ ET. xo is sufficiently small so that.2.(E) as in lemma 4.2.2.5 Exponential Stability Theorem If the original system (4. for fixed T and some h' < h. that x(t) e Bh. TIE). . TIE] and Proof of Theorem 4.Xav) z(O) . It is important to remember that. 'Y (T). and for all t e [ 0. theorem 4.4 does not state that (4.21) I I I z(t).. then the error is of the order of E. (AS) Xav(t) e Bh' for all t e [ 0.h '.2.(E)Izl forE ~ E1.t..2. :::. Consider first the following assumption.22) Using the BellmanGronwalllemma (lemma 1. and x = 0 is an exponentially stable equilibrium point of the averaged system the equilibrium point x = 0 of the original system is exponenThen tially stable for E sufficiently small. for all Xav E Bh. although the intervals [ 0.2. E) (4. TIE].Xav(t )I lx(t)z(t)l + lz(t)Xav(t)l f(E)IXavU)I +(1 +lJ. Let ET (with O<ET ~Ed such that Y.4 establishes that the trajectories of the original system and of the averaged system are arbitrarily close on intervals [ 0.2.2.5.2. Z. we have that d dt (z. x) has a bounded integral (i.21) 0 for all t E [0. and the order is related to the order of convergence of 'Y (T).23) := lJ.1) and the averaged system (4..24) By assumption. Proof of Theorem 4. a4 such that.
Since the averaged system is autonomous.3) . Considering v (z ).5. with z replacing Xav· The derivative of v(z) along the trajectories of (4.28) are still verified. using previous inequalities (including those from lemma 4.. provided that E is sufficiently small. 0 In theorem 2. such that to+ 6 a2/ ~ J w(r) wT(r)dr to ~ a 1I for all t 0 ~ 0 (4. 4. 1 +HE) 1 .2.3. however. and define t 2 = min a () E .2. providedthat w is persistently exciting.3. (4.27). b) The proof of theorem 4. and any such function will provide a bound on the rate of convergence of the original system forE sufficiently small.2.2.30) implies that v(z) !S. the error x . inequalities (4. v(z(to)) e.34) for all t ~ to ~ 0. system (4.28).2. The proof assumes the existence of a Lyapunov function satisfying (4. Denote Let t2' be such that a 3 .17) !S.1 7) is exponentially stable. which is the bound on the rate of convergence of the averaged system that one would obtain using (4.2.33) Since a( t) > 0 for all E !S. but does not depend on the specific function chosen.2. (2.5). e !S.2.4.Ht) r~J+Jx(to)l ew<•><tto> a 1 ( 4 . and provided that all assumptions are valid globally. we found that system (4. nonautonomous system. 2 Using (4.26)(4.4. t2. Yet theorem 4. This is a useful conclusion because it is in general very difficult to obtain a guaranteed rate of convergence for the original. c) The conclusion of theorem 4.  WJI Xavl 2 (4.5 gives a useful bound on the rate of convergence of the original system. Since both x and Xav go to zero exponentially with t. it is usually easier to find such a function for it than for the original system.3 APPLICATION TO IDENTIFICATION (4.5 is a local exponential stability result.31) Comments a) Theorem 4. i.16).2.2) lx(t)l !S.5 to obtain a uniform approximation result. t 2'). to combine theorem 4. As E tends to zero.2.2 Averaging TheoryOneTime Scale 175 I (4.2._ . a3.27) is to be taken along the trajectories of the averaged system (4. a 2.3) v(z)l (4. a4j Xavl The derivative in (4.Xav also goes to zero exponentially with t.5) + [!V](Ep(t.Z.17) = v(z)l (4.27) (4.174 Parameter Convergence Using Averaging v(Xav) Chapter 4 Section 4.2. The original system will be globally exponentially stable if the averaged system is globally exponentially stable.5 is quite different from the conclusion of theorem 4. the original system is exponentially stable. with rate of convergence (at least) E a (E).2.  WJizi2+W4lf(E)Izl2 (4.2.30) for all t !S.1) The evolution of the parameter error is described by cP(t) = gw(t)wT(t)<P(t) <P(O) =<Po (4. it follows that cP(t) = ge 1(t)w(t) <P(O) =<Po (4.2.If (E2' )a 4 > 0.2.2. with an estimate similar to (4.3.17).2. The function v is now used to study the stability of the perturbed system (4.2.15).2.28) Ia~:v I !S.25). t 1· (Et. It is possible.2) is exponentially stable.3.2.E)) z (4.17) is given by Ji(z)l (4. and x(t 0 ) sufficiently small that all signals remain in Bh· In other words.2.2.2.26) and (4.2. In other words.2.4 and theorem 4. we will study the case when g = E> 0 and the update law is given by (cf.1. the proof provides a bound on the rate of convergence.2.2. w (E) tends to EI 2 a 3 I a 2.2.2.1 )) and (4. E .2.\f(t)a4 2a2 Consequently.2w(•)(t to> (4. where z (x) is defined by (4.29) and.2.2. and this bound gets arbitrarily close to the corresponding bound for the averaged system.32) To apply the averaging theory to the identifier described in Chapter 2.5) !S.5 does not relate the bound on the error to e. there exist constants a 1.2. o> 0.26)(4.e.2.
(4.3.(dw) r ] S.13) 2 Re [ Hw.3. Recall that Rw(l) may be expressed as the inverse Fourier transform of the positive spectral measure Sw(dw) Rw(l) = Convergence Analysis When w is persistently exciting. we have that Rw(O) = (4.b. I ¢avl 2 1 T 2 <Pav<Pav (4. in terms of rates of parameter convergen. (dw) is greater than or equal to 2n points (the dimension of w = the number of unknown parameters = 2n)..8) may be replaced by a summation Rw(O) oo Since S.3.3 Application to Identification 177 On the other hand. Rw(O) is also given by Rw(O) = 1 11' = ~ k rl 2 Re HwrUwk) HwrUwk) [ A 0 AT ) (4.8) then the integral in (4.3.. w is persistently exciting (PE) if and only if Rw(O) is positive definite.17)) Hw.3.3. w(r)wT(l+r)dr e IR.6) Therefore. Our interpretation is. it could be used to determme the spectral content of the reference input that will optimize the rate of convergence of the identifier. then w is also stationary. With knowledge of the plant. . given the physical ~on.(Jw)S. 7. the averaged system corresponding to (4.(Jw) AO oo It was shown in Section 2.. S~ch a procedure is very reminiscent of the proce?ure 10~1cate? 1~ Go?dw1~ & Payne [ 1977] (Chapter 6) for the design of mput s1gnals m 1dent1ficat10n.3.9) lim Too ~ J to to+ T This system is particularly easy to study.ce of the averaged system rather than in terms of parameter error covanance in a stochastic framework. since it is linear. the minimum and maximum eigenvalues of Rw(O). Equation (4. AT . the rate of exponential convergence of the averaged system is at least gAmin (Rw(O)) and at most g Amax (Rw(O)).3. (2.11) to determine the rate of converge~ce is limited.straint~ on r. 7.4) where we used the notation of Section 1. 7 (proposition 2.2.2.5) and gAmin (Rw(O)) I <Pav 1 2 ~  v(¢av) ~ .3.(Jw)Hw. We can conclude that the rate of convergence of the original system for g small enough is close to the interval [gAmin (Rw(O)).11) Further. A natural Lyapunov function for (4.7) where Amin and Amax are. respectively.6 for the autocovariance of w evaluated at 0.3.2) is simply Since the transfer function Hwr depends on the unknown plant being identified.3. ] (slAt 1 !Jxfi(s) 1 (4. while a sinusoidal component contributes two points (at + w and .8) gives an interpretation of Rw(O) in the frequency domain.(jw)H'[.3.w). and also a mean of computing an estimate of the rate of convergence of the adaptive algorithm.(Jw)S.(Jw) Hw. 134 ). . If the input r is periodic or almost periodic r(l) = ~rk sin(wkt) k and. Note that a DC component in r(t) contributes one point to the support of S.12) 2~ OOJ J fi. With these definitions. however. w is the output of a proper stable transfer function fi wr given by (cf.3.(s) = [ (slA).2) that this is true if the support of S. the averaging theory presented above leads us to the limit Rw(O) := rP av =  g Rw(O) <Pav <Pav(O) = ¢o (4. using (4. Rw(O) is a positive definite matrix.g A max (Rw(O)) I <Pav 1 2 (4. (dw ).7).(dw) (4.5) and (4. The autocovariance matrix defined here is similar to the average information matrix defined in Goodwin & Payne [ 1977] (p. g AmaxCRw(O)) ].3. the use of (4. Its spectrum is related to the spectrum of r through Sw(dw) = Hw.16)(2.1) that when w is stationary.9) is v(¢av) = .3.(dw) is an even function of w. Thus.176 Parameter Convergence Using Averaging Chapter 4 Section 4.nxn (4. given the spectral content of the reference input.10) 2~ J eiwtSw(dw) oo 00 (4.(dw) 00 A. It followed (proposition 2. if the input r is stationary.3.3.
error first moves. Example To illustrate the conclusions of this section.9). We consider an input of the form r = r 0 sin(w0 t). We found. (1 2 .3. (4. we find that the estimates provided by direct proof and by averaging are essentially identical for g = E small.4 show the plots of the para~eter erro.3). 0. and with two different adaptation gains g = 1.1 was based on the Lyapunov function of theorem 1.1. so that We notice the closeness of the approximation for g = 0.3).1. kp s + ap (4.o.15) are computed to be . Since the number of unknown parameters is 2.6 are plots of the Lyapunov function (4. Figures 4. that the estimate of the convergence rate tends to ga 1 I~ when the adaptation gain g tends to zero.25 cp1 0. It is also interesting to compare the convergence rate obtained through averaging with the convergence rate obtained in Chapter 2.5. (4.191 g. it slowly moves along the direction correspondmg to the small eigenvalue.3. the gain I 1 can easily be taken into account. and g = 0. ap = 1. The constants a 1 .3. we averaged the differential equation itself and found that the norm becomes a Lyapunov function to prove exponential stability.0. using a logarithmic scale.5. 309 g. 0.9). / 2 = 10 are arbitrarily chosen such that I~{} 1)1 = 1).3) and (4. It shows the two subspaces correspondmg to the small and large eigenvalues: the parameter . parameter convergence will occur when the support of S. we consider the following example A c/>T(O) = (0. one as a funct~on of the other when g = 0..15) with cl>av(O) = c/>o.3 + YS g 4 3 T .14) Tlme(s) 10.1. Figures 4. When ro = 1. corresponding to the two eigenvalues. fast along the direction of the eigenvector correspondmg to the la.3 Application to Identification 179 Note that the proof of exponential stability of theorem 2. The averaged system can be found by using (4. Although~ is not monic.4.0 The filter is chosen to be ~(s) = (s+/ 2 )11 1 (where / 1 = 10. .375 0. We observe the tw? slopes.5 and 4.3. kp = 2. ~ resulted from the PE condition (2.1. Then. the eigenvalues of the averaged system (4.1.rge eigenvalue. We let 0{0) = 0. Figure 4.1 to 4.3.rs c/> 1 and c/>2.125 .ap) I / 1).05.125 \Averaged P(s) = .1 that was an average of the norm along the trajectories of the system. By comparing (4. wo = 1.5 0. The closeness of the estima~e of the convergence rate by the averaged system can also be appreciated from these figures. i.3. Y5 rr 4. In this chapter.5. and .e. (dw) is greater than or equal to 2 points..3. in the proof of exponential convergence of theorem 2.199.4 AVERAGING THEORYTWOTIME SCALES We now consider a more general class of differential equations arising in the adaptive control schemes presented in Chapter 3.o 0.g = . for both the original and averaged systems.3.4).13) Figure 4.7 represents the two components of c/>..3.1: Parameter Error c/> 1 (g = 1) apkp cP av g r6 It 2 1}: +wfi wfi+aj apkp w2+a2 0 p k2 p wfi+aj ' cPav (4. so that the support consists of exactly 2 points.10) for g = 1 and g = 0.178 Parameter Convergence Using Averaging Chapter 4 Section 4.1. The nominal parameter 4 (kp lit.
J__ __l__ _.45 ln(v) 3.LL='.0 c/>2 0.75 0.0 2.Lyapunov Function (g = 1) .l.90 0.3: Parameter Error ¢ 2 (g = 1) Figure 4.45 0.4: Parameter Error ¢2 (g = 0.75 25 50 75 100 0.L.15 0.50 L _ _..1) 0..60 0.25 7.225 0.. 0.4 Averaging TheoryTwoTime Scales 181 Averaged 0.0 6.5 10...L.0 7.25 0..1) Figure 4.075 0.30 2.5: Logarithm of the .0 7.0 Tlme(sl Tlme(sl Figure 4.15 1.0 0.0 2.15 0.225 0.5 10.2: Parameter Error = 0.50 c/>2 0.60 5.30 Averaged c/>1 0.180 Parameter Convergence Using Averaging Chapter 4 Section 4.0 0.0 0.15 0.5 5.0 25 50 75 100 Tlme(sl Tlme(sl ¢ 1 (g Figure 4.0 /Odgiool 0.00 \ Averaged 0.75 0.075 0..5 5.90 0..
3) Tlme(s) and the system Figure 4.y) A(x)y + Eg(t.0 c/>1 0.1 Averaging TheoryTwoTime Scales 183 Separated Time Scales We first consider the system of differential equations E/(t.7: Parameter Error cp 2 (cp 1) (g = 0. 0) dT to (4. q 1 !.2) is linear in y. q 2 > 0.1 0.7L::::. there exists P(x) satisfying p 1 !5. p 2.1) Xav(O) = Xo (4.4. We make the following additional assumption.n.2). x e IR. whose dynamics are of the order of E with respect to the fast dynamics. andy e IR.1) (4. The state vector is divided into a fast state vector y and slow state vector x.182 Parameter Convergence Using Averaging Chapter 4 Section 4. y) 0. As previously.45 0. X. It is also possible to show that the definition is equivalent to requiring that there exist p 1 . . We will make the following assumptions..4.50 0.75 0.0 0.m. The dominant term in (4.2) Jn(vl where x(O) = Xo.. if there exist m.AT(x)P(x) +P(x)A(x)5. x.J___j__ 25 50 _J__ _.4 4.4.4.p 2 I.1) Comments This definition is equivalent to require that the solutions of the system y = A (x) y are bounded above and below by decaying exponentials.1)(4.0 (4. such that for all x e Bh and t . X.4.x.J__ _L __ ____j fav(X) = 75 100 roo lim ~ to+ T J j(T.4. q 1 . m'.2 Figure 4. we define a 7.4.::: 0 (4.4) is the averaged system corresponding to (4. and q 2 I 5. . independently of the parameter x. but is itself allowed to vary as a function of the slow state vector.4.P(x)5.4.5) 0.90 0. y(O) =Yo. X'> 0.15 0.0 . such that for all x e Bh.2 0. Definition Uniform Exponential Stability of a Family of Square Matrices The family of matrices A (x) e lR m x m is uniformly exponentially stable for all x e Bh.1 0.6: Logarithm of the Lyapunov Function (g = 0.60 0.30 c/>2 0.
x.4.4. 0) = 0 for all t :::::: 0.E)+EP2(t.4.. lav· Proof of Lemma 4. x) = f(t.4 Then Averaging TheoryTwoTime Scales 185 Assumptions For some h >0 (B1) The functions f and g are piecewise continuous functions of time. E)i ::::.4. This assumption is technical.:::: 0 there exist functions w.fav(x). given T :::::: 0 (B4) convergence function 'Y (T).4. HE)izl Under the transformation.Z. so that d(t.4.x2.2.3 is an exponential stability theorem. k21YI (4. x) has zero average value. so that for some lav .xi.2) and the averaged system (4.2. 0) has continuous and bounded first partial derivatives with respect to x.5.(t. TIE]. HE) as in lemma 4. Jx(t)Xav(t)l ::::. fav(O) = 0. we first obtain the following preliminary lemma. x.4..4.(E) as in lemma 4.4. / 2 . 1/I(E)br (4. Theorem 4.1 )( 4.4) satisfy assumptions (B 1)(B3) .11) where (B2) IPI(t.4. k2 depending on /1 .4. Yo e Bho (where h'. 0.YI) f(t. x) has zero average value.1 )( 4. g(t.2 and E1 >0.:::: 0 (4.. 7) (B3) Z = Efav(Z)+EPI(t.2).(t. z.Y2)l::::. and x e Bh.4. and for some / 1 .E)l ::::.1)(4. when E tends to zero. IIIxix21 + 121YIY21 Jg(t. y 1.2 is an approximation theorem similar to theorem 4. We are now ready to state the averaging theorems concerning the differential system (4.184 Parameter Convergence Using Averaging Chapter 4 Section 4.z. Theorem 4.1) becomes (4. for all t : : : 0. l3lxix2l + 14IYIY21 (4. 0) = 0. HE)kiizl IP2(t. Er.Y2)l::::.3 Exponential Stability Theorem If the original system (4.4. for some ka:::::: 0 II a~~) II : :. Moreover.2.. and fav(x) ~1as continuous and bounded first partial derivatives with respect to x. and continuous functions of x and y.x 2 e Bh· Let d(t.4) satisfy assumptions (B1)(B5).4.4. /2. and that ad~. similar to theorem 4.2. such that the transformation (4. Assume that the convergence function can be written as 'Y(T) Ixi.4. xI> x 2 e B h.10) lf(t. the function fav(x) has continuous and bounded first partial derivatives in x. x). I XavU )I e B h • on the time intervals considered.4.4. and 1 Jxzl ::::.4.4..4.4.x2. and x = 0 is an exponentially stable equilibrium point of the averaged system Then the equilibrium point x = 0.1 in Appendix. Theorem 4./4. and will allow us to guarantee that all signals remain in B h • As for onetime scale systems. f(t.12) forallx 1 .. system (4. y. Also assume that f(t. (BS) ka for all x e Bh (4.3. Theorem 4. y = 0 of the original system is exponentially stable for Esufficiently small. 0) has average value fav(X ).2) and the averaged system (4. y 2 e Bh..4. 0). similar to lemma 4.1 Perturbation Formulation of AveragingTwoTime Scales If the original system (4.4. E . with for some k 1.2 Basic Averaging Theorem If the original system (4.z) is a homeomorphism in Bh for all E ::::.E) z(O) = Xo (4. and E ::::. Er > 0 and for all t e [ 0.4.9) X=Z+EW. ho are constants to be defined later).6) for all t : : : 0.4) satisfy assumptions (B 1)(B5) Then there exists 1/. and for some h 0 .4.xi.2) and the averaged system (4.13) for some br :::::: 0. Lemma 4.3 such that. x.y.1)(4. A(x) is uniformly exponentially stable for all x e Bh and. Moreover. /3. The function f(t.4 and guarantees that the trajectories of the original and averaged system are arbitrarily close on compact intervals.8) For some h' < h.2.YI) g(t. 0.. for all x e Bh.z.
y + v(t.4. x) The averaged system is obtained by averaging the righthand si~e of (4.2 Mixed Time Scales We now discuss a more general class of twotime scale systems. := JeA(x)(tT)h(r.x)) ax + g'(t. x) (4.5.4.17) x)] dr (4.x.x)) ax (4.28).20) is of the form of (4.1).4.4.4.v(r.14 )( 4.x.3 gives a useful bound on the rate of convergence of the nonautonomous system..15). i. X. and II ah ~. 0) = 0.4.x)f'(t. we apply the previous results to the output error direct adaptive controlalgorithm for the relative degree 1 case.4. x.4.x.2 and 4. ~ 0.4. 0) = 0 for all t t~O.15) We will show that system (4.18) Since v (t.4.4. y') A(x)y + E ( av(t. it is usually easier to obtain such a Lyapunov function for the averaged system than for the original nonautonomous system.4.21) + g'(t.4. and (A4.x. x) A (X) V (t.xeBh.2) when f(t.X) + ~ [eA(x)(tT)] h(r. we req~ir~ Ass~mpti?ns (B1):(B5) to be satisfied.2).4. x) + eg'(t.14).4. Consider now the transformation y = and using the fact that eA<x)(tT) and j__ eA(x)(tT) are bounded by ax exponentials (( 4.e.3) (w1th f given by (4.4.y+v(t.4. y') A(x)y' + h(t. g'.14). so that the definitions (4.14). x) satisfies a 7ii v(t. y'. 6 f [eA(X)(/T) ah(r. x) replacing y' in (4. v (t.4. lim Too ~ lofT J f'(r.0. X) v (t' 0) 0 (4. x) represents the steadystate value of the variable y' with x frozen and e = 0 in (4. x) (B6) h (t. Since the averaged system is autonomous.21)) agree. x is a slow variable.v (t. the rate tends to the bound on the rate of convergence of the averaged system that one would obtain using the Lyapunov function for the averaged system. The averaged system corresponding to (4.4.4.x) dr 0 I (4. In this case.4.3). x. X) + h (t.5 with (3.av(t.4.16) II av ~.4.4 Averaging TheoryTwoTime Scales 187 Comments As for theorem 4.4. ax. In particular. x))) (4.5 APPLICATIONS TO ADAPTIVE CONTROL For illustration.4. 4. we have that e(t) = Ame(t) + bmr/JT (t )wm(t) + bmr/JT (t) Q e(t) ' . Then.4. arising in adaptive control ef'(t.y+v(t. and assume that the following limit exists uniformly in t and x faix) = av(t. X.1 )( 4. (4.4. and the following assumption on h (t. (4.4.x) f'(t. but y' has both a fast and a slow component.x))dr lo (4. x) II is bounded for all This new assumption implies that v (t.x)) . f is averaged with v (t.186 Parameter Convergence Using Averaging Chapter 4 Section 4.x)) (4. and conclusions about its exponential convergence can be applied to the nonautonomous system for e sufficiently small.x.2.17).4.4. y) f'(t. since ax.5).30) in the proof of theorem 4. and ( 4.20) so that (4.21) with y = 0.y+v(t. we assume Similar Lipschitz conditions on f'.23) Intuitively. To apply theorems 4.x. (4.x) ax.4.14) (4.y) g(t.15) is obtained as follows. As E .4.3.4.15) can be transformed into the system (4. x e Bh.y+v(t.22) (4.5.19) 4. the proo" of theorem 4. Define the function v (t. We established the complete description of the adaptive system in Section 3. x) II It also implies that is bounded for all t = ~ 0.
5.5. then tum to the nonlinear equations.22): Note that the.1.1 Output Error SchemeLinearized Equations The linearized equations. are (sf . v(t.. Wmf(t) ~ .6) Wmf(t) = Jc~eAm(IT)bmWm(r)dr 0 I _!_ M(wm) (4.5. c/> = 0.6. (4.m.A).4) (0) = .11)) Since r is a scalar.(jw)H'!.A). 211'Co ..4._ T J 10 Wm(t) W~f{t) dt (4.7) and (4.14).5.7) co r/>(t) (4.9) Frequency Domain Analysis To derive frequency domain expressions. we may use (4.1 bJr 1 M . S.M which is a stable transfer function.1bm We let 1 = co M(s) is stable.1b>. signal WJ was also used in the direct proof of exponential convergence m Chapter 2 (cf. lm ( ii.34)). ¢) defined in (4.5.6.8) to obtain a frequency domain expression for Rwm wmr (0) as R (4.lim [ieA. while the second in antisymmetric.¢) and !av is given by . They describe the adaptive control system.5. 4.5.1) is a set of linear time varying differential equations. lo which was called the cross correlation between Wm and Wmf (evaluated at O) in Section 1.3) to+ T Rwm wmf (0) = T+00 lim . (2.1) Section 4.1Wm Wmf • 00 J ii.5 Applications to Adaptive Control 1 189 gc~e(t)wm(t).(jw)H'!.4.5.(jw)) Re M(Jw) S.(dw) m (4. Consequently. we assume that r is stationary.. linearized around the equilibrium e = 0. describing the adaptive system for small values of e and ¢.16) is now where the last equality follows from (3.5) where the transfer function from r. Recall that Am is a stable matrix.4.5.(jw)H'!. We first study these equations. Define a filtered version of Wm to be (4.wm is stable.5. it is easy to see that (4.gc~e(t)Q e(t) where g is the adaptation gain.(dw) .5.5. The spectral measure of wm is related to that of r by (4.M (sf . Since the transfer function from r.m. Since CmT(sf Am).m. .6). •7 7 Re ( ii. Section 1.m. With the exception of the last terms (quadratic in e and c/>).(jw)M(jw)S.Wm is given by (using (3.oo m .<Hbmw~(r)dr T+oo ]• (4. this implies that wm is stationary.15) with the functions f' and h satisfying the conditions of Section 4.5.(jw)] lm M(Jw) S.!.5. This may be used to show that 211'Co. lS statwnary. The function v (t.2) is of the form of (4.5. .oo + .8) ~ to+ T J Wm(t)c.5.(dw) 211'Co oo and that the first matrix in the righthand side is symmetric.4.(dw) is real and consequently an even function of w (cf.188 Parameter Convergence Using Averaging d>(t) = Chapter 4 (4.2) Since Wm is bounded. (4.
1 (0)] rf>av . w 1 (0) + RJ'.20) m (a~+ w6) . so that its eigenvalues need not be real:'" ifowever.r(}wk )Hw . (4. the real .14) amkm r6 kp r/> av = .5./0) is a positive semtdefimte matnx.16) V(cf>av) = gcf>J:v [Rw. ~T ~. when no unmodeled dynamics are present.10) Example As an illustration of the preceding results.12) To guarantee persistency of excitation. R (0) need not be symmetric. we use a sinusoidal input signal of the form r(t) = rosin(wot) The matrix in parentheses is symmetric positive semidefinite.15) We will choose values of the parameters corresponding to the Rohrs examples (Rohrs et a/ [ 1982]. The adaptive process is to adjust the feedforward gain c0 and the feedback gain do so as to make the closedloop transfer function match the model transfer function M(s) (4. and a natural Lyapunov functiOn IS agam (4.. As previously. r(t) = ~rk sin (wkt) (4.co do(t) do (4.4) shows that the averaged system is a LTI system rf> av = .w. it is positive definite if Wm is PE.wfi) (a~+ w5) 2 amk~ (a~ +wfi) 2 rf>av  (a~+ wfi) k2 (4.5.10) is now (4.8).5.e.5.. that the averaged system defined in (4.(t)r(t) + cf>y(t)ym(t)) ge 0(t)r(t) an expression for Rwm wmf (0) is 1 r( . (4.5.2 km k~ (a~.17) (4. ~ [ T ~ Co k Re ( Hw.14).3.5..5.5.. w.g Rwm w...ictly positive real.7) and (4..5.5.190 Parameter Convergence Using Averaging Chapter 4 Section 4.5.5..18) rJ>y(t) (4. using (4. the matrix Rw.5.1 (0) rf>av rf>av (0) = r/>o (4.g .2) . we consider the following example of a first order plant with an unknown pole and an unknown gain Convergence Analysis Sin~e M(~) is str. (4.5.5.11) and P(s) s +ap (4. see also Section 5.(jwk) m ~• .Pa~s ar~ guaranteed to be positive.19) It can be checked.5 Applications to Adaptive Control 191 With (4.2) becomes eo(t) am eo(t) + kp( r/>. i.5.(t) ) ~ y(t) where r/>. Unlike the matrix Rw(O) of Section 4.5. When the reference input r is periodic or almost periodic..(t) coU) .13) Thus.
10: Parameter Error c/Jy (¢.0 0.2 0.0553 ±} 0. with three different frequencies (w 0 = 1..9 and 4..2 0.2 1.(t) Am e(t) + bm cpT (t) Wm(t) + bm cpT (t) Q e(t) gwm(t)c.1.8: Parameter Error c/Jy(c/J. ro = 1.. This explains the oscillatory behavior of the original and averaged systems observed in the figure.5 cpr Figure 4.5.21) Figure 4.0 1. the two eigenvalues of the averaged system are computed to be .0 0.) (r = sint) Figure 4.. 3.4 0.5537 g.8.10 corresponds to a frequency of the input signal w0 = 5. 4.) (r = sin 5t) From (4.25 1.05076 )g.0.4.192 Parameter Convergence Using Averaging Chapter 4 Section 4. km = 3.) for the original and averaged system. which did not exist in the previous examples of Section 4. 1). and are both real negative.5 0.3.0.0. g = 1. We let 8(0) = 0. kp = 2.6 Parameter Error c/Jy(c/J.8 0.0 1.4 1.5._ __L_ __l__~ cpy 1.0 o.10 show the plots of the parameter errors c/Jy(c/J.5 0. nonlinear differential equations cpy 0. so that cpT(O) = (.) (r = sin3t) /Averaged 1.5 1. The nominal parameter (J*r = (kmlkp.6 Figure 4. (ap.e(t) gQe(t)c.5 e(t) J.5 0.o L _ 1.4 0.8 cflr 0.3 0.6 0.9: 0. such that the eigenvalues of the matrix Rwm wmf (0) are complex: (.45) . c/>y 1. wo = 1. 5).5 Applications to Adaptive Control 193 With am = 3. ap = 1.0 0..L. 4.5 __L_ __L__ __t_ __L_ _..5.am)/kp).e(t) cpr (4.0 0. Figures 4.2 Output Error SchemeNonlinear Equations We now return to the complete.0163g and .9 0.
bm ¢T Q)l ~ bm ¢ T Hwmr(jw) ) S.5.5.e.1. w.5.5.¢)]dt (4.5 (cf. whose m poles are the zeros of det (sf Am).23) where !av is defined by the limit !av(¢) =  lim Too ~ J to [wm(t)c~v(t.5. . This means that if the controller parameters are frozen at any point of the trajectory the resulting time invariant system must be closedloop stable..25) and write the transfer function (4. as n(s) Xm S ~ the ratio of a vector polynomial where S.e.8). w is related to r through a linear time invariant system.5.5.+b.r+iTw.5.5.J iJ.27) so that the averaged system is cPav = gfav(cfiav) to+ T ¢(0) (4.30) H..5 Applications to Adaptive Control 195 Je(A.32) xm(s) Denote¢. (4. so that ¢Tw = ¢. Frequency Domain Analysis The expression of fav in (4.29) fav(¢) = 1 7r J [fiw.28) so that using (4. (4. Q)(tT)bm¢Twm(r)dr 0 I T . i.e.11)) that W = fl(s) (r + _!_ cpTW] co (4. but some manipulations will allow us to obtain a more interesting result. rp) m Recall that w is related to r through the transfer function iJw . precisely because w depends on e.24) can be translated into the frequency domain.25) can be replaced by (4. and a characteristic polynomial Xm(S ).(jw. "' by noting that w(t) = wm(t) + Q e(t). wm is an exogeneous signal. related to r through a stable transfer function.= cac 0.(dw) is the spectral measure of r.5. If det (sf Am .21) was obtained from the differential equation n(s ). Aav( ¢) =  1 .(dw) Xm(s) = det(s/ Am) (4. when ¢ is constant. assuming that the slow variable (i.22) Section 4.5. we can write (4.32) becomes . with a transfer function depending on ¢.26).5.27).24) is uniform in the sense of (B3) and provided that the matrix Am+ bm ¢T Q is uniformly exponentially stable for¢ E Bh..bm¢T Q ) is Hurwitz (as we assume to apply averaging).}w)) [ cJ.5. iJ_ (s) wmr = ( ) (4.31) We found in Section 3. e) is averaged.5.(jw)+Q(jwlAmbm¢TQ).5.1 (0) in (4. but now depends nonlinearity on¢. Therefore. but depends on internal variables.¢) = Chapter 4 (4.(dw) 27rCo _ 00 00 (4. ¢) H !. In general. The signal w is not an external signal.q. the differential equation describing the fast variable (i. Recall that (4.24) + The assumptions of the theorems will be satisfied if the limit in (4.5. noting that Wm is related to r through the vector transfer function Hw . However..5. In the context of averaging. assuming that ¢ is fixed. On the other hand.io~ in (4.1 2 oo 00 Explicit Expression for Hwr(s. (3. (4. ¢) M(Jw) S.g w(t)c~ e(t) (4. this transfer function is stable.¢)c~v(t.5.27) is of limited use. (}w/ Am.5.5.. ¢) is constant.(Jw. Note that fav can be factored as (4.¢) Qv(t..5.26) where Aav: IR 2nIR 2nx 2n is similar to Rw w /0) in Section 4.25) is more complex than in the linear case. Assuming that¢ is constant.v m. The expres. i.194 Parameter Convergence Using Averaging v(t.5.9). (3. Let bm ¢T Hw.5. with an expression similar to the expression of Rw.e..r(.5.
I [ IR.5 Applications to Adaptive Control 197 0 B(r/>) := 0 n(s) ~·I E IR 1 x 1 IR I x 1n . r/>.196 Parameter Convergence Using Averaging Chapter 4 Section 4. B(r/>) E IR2n x 1n Denote Xq.(s) X. A av can be written Xm Xm (.5. Therefore.5.p(S) m (4.38) co (4. the poles of the model transfer function with feedback r/>. Xq.p AavC1>) = On the other hand 21rco r j I xmUw) 1\I +B(¢))H.(S) = det (sf Am bmr/>TQ) The vector transfer function Hwr can therefore be expressed in terms of the vector transfer function fi wm.39) Xq.35) and.(jw) oo xq.37) can be written := Xm (s). X.m ) + w A A Xq.e.5._!__ co ¢T fi(s) (4.41) With (4. w m l v(¢av) = I rPavl 1 = rP~vrPav· Note that by (4.2n .42) Consider now the trajectories of the averaged system and let ~ m (r) _ ~ m A Xq.(jw)(l + B T(¢))M(jw) S.~~fi CoXm A · if[ . Define .5.38).5. r+r * Co l (4.p _ Co .(dw) "' ] · (r) 1> ~T (4.40).Uw) Xm Xq.5.5.. (4.36) Hwr(S. (s) is also given by Xq.. 0) Convergence Analysis Hw . as expected (4.33) i..5.33) can be written X m w = ·Hwmr A A [ Xq.5.36)(4.5.34) W = [~) = !:[:m]+!: B(rJ>)·[:m] (4..5. giving the poles of the adaptive system with feedback 8.I x I IR.5.37) Denote R(¢av) .(s) "' (4.2n .5.5. by Xm(S) (/ +B(rJ>))Hw .43) (4.5.(s) so that (4.40) With this notation. ll. it follows that for all¢ (4. X.(s) is closedloop characteristic polynomial.5. that is.I x 2n  I l (4.
r5 .g e0(t) r(t) 199 · ff[ . By theorem 4. . Example We consider the previous two parameter example. Using this result. R(rl>av) is at least positive semidefinite.45) .47) which is identical to the expression for the linear case (4. + 2 kp am km ] 2 2 rp.3.the result is the. The adaptive system is described by 1/>T c/> w5 + . then averagmg 1t (cf. .kp 1/>y) amkm 2 T km k~ (a~.(dw) m It follows that the derivative of v is given by (4. We can also verify the expresswn usmg (4. we obtain.w5) (w5 +a~ ) 2 .5. Rates of convergence can also be determined.22) . In fact.5.5. sa~e as the result obtained by first linearizing the system. 4> y wo+am = (4 5 49) . After ~?m~ manipulations.5 ~ . The averaged sy. using (4.g k2 p  w5 +(am.V = 2g a~ 2 ] r6 kp [ wo +(am.44) Section 4.4.stem ~an be computed using (4.I 2 ~1 amk~ 2 2 wo+am (4. It is interesting to note that.198 Parameter Convergence Using Averaging Chapter 4 (4.5. for the averaged system (droppmg the av subscripts for simplicity) 1/>. kpam km w5+ a~ + [a~~ 2 2 wo+am k+.5.e. or using (4. provided that R (1/>av) given in (4. This indicates that the adaptive control system may not be globally exponentially stable. using the Lyapunov function v( 4> av).9).20)). it can be seen.5. Using the Lyapunov function v(rl>av).50) w5+ a~ (wB +a~ )2 It can easily be checked that when the first term in brackets is equal to 1 (i.5.5.5. from the expressions of the averaged systems .geo(t)ym(t) (4. with cf>y replaced by zero).. The term I Xm I X<1>l 2 is strictly positive. for g sufficiently small. given any 4> bounded.1 only by the scalar weighting factor I Xm I x</>1 2 • Recall that xm(s) defines the nominal closedloop poles ( i.(t) ~y(t) Applications to Adaptive Control . when 4> = 0. Since M(s) is strictly positive real.kp cl>y) 1/>.12). and it approaches unity continuously as 4> approaches zero.42)(4.45).geJ(t).5.kp c/>y) + kp c/>. (4. lrl>avl increases. it is positive definite if Wm is persistently exciting.(Jw)M(jw)S. As in the linearized case. as Amin(R(rl>av)+R T(rl>av))) tends to zero in some directions.5. the exponential stability of the original system is also guaranteed for g sufficiently small.5. c/>T w5+a~ k~ amk~ 4> (4. so that Consider the case when r = r 0 sin(w0 t).5. It is remarkable that this result differs from the expression obtained by linearization followed by averaging in Section 4.48) c/>y r6 gkpT [ kmcl>r + c/>y w5 +(am. while X<1> defines the closedloop poles with feedback gains 0 = 4> + 0 .kp r/>y) 2 [(am.5.5. we find that for v and the guaranteed rate of parameter convergence of the averaged adaptive system is g a.47) and the definition of the averaged system (4. The rate of convergence of the original system can be estimated by the same value.e.42). this argument itself constitutes a proof of exponential stability of the averaged system.44) replaces Rwm wm/0) given in (4.5.43).
The first set of figures is a simulation for initial conditions ¢.38) and (4. averaged.3 Original l 90 120 Figure 4.5.5. and linearizedaveraged systems (the minimum slope of the curve gives the rate of convergence).4 30 60 90 120 Linearized.26). and (4.5. and cf>y· Figure 4. km = 3. kp = 2. g = 1.5.5. Figure 4.12 and 4. since parameters eventually get arbitrarily close to their nominal values. Due to the change of sign in cf>y(O).5.5. The slope for the linearizedaveraged system is asymptotically identical to that of the averaged system. ln(cjl cj>} 0. 0 0.bm cf> T Q) is Hurwitz). that the system obtained by linearization followed by averaging is identical to the system obtained by averaging followed by linearization.2 It shows the close approximation of the original system by the averaged system.5.5. wo = 1.12: Parameter Error ¢. but initial conditions ¢. 200 Parameter Convergence Using Averaging Chapter 4 Section 4. Also.(0) = .9).42) ).11 represents the time variation of the function ln (v = cf> T cf>) for the original. r0 = 1.averaged Time(s} c/>y 0. ¢y(O) = . (4.1 0 0 30 60 Time (s} Figure 4. given any prescribed Bh (but such that det (sl Am ..5.0.11: Logarithm of the Lyapunov Function 0.0.2 0.13: Paratp.50) can be used to obtain estimates of the rates of convergence of the nonlinear system. (4.13 show the approximation of the trajectories of¢.eter Error cf>y .I c/>r 0.23) with (4. ap = 1. Figures 4.. We reproduce here simulations for the following values of the parameters: am = 3.6 0. the rate of convergence of the nonlinear system is less now than the rate of the linearized 0.3 Time(s} Figure 4.5 and ¢ y(O) = 0.(0) = 0.1 0) with (4.5 Applications to Adaptive Control 201 (( 4.14 represents the logarithm of the Lyapunov function for a simulation with identical parameters.1 0 O.
due to the nonlinearity of the differential system.15 shows the statespace trajectory rJ>y(r/>.)(r = Time(s) sin5t) Figure 4.).0 1.e.5.28)) (4. For simplicity.15 shows the distortion of the trajectories in the statespace.5 Applications to Adaptive Control 203 system. and it should be noted that this is achieved despite an adaptation gain g equal to 1. ""~LinearizedAveraged 1.5 1.5. 4.51) so that the averaged system is again of the form ~av = gAav(rPav)cPav (4.3 Input Error Scheme An expression for the averaged system corresponding to the input error scheme may also be obtained.202 Parameter Convergence Using Averaging Chapter 4 Section 4. corresponding to Figure 4. r/>) Recall that (cf. but for values which are often practical ones. Within the framework of aver~g ing.0 0. not necessarily for very slow adaptation).3 0.5 60 90 120 cpr Figure 4. we neglect the normalization factor in the gradient algorithm and the projection for the feedforward gain c0 • g v(t)vT(t)rJ>(t) (4.0 0.5.14: Logarithm of the Lyapunov Function The equation describing the parameter update is then simply ~(t) =  These simulations demonstrate the close approximation by the averaged system.5 0.10)) . and parameters as above except w0 = 5. (3.52) where Aav(cPav) is the autocovariance of the vector v at t = 0: It depends on cPav because v is obtained from a closedloop system w1th feedback depending on the parameter error cPav.5.5.6 0.5 0 30 1.l.r/>y(O) = 1.53) Explicit Expression of Hvr(s.2 cpy 0.9 0. Figure 4. that is with initial conditions r/>.(0) = . Figure 4. however. We consider the scheme for arbitrary relative degree.15: Parameter Error r/>y(r/>. while it was larger in the previous case. so that we may wnte (as in (4.5. the system is a linear time invariant system.5. This shows that the averaging method is useful for values of g which are not necessarily infinitesimal (i.10.
XI = L.36)( 4.204 Parameter Convergence Using Averaging Chapter 4 Section 4. i. co coxmliT . the oscillatory behavior of the output error scheme is not observed for the input error scheme.5. (4.58) XI x2 i.l(jw)l2 IXm(}w) I2 fi*wmr(Jw) .54) Since the controller is the same as for the output error scheme.<<P) in the frequency 1 =  Aav(<P) 27!' 1] o  co 2 oo c* oo J Ii.60) v i . 56) (ML ) AA I = (s+am)ll km (s + /2) =  ll am12 lt + km km S + /2 (4.5.5.5. parameter convergence rates may be estimated from the preceding expression.37).r + <P Co 1 T w co co . = co_ co so that w and = ~ Co Xm A (w ) X<P (4. 5.w * • ""' • Co Co X<P Co m l (4.(s) Co xis) m A (4.e. Therefore..5.5.amYm + kmr cor+doYp . and we choose£= (s+/ 2 )!1 1• Notethat(Mit 1 maybe expressed as . ' .5.I (z) = i I [ ~ l • = i I I m r +. Example We consider once again the example of Section 4. Convergence Analysis Using the foregoing result we may domain as ' express Aa.5.5.(U) x2 i.61) With (4. Co 0 LA I AX<P Xm ( A Wm) (4.37). the symmetry of Aav(<P) guarantees that around the equilibrium.5.5.aPyP + kpu .am) kp X<P jw A ( • ) Again.<PTw 1 C~ J (4.H w .5.36) using </1.e.57) ~ Xm (r) co Therefore x<P Yp Ym u .2. but for the input error scheme.55) Note that the matrix Aav(<P) is now symmetric and is a positive semidefinite matrix for all <fi.1(Yp) A I l2X2+liYp  CoiI()Xm(s) s . Although the convergence properties are quite similar. rewrite (4.5. we may use (4.5.56) simply becomes The equations describing the overall adaptive system with the input error scheme are (4.5 Applications to Adaptive Control 205 (4. First.d 0 = • d 0 (ap.r + . we neglected the normalization factor and the projection in the update law for simplicity. A • <Py = d 0.59) which is the equivalent of (4. the linearized system is described by a linear time invariant system with only real eigenvalues. The model transfer function is M = km I ( s + am). Again./ 2 x 1 + f 1u c and .40) for the input error scheme. It is positive definite if the input r is sufficiently rich.
0..25.206 When r Parameter Convergence Using Averaging = r 0 sin (w 0 t ). ¢y(0) = 0.62) 0. reflecting the fact that the matrix above is symmetric and.075 0. l2 = 10.16: Parameter Error ¢y(¢. am = 3. has only real eigenvalues.(0) =. therefore.4 0. The analysis had the interesting feature of considering nonlinear differential equations as well as linear ones.225 c/>y 0. 4.3 [::] a~+ w6 a~+ w6 (4.15 When w0 = 5.5.2 Figure 4. Figure 4. ¢.2 o.2 Figure 4. Figure 4.05. 0. 0. but extended to all adaptive systems considered in this work.2 0. ap = I.0. Therefore. the .16 shows that trajectories of the output scheme exhibit an oscillatory type of response. In this chapter. Chapter 4 Section 4. 0. The approximation of parameter convergence rates using averaging was justified by general results concerning a class of systems including the adaptive systems described in Chapters 2 and 3. w0 = 5.)Output Error Scheme The parameters are km = 3. It= 10.17: Parameter Error ¢y (¢.375 0.17 shows the response for the input error scheme under comparable conditions. As may be observed.5 Applications to Adaptive Control 207 the averaged system is [ : :] = g r6 2 It w6 +I} 0. The application to adaptive systems included useful parameter convergence rates estimates for identification and adaptive control systems. The rates depended strongly on the reference input and a frequency domain analysis related the frequency content of the reference input to 0.6 CONCLUSIONS Averaging is a powerful tool to approximate nonautonomous differential equations by autonomous differential equations.)Input Error Scheme trajectories do not exhibit oscillatory behavior.4 0. including adaptive control systems. g = I.5. we introduced averaging as a method of analysis of adaptive systems. kp = 2. the application was not restricted to linear or linearized systems. r0 = I.
whlle for the strictly positive real (SPR) error equation it had possibly complex eigenvalues . Typically. Besides requiring stationarity of input signals. For example. ws. Even if a detailed highorder model of the plant is available. not necessarily l~fimtesimal) a~d for acceptable time constants in the parameter variations. then.nderstan?mg of the dynamical behavior of the parameter error is conSid~rably mcreased using averaging. . even in the nonlinear adaptive control case. corrupted by noise. averaging also We showed however. r~qUlre~ slow parameter adaptation. For frequencies below wH. it was found that the t:aJectory of the para~eter error corresponding to the linear error equatiOn could . These resu. What is available. in a framework similar to Doyle & Stein [ 1981 ].f the adaptation gain that :vere close to 1 (that is. Bode diagrams of the form shown in Figures 5. that the approximation by the averaged system was good for ~alu~s o. and so on. and only an "envelope" of the magnitude response beyond wH· The dashed lines in the magnitude and phase response correspond to the approximation of the plant by a finite order model.208 Parameter Convergence Using Averaging Chapter 4 the convergence rates.lts. They have t~e hr!utatwn of depending on unknown plant parameters. 209 . They may also correspond to the highorder dynamics that one wishes to neglect. w 3. We begin discussing the representation of such uncertainties in plant models.1 and 5. assuming that there are no dynamics at frequencies beyond wH. poles in the neighborhood of w 2 . An inspection of the diagrams shows that the data obtained beyond a certain frequency wH is unreliable because the measurements are poor.be approximated by an LTI system with real negative eigenvalu~s. through SimulatiOns. but an approximat. the designer does not have a detailed statespace model of the plant to be controlled. and obtained for example by performing inputoutput experiments. and complex pole pairs in the neighborhood of w4. so that part of the plant dynamics must be neglected. Consider the kind of prior information available to control a stable plant. such as sinusoidal inputs. it is easy to guess the presence of a zero near w 1. or because its dynamics are not completely understood. is essentially no phase information.2 are obtained.wn of the complete parameter trajectory is obtained and the u.1 STRUCTURED AND UNSTRUCTURED UNCERTAINTY In a large number of control system design problems. are. CHAPTER 5 ROBUSTNESS 5. either because it is too complex. 1t appeared that a basic condition is simply that parameters vary more slowly than do other states and signals of the adaptive system. In fact. it is usually desirable to obtain a reduced order controller. useful for the optimum design of reference input.
208
Parameter Convergence Using Averaging
Chapter 4
the convergence rates, even in the nonlinear adaptive control case. These resu.lts. are. useful for the optimum design of reference input. They have t~e hr!utatwn of depending on unknown plant parameters, but an approximat.wn of the complete parameter trajectory is obtained and the u.nderstan?mg of the dynamical behavior of the parameter error is conSid~rably mcreased using averaging. For example, it was found that the t:aJectory of the para~eter error corresponding to the linear error equatiOn could .be approximated by an LTI system with real negative eigenvalu~s, whlle for the strictly positive real (SPR) error equation it had possibly complex eigenvalues . . Besides requiring stationarity of input signals, averaging also We showed however, through SimulatiOns, that the approximation by the averaged system was good for ~alu~s o.f the adaptation gain that :vere close to 1 (that is, not necessarily l~fimtesimal) a~d for acceptable time constants in the parameter variations. In fact, 1t appeared that a basic condition is simply that parameters vary more slowly than do other states and signals of the adaptive system.
r~qUlre~ slow parameter adaptation.
CHAPTER 5 ROBUSTNESS
5.1 STRUCTURED AND UNSTRUCTURED UNCERTAINTY In a large number of control system design problems, the designer does not have a detailed statespace model of the plant to be controlled, either because it is too complex, or because its dynamics are not completely understood. Even if a detailed highorder model of the plant is available, it is usually desirable to obtain a reduced order controller, so that part of the plant dynamics must be neglected. We begin discussing the representation of such uncertainties in plant models, in a framework similar to Doyle & Stein [ 1981 ]. Consider the kind of prior information available to control a stable plant, and obtained for example by performing inputoutput experiments, such as sinusoidal inputs. Typically, Bode diagrams of the form shown in Figures 5.1 and 5.2 are obtained. An inspection of the diagrams shows that the data obtained beyond a certain frequency wH is unreliable because the measurements are poor, corrupted by noise, and so on. They may also correspond to the highorder dynamics that one wishes to neglect. What is available, then, is essentially no phase information, and only an "envelope" of the magnitude response beyond wH· The dashed lines in the magnitude and phase response correspond to the approximation of the plant by a finite order model, assuming that there are no dynamics at frequencies beyond wH. For frequencies below wH, it is easy to guess the presence of a zero near w 1, poles in the neighborhood of w 2 , w 3, and complex pole pairs in the neighborhood of w4, ws. 209
210
Robustness
Chapter 5
Section 5.1
Structured and Unstructured Uncertainty
211
model output YmU) in response to reference signals r(t) driving the model. This is shown in Figure 5.3.
+
Figure 5.3:
Model Following Control System
Figure 5.1:
B Jde Plot of the Plant (Gain)
The controller generates the input u(t) of the plant, using YmU ), Yp(t) and r(t) so that the error between the plant and model output e0 (t) := Yp(t) Ym(t) tends to zero asymptotically. Two options are available to the designer at this point. NonAdaptive Robust Control. The designer uses as model for the plant the nominal transfer function P•(s)
P*(s) (s + w2) (s + w3)( (s + 114)2 + (w 4)2 )( (s + vs) 2 + (ws) 2)
(5.1.1)
WH
fog
W
Figure 5.2:
Bode Plot of the Plant (Phase)
The gain kp in (5.1.1) is obtained from the nominal highfrequency asymptote of Figure 5.1 (i.e. the dashed line). The modeling errors due to inaccuracies in the polezero locations, and to poor data at high frequencies may be taken into account by assuming that the actual plant transfer function is of the form
fo(s)
To keep the design goal specific and consistent with our previous analysis, we will assume that the designer's goal is model following: the designer is furnished with a desired closedloop response and selects an appropriate reference model with transfer function M(s). The problem is to design a control system to get the plant output Yp(t) to track the or
P(s)
= P*(s) + Ha(s)
(5.1.2)
(5.1.3)
where Ha(s) is referred to as the additive uncertainty and Hm(s) as the
212
Robustness
Chapter 5
Section 5.1
Structured and Unstructured Uncertainty
213
multiplicative uncertainty. Of course, I Ha(jw)l and 1 Hm(jw)l are unknown, but magnitude bounds may be determined from inputoutput m~asurements and other available information. A typical bound for I Hm(Jw)l is shown in Figure 5.4.
Magnitude
Adaptive Control. The designer makes a distinction between the two kinds of uncertainty present in the description of Figures 5.15.2: the parametric or structured uncertainty in the pole and zero locations and the inherent or unstructured uncertainty due to additional dynamics beyond wH· Rather than postulate a transfer function for the plant, the designer decides to identify the polezero locations online, i.e. during the operation of the plant. This online "tuneup" is for the purpose of reduction of the structured uncertainty during the course of plant operation. The aim is to obtain a better match between M(s) and the controlled plant for frequencies below wH. A key feature of the online tuning approach is that the controller is generally nonlinear and timevarying. The added complexity of adaptive control is made worthwhile when the performance achieved by nonadaptive control is inadequate. The plant model for adaptive control is given by
(5.1.5)
<
,l
. I
! I
Figure 5.4:
Typical Plot of Uncertainty
I HmUw)l
and
or
IHmuUw)l
Given the desired transfer function M(s), one attempts to build a linear, timeinvariant controller of the form shown in Figure 5.5, with feedforw~rd compensator C(s) and feedback compensator fr (s), so that the nommal closedloop transfer function approximately matches the reference model, that is,
P{s)
(5.1.6)
P*(s) C(s)
[I
+ fr(s)P*(s) C(s))
I
M(s)
Yp
(5.1.4)
Figure 5.5:
Nonadaptive Controller Structure
<:ver the fr~quency range of interest (the frequency range of r). Further, C(s) and F(s) are chosen so as to at least preserve stability and also reduce sensitivity of the actual closedloop transfer function to the modeling errors represented by Ha, or Hm within some given bounds.
where Po.(S) stands for the plant indexed by the parameters o· and Hau(s) and knu(s) are the additive and multiplicative uncertainties respectively. The difference between (5.1.2)(5.1.3) and (5.1.5)(5.1.6) lies in the online tuning of the parameter o• to reduce the uncertainty, so that it only consists of the unstructured uncertainty due to highfrequency unmodeled dynamics. When the plant is unstable, a frequency response curve as shown in Figures 5.15.2 is not available, and a certain amount of offline identification and detailed modeling needs to be performed. As before, however, the plant model will have both structured and unstructured uncertainty, and the design options will be the same as above. The difference only arises in the representation of uncertainty. Consider, for example, the multiplicative uncertainty in the nonadaptive and adaptive cases. Previously, Hm(s) was stable. However, when the plant is unstable, since the nominal locations of the unstable poles may not be chosen exactly, Hm(s) may be an unstable transfer function. For adaptive control, we require merely that all unstable poles of the system be parameterized (of course, their exact location is not essential!), so that the description for the uncertainty is still given by (5.1.6), with Hmu(s) stable, even though P.(s) may not be. 8 A simple example illustrates thi~: consider a plant with transfer function
214
Robustness
Chapter 5
(5.1.7)
Section 5.1
Structured and Unstructured Uncertainty
215
fi(s)
m
(s  1 + t) (s + m)
with E > 0 small and m > 0 large. For nonadaptive control, the nominal plant is chosen to be 11 s 1, so that
 s 2 + s  E(s + m) (s  1 + t)(s + m)
for one application, and not for another. A key set of observations made by Rohrs, Athans, Valavani & Stein [1982, 1985] is that adaptive control algorithms which are proved stable by the techniques of previous chapters can become unstable in the presence of mild unmodeled dynamics or arbitrarily small output disturbances. We start by reviewing their examples.
5.2 THE ROHRS EXAMPLES Despite the existence of stability proofs for adaptive control systems (cf. Chapter 3), Rohrs et a/ [ 1982], [ 1985] showed that several algorithms can become unstable when some of the assumptions required by the stability proofs are not satisfied. While Rohrs (we drop the et a/ for compactness) considered several continuous and discrete time algorithms, the results are qualitatively similar for the various schemes. We consider one of these schemes here, which is the output error direct adaptive control scheme of Section 3.3.2, assuming that the degree and the relative degree of the plant are 1. The adaptive control scheme of Rohrs examples is designed assuming a first order plant with transfer function
P 0·(s) =  s + ap
A
(unstable)
(5.1.8) is chosen
For adaptive control on the other hand, with
P.(s) = 11(s + O*) 0
s s+m 
(stable)
(5.1.9)
and 0• =  1 + E. In the preceding chapters, we only considered the adaptive control of plants with parameterized uncertainty, i.e., control of P•• 0 Specifically, we choose P• of the form kP fip I dp, where fip, dp are monic, 0 coprime polynomials of degrees m, n respectively. We assumed that {a) The number of poles of P., that is, n, is known. 0 (b) The number of zeros of P., that is, m.::;; n, is known. 0 {c) The sign of the highfrequency gain kp is known (a bound may also be required) . is minimum phase, that is, the zeros of rip lie in the open left half plane (LHP). I_! is important to nAote that the assumptions apply to the nominal plant Po·· In particular, P may have many more stable poles and zeros than P8•• Further, the sign of the highfrequency gain of P is usually indeterminate as shown in Figure 5.1. The question is, of course,: how will the adaptive algorithms described in previous chapters behave with the true plant P? A basic desirable property of the control algorithm is to maintain stability in the presence of uncertainties. This property is usually referred to as the robustness of the control algorithm. A major difficulty in the definition of robustness is that it is very problem dependent. Clearly, an algorithm which could not tolerate any uncertainty (that is, no matter how small) would be called non robust. However, it would also be considered non robust in practice, if the range of tolerable uncertainties were smaller than the actual uncertainties present in the system. Similarly, an algorithm may be sufficiently robust {d)
kp
(5.2.1)
and the strictly positive real (SPR) reference model (5.2.2) s +am s+3 The output error adaptive control scheme (cf. Section 3.3.2) is described by
P• 0
km 3 M(s) =   =  A
u
eo
c0 r + doYp Yp Ym ge 0 r geoYp
(5.2.3) (5.2.4) (5.2.5) (5.2.6)
co
do
As a first step, we assume that the plant transfer function is given by (5.2.1 ), with kp = 2, ap = 1. The nominal values of the controller parameters are then
co
1.5
(5.2.7)
the second term called the unmode/ed dynamics. Figures 5. they would be considered rather mnocuous.3. n(t) = 0.2. at low frequencies. and Controller Plant Output (r = 4. On the other hand.am k p = . the measured output Yp(t) is also affected by a measurement noise n(t ).1 t . but eventually diverges to infinity. the output error initially converges to zero. along with the controller parameters c 0 and do.8 show a simulation with r(t) = 4. it diverges to infinity. 2 229 s+ 1 s 2 + 30s + 229 (5. 60 40 20 20 0 2 s+1 229 s2 +30s+229 2. . the adaptive . n(t) = 0 ( c0(0) = 1. 0 = ap.65 and other initial conditions are zero). three different mechanisms of mstab1hty appear: (R3) With a reference input having a small constant component and a large high frequency component.2.6: Rohrs ExamplePlant. After staying in the neighborhood of zero for an extended period of time.1 (5. ~~wever. In Rohrs examples. 7 and 5. stable modes.8) The behavior of the adaptive system is then studied.cheme is stable and the output error converges to zero. the output error diverges at first slowly.5 10.0.7 (R2) Figure 5.15 ± j 2.9) 80 In analogy with nonadaptive control terminology.65. The poles of the unmodeled dynamICS are located at . and.0 7. as predicted by the stability analysis.14. the output error initially converges to zero.0.14.85 sin 16. d 0(0) = . until they suddenly diverge to infinity. The actual plant with the reference model and the controller are shown in Figure 5. that illustrates this behavior ( c0(0) = 1. ~s . and other initial conditions are zero). and then more rapidly to infinity.9 and 5. Reference Model. the controller parameters c0 and d0 drift apparently at a constant rate. Rohrs showed through simulations that without measurement noise or unmodeled dynamics.3. Figures 5. The actual plant is only approximately a first order plant and has the third order transfer function P(s) With a large.5 5. At the outset.10 show a simulation with r(t) = 0. this term is approximately equal to I. with unmodeled dynamics. n = 0) An important aspect of Rohrs examples is that the modes of the actual plant and those of the model are well within the stability region. assuming that the actual plant does not satisfy exactly the assumptions on which the adaptive control system is based. constant reference input and no measurement noise. along with the controller parameters c0 and d 0 .2 (Rl) The Rohrs Examples 217 d. d 0(0) = . Moreover.216 Robustness Chapter 5 Section 5. ' .6.0 Tlme(sl Figure 5. t?~ unmodeled dynamics are welldamped. With a moderate constant input and a small output disturbance.3 + 1. From ~ traditional control design standpoint.
11 and 5.8 Controller Parameters (r = 4.3 ROBUSTNESS OF ADAPTIVE ALGORITHMS WITH PERSISTENCY OF EXCITATION Rohrs examples show that the boundedinput boundedstate (BIBS) stability property obtained in Chapter 3 is not robust to uncertainties.218 Robustness Chapter 5 Section 5.5 5.2 The Rohrs Examples 219 15 10 5 OL5 10 200 15L~L~ 0 2.5 Yp Figures 5.12 show a simulation with r(t) = 2. n = 0) 5. In .85 sin 16. n = 0) Figure 5.5 sin 16.9 Plant Output (r = 0.14.3 + 1.1 t . d 0 (0) = . 0 5 10 15 20 Time (sl Figure 5. while in adaptive control.3.0 0 5 _____ L_ _ _ __ L_ _ _ _J __ _ _ _~L~ 10 15 20 Tlme<sl Tlme<sl Figure 5.1 t .0.3 + 1.5 10. and other initial conditions are zero). Instability is also observed with other frequencies of the disturbance. Rohrs examples stimulated much research about the robustness of adaptive systems. Examination of the mechanisms of instability in Rohrs examples show that the instabilities are related to the identifier. This justifies a more careful consideration of robustness issues in the context of adaptive control. simulations show that when smaller values of the output disturbance n(t) are present.0 7. such instabilities involve computed signals. but after a longer period of time.85 sin 16. variables associated with the plant are also involved.0 1. Although this simulation corresponds to a comparatively high value of n(t ). n (t) = 0.10 Controller Parameters (r = 0. The controller parameters simply drift at a slower rate.1 t ( c 0 (0) = 1. In identification.5 3. instability still appears. n = 0) 4.65. including a constant n(t ).
1 t) . the signal u will be considered either a disturbance or an input. ..X E IR n .5 25. n = 0. provided that the nominal adaptive system is exponentially stable (guaranteed by a persistency of excitation (PE) condition). Our presentation follows the lines ofBodson & Sastry [1984].1) and the unperturbed system (5.3.1 Small Signal 1/0 Stability Consider the perturbed system (5. 5. . We restrict our attention to solutions x and inputs u belonging to some arbitrary balls Bh E IRn and BeE IRm. Thus. we consider properties of a socalled perturbed system Figure 5.1t) x X so 120 = f(t. u) x(O) = xo (5. Let f be Lipschitz in u.11 Plant Output (r = 2. 0) x(O) = xo (5. n = 0. we will obtain robustness margins. x E Bh. x.2) where t ~ 0.220 Robustness Chapter 5 Section 5. Yp 12. . The main difference from classical linear timeinvariant (LTI) control system robustness margins is that robustness does not depend only on the plant and control system.5 0~~. the practical notion of robustness is that stability should be preserved in the presence of actual disturbances present in the system. Let u E L 00 • If x = 0 is an exponentially stable. 0.3. but also on the reference input. bounds on disturbances and unmodeled dynamics that do not destroy the stability of the adaptive system.3.3 Robustness with Persistency of Excitation 221 12.3.3. 0) = 0. In this section. Theorem 5. i.3.1) and relate its properties to those of the unperturbed system = f(t.2). in the sense that exponentially stable systems can tolerate a certain amount of disturbance. an arbitrary small disturbance can destabilize an adaptive system. Of course.0 37. without disturbances or unmodeled dynamics). x. we will show that the property of exponential stability is robust.1 Exponential Convergence and Robustness In this section. Let x = 0 be an equilibrium point of (5. Let f be piecewise continuous in t and have continuous and bounded first partial derivatives in x.e. for all t ~ 0.12 Controller Parameters (r = 2. u E Be.2). with Lipschitz constant lu.3. for all t ~ 0.0 0 20 40 60 80 Time(sl some cases. equilibrium point of the unperturbed system 160 2000~~~L~L~L_~~_j 20 40 60 80 Tlme<sl Figure 5. .5 50. f(t. that is.5 sin 16. for all t ~ 0. u E IR m.5 sin 16. Robustness margins must include actual disturbances for the adaptive system to be robust in that sense. Depending on the interpretation. which must guarantee persistent excitation of the nominal adaptive system (that is. which is otherwise proved to be BIBS stable. x E Bh. u E Be.
Then. o.3. 0 <II ull 00 < c 00 implies that x(t) converges to a B6 ball of radius o = 'Y 00 II ull 00 < h.1 0) 1'00 'Yoo o .2) satisfies the conditions of theorem 1.8) for all t ~ T. since m ~ 1). since the derivative is now to be taken along the trajectories of (5. inequalities (5.1) mo ] (5.9) can now be written dv (t ' x) dt I :.1 relates internal exponential stability to external input/output stability (the output is here identified with the state).3. Part (b) further extends it to non zero initial conditions.3.3. av(t.3.x.O)] I ± (5. u) 1= 1 ax. 5.1 The differential equation (5.3 Robustness with Persistency of Excitation 0 223 (a) The perturbed system is smallsignal L 00 stable. If we consider the same function to study the perturbed differential equation (5.1 ). by continuity of the solutions. that is: for all E > 0. it is known that simple stability and even (non uniform) asymptotic stability are not sufficient conditions to guarantee 110 stability (see e. x e Bh. 11 ull 00 (5.3. x.. a3lxl 2 dt (So3o2) I This inequality is the basis of the proof.1) starting at x 0 (b) There exists m ~ 1 such that.3.3. c00 > 0.3. Ex. there exists T ~ 0 such that lx(t)l S:(l+t)o (5.3.7) . so that there exists a Lyapunov function v(t.. 379).4) I <50301 > = av(t. x) + at 'i.3. lx(t)l <h.x)l :.u)f1(t. Also.4.3.3) for L 00 stability.3. Although lack of exponential stability does not imply input/output instability. T 1 (T 1 > To~ 0).6) dv(t .3. dv (t' X) + av (t' X) dt (5o3o2) i = I ax. .5)(5. We show that this implies that x(t) e Bo for all t ~ 0 (note that oI m s..12) m := [ :: lt ~ a 3I xI [ I xI  Inequality (5.3. for all lxol <him.3. we require a linear relationship between the norms in (5. . there exist 'Y 00 .g.3.(t.3.3) 0.13) v(t. such that II u II 00 < c00 implies that II xll 00 s: 'Y 00 II ull 00 < h = (5. there would exist To.3.9) Define (5. azlxl 2 (5.3.5. x) /. and for all t ~ 0.. Theorem 5 . p.3. while (5. that it were not true.3. Til: lx(t)l ~ olm. a3lxl dt (5. x) satisfying the following inequalities a1lxl 2 :. a41x1 (5. x) dt where x is the solution of (5.1 is a direct extension of theorem 1 of Vidyasagar & Vannelli [1982) (see also Hill & Moylan [1980)) to the non autonomous case.5) and (5.1. for the sake of contradiction. · [f1(t.x. inequality and for all t e [T0 .3.. for some strictly positive constants a 1 ° • a 4 .5) (5. and the Lipschitz condition on f dv(t. The two derivatives are related through dv(t.1) starting at x 0 .2). Comments Part (a) of theorem 5.7). (5.3. In contrast with the definition of BIBS stability of Section 3. 7) still hold.. that is.x) s.3.11) (5.3. along the solutions of (5.. Suppose. 1av~~x) 1 :.1) instead of (5.3. Part (a) Consider the situation when I x 0 1 S: oI m (this is true in particular if x 0 = 0). Using (5.. Kalman & Bertram [ 1960]. Proof of Theorem 5. such that lx(To)l =olm and lx(TI)I>o Consequently.3.3.x) :. for all t ~ 0.1) 2 +a41xllullull 00 (5..222 Then Robustness Chapter 5 Section 5.6) is modified.
using an argument similar to (a).13) v<a 3 (olm) 2 (! +f)f which is a strictly negative constant. when the rate of exponential convergence is increased. Suppose it was not true. and for a wide category of nonlinear systems (bilinear systems for example).3. where m is defined in (5.3.2. In Section 3. This is also guaranteed. This follows directly from (b1). in itself. this contradicts the Section 5.3.224 Robustness Chapter 5 However.28)) (5.5. we showed that the overall output error adaptive scheme for the relative degree 1 case is described by (cf. when the relative degree of the plant is 1.14) implies that o<h. (3. by the disturbance acting on the system). Finally. will vary with them.2. it can be expec~d that c 00 will remain bounded despite the freedom left in the choice of h and c. While the first term is proportional to the norm squared.3. and v(Tt>x(Tt))> a1 o2 Part (b) Assume now that Ix0 1> oI m. Note that if we used the Lyapunov function satisfying (5. provided that lxol <him and II ull 00 <c 00 . (bl) for all f > 0. a 4 1 lu II ull xi acts like a disturbing force.2 Robustness of an Adaptive Control Scheme For the purpose of illustration. Be. given any balls Bh. using equations (5. In the example of Section 5. This example contains the specific cases of the Rohrs examples. independent of the size of the input.3. we consider the output error direct adaptive control algorithm of Section 3.e. but the values of 'Y 00 (the L 00 gain) and c (the stability margin). Yet. 0 .3. 7) to obtain a convergence rate for the unperturbed system. for all t ~ 0. Then. with other parameters remaining identical. Xo) ~ a2lxol 2 < and v(t. However. a result for non zero initial conditions.5. the second is only proportional to the norm. from (5.3. hly 00 ) (5. so that when I xl is sufficiently large. then the results are valid globally too.12). To see this.13).3. consider (5. In general. The system remains stable and has finite 110 gain.3. u (t) e Be. the size of the ball B61 m involved decreases when the amplitude of the input decreases. and lxol<hlm~h implies that lx(t)l ~ m I xol < h for all t ~ 0. fact that v ~ 0. in [To. for some f > 0 and for all t ~ 0 lx(t)l >(olm)(! +f) and.x(t))~O h2 a2  m2 for all t ~ 0. c) Explicit values of 'Y 00 and c00 can be obtained from parameters of the differential equation. while the size of Bh 1m is independent of it.14). there exists T ~ 0 such that I x (T)I ~ (o I m) ( 1 +f).a 3 1xl 2 acts like a restoring force bringing the state vector back to the origin.3 Robustness with Persistency of Excitation 225 (5.3.14) Additional Comments a) The proof of the theorem gives an interesting interpretation of the interaction between the exponential convergence of the original system and the effect of the disturbances on the perturbed system.3. Note that although part (a) of the proof is. the Lipschitz condition is not verified globally. using an argument identical to the one used to prove (a). We show the result in two steps. it can be verified that. this contradicts the fact that v(O.13) shows that. and c 00 := min(c.3.3. the system satisfies a Lipschitz condition with constant lu depending on the size of the balls (actually increasing with it). pulling the state away from the origin. we see that this happens when IxI = oI m = 'Y 00 I m II u II 00 • b) If the assumptions are valid globally. recall that the assumptions require that x(t) e Bh.1 0) and (5. Thls term is caused by the input u (i. I x(t)l ~ o(1 +f). The balls Bh . so that the 110 stability will only be local.9): the term . Note that an upper bound on Tis a h2 T~2a3o2f (b2) for all t ~ T. the L 00 gain is decreased and the stability margin c00 is increased. 5. the restoring force equilibrates the disturbing force.3. This term originates from the The term exponential stability of the unperturbed system. this rate would be a 3 I 2a 1• Therefore.5)(5. Til. In the form (5. Be are consequently arbitrary in that case.
3. (5.18)(3.n.16) := [:] € IR.3.3.n1) = IR.3. we return to the derivation of (5. IR.3.e.17) and the definition of (5.w}.gn (t) Wm(t).gc~ e(t) Q e(t) IR. where X x(O) = x0 (5.3.3.Sn2 (5.3. (5. Using the definition of Am. with the definition of x in (5. 0) e (IR. b[) e (IR. IR.24) where we define q[ = (0. bp. in Section 3. 1.gc~e(t) wm(t). wO>r.16) can be written f in .gc~ e(t) wm(t).3. 0. as in Figure 5. w<W ). Denote by the output of the plant P·(s) (that is the output without measurement noise) and by yp(t ).3.)> y. IR.1.15) is a nonlinear ordinary differential equation (actually it is bilinear) of the form .15) (i.n. gc~ew gnw (5.22) Recall that we also found.> (5.8.25) As previously.3. 0.19).Yp = + bm cf>T(t)qnn(t) + bnn(t) (5.20) .3.gn 2(t) qn A w< 2> + b>cJ Xp + b>. we represent the model and its output by and c/>(t) e IR 2n. IR.u = Ame(t) + bmct>T(t)wm(t) + bmct>T(t)Qe(t) A w< 2> + b>. The plant P• has a minimal statespace 0 representation [Ap.28)) in Section 3.5. The update law is given by Consider the case when the measured output is affected by a measurement noise n(t). so that (5. bm and Cm in (3.5.x) which is of the form (5.23) and the regressor is now related to the state e by 0 w w< 1>.15)(5.17) Ym C~Xm (5. Chapter 5 Section 5. the adaptive system with measurement noise is described by e(t) • Yp CTXp p (5.Xm.6.3.2). Am is a stable matrix.15) (that is (3.gc~e(t)qnn(t) .16)) is exponentially stable in any closed ball.3. we let xJw = (xJ.Jn2.5.21) where we define b[ = (0. provided that Wm is PE.18) To find a description of the adaptive system in the presence of the measurement noise n(t).226 Robustness Ame(t) + bmct>T (t) Wm(t) + bmct>T (t) Q e(t) . the 0 measured output. cJl such that y.19) The observers are described by AwO>+b>. IR. Using these results.gc~e(t) Qe(t). Robustness to Output Disturbances and we let e = Xpw.2.3. and where e(t) e 2 wm(t) e IR n is bounded for all t : : :.15) (5. that (5. affected by noise.3 Robustness with Persistency of Excitation 227 controller is now (5. 3n.n + 1) = IR.n1.5.X= f(t.3.2n. the description of the plant with which.n U and the COntrol input is given by = (} T W = cf>T W + (}*T W. (5.3.Ym w< 2>. As previously.3.w}.gn (t) Qe(t).3. 0.
gn(t)qn c'/:z.2 Since Wm is PE.16)) is exponentially stable in any Bh by theorem 3.29) The actual output is modeled as the output of the nominal plant. If Wm is PE Then there exists 'Y n> Cn > 0 and m ~ 1.e.1 can be applied with u containing the components of P 1(1 ).2.30) The operator Ha represents the difference between the real plant. and provided the initial error is sufficiently small. P2(t).3. satisfying the 8 assumptions on which the adaptive control scheme is based. is an example of instability when the persistency of excitation condition of the nominal system is not satisfied. For this reason. within th1s bound. the stability of the adaptive system will be preserved. The results on p 1 ).28).: :.3.3. and their bounds can be made arbitrarily small by reducing th~bounds on the disturbances.3. given a bound on n. with lx(t)l . and the idealized plant P(s). The following theorem guarantees the stability of the adaptive system in the presence of unmode1ed dynamics satisfying (5.27) Note that if n e L 00 .3 Robustness with Persistency of Excitation 229 where p 1(t) e IR 5n.3.31) for all t ~ 0. Robustness to Unmodeled Dynamics We assume again that there exists a nominal plant P·(s ).2and P2(t) e IR.3. there exists kn ~ 0 such that robustness margin of the adaptive system to output disturbances. knll nll 00 (5.26) Section 5. f3a may include the effect of initial conditions in the unmodeled dynamics and the possible presence of bounded output disturbances. the perturbed system (5.2 Robustness to Disturbances Consider the output error direct adaptive control scheme of Section 3. the unperturbed system (5.3. and we define the output of the nominal plant to be (5. where u contains the components of p 1 and P 2• Although p 1(t) depends quadratically on n. Theorem 5.e.2are given by = P1U) PAt) = 0 [ . since it is the purpose of the adaptive scheme to reduce to zero the structured or parametric uncertainty.3.18).1).15) (i. we deduce the following theorem.25) (i. b) The deviations from equilibrium are locally at most proportional to the disturbances (in terms of L norms).2. We assume that Ha: L ooe L ooe is a causal operator satisfying (5.28) From these derivations.26)) is a special case of the general form (5. c) The L 00 gain from the disturbances to the deviations from equilibrium can be reduced by increasing the rate of exponential convergence of the unperturbed system (provided that other constants remain identical). using (5.3.26) is a special form of system (5.3.31). plus some additive uncertainty represented by a bounded operator Ha (5.3. Proof of Theorem 5.2.3.3. The perturbed system (5.3. (5.2x sn. where n e L 00 • Let h > 0. IIP1II 00 +II P2ll 00 . d) Rohrs example (R3) of instability of an adaptive scheme with output disturbances on a non persistently excited system. m lxol <h for all t ~ 0. We refer to it as an additive unstructured uncertainty.8. Sn. 0 Comments a). assuming that the relative degree of the plant is 1. Assume that the measured output Yp of the plant is given by (5.228 Robustness f(t. (5. A specific bound Cn on II n II 00 can be obtained such that.1 ). x) + P1(t) + P2(t)x(t) Chapter 5 (5.gn(t)Q bmn (t)q'[l 0 (5.3. then p 1 and P 2 e L 00 • Therefore. such that II n II 00 < Cn and I x (0)1 < hIm implies that x (t) converges to a B6 ball of radius o = 'Yn II nil 00 .3.3. Cn is called a .3.3. and it constitutes all the uncertainty. P 2(t) can be translated into similar (1 results involving n(t ).:::. so that theorem 5.
The importance of the result is to show that if the unperturbed system is persistently excited.3 gives some margins of unmodeled dynamics that can be tolerated without loss of stability of the adaptive system.31) and is such that trajectories of the adaptive system are continuous with respect to t.2.35) and (5.3. but we saw in Chapter 4 how some estimate could be obtained. Such an analysis can be found in Astrom [ 1983].3.33) Since x E Bh.36) where Cn is the constant found in theorem 5. the input u is given by OT W o·T Wm = (}*T W + ¢T W + O*TQ e + (} . 0 . and its success relies mainly on the separation of time scales between the evolution of the plant/observer states.6. 'Y a• f3a sufficiently small. where P• satisfies the assumptions of Section 3.34) where ha(r) is the impulse response of Ha· The constant f3a depends on the initial conditions in the unmodeled dynamics.6).35) < Cn (5.3. by assumption = Ha(u ).3. A heuristic analysis of the Rohrs examples gives additional insight into the mechanisms leading to instability. T]. The theorem would then apply.24). observer states. or conversely that an arbitrary small disturbance cannot destabilize the system. Thisseparation of time scales is especially suited for the application of averaging methods (cf.3. and using (5. such as in example (R3). by continuity of the solutions. Let 'Y a• f3a sufficiently small that 'Ya 'Yu < f3a + 'Ya f3u 1'Ya 'Yu (5. This convergence depends especially on the parameter convergence.3. and x(T) =h. Following Astrom [ 1983]. Then.36). T]. T]. since it includes possible nonlinearities. Define n that.2. Since none of the constants 'Ya. 'Yn and f3n is dependent on T.3 Let T > 0 such that x(t) ~ h for all t e [0.3. f3a. 5. and the adaptive parameters. and suggest practical methods to improve robustness. Chapter 4).:::: 0 such that I for all t E U r II 00 .2.32). boundedinput boundedoutput operators.:$ 'Yall Urll 00 + f3a E [0. so (5. (5.3. and therefore on conditions on the input signal r(t ).2.3.rqnn + ¢ T Wm + ¢ T Qe + + ¢ T qnn (5.3.29)(5. we showed in Section 5. Indeed. The most difficult parameter to determine is possibly the rate of convergence of the unperturbed system. T]. suppose it was not true.3.By theorem 5. T]. including the plant states.3 the importance of the exponential convergence to guarantee some robustness of the adaptive system.3.3. resulting in a contradiction since 1x(T)I <h.3.3. it will tolerate some amount of disturbance. there would exist a T > 0 such that Ix(t)l ~ h for all t E [0. we will ~how that instability in the Rohrs examples are due to one or more of the following factors [0.37) Proof of Theorem 5. and it is likely that the estimates would be conservative. the state trajectories of the adaptive system remain bounded. The proof of theorem 5.3. One can use 00 for Xo.230 Robustness Chapter 5 Section 5. Applying the small gain theorem (lemma 3.3. under the conditions of averaging.24) is very general.3 Robustness to Unmodeled Dynamics Consider the output error direct adaptive control scheme of Section 3. there exist 'Yu· f3u. Using (5.:$ 'Y u II n til 00 + f3u (5.3. this implies that 1 )I < h for all t E [0. If Wm is PE Then Comments Condition (5. Assume that the nominal plant output and actual measured plant output satisfy (5. Ha 8 satisfies (5. provided that they can be represented by additive.4 HEURISTIC ANALYSIS OF THE ROHRS EXAMPLES By considering the overall adaptive system.3. Needless to say the expression for these robustness margins depends in a complex way on unknown parameters. lx(t)l <h for all t .30). and the evolution of the adaptive parameters. assuming that the relative degree of the plant is 1.32) II for all t U nrll 00 .3 Robustness with Persistency of Excitation 231 Theorem 5. If the operator Ha is linear time invariant.3. and so on. the stability condition is a condition on the L 00 gain of Ha. it x(t follows that II ntll 00 < Cn.3.:::. 'Ya = llhalll = Jlha(r)ldr 0 (5.o. unmodeled dynamics. Given 'Y a• f3a it is actually possible to compute these values.
It is the accumulation of updates based on incorrect information that leads to parameter drift.0. but c .17. leaving e 0 = Yp. When r(t) = 2 sin t. until it reaches the limit of stability of the closedloop plant with unmodeled dynamics. In particular. These signals cause the adaptive loop to try to get the plant loop to match the model at high frequencies.0 as t .Ym at zero. Conversely.2) will lead to Yp. note that instabilities of this type can be obtained for systems of relative degree greater than two even without unmodeled dynamics and can lead to the socalled bursting phenomenon (cf.Ym ) .4. against bursting is persistent excitation.4.4.11 and 5. *( • gyp Yp . or from output disturbances. when e0 . as would be the case if the original system was exponentially stable. The output error e0 eventually converges back to zero. consider the case where n = 0. b) Consider now the case when the reference input. c) a large reference input with a nonnormalized identification (adaptation) algorithm and unmodeled dynamics.gymn .gn 2 . the actual update law for d 0 is not (5. the parameter slowly drifts. resulting in a closed loop unstable system and a large increase in eo.1) for DC signals. b) the presence of significant excitation at high frequencies.3) the lack of sufficiently rich inputs to • allow for parameter convergence in the nominal system.232 a) Robustness Chapter 5 Section 5. the controller parameter does not necessarily converge to its nominal value.4.2. This repeats at random instants. where do slowly drifts in the negative direction.4) Analysis Consider now the mechanisms of instability corresponding to these three cases.2d0 = 1 (5.5 and d 0 :::: .4. so do 0. is Yp f 2co s + 1. since it is indistinguishable from the output noise n(t ). without unmodeled dynamics.gn 2 is not counteracted by any restoring force. so that the assumption that c0 . and eventually to instability. If an output disturbance n (t) enters the adaptive system. the parameter d0 deviates from d0 but remains close to the nominal value.2). The error converges to a neighborhood of zero. • prevent the drift of the parameters due to unmodeled dynamics or output disturbances.13 shows the evolution of the parameter do in a simulation where r(t) = 2 and where r(t) = 2 sin t.03. d 0 are fixed is justified.1 sin 16. which will tend to make do slowly drift toward the negative direction. leading eventually to instability. resulting in a closedloop unstable plant. Figure 5. As before a safeguard.Ym. The presence of noise in the update law leads to drift in the feedback coefficients to a region where they are large. In the nominal case.4 Heuristic Analysis of the Rohrs Examples . but which is unstable for do ~ 0. When the PE condition is not satisfied. a) Consider first the case when the input is not sufficiently rich (example (R3)). and is referred to as bursting. Consider Rohrs example.458d 0 2 (5. resulting in the instability of the identification algorithm.2) Any value of c 0 . This implies the condition that 2c 0 1.gn 2 where we find the presence of the term . unmodeled dynamics are present.4. The result is observed in Figures 5.6) anymore. but a large 'blip' is observed.oo for a constant reference input. do = 233 (5. contain a large component at a frequency where unmodeled dynamics are significant (example (R2)). originating either from the reference input. so that the transfer function from r . Finally. In terms of the discussion of Section 5. we see that the constant disturbance .4.Yp is in fact given by Yp f = 458c 0 (s + 1)(s + 30s + 229). only the DC gain of this transfer function must be matched with the DC gain of the reference model. If a constant reference input is used. This instability is called the slow drift instability.2d0 (5.yP. and d0 . assuming that c0 and d 0 are fixed. In example (R3).12. or the output disturbance. The closedloop transfer function from r. d 0 satisfying (5. so that PE conditions are satisfied. and the signal available for parameter update is very small and unreliable.3. note that when output disturbances are present. it can cause the parameters c0 .1) which is identical to (5. but to a value such that the closedloop transfer function matches the model transfer function at the frequencies of the reference input. Anderson ( 1985]).1 t. For example. In the first case. d 0 to move along the line (more generally the surface) defined by (5. the output error tends to zero.
there are still unique values of c 0. Figure 5..4 Heuristic Analysis of the Rohrs Examples 235 wo 0. d 0+ close to c0 d0. . Astrom [1983]).4. and the .234 Robustness Chapter 5 Section 5.1t) Let us return to Rohrs example. and also on the reference input r. and thereby to an unstable overall system.31 1.8 1.2 It may be verified that these values are identical to c0 . the unmodeled dynamics. This is the instability observed in example (R2). 1.6 1 2 5 10 1. P. d6 depend on M.r = d 0+ quickly decreases for w0 > 5.6 (s + 1)(s 2 + where wo is the frequency of the reference input.3 + 1.03 and d 0+ :s. that is where the sinusoidal component of the input is at a frequency where model matching is possible.4 eo. 1. d 0 defined + earlier. Therefore.8 1. these would be the nominal c0 d0.::: 16. On the other hand. the adaptive control system updates the parameters.72 7.2 0.13 Controller Parameter d 0 (n = 0.14 Controller Parameters (r = 0.14 shows a simulation where r = 0..5) 30s + 229). (cf.53 1. Note that the parameter d 0s.85 sin t. the 0 50 100 150 200 Time(s) Figure 5. Note that the values of c0 . In contrast.17.69 1. but now they are the values c0+. d 0 such that the transfer function from r. 0 0.17. when w0 .1 sin 16.Yp matches M at the frequency of the reference input w0 • Without unmodeled dynamics.3 + 1.. With unmodeled dynamics.03.09.2 0 25 50 75 100 Time<s> Figure 5.26 1. Therefore.458d0+ closedloop system is unstable when d 0 :s. by attempting to match the reference model at a high frequency. d0+ given by (5. . trying to match the closedloop transfer functionincluding the unmodeled dynamicsto the model reference transfer function.. 0. it may be verified through simulations that the output error tends to zero and that the controller parameters converge to the following values c0" and do. n = 0) Then..67 1.04 do. the adaptive system leads to an unstable closedloop system. . .85 sin t. On the other hand.44 2. the parameters converge to values c0+. with a sinusoidal reference input r(t) = r 0 sin (w 0 t ).
4.5 AVERAGING ANALYSIS OF SLOW DRIFT INSTABILITY As was pointed out in Section 5. 5. decrescent function v(t.5. Astrom [1983] for a simple rootlocus argument).5.5. By definition.5 Averaging Analysis of Slow Drift Instability Xav = E 237 (5. c) Consider finally the mechanism of instability observed with a large reference input (example (Rl)). the additional assumptions (i)(iv) of the theorem guarantee that the averaged system (5.5. we use X One Time Scale Systems Recall the setup of Section 4. The reader may wish to review those assumptions before proceeding with the proof of the following theorem. We will not repeat the definitions and the assumptions (Al)(A5) of Section 4.5. z) satisfying. x) I~ v(t. This mechanism will be called the highgain identifier instability.5.2) adaptive system remains stable.2) satisfies I ~ E k 2 j xj 2 E (5. E) (5.Yp and e0 by 2 and therefore is equivalent to multiplying the adaptation gain by 4.236 Robustness Chapter 5 Section 5.1) and the averaged system (5. In Section 5.5. despite the unmodeled dynamics. or the presence of significant excitation at high frequencies. x) k 1 j xj for some k 1 > 0 (iv) the derivative of v(t.1) is unstable for (5. .1. x) along the trajectories of (5. we develop general instability theorems for averaging of one and two time scale systems.4.2. A simple fix to these problems is to replace the identification algorithm by a normalized algorithm.5. 7) Roughly speaking. where we considered differential equations of the form X = E f(t' X. the first step is to use the transformation of lemma 4.1) = Z + E W.5.1 As' in Chapter 4.4) sufficiently small. to separate the evolution of the plant/observer states and the adaptive parameters. a system is unstable if it is not stable. we make the analysi!) more rigorous using the averaging framework of Chapter 4.3) do .2) satisfy those identical assumptions.5.5.5. Theorem 5. for some HE) e K .1 ).5.6) fav (Xav) where fav(X) = lim Too ~ J to lo+ T f(r. O)dr (5. multiplying r by 2 means multiplying Ym.2. Proof of Theorem S.2) satisfy assumptions (Al)(A5). meaning that there exists a neighborhood of the origin and arbitrarily small initial conditions so that the state vectors originating from them are expelled from the neighborhood of the origin.(t. Z) ' (5.> phenomenon was already given in the preceding section. 0) = 0 (ii) v(t.2.5.5. the function fav (x) has continuous and bounded first partial derivatives in x.3 to transform the original system. (5. In this section. the original system (5.5.1 assuming that the limit exists uniformly in t 0 and x.1 Instability Theorem for One Time Scale Systems If the original system (5. Although we do not have explicitly a high adaptation gain g. (t. The phenomenon under study is the socalled slow drift instability and is caused by either a lack of sufficiently rich inputs. x) such that (i) v(t.but we will assume that the systems (5. recall that the adaptation law is given by (5. In Section 5. Then for some k 2 > 0. X. and there exists a continuously differentiable. originating either from the reference input or output disturbances. A heuristic analysis of thi.5) and their averaged versions with w. we apply these results to an output error adaptive scheme. Thus.g eoYp (5. x) > 0 for some x arbitrarily close to 0 (iii) Iav~. Astrom [ 1983] introduced an analysis of instability based on slow parameter adaptation.2) is unstable. Our treatment is based on Riedle & Kokotovic [ 1985] and Fu & Sastry [ 1986].5.2) Instability Theorems Using Averaging Remark By an instability theorem of Lyapunov (see for example Vidyasagar [ 1978]).4.5. 5. The instabilities obtained for high values of the adaptation gain are comparable to instabilities caused by high gain feedback in LTI systems with relative degree greater than 2 (cf.2.
z. E) where p(t.5.5.1 ).5.1) into (5. (B3) and (B5).15) and the averaged system (5. Theorem 5. consider the derivative of v (t.14). the system (5.P(E)Izl (5.1 can be found prescriptively.12) and if A has at least one eigenvalue in the open righthalf plane.4 before following the next theorem.5.15) is now assumed to be constant and stable rather than a function of x which is uniformly stable.z> EP (t .5.4.5.5.238 Robustness Chapter 5 (5.16) where fav (x) is defined to be (5.1).5. decrescent function required by theorem 5.5.5.E > az I (5. the original system (5. It is also easy to verify that the conditions (iii).2) Using the inequalities (5.z) v(t.8) is positive definite.5.7) Two Time Scale Systems We now consider the system of Section 4. that is (5. and consequently so is the original system (5.5. y e IR. y) Ay + Eg(t.9).1.5.5. x) > 0 for some x arbitrarily close to 0 (iii) I av~. I~ x) k 1 xl for some k 1 > 0 1 (iv) the derivative of v(t.5.9) for some .4.8) Ef(t. x) along the trajectory of (5. then it is clear that v(t' z) I (5.15) (5.8) (5. namely with x e IR.4 namely Efav(Z) + Ep(t.P(E)kd zl 2 (5.1 0) Xav = Efav (Xav} 10 + T (5.. we advise the reader to review the results of Section 4. but no eigenvalue on the jwaxis.15) and the system (4.8}.14}.P( Eo) k I > 0. if the averaged system is linear.5. The only difference between (5.1 are also satisfied by v(x).x. z) (5.5. z) along the trajectories of (5. (5.5. x) such that (i) v(t.5. (B2).5. . E) satisfies IP(t.5 Averaging Analysis of Slow Drift Instability 239 some directions (and arbitrarily close to the origin).x.y) (5.5. (5.14) (5.14).5.O)dr to (5.5.5. z) I ~ HE) to transform (5.5.2 Instability Theorem for Two Time Scale Systems If the original system (5.X)I (5. 0 Comments The continuously differentiable.5.16) satisfy assumptions (B1). (B3) and (B5) (only assumption (B4) is not necessary).E)I ~ . In this case.11) I ~ Ekzl zl 2  E. the function v can be chosen to be v(x) = xTP x The functions f.5.P(E) e K.5. so that v(x) takes on positive values in Then for some k 2 > 0.n.5. By the same Lyapunov instability theorem as was mentioned in the remark preceding the theorem. it follows from (5.4) and (5.5.Z.~. (5.5.z >I + avu. Now.2) of Chapter 4 is that the A matrix of (5.11) that for E ~ Eo.5.m. 0) = 0 (ii) v(t.< = vt. g satisfy assumptions (Bl).17) If Eo is chosen so that kz . (4. Z.5.8) is unstable. IE aw.5. z.16) satisfies (5.5.6) Section 5.13) where P satisfies the Lyapunov equation ATP+PA =I V(t.5.8) fav(X) =lim~ T+oo J f(r.18) sufficiently The Taussky lemma generalized (see Vidyasagar [ 1978)) says that P has at least one positive eigenvalue. we have that v(t.5. (B2}. along with the assumption that there exists a continuously differentiable decrescent function v(t.5. X. As in the case of theorem 5. The averaged system is .16) ~ EkziXI 2 E (5. (iv) of theorem 5.15) is unstable for small.
z..5. x) is defined to be I ATP+PA =I Using the transformation of lemma 4.16). The controller is.22) (5.5.2 and 4. (5.z. y) values arbitrarily close to the origin (let y = 0 and use assumption (ii)).5.5. y.15). 5. v(r. the total state of the plant and observers 3n . (5.28) j'(t.z)l + k3IYI .4.5 Averaging Analysis of Slow Drift Instability 241 Proof of Theorem 5.15)as in the proof of theorems 4. 21) (5. As we saw in Section 4.5.2 To study the instability of (5. (5.5.2 Application to the Output Error Scheme Tuned Error Formulation with Unmodeled Dynamics We will apply the results of the previous section to an output error adaptive control scheme (see Section 3. x is the slow variable but y' has both a fast and a slow component. D v'l(t.4.5..5.5. In analogy to (3.3 into v(t.3.5.4.16) will exist if the following limit exists uniformly in t 0 and x !av<x) = lim Too ~ to+ T to J f'(r.z.20).15) is unstable for Esufficiently small.5. Now.4.25) can be transformed into the system (5.y)l 2 2:: w(E)jzj 2 +q(E)jyj 2 (5.5.28) is precisely (3. y) is a positive definite function along the trajectories of (5.5.14).2 is applicable with the additional assumption (B6) of Section 4. having the form E Yp [ cJ 0 0] (5. E) and P2(t.5.5.5.cJ 0 A Xp wO> w<2> Xp w<l> w<2> + bp b).5.5.14).5.5.4. x. E)+ EP2(t.5. E) Ay + tg{t. Hence.y)l Xp wo> w<2> Ap 0 0 0 A 0 b>.21) and consequently the original system (5. y) where p 1(t. .2) designed for a plant of order n and relative degree one.240 Robustness Chapter 5 Section 5. a more general class of twotime scale systems arises in adaptive control.x)dr 0 (5.21) The averaged system of (5.v(t.x) := JeA<tT)h(r.5. applied to a plant of order n + nu . x. Thus v1(t.25) of the form of (5. consider = v(t.24) (5.24). we consider another decrescent function v t.20.23) The instability theorem of 5.5. for simplicity. (5.3 (the reader may wish to follow through the details). (5.19) where P is the symmetric positive definite matrix satisfying the Lyapunov equation In (5.2tk 3 yTPg(t.24). y)>O for some (x.5.24).5. z.5.5. E) satisfy (5. z.5. Clearly.20) (5.5.14).. x(z). 21) for some a(E)k 2 and q(E)k 3 as E0.20.21) through the use of the coordinate change y = y' .5.E)j :$ :$ HE)k4jzj (5. Xp E IR n + nu rather . the system (5.E)I IP2(t. y') Again.y. it may be verified that v·1(t.20) Using exactly the same techniques as in the proof of theorem 4.5. the system (5. x) + Eg'(t. v 1(t.26) where v(t.16) with the difference that to IR. x))dr IPI(t. however. 0 u Mixed Time Scales As was noted in Chapter 4.25).20.5.z. z. X. y. 21).2 + n . z.5.5.27) E/aiz) + EP! (t. (5.1.y) (5.z.5. (5. x) (5.5.z. where the extra nu states represent the unmodeled dynamics.n. (5. (5. Xpw e IR " satls fi es t he equatiOns k5IYI and HE) e K.20).25) Ay' + h(t. we let the matrix A be a constant matrix (we will only consider linearized adaptive control schemes in the next section). (5.5. we may transform (5. y') (5. z.4.
w can be written as w = w. 0 = 0 is an equilibrium of the system.e + b. (3. The output error parameter update law is Averaging Analysis To apply the averaging theory of the previous section.Ym = Yp.30) As in the ideal case.Yp• we may subtract equation (5. T c.36) may be combined to yield b. is stable.5.o w (5.32) This is similar to (3.28) for the adaptive system.5.5. = 0. If the tuned error e.5. E IR 2n having the obvious interpretation). (} A.5. defining e = Xpw. An error formulation may now be derived. + b.5.2 and c..Ym can be decomposed as eo = Yp.w!'o (5. (5.30) and (3.5.25) so that averaging may be applied. is small) and focus attention on the linearized and simplified system e e .5.29) Now. Consequently.37) with the corresponding equation (3.5. = [ ~ l e E IR3n + n.31) Typically.5.O Qe .Yp• + e. The averaged parameter error .T e + e. for which the tuned plant transfer function approximately matches the model transfer function at low frequencies (at those frequencies. the effect of unmodeled dynamics is small). (5.5. bxdo•cJ A+ bxcl bxdl 0 A A.w. The first of the new terms is an exogenous forcing term and the second a term which is linear in the error state variables.242 Robustness Chapter 5 Section 5. one sees the presence of two new terms in the second equation.35). and refer to it as the tuned value o.Qe (5.37) reduce to (3.2 ' (} .5.5.Ym (5.5. = Yp•. Let 0 := (}. and we define the tuned plant as Xpw• Yp• where A. We define . Xpw (5. + Qe (with w.33) Now.xpw• T g (c.31) to get A.o.co•r c.35) along with the update law "' (} = . A. the terms disappear and the equations (5. Since A. We will also treat the second term ge. which are chosen as tuned values correspond to those values of 6.34) Ap + bp d0 .Qe as a small perturbation term (which it is if e.5.w. with respect to the tuned system instead of the model system of Section 3. (5.5 Averaging Analysis of Slow Drift Instability 0 = fJ = . and rewrite (5. it may no longer be possible to find a o• E IR 2n such that the closed loop plant transfer function equals the model transfer function. the tuned closedloop matrix (cf.gc[eQe.28) as Comparing the equations (5.5.geow 243 (5. a small parameter.5.3 7) ~T ~T Note the analogy between (5.5.gc.ge.O w.5.30) from equation (5. the origin = 0.ge..gc!'ew.36) (5.38) .ew.5.5.19)..24).38) is of the form of the mixed time scale system (5. the values o. ~ In tum the output error e0 := Yp.Xpw• + b.Xpw• and eo = Yp. Without the term e.)w (5. and the error between the tuned plant output and the model output is referred to as the tuned error e. the transfer function of the tuned plant is not exactly equal to the transfer function of the model.. it is easy to see that the system (5.20). we set the gain g = E.5. Instead.e + T~ b.5. Now.5. represent the parameter error with respect to the tuned parameter value.5.. we will assume that there is a value of fJ which is at least stabilizes the closed loop system. so that (5. bp bx 0 E IR3n + nu.e + b.18)).5.5.cJ bxcT bp dl (5.28). we will drop this term for the sake of our local stability/instability analysis.  ~T We will call A.
is Hw.(jw) > 0 (the reason that this is heuristic rather than precise is because the columns of Hw.38) is exponentially stable.1M.(jw)S. we can make the following observations a) If the simplified.)d. e.42) Frequency Domain Analysis To derive a frequency domain interpretation.39) where fav (0) is defined as f. Also.rUw) may not be linearly independent at every set of 2n frequencies).(s) is close to the model transfer function M(s) at least for low frequencies (where there are no unmodeled dynamics).(s ).44) so that 1 (sf.3.(ii) "  )~~ ~ 'T w.(s) ~ (5. even if M(s) is so.9) for averaging in the ideal case.5.5.37) is locally stable in the sense that if the tuned error and the initial conditions on 0 are sufficiently small.5.5. Consequently Rw./0) =  ! (5.43) where the transfer function from r to w. The chief difference is the presence of the tuned plant transfer function M.5..41) Thus (5.6). then the system (5. the reference input is required to have at least 2n points of support in the frequency range where ReM.40) may be written as (5..w.) 1 b. The proof of this follows from theorem 5.45) The crosscorrelation between w. for Rw. From the form of the integral in (5.5. is then given by 1 It is easy to see that if all the eigenvalues of Rw. trajectories will eventually be confined to a neighborhood of the origin (the size of the neighborhood depends on e. (0) may not be a positive semidefinite matrix.5.(s).(dw) 27rCo• .5. one may deduce that a necessary condition for Rw. the (simplified) overall system (5.w[(.46).46) and (4.rUw)M. it follows that to keep the adaptive system stable. then Rw•w (0) may fail to be positive semidefinite.5.5..5 Averaging Analysis of Slow Drift Instability 1 ~. linearized system (5.(t) ci [ e'"<<>b.5. M.5. we assume that r is stationary.38) for E small is easily determined from Lhe eigenvalues of the matrix in (5.A) 1bAP .5..5.5.oo 00 245 0av satisfies Rw. .45). then the original ~ystem (5. '! .5. ~T ~ Hw.(t )I goes to zero).38) is unstable (in th~ sense of Lyapunov).w.46) (5.{jw) is negative.w.38).5. (sf A.= Jc.31) is concerned./0) to have no negative eigenvalues.rUw)Hw. The spectral measure of w./0) to have no zero eigenvalues is for the reference input to have at least 2n points of support.5. (0) are in the right half plane.(t T 0 I r)b ( .(dw) ~ • ~ T (5. It is important at this point to note that all of the analysis is being performed on the averaged version of the simplified linearized system (5. sufficient excitation at lower frequencies is required.5. and goes to zero as I e.w.5. Since the tuned plant transfer function M. As previously.40) may be written as the crosscorrelation at 0 between w.] dt l ii (5.(s) in place of the model transfer function M(s).{jw) S. Heuristically speaking.w.36). and w.e A.5.(s) may not be strictly positive real. .40) Note that fav(O) is a linear function of 0 so that the stability/instability of (5.(t) and W•f (t) Note the similarity between (5. In fact.1. (5.5. is related to that of r by Sw. It is also important to see that the stability/instability criterion is both signaldependent as well as dependent on the tuned plant transfer function M.w. if even one of the eigenvalues of Rw. As far as the original system (5.38) is globally asymptotically stable for E small enough. (4. 1f a large part of the frequency support of the reference signal lies in a region where the real part of M. = M. (5.r )dr (5. heuristically speaking. the matrix in (5.(Jw) Hw.w.5. (0) lies in the left half plane.(dw) = Hw. This transfer function is obtained by denoting the transfer functio~ of the tuned plant T Co• c.244 Robustness Chapter 5 Section 5.
8 (5.. In this instance. if Rw.L 0~ p(t) . and negative for 0. corresponding to the tuned error is small. the first input results in an unstable system and the second in a stable one. do.w. However.w. and is easily computed to be .w. where the reference input is only a DC input: e...r(t) J. the parameter error vector ii driven by e.37) since there are no adaptive parameters in the feedback loop. we discuss an example from Riedle & Kokotovic [1985].1 J.246 Robustness Chapter 5 Section 5. For example. The adaptive controller is designed assuming a first order plant with relative degree 1.w. the averaging analysis may be inconclusive if the averaged system Rw./0) Rw.0 0. we assume that the 'nominal' plant is of the form 0.L > 0 is a small parameter. if Rw.5.1 0 .47) e where J.:.2 0.4sint. Thus. cAW. This is what happens in Rohrs example (R3).LW (1 + J. These results are borne out in the simulations of Figure 5. 1.4 0. but its place is taken by the output disturbance which causes the parameters to drift away from their tuned values.15. The result of this section also makes rigorous the heuristic explanation for the instability mechanism of the example in (R2) where a significant high frequency signal is present in a range where the tuned plant transfer function is not strictly positive real.0 0. kp is chosen to be 1. drifts away from the origin in the presence of noise. 1 L) . since the closed loop plant matches the model at low frequencies.38) and for (5.1. 00 R w. The model is of the form 1/(s + 1) and we set the tuned value of c0 . _. including soml at zero.5.Y.37) is also unstable.w.6 0. namely c0 • to be 1I kp for the analysis. using arguments from theorem 5. ] + [ ~~g ] r(t ) e.Y. Rw. The error system is 0. Of course.5. if J.L = 0.0 0.. then the original system (5.2 0.f (0)  J 00 _ ( 1 .49) 0 0 Note that (5.2 1.4 sin t + sin 5t Rw.5.5. = 1 identical to 0*.5.046 0. then the original system is unstable.026 sin5t 0. We consider the plant Pes) .J.L 1 [.48) 0./0) has at least one eigenvalue in the open left half plane.49) is simpler than (5. Example In this section.w.4 P8• = S+T kp (5.w./0) r 2(t) Thus. A tuned plant is easily obtained by removing the unmodeled poles at .5.37).5. (0) has all its eigenvalues in the closed right half plane.6 Figure 5. Simulations seem to suggest that.15 StabilityInstability• Boundary for r 1(t) and r 2(t) = r 1(t)+0.15 ± j 2 to get the tuned values Co• = 1. For the example.5./0) has some zero eigenvalues.1 and r1(t) I wl < 1./0) is a scalar. the averaging is inconclusive for (5.5.(t) (5. in this case..LW 2 )2 1 2 + J.2 with kp unknown.5 Averaging Analysis of Slow Drift Instability 247 b) If the simplified linearized system is unstable.\ 5'.L )2 w2 S (dw) ' (5 5 50) · · Note that the integrand is positive for I wl > 1.
Thus. Choice of the frequency range of interest Plant Order Determination Considerations Frequency range over which tracking is desired Modeling of the plant dynamics in the frequency range of interest Low Filter high frequency components (unmodeled dynamics range) Sufficient richness Spectrum within frequency range of interest 3. it is the frequency range over which accurate tracking is desired.6. it is important to monitor the signal excitation in the identifier loop and to tum off the adaptation when the excitation is poor.e. an incorrect plant model resulted in an unstable loop. On the other hand. is liable to be illconditioned. partially left to the designer. Tum off parameter update when not rich enough . check step 5 5. into a region of instability. noise caused parameter drift.1 Robust Identification Schemes An important part of the adaptive control scheme is the identifier. When only parametric uncertainty is present. and may be influenced by a careful design. Steps of Robust Identification Step 1. so that the spectral content of the input becomes important. and b) the output had energy in the frequencies of the unmodeled dynamics (a DC signal and a highfrequency sinusoid).248 Robustness Chapter 5 Section 5. if the plant is of high order. resulting in slow parameter convergence along certain directions in parameter space. In summary. The robustness of the identifier is fundamental to the robustness the adaptive system. the covariance matrix Rw(O) (see Chapters 2 and 3) measuring the extent of persistent excitation. 2. From a practical viewpoint. Filtering of the plant input and output signals is achieved by the observer. so as to exclude the contribution of data from frequency regions lying outside the range of frequencies of importance to the controller (i. In an adaptive control context. low pass filtering with a cutoff somewhat higher than the control bandwith). the only cure is to inject extra perturbation signals into the reference input so as to provide excitation for the identification algorithm. The examples of Rohrs consisted of scenarios in which a) the input was not rich enough (only a DC signal). The order of the plant model must then be selected.6 METHODS FOR QUALITATIVE DISCUSSION Adaptive systems are not magically robust: several choices need to be carefully made in the course of design. 5. In the first case. In such an event. and they need to explicitly take into account the limitations and flexibilities of the rather general algorithms presented in earlier chapters. Parameter convergence is not guaranteed in general. with asymptotic tracking. Then.6 Methods for Improving Robustness 249 IMPROVING ROBUSTNESS5. adaptive schemes are proved to be stable.·y range of interest. This includes the case when the level of excitation is so low as to make it difficult to distinguish between the excitation and the noise. We summarize this discussion in the form of the following table for a robust identification scheme. An initial choice of the designer is the frequency range of interest. In the presence of unmodeled dynamics and measurement noise. We begin with a qualitative discussion of methods to improve the robustness of adaptive systems. consistent with a good low frequency model of the plant. To reduce the effect of noise. It is also clear that if the excitation is poor over periods of time where parameters vary. it is important to choose a low enough order plant model capable of representing all the plant dynamics in the frequen. The spectrum of the reference input is another parameter. drift instabilities may occur. The order should be sufficient to allow for modeling of the plant dynamics in the frequency range of interest. Regressor Filter Selection Reference Input Selection If not. The presence of a large parameter vector in the identifier may also cause problems of numerical conditioning in the identification procedure. the parameter identification will be ineffectual. with a bandwith determined by the filter polynomial (denoted earlier X(s)). A review of a few specific update law modifications is given in the next section. a great deal of excitation (a number of independent frequencies) will be required. but is not necessary to achieve stability. and is usually limited by actuators' bandwith and sensor noise. it is important that the input signal: a) be rich enough to guarantee parameter convergence. and b) have energy content in the frequency range where the plant model is of sufficient order to represent the actual plant. Recall that the identifier identifies the portion of the plant dynamics in the frequency range of the input spectrum. or adaptation algorithm. In the second. it may be reasonable to further filter the regression vectors in the identification algorithm. 4.
add perturbation signal Limit perturbation to plant 5. be too fast since this would involve an inconsistency of parameter values and sensitivity to noise. In order to use the schemes in the form presented so far. The choice of reference input is also one of the choices in the overall control objective. However.7 ROBUSTNESS VIA UPDATE LAW MODIFICATIONS In the previous sections. persistent excitation in the correct frequency range for identification may require added reference inputs not intended for tuned controller performance. The update law with deadz~>ne is given by . In these instances the estimator needs to converge much faster than the rate of plant parameter variation. This. In practice.1 Deadzone and Relative Deadzone The general idea of a deadzone is to stop updating the parameters when the excitation is insufficient to distinguish between the regressor signal and the noise. one is often confronted with systems which are fairly well modeled. since the input is required both for generating the reference output required for tracking. In certain instances. the estimator needs to discount old inputoutput data: old data should be discounted quickly enough to allow for the estimator parameters to track the timevarying ones. adaptive control is most useful in scenarios involving slowly changing plant parameters.6. DESIRED MODEL REFERENCE BANDWIDTH PARAMETER VARIATION UNMODELED DYNAMICS FREQUENCY SPEED OF ADAPTATION REFERENCE SIGNAL BANDWIDTH Figure 5. however. except for a few unknown and uncertain components which need to be identified.6 Methods for Improving Robustness 251 If the excitation is not rich over periods of time where parameters vary. all of the prior knowledge needs to be completely discounted. More specifically. 5. however. 5. we reviewed some of the reasons for the loss of robustness in adaptive schemes and qualitatively discussed how to remedy them. however.4 Time Variation of the Parameters The adaptive control algorithms so far have been derived and analyzed for the case of unknown but fixed parameter values.1 in the next chapter. resulting in extra requirements on the amount of excitation needed. Further. In particular.3 The Usage of Prior Information The schemes and stability proofs thus far have involved very little a priori knowledge about the plant under control. as well as furnishing the excitation needed for parameter convergence (this dual role is sometimes referred to as the dual control concept). inspired from Johnson [ 1988].7. The discounting should not. We conclude this section with Figure 5. From a robust control standpoint.16. the reference model should not have large gain in those frequency regions in which the unmodeled dynamics are significant. control should only be attempted over a frequency range where a satisfactory plant model and controller parameterization exists. the insertion of perturbation signals into the reference input can result in undesirable dithering of the plant output. We indicated above how important the choice is for the identification algorithm. In some applications (such as aircraft flight control).250 Robustness Chapter 5 Section 5. The reference input in adaptive systems plays a dual role. Therefore. indicating the desired ranges of the different dynamics in an adaptive system.6. In this section. increases the order of complexity of the controller. In practice. 5. the adaptation is turned off when the identifier error is smaller than some threshold. the control objective (or reference model choice) should have a bandwidth no greater than that of the identifier.6. 5. consider the input error direct adaptive control algorithm with the generalized gradient algorithm and projection.2 Specification of the Closed Loop Control ObjectiveChoice of the Reference Model and of Reference Input The reference model must be chosen to reflect a desirable response of the closedloop plant. the problem of incorporating prior information can be solved in a neat and consistent fashionfor this we refer the reader to Section 6. Thus.16 Desirable Bandwidths of Operation of an Adaptive Control System. we present modifications of the parameter update laws which were recently proposed as robustness enhancement techniques.
1.6) where flo is a prior estimate of fl (for this and other modifications see Ioannou ( 1986] and Ioannou and Tsakalis [ 1986]).1 which is not determined by e 1 or e 2 alone. Sastry [1984]. the deadzone threshold.7.1 from observing the noise floor of the parameter update variable e 1 (with no exogenous reference input present). From a practical standpoint. It is claimed (cf.1. relative identification error). resulting in undesirable closedloop performance.5). Ortega. K.7. 0 if I e t1 ::. It is also possible to modify it online depending on the quality of the data.1 which follow from their calculations are. one would choose .1 in equations (5.6) and (5.7.252 (} Robustness Chapter 5 Section 5. though this claim has not been fully substantiated.7. the adaptive law can only guarantee that the relative (or normalized) identification error becomes smaller than the deadzone eventually. Equation (5.7.1 (5.7. 5. for the direct output error scheme (5.7. The logic behind this socalled relative deadzone is that if the regressor vector is large. Ortega. then the identification error may be large even for a small 5.2) represents the size of the deadzone.7. Also. Kreisselmeier & Anderson [1986]. Narendra & Annaswamy [1986]) have suggested different techniques for the choice of the deadzone . If the deadzone .5) where u is chosen small but positive to keep fl from growing unbounded. . the output error direct adaptive control algorithm with gradient algorithm is modified to _.7) may be less susceptible to bursting than (5.2) and e 1 in (5. . Peterson & Narendra [1982].1 is.7) is that it retains features of the algorithm without leakage (such as convergence of the parameters to their true values when the excitation is persistent).1) (5. Praly & Landau [1985] and Kreisselmeier & Anderson [ 1986] is somewhat different in that it involves a deadzone size .7.7) attempt to capture the spirit of (5. It is easy to see how the other schemes (including the leastsquares update laws) are modified.1).7..7. The most critical part in the application of these schemes is the selection of the width of the deadzone ..1) and (5.7) tries to tum off the drift towards 0 when I e 11 is small. Thus.3) (5. as before.7. The bounds on . The chief advantage of the update law (5.7.4) will not tend to zero. it is then important to prove that the regressor vector is bounded. The details of the relative deadzone are somewhat involved (the three papers referenced above are also for discrete time algorithms). then set 0 = 0.6.7 .3 Regressor Vector Filtering The concept of low pass filtering or pre!conditioning the regressor vector in the parameter update law was discussed in Section 5. but by how large the regressor signal in the adaptive loop is (the deadzone acts on a suitably normalized. (5. However. It is usually accomplished by low pass filtering of the process input and output in the .2) (} and as before. Similarly. but will only be asymptotically bounded by a large . e 2 in equations (5. if c0 = c min and 0 < 0. Praly & Landau [1985].7. these results are to be interpreted as mere existence results.7.7.7) Both (5. To complete the proofs of stability with relative deadzones.7. if the closedloop system were unstable.7. Two other interesting modifications in the spirit of (5.1. and one suggested by Narendra and Annaswamy [1987] (5.7.1 and establish that the tracking error eventually converges to the region I etl ::.7.3).1.7." unlike the absolute deadzone approach.1 where . the algorithm (5. extremely conservative. Samson [1983].6) tries to bias the direction of the drift towards flo rather than 0 and (5. 7 Update Law Modifications 253 (5. (5. The parameter .5) without its drawback of causing flo if e 1 is small. the absolute identification error could be unbounded._ (} (} c c ge1v if I etl > .4) transfer function error due to unmodeled dynamics.reisselmeier & Anderson [ 1986]) that the relative deadzone approach will not suffer from "bursting.2 Leakage Term (uModification) Ioannou & Kokotovic [ 1983] suggested modifying the parameter update law to counteract the drift of parameter values into regions of instability in the absence of persistent excitation.1 is too large.5) are (5. The approach taken by Peterson & Narendra [ 1982]for the case when the plant output is corrupted by additive noiseand by Praly [1983] and Sastry [1984]for the case of both output noise and unmodeled dynamicsis to use some prior bounds on the disturbance magnitude and some prior knowledge about the plant to find a (conservative) bound on . The approach of Samson [1983]. Praly [1983]. A number of recent papers (for example. The original form of the modification is. however.7.7. Practically.
8 CONCLUSIONS In this chapter. Averaging and Hybrid Update Laws A key characteristic of the parameter update laws even with the addition of deadzones. 5. slow down the rate of adaptation or use a hybrid update algorithm (this is rather like a variable time step feature in numerical integration routines). Another modification of the parameter update law in the same spirit is the socalled hybrid update law involving discrete updates of continuous time schemes. the tk refer to parameter update times. First. Also. In practice. In fact. they are incapable of tracking fast parameter variations. If it appears that the plant parameters are not varying very rapidly.8 Conclusions 255 identification algorithm and is widely prevalent (cf. they result in undesirable transient behavior if the initial parameter estimates result in an unstable closed loop (since stabilization is slow). yet without excessive parameterization.8) relies on the averaging inherent in the integral to remove noise. if the identification error momentarily becomes large. the parameter update operates instantaneously.254 Robustness Chapter 5 Section 5. A great deal of the preceding discussion should serve as design guidelines: the exact design tradeoffs will vary from application to application. implement the adaptive law with a deadzone whose size is determined by observing the amount of noise in the absence of exogenous input. we studied the problem of the robustness of adaptive systems. together with recently proposed update law modifications. such as leakage. b) Choose a reference model (performance objective) whose bandwidth does not extend into the range of unmodeled dynamics. The general message is that it is perhaps not a good idea to treat adaptive control design as a "black box" problem. but rather to use as much process knowledge as is available in a given application. their ability to maintain stability despite modeling errors and measurement noise. Second. Then. the best way to use them is after the initial part of the transient in the adaptation algorithm or a short while after a parameter change. the remarks in Wittenmark & Astrom [ 1984]). leakage and regressor vector filtering is their "impatience. and a heuristic analysis gave additional insight. Further explanations of the mechanisms of instability were presented.8) lk In (5. The results are not yet in final form in the literature and estimates accruing from systematic calculations are conservative and not very insightful. we studied in great detail the averaging effects of using a small adaptation gain on the parameter trajectories. tk + 1 ]. The result indicated that the property of exponential stability is robust. If prior information is available. inject extra excitation into the exogenous reference input as a perturbation signal. 7. We did not give a formal statement of the convergence results for all the adaptation law modifications. Finally. that is.8). Slow adaptation and hybrid adaptation laws suffer from two drawbacks. Other modifications. The mechanisms of instability found in the Rohrs examples were discussed in view of the relationship between exponential stability and robustness. may be added . A possible cure for this impatience is to slow down the adaptation. BIBS stable systems can become unstable in the presence of arbitrarily small disturbances). The complexity of the relationship between the robustness margins and known parameters. Anderson & Bitmead [ 1984].7. illustrating several mechanisms of instability. c) In the course of adaptation. and the dependence of these margins on external signals unfortunately made the result more conceptual than practical.7.7. transient noise. which the "impatient" algorithms are better equipped to handle. perhaps for reasons of spurious. Some formalization of the concept and its analysis is in the work of Johnson. and the controller parameters are held constant on [ tk. various methods to improve robustness were reviewed. The logic is that low pass filtering tends to remove noise and contributions of high frequency unmodeled dynamics. We first reviewed the Rohrs examples. a reduction of the effect of additive noise (by averaging) is also observed. One such modification of the gradient update law (due to Narendra. the amplitude of the disturbances should be checked against robustness margins to determine if stability is guaranteed. we derived a general result relating exponential stability to robustness. Consequently. using an averaging analysis. 5. monitor the excitation and turn off the adaptation when the excitation is not rich over a large interval of time. If necessary. The law (5. We have attempted to sketch a sampling of what is a very new and active area of research in adaptive systems. A guideline for design might run as follows a) Determine the frequency range beyond which one chooses not to model the plant (where unmodeled dynamics appear) and find a parameterization which is likely to yield a good model of the plant in this frequency range." Thus. while examples show that the BIBS stability property is not (that is. Khalifa & Annaswamy [1985]) is lk +I 8(tk+ 1) = 8(tk)  Jge vdt 1 (5.4 Slow Adaptation. use it (see Section 6.1 for more on this). In Chapter 4.
The representation (6. and Dj are known.1.1. antialiasing filters (for digital implementations).256 Robustness Chapter 5 d) as desired. sampling intervals (for digital implementations). and Bai and Sastry [ 1986]. CHAPTER6 ADVANCED TOPICS IN IDENTIFICATION AND ADAPTIVE CONTROL The guidelines given in this chapter are for the most part conceptual: in applications. the difference an adaptive controller makes in a given application is chiefly due to the art of the designer! 6.1. {3j are unknown.1 USE OF PRIOR INFORMATION 6. Consider for example the circuit of Figure 6.1) is general enough to model several kinds of "partially know~!" systems. real parameters.~ {3jDj(s) j =I n (6. resetting. {3j from inputoutput measurements of the system. Dasgupta [ 1984]. Examples Network functions of RLC circuits with some elements unknown.1) where N. stable rational transfer functions and a. Implement the appropriate startup features for the algorithm using prior knowledge about the plant to choose initial parameter values and include "safety nets" to cover startup..1. shutdown and transitioning between various modes of operation of the overall controller.1 Identification of Partially Known Systems We consider in this section the problem of identifying partially known singleinput singleoutput (SISO) transfer functions of the form No(s) +~a. The identification problem is to identify a. proper. with the resistor R unknown (the circuit is drawn as a two port to exhibit the unknown a) 257 . Even with a considerable wealth of theory and analysis of the algorithms. controllerarchitecture featuring several levels of interruptability.. questions of numerical conditioning of signals. and so on are important. N. The problem was recently addressed by Clary [ 1984].(s) P(s) i =I m D 0 (s).
stable rational transfer functions and a. controllerarchitecture featuring several levels of interruptability. The representation (6. antialiasing filters (for digital implementations). resetting.(s) P(s) i =I m D 0 (s). sampling intervals (for digital implementations). CHAPTER6 ADVANCED TOPICS IN IDENTIFICATION AND ADAPTIVE CONTROL The guidelines given in this chapter are for the most part conceptual: in applications. Implement the appropriate startup features for the algorithm using prior knowledge about the plant to choose initial parameter values and include "safety nets" to cover startup. {3j from inputoutput measurements of the system..1.~ {3jDj(s) j =I n (6.1. Even with a considerable wealth of theory and analysis of the algorithms.1) where N.256 Robustness Chapter 5 d) as desired.1 USE OF PRIOR INFORMATION 6. The identification problem is to identify a. and Bai and Sastry [ 1986].1. The problem was recently addressed by Clary [ 1984]. proper.1) is general enough to model several kinds of "partially know~!" systems. N. with the resistor R unknown (the circuit is drawn as a two port to exhibit the unknown a) 257 . shutdown and transitioning between various modes of operation of the overall controller. Dasgupta [ 1984]. and Dj are known. questions of numerical conditioning of signals. Examples Network functions of RLC circuits with some elements unknown. real parameters. Consider for example the circuit of Figure 6. the difference an adaptive controller makes in a given application is chiefly due to the art of the designer! 6.1 Identification of Partially Known Systems We consider in this section the problem of identifying partially known singleinput singleoutput (SISO) transfer functions of the form No(s) +~a..1. {3j are unknown. and so on are important.
s"'.4) {31 with m ~ n and a 1.. +a 1 s"+/3nsn1+ .1) by choosing No(s) Do(s) Figure 6..1. +{31 Sensor which is of the form of (6.258 Advanced Topics in Identification and Control Chapter 6 Section 6..1 is i1 [ i2 (6.1 Two Port with Unknown Resistance R 0 snl~(s) N1(s) Dj(s) s1 . . with a plant P(s) known.1 I~ (s) . . which is of the form ( 6. as above The closedloop transfer function.. . . A simple example of this is shown in Figure 6.1 Pes) Use of Prior Information a. namely P1 1 + k P is of the form of (6.2 Plant with Unknown Feedback Gain ~ {Jj j = 1 sj. .1) by choosing. Sn+fJnsn1+ .. If P is unstable.2.1.1: 1) if P is stable. but known actuator and sensor dynamics (with transfer functions Pa(s) and Ps(s) respectively).5) 1 [Y (s) Y21(s) 11 Y ds) ] Y22(s) · [v ] = v~ (6. n If the short circuit admittance matrix of the two port in Figure 6. c) Classical transfer function models. with unknown plant.7) where ~ (s) is (any) Hurwitz polynomial of order n. .1.1_n_ __ (6.1 which is of the form (6.3.1 + ··· +a 1 259 resistance).1 1~(s) N1 (s) Dj (s) i = 1.1. Circuits with more than one unknown element can be drawn as multiports to show that the admittance function is of the form of (6. m }=1.2) then a simple calculation yields the admittance function 1 +RY22 where ~ (s) is (any) Hurwitz polynomial of order n. plants of the form studied in Chapter 2 No(s) Do(s) 0 sn I~ (s) s 1 . . that is. sl1 Pes) Ps(S) .1.3) Actuator Plant amsm1 + ..n (6.. then by writing p "" Np 1Dp as the ratio of two pro~er s!able rational transfer functions.1 I~ (s) i=l.1 Pa(s)Ps(s)l~(s) sj.1.1 )... and a feedback gain k unknown.1. d) Systems with some known poles and zeros. Consider the system of Figure 6.. ___:1_=_.. .1 ).. {Jj unknown.3 Unknown Plant with Known Actuator and Sensor Dynamics The overall transfer function is written as m ~a. m }=1.. the closedloop transfer function is NP I Dp + k Np.1. Interconnections of several known systems with unknown interconnection__ gains. v1 (6.1.1. + (6.6) sn + Figure 6..1. can be stated in terms of the set up of (6.1 ).. . b) Figure 6.sj.
In analogy to the expression of the plant equation (6.. Using (6. }Vn+m(t)) E IR.1.1) (2. the signals zp(t)..18) or. the parameter error wlll co~verge . It is easy to verify that in the instance in which no prior information about the plant is available (case c) above).m }=1. ' f3n) E Equation (6.n (6.1..1.11) p g>O (6.1.17) or the leastsquares algorithm (} (6. .16) . for example. The purpose of the identifier is to produce a recursive estimate 0(t) of the parameter vector (}•. usage of the particular structure embodied in (6.1.14) in Chapter 2.12) along with covariance resetting.13) We also define the parameter error </>(t) j =I ~ n f3JDJYp + i =I ~ m aiNJ (6.15) so that. The identification problem is to determine the unknown parameters ai. . {31.8) (6. equation (6. with a reduction of computational requirements and usually faster convergence properties.(}. so that all the update algorithms and properties of Cha~ter 2 can.19) Note the close resemblance between (6.1. These decaying terms will be neglected for slmphcity.12) above is a reformulation of equation .<expone~tlal~y) to zero for the gradient or the leastsquares algonthm w1th resettmg 1f the vector w is persistently exciting. However.12). Yp(s) denote the input and output of the plant. if there exist a I> a 2. Since r and Yp are available.1) and identify the whole transfer function using one of the procedures of Chapter 2. p~ope:. One could. {31 from inputoutput measurements. · · ·.9) and the nominal parameter vector fJ• (J•r := (al.n+m a 2/ ~ J }V(T)}VT(T)dT to • ~ a1/ for all t 0 ~ 0 (6. i.ge 1w IR... (6.1.1. are appr~priate parameter upd~te laws.8) as g>O (6. for the analysis Wj := Njf Wm+j := DJYp i=1.e.14) Defining the signals and the identifier error e1(t) Zj(t).1. Let f(s). o> 0. As in Chapter 2.2.2) of Chapter 2. e1(t) = <J>T(t) }V(t) (6.1..1.1.1. .1. }V(t) are obtainable through stable Afiltering ~f r and Yp• up to some exponentially decaying terms (since Ni and D1 are stable.1 ). of course. ' am.1.Zp(t) (6. we define the output of the identifier as Zj(t) := (JT(t) }V(t) fJ(t). .10) we may rewrite (6. the gradient algonthm fJ = ..1.1.14).2. be used verbatim in this context.1 Use of Prior Information 261 Identification Scheme We now return to the general system described by (6.n+m (6. rational functions).1.12) and the plant parameterization of (2..1) will result in the identification of fewer unknown parameters.260 Advanced Topics in Identification and Control Chapter 6 Section 6. Thus. neglect the prior information embedded in the form of the transfer function ( 6. such that t0 + 6 where }VT(t) :== (}VI(t).1.16) is now exactly the same as (2.1. in the time domain Zp(t) = (J. .r}V(t) (6.1.3.
1. .1) with Ni. w is persistently exciting if and only if r is sufficiently rich of order n + m.1 shows that the identifier parameter will converge to this value. .1. ' Zp(JWn+m)) = o·Tuiw.1.25) It is important to note that (6. . b) It is difficult to give a more concrete characterization of identifiability. Ni.24) 0. (6. m I ADj(}w)l I<DaJ. Thus.1. • .1. Dao.Ni) (Jw)l < for all w. .Zp(t) . !51..20) and (6.9) are now replaced by (zp(Jw1).11). Wam+j := Da}Yp }= 1. the signals of (6.21) The identifiability condition implies that (6.. Nai . . we would have i = 0. since the signals w.21) has a unique solution for o·. In particular. which we will denote Effect of Unmodeled Dynamics The foregoing set up used transfer functions of the form ( 6.17). We will assume that I ANi(jw)l := I (Nai.1..12) is still valid. the nontrivial condition comes from the lower bound in (6.1. the transfer functions used to approximate the Ni and !51 will generally be low order.1. ' HwrUwn+m)) i=1. we simply assume: (AI) Roundedness of the Regressor The plant fi(s) is stable.1. j = < t (6.1 similar to the proof of theorem 2. The identifier. it was shown that the identifiability condition holds if the numerator and denominator of the plant transfer function are coprime polynomials. t (6. • wn + m• the vectors H wrUw i ) E <r n + m (i = 1 ' .1. . The parameter update laws of (6. uses the signals Wai (t) so that the identifier output (6.1 Under assumptions (A 1) and (A2)..!51 )(Jw>l for all w..1. 7.1.1.22). .1 6.1.1. . In practice.23) Proposition 6. ' n + m) are linearly independent. In that case.22) (A2) Identifiability The system is assumed to be identifiable. DaJ are approximations of the actual transfer functions N 0 . Proof of Proposition 6. . the identifier model of the plant is of the form Nao(s) + ~ ai Nai (s) i =1 m Dao(s) _: ~ f1}Daj(S) j =1 n (6. the Ni and !51 are only known approximately.m (6. Zp pertain to the true plant. n (6. while e1 (t) = Z. .1..Uw1).1.. An exception is the case of example c) and discussed in Chapter 2. proper stable transfer functions (neglecting highfrequency dynamics and replacing near polezero cancellations by exact polezero cancellations). . n The identifier uses the form (6.1. meaning that for every c~oice of n + m distinct frequencies w 1 . however.1.262 Advanced Topics in Identification and Control Chapter 6 Section 6.. Comments a) From (6. . it follows that if an input with (n + m) spectral lines is applied to the system. and !51 known exactly.18) are modified by replacing w by wa.2 Use of Prior Information 263 Frequency Domain Conditions for Parameter Convergence The techniques of Chapter 2 can be used to give frequency domain conditions on r(t) to guarantee (6. As usual. since the components of if wr(s) are proper stable rational functions of different orders. where Pa(s) is a proper stable transfer function. Do. while the true plant fi(s) is accurately described by (6.19). and Nao.1)..26) where O(t) is the parameter estimate at time t. .. . To relate this condition on w to the frequency content of r.1. To guarantee the upper bound. . we will need a new identifiability condition on the transfer function from r to w.(t).1.1.2. and the reference signal r is piecewise continuous and bounded. Consequently. while proposition 6.19).
:1w Wa T ~ to+ 6' J to 2 [xT(w(r) + .1.1.1.1. when w persistently exciting and .28) and gPO* .23). Thus.:1w shrinks (or. Dj) is small enough and w is persistently exciting. It is plausible that if E (the extent of mismodeling of the N.1.28).35) 0 The conclusion nowfollows readily from (6.30) in (6. (6.1.w(t) a 2/ ~ J w(r)wT(r)dr to ~ a1/ for all to~ 0 (6. ( 6.1.31) Let o' = o. condition (6.1. w Wa A l.30) and (6.1.1.gPwa w'[ P (6.1) can be used to guarantee the convergence of the parameter error to a ball around the origin.1.:1w is also persistently exciting. the gradient algorithm is described by lj> = (J T . ~ a {'. It is further readily obvious from the estimates in the statement of the theorem that the size of the ball goes to zero as . then Wa is also persistently exciting.e.28) Then s~p o·T .28).gwa Wa lj>. simply because .264 Advanced Topics in Identification and Control (J T(t) W (t) . When next lemma. E in (6.24) iS Small enough. This is established in the following two lemmas.1.29) Proof of Lemma 6.2 w + . (6.19) holds with w replaced by wa. For the lower bound.1. a' 2 o' > 0. Thus. we see that Wa is guaranteed to be persistently exciting. The upper bound of the integral in (6. (6.:1w is bounded and w satisfied a similar inequality. we use the triangle inequality to get Note that .1. such that Equations (6.1.e.1.:1w(r)l T (6. the terms in (6. i. i.:1w is sufficiently small.0 ] .33) and the leastsquares algorithm by lj> = (J w + .1.3.1.:1w(r ))} dr ~ a' 1 (6.:1w is persistently exciting if there exist a' 1 . if the systems of (6. there exist a I> a 2.31) are bounded.(J*T W(t) a lj>T(t)wa(t)+O*T(wa(t)w(t)) Chapter 6 Section 6. o> 0 such that 10 + 6 Define .1. a' 2 .. .27) If Persistency of Excitation under Perturbation the signal w (t) e IR n + m is persistently exciting.1.29) with resetting are exponentially stable.:1wP is bounded.29) are exponentially stable in the absence of the driving terms..1. since it is the difference between the outputs of two proper transfer functions with a bounded reference input r.29) (6. It is easy to see that this is guaranteed if wa is persistently exciting.1.34) is automatically verified.34) (6.0 1 112 sup l. the robustness results of Chapter 5 (specifically theorem 5.33).2 (6. the inaccuracy of modeling decreases). P . equivalently.:1w(t)l < a1 [.:1 w(t) e IRn+ m satisfies (6.1.:1w(t) = Wa(t). It therefore remains to give conditions under which the undriven systems of (6. follOWS from the .:1w iS Small.1.28) and ( 6.g and the signal . Consequently.. for all x e IR n + m of unit norm. The claim that .32) Consequently.1.1 Use of Prior Information 265 Lemma 6.29) have a similar form as without unmodeled dynamics with the exception of the extra forcing terms in (6.1.
+ {'JI (6.1. Then. 2 fN sup T I r(T)I (6. a nonadaptive controller is designed parametrically.37) where N is the maximum order of the ANi. indirect adaptive control scheme for a SISO continuous time system using an identifier in conjunction with an arbitrary Figure 6. Clp(s) monic.1.1) with fip(s).2.5. Vidyasagar [ 1985]). we will specialize our scheme to a pole placement adaptive controller.37) and applying theorem 5. We also indicated how it could be replaced by a pole placement algorithm for nonminimum phase systems in Section 3. and the definition of AW. we considered the specific case of a model reference indirect adaptive control algorithm. The degree of JP is n.1. A reason for the popularity of the indirect approach is the considerable flexibility in the choice of both the controller and the identifier. We will not need any conditions calling for the parameters to lie in a convex set or calling for lack of unstable polezero cancellations in the identifier.266 Lemma 6.4 Basic Structure of an Indirect Adaptive Controller The unknown.4. 6. we consider more general controller designs..1 Indirect Adaptive Control Scheme The basic structure of an indirect adaptive controller is shown in Figure 6.1.I (consequently.1 + · · · + a 1 Sn+f3nsnl+ . We discuss a general. we will need sufficient richness conditions on the input that we were not assumed previously to establish global stability.2 Global Stability of Indirect Schemes 267 If Then g(s) is a proper. the controller parameters are written as functions of the plant parameters. the input to the identifier is also sufficiently rich to cause parameter convergence in the identifier. In this section. Aw(t) is small enough for f small enough. Global stability of indirect schemes was shown by several authors in the discrete time case (Goodwin & Sin [ 1984]. Thus. roughly speaking. In Section 3.1. that is.3. the scheme is made adaptive by replacing the plant parameters in the design calculation by their estimates at time t obtained from an online identifier. the adaptive system consists of a fast parameter identification loop and a slow controller update loop. However. Using the estimate (6. stabilizing controller. for example. strictly proper plant is assumed to be described by an sn. robust controlsee. following the approach of Bai & Sastry [ 1987]. and others). it is easy to verify that I Aw(l)l :::.2. A Dr Thus. as well as to a factorization based adaptive stabilizer (of a kind that has attracted a great deal of interest in the literature on nonadaptive.1. some of the a i 's may be zero).. The compensator is a proper.3. and that of np is less than or equal to n. From lemma 6.3 see Doyle [ 1984]. m th order compensator of the form . and the persistency of excitation of w guarantees that of Wa. For application of the results. 2 n sup I g(jw)l w (6. nth order rational function with corresponding impulse response g(l) the L 1 norm of g(t) can be bounded by 00 II g lit = J Ig(T)I dr 0 :::.1.3. and Kreisselmeier [ 1985. we see that the parameter error will converge to a ball of radius of order f.. 6.36) Proof of Lemma 6. Cristi & Das [ 1985] used random parameter update times for proving convergence. coprime polynomials.3. Elliott.. We will show that when the reference input is sufficiently rich. First.3.3 Advanced Topics in Identification and Control Chapter 6 Section 6. stable. The controller is updated only when adequate information has been obtained for a "meaningful" update. In a continuous time context. Anderson & Johnstone [ 1985]. 1986] assumed that the parameters lie in a convex set in which no unstable polezero cancellations occur.2 GLOBAL STABILITY OF INDIRECT ADAPTIVE CONTROL SCHEMES The indirect approach is a popular technique of adaptive control.
. The design of the compensator is based on the plant parameter estimates 8 ( t ). (6.1 where 8*T = Consider the timevarying system (a..2.2. and therefore the compensator converges asymptotically to the desired one.(t) = (6. i. such that to+ 6 The adaptive scheme proceeds as follows: the identifier obtains an estimate of the plant parameters.1 ~ u Yp ·~· ..1) that the parameter error ¢ ( t) is bounded. This is measured through the "information matrix" I.8).268 Advanced Topics in Identification and Control Chapter 6 Section 6. model reference.. i. even if Yp( t) may not be.2.9) Update Sequence for the Controller Although the update law (6. it has been shown in Chapter 2 that under the condition that fip.2.. as t . It is shown in the Appendix (lemma A6.) is performed.e. . an. J w(r)wT(r)dr I.{3. Lemma 6. we will use the leastsquares with resetting ¢( t) P(t)w(t)e 1(t) P(t)w(t)wT(t)P(t) forallt>to (6.2..4 can be understood as a timevarying linear system which is asymptotically timeinvariant and stable..2. to be specified hereafter. Further. a2..2 Global Stability of Indirect Schemes 269 ¢ ( t).2. Then u su sn. The hope is that. A.9) may be shown to be asymptotically stable when w satisfies (6.1 0) if u is sufficiently rich. We will assume that there exists a unique compensator of the form (6.8) i>(t) with P( t/) = t # t. where t. (6. Therefore. the identifier identifies the plant correctly.(t)yp(t) has the form (6.2) for every value of the plant parameter estimates.. there exist m.0 asymptotically. it is of practical importance to limit the update of the controller to instants when sufficient new.4) Hurwitz polynomial of Note however that the upper bound in (6. then the timevarying compensator will converge to the nominal compensator and the closedloop system will be asymptotically stable. i.2.11) The identifier output is 8T(t)w(t) (6.. Anf3n) Y.3) the form (6. a> 0 such that the state transition matrix ~ ( t. The compensator design (pole placement.7) For the identification algorithm. o>O.X = (A + AA(t))x (6.oo. +6 k 0 I > 0.8* as t .. then the identifier error e 1(t) = y.1 0) (in fact only the lower bound).e. Further.2. Define w(t) e IR. sn + An sn. the system of Figure 6.e.2.2.6) where A is a constant matrix with eigenvalues in the open LHP and II AA ( t) II is a bounded function oft converging to zero as t .2. information has been obtained. . assuming that the plant parameter estimates are the true parameter values (certainty equivalence principle). is the sequence of resetting times. if there exist C(s) (6. ·· sn.to)ll:::.I+ a monic + A I.2. If ¢ ( t) is the parameter error ¢(t) = 8(t)8*.oo.2..2.oo. mea(tto) where 8( t) is the estimate of 8* at time t. if the support of the spectrum of u has at least 2n points (assuming that u( t) is stationary). ·· with A(s )..2.2n with Laplace transform a2I ~ J w(r)wT(r)dr to ~ a 1I for all t 0 ~ 0 (6.2.2) a 1 . . t 0 ) of A + AA (t) satisfies ll~(t. dp are coprime polynomials. We first design an identifier as in Chapter 2. etc.2. The following lemma guarantees the asymptotic stability of the linear time varying system. if w( t) is persistently exciting.11) is asymptotically stable.2.5) .1 Yp ~ (6. It would appear intuitive that if 8 ( t). w satisfies the lower bound condition (6.1 0) is not needed for the specific algorithm chosen here. Then (6.10) [ ~·~· .
First. satisfies I.. we still have that rfi(t)0 as t .oo as i . For the adaptive control situation. +Cl 0.2..oo as i . If scenario (b) happens.. snldp. + o. is bounded.. ...4. C are constant matrices. )T (6. Using (6.12) Hw. Further. the identifier parameter error converges to zero exponentially as too.2.2. as in Chapter 3) where A. := argmin Cl J w wT dT I. = oo or o. is an unbounded sequence. Thus. we see from the arguments of Chapter 2 that w is persistently exciting. ) and fie( t. Even in this (worstcase) situation. contradicts the hypothesis that o. with the plant and compensator as in lemma 6.2.2 Convergence of the Identifier Consider the system of Figure 6. Consequently. 0 Theorem 6. we choose the update sequence t. however.13) and Since the degree of fi is (n + m ). there exists 0 < ~ < I such that (6.3 Asymptotic Stability of the Indirect Adaptive System Consider the system of Figure 6.oo.I1C{t) and 11b{t)O as rfi(t)0. and t. w ( C + 11 C ( t ) )z Proof of Lemma 6.2. + 1 . If Then the inpuf r is stationary and its spectral support contains at least 3n + m points the adaptive system is asymptotically stable. notice that when the plant parameters are known. we will establish this fact.t.(S) = ~ "Adpd A fj [A . We proceed by contradiction.t'from lemma A6. Now. If this happens the system becomes time invariant after t.2. ~ yi (6.12). by t 0 t. Then the persistency of excitation of w( t) follows as a consequence of the result of lemma A6. ) are the denominator and numerator of the controller at time l.2.2 it follows that Wm( t) approaches w( t) for t large enough. then one of two following possibilities occurs (a) (b) There exists an i < oo such that o.2 and the fact that wm( t) is persistently exciting. More precisely. the compensator parameters are held constant between t.15) Then. where o. + 1 = t. and o.14) where de( t. from lemma A6. If Then the input r is stationary and its spectral support contains at least 3n + m points. kpfipsnl A dp. however. we may write the transfer function . kpfip.2.oo .2 The proof uses lemmas which are collected in the Appendix. This fact. one can find the transfer function from r to u to be :=fj where 11A ( t).2. a (6.13) if the sequence o. This.2. Lemma 6. since the controller is not updated. no more than (n + m) of the spectral lines of the input can correspond to zeros of the numerator polynomial.270 Advanced Topics in Identification and Control Chapter 6 = Section 6. If o. rfi ( t) :f..2.2.2.2. Consider the scenario (a) first. consider scenario (b). = oo. + 1 • We will assume that the compensator parameters are continuous functions of 8.14).2.1 shows the conclusion (6.9) and resetting times identical to the controller update times given in (6. b is a constant vector and A is stable. 0.1. .2 from r to w to be Global Stability of Indirect Schemes 271 Thus.2. given 'Y > 0. 11 b( t) are continuous functions of rfi ( t) and 11A{t). (6. In particular. 11 C( t).2. is a bounded sequence.4 with the leastsquares update law (6. contradicts the assumption that o. We may now relate the richness of the reference signal r( t) in Figure 6. . but we may still write the following equation relating r( t) to w( t) (A+ 11A(t))z + (b + 11b(t))r o. .8). the closed loop system is timeinvariant and stable so that we may write the equation relating r( t) to the signal wm( t) ( wm( t) corresponds to w( t) in the case when rfi ( t) = 0.2. := t. lemma A6.4 to the convergence of the identifier..
dp.22) from ~ ml O(t. has the interpretation of being the nominal compensator parameter vector.21)). ).) fip( t.)l I Oc ( t.3 in the Appendix that (6.) and dp( t. dp( t.2.+ 6. The compensator C is chosen so that the closedloop poles lie at the zeros of a given characteristic polynomial dc1 (s ).21) is satisfied by guaranteeing that the smallest singular value of the matrix in lemma A6.4 The first half of the theorem follows from lemmas 6.2.2a slight modification of the arguments of lemma 6.1 points all signals in the loop are bounded and the characteristic polynomial of the closed loop system tends to dcl (s ).2. . it is easy to verify that if kp( t. the "adaptive" pole placement scheme is mechanized by using the estimates fip( t.2. For the second half.)0*1 I Oc(t.2. Let (6. d we see that (6.21) Since J wwTdt I.2.2.3 in the Appendix) for any arbitrary dc1 of degree (2n. + o.)O.2. with the least squares identifier of (6. + o. de each of order n .2.2.) are coprime I Oc(t.2.21 ).2 is needed to account for the new condition (6. so does Be ( t. + o.))A(O*)II Further.o.3 in the Appendix that there is an m > 0 such that a IIA(O(t.2. I  0 exponentially.16) may be solved (see lemma A6.25) for some m 1• Since 0( t.) of the numerator and denominator polynomials.3 of the Appendix (6. In analogy to the plant parameter vector 0.) fip( t. we need to modify the definitions of the update times.j ~ miiA.2.3 follows readily from lemmas 6. equation (6.:::: 'YI kp( t.2. ).24) More precisely (6. with fie. Using lemma A6.2. dp( t.2.2.) are coprime at the instants t.) to .23) o. 6. 6.) fip( t. there exist unique fie( t. When the plant is unknown. to guarantee that kp( t.19) where measuring the coprimeness of kp( t.2. Moreover.2.2.2.272 Advanced Topics in Identification and Control Chapter 6 Section 6.2 Global Stability of Indirect Schemes 273 Proof of Theorem 6.I such that (6.2.2. Further.) fip( t.2. we see from lemma A6.2.2) with m = n.I.2.1 (0*)111 O(t. If Then the input r( t) is stationary and its spectral support contains at least 4n . subtracting (6.2.) are coprime.) converges to 8* exponentially. + o.21) in the update time.2.2 Indirect Adaptive Pole Placement Pole placement is easily described in the context of Figure 6.16) When fip. )exceeds a number u > 0. but this is easy because of the convergence of the identifier and the coprimeness of the true fip. (6.I) (6.2.3 again. de need to be found to satisfy (6.2. . dp( t. we have the parameter vector of the compensator (cf.19)(6.4.18) As usual.2.) follow from the plant parameter estimates 0 (t) of the identifier.1 and 6. is the smallest real number satisfying I.20) (6.) fip( t.4 Asymptotic Stability of the Adaptive Pole Placement Scheme Consider the adaptive pole placement scheme.) and dp( t. In other words.1. defined by (6. Proof of Theorem 6. Thus (6.) I is bounded (by (6.) and dp( t.)0*1 o.2.1). + o.9) and the update sequence t. 0 Theorem 6. I Oe ( t).2. we get that (6.2.2. ) of order n . dp are coprime.) and de( t. ). typically of degree (2n . fie.8)(6.22) with d the vector of coefficients of the polynomial cl· It follows from lemma A6.17) The estimates for kp( t.1).
the ring of proper. Dp. or.28) in terms of the coefficients of NP. We assumed that P and C are factored coprimely in R (not uniquely!) as C(s) (6.5 is BIBO stable if and only if each of the four elements of (6.5 is BIBO stable if and only if CNpNc + DpDc).2. Adaptive Stabilization Using the Factorization Approach The objective is to design the compensator C(s) adaptively based on the estimate 0 of the plant parameters so that the closed loop system is asymptotically stable with all signals bounded.2.2. NpNc + DpDc is a unimodular element of R .2.19)(6.cJ(si Apt).27) From Vidyasagar [ 1985]. If f e JR.2. n and I e JR. The ring R is a more convenient ring than is the ring of polynomials for several reasons (see for example.2. u 2 is given by Heu A JT (si Apf). However. Vidyasagar [ 1985])including the study of the robustness properties of the closed loop systems.30) and further that all solutions of (6.2.2.1bp (6.28) may be written as Figure 6. bp. we can state that a compensator stabilizes the system of Figure 6.1/ + Q(s) [ 1. bp.29) (6.21).. let (Ap.274 0~. it may be shown (see Vidyasagar [1985]) that a right coprime fraction of Pis given by cJ(si Ap1 ).26) is stable. cJ) be a minimal realization of P(s ).1/ Q(s)c'J(si Apf). in fact.lcJ (such a choice is possible by the minimality of the realization of Ap.5 if and only if .2.2. stable. Set u 2 (t) = 0. In this chapter.32) (with Q(s) = 0) lies in calculating The system of Figure 6.1bp (6. We will be interested in a parameterization of all solutions of (6.2. The optimal choice of Q(s) is the topic of the socalled H 00 optimal control systems design methodology.n are chosen to stabilize Apf := Ap.2). cJ).1bp] = 1 [1 P] A A (.2. rational functions. obtaining the desired closed loop transfer function. then.2.JT (si Apf). i. The identifier and compensator update sequence { t. and use a fixed Q rather than one whose calculation depends on (s) the current estimate of plant parameters. The only difficulty in mechanizing the n th order compensator of (6. we will not concern ourselves with anything more than stabilization.31) (6.26) with Q(s) e R an arbitrary element chosen to meet other performance criteria (such as minimization of the disturbance to output map. Advanced Topics in Identification and Control 0 Chapter 6 Section 6.2. optimal desensitization to unmodeled dynamics.3 Indirect Adaptive Stabili~ationThe Factorization Approach The Factorization Approach to Controller Design We will briefly review the factorization approach to controller design (the nonadaptive version).1 belongs to R .32) The plant Pis as defined in (6. belongs to R . The optimal choice of Q(s) depends on the plant parameters.2. 1 +PC 1 (6. For simplicity we will. Without loss of generality.28) parameterizes all stabilizing compensators. it follows that the system of Figure 6.e.2. such a choice of Q(s) may not be unique or depend continuously on plant parameters.1bp 1. e 2 to u 1 . The transfer function relating e 1 . Consider the controller structure of Figure 6.2. ) are specified in (6.2 Global Stability of Indirect Schemes 275 (6. equivalently.).5. fix Q(s) = 0 in the adaptive stabilization which follows. For this purpose.2.5 Standard One Degree of Freedom Controller 1 +cJ(si Ap1 ). etc.28) 6.1) and the compensator Cas in (6.31) and (6. Equation (6.bp/T and Apt := Ap.
. and Vidyasagar [ 1985].1 (t.9). Ap( t. ) a 2( l.5 Asymptotic Stability of the Adaptive Identifier Using the Factorization Approach Consider the set up of Figure 6. we consider the controllable form and observable form realizations of the plant estimate (6.. (6. To that effect.2.)[pl+{31(l. The issue of the parameterization of the adaptive controllers becomes a dominant problem. ) 0 (6. If a 1( l.) bp(t. consider any stable polynomial S 11 + p11 s 11 .2 Global Stability of Indirect Schemes 277 f ( t.)jT = (6.)M (t. the.) = [P1+{31(t.) = De.) . These results will help us to establish the parameterization of MIMO adaptive controllers in Section 6. transfer function properties which are easily established for SISO systems are much more complex for MIMO systems.1 + · · · + p 1 • Then it is easy to see that JT(t.) {31(0 {32(1.) Nc( l.2.34) M( l.37) Then.3.2. )cJ< l.3. In Section 6.2.2.) cJ(t.) bp(l.) and Ap ( l.2. Indeed.) and I ( l.8).. As previously.3.t 1 Note that M( l. .) is the only calculation that needs to be performed.) E IR" to stabilize Ap( t.2.. Elliot.} of (6.2.) have their eigenvalues at the zeros of s" + p11 S11 I + · · · + p 1 • The compensator of (6. with the least squares identifier of (6.4). the update sequence { l.) = [o ··· 011 Let the transformation matrix M( l. . Elliott & Wolovich [ 1982].) and f(t.276 Advanced Topics in Identification and Control Chapter 6 Section 6. See also the review/survey papers by Dugard & Dion [1985] and Elliott & Wolovich [1984].35) Therefore Ap ( t. ) .) ApM( l. 4n points. Johansson [ 1987]).2.design of adaptive schemes will follow in a more straightforward manner from SISO theory (Section 6. the estimate of Ap based on the current plant parameter estimate.) cJ (t.4.) Then the adaptive system is asymptotically stable.31).2. Alternate adaptive control schemes may be found in the . Goodwin & Long [ 1980]. Wolovich & Das [ 1984]).2.5 follows as the proof of theorem 6. we briefly review some basic results. we will concentrate our discussion on a model reference adaptive control scheme and follow an approach parallel to that of Chapter 3. ).2.) .19)(6. ) · · · an( l.bp/T ( l.37). (6.) I 0 = bp I f3n(ti) = [.) cJ<td Ap <t.)M(t. and quadratic optimization approaches (Borisson [ 1979].32) with Q(s) = 0 can be made adaptive by choosing C(t.36).3 MULTIVARIABLE ADAPTIVE CONTROL a 11 ( l.) f3n(t.) bp( l. p11 +f3 11 (t. Once the parameterization is established.5.2.3. Indeed.34) the input is stationary and its spectral support is not concentrated on k ::.2. 6.( l. Now f( t. The understanding of parameterization issues benefited significantly from progress in the theory of nonadaptive MIMO control theory. Koivo [ 1980]).2.) with A I A 6. (6.21). (6.33) to (6. ) ] Theorem 6.) I cJ(t.. I ( t. as before we have the following theorem: 0 0 1 0 0 Ap( I.2.33) [a 1( l. Proof of Theorem 6. the pole placement approach (Prager & Wellstead [ 1981 ].2.36) (6.) are easily read off.2.)f M.. Callier & Desoer [1984].).) and 0 0 0 {31(!.1 Introduction The extension of adaptive control algorithms for singleinput singleoutput systems (SISO) to multiinput multioutput systems (MIMO) is far from trivial.) I ( l..3.2. and the compensator of (6. Several MIMO adaptive control algorithms were proposed based on the model reference approach (Singh & Narendra [ 1982]. More details may be found in Kai1ath [1980]. Pn+!3n(l.) relate the realization (6.
6. NL) is called a left fraction (!. It may be shown that these definitions correspond to the similar definitions for a minimal statespace realization of a proper hs1. DR) are unimodular. ~ = NRDRa pair ~ ~ I (DL.2 Preliminaries 6. Indeed. Extracting the gcrd R of a right fraction (NR 1. NR. DR) are called right coprime if their gcrd is unimodular. a unimodular matrix is nonsingular.1. Greatest common left divisors are defined by making the appropriate transpositions. • NR. A necessary and sufficient condition is that detDR(s) is a real number different from 0.1 Factorization of Transfer Function Matrices Right and Left Fractions Consider a square transfer function matrix Pes).f) of Pes) E JRP XP(s) if • NL.278 Advanced Topics in Identification and Control Chapter 6 Section 6. i. DL e JRP xp[s] • DL nonsingular. It may be shown that for all P (s) e JRP x P(s ). Given (DL. Given a pair (NR. except for a constant real factor. Two matrices R 1 and R 2 are said to be right equivalent if there exists a unimodular matrix R such that R 2 = R 1R.f) of P(s) e JRPXP(s) if Similar definitions follow for a left fraction (DL. Right and Left Coprime FractionsPoles and Zeros A polynomial matrix DR(s) is called unimodular if its inverse is a polynomial matrix. it may be shown that all gcrd's are so related. The set of polynomial matrices of dimension p x p is denoted IIV xp [s] (note the slight difference in notation). there exists a right coprime fraction of P (s ). but all gcrd's are equivalent in a sense to be defined hereafter. i.2. that is. another right fraction may be obtained by multiplying NR and DR on the right by a nonsingular polynomial matrix R (s) e JRP x P[s ]. n =order of the system = adet(DR(s)). The new pair (NR = NR 1 R. i. DR 1 ) is pretty much like extracting common factors in a scalar transfer function. Further. Similar definitions are found for left coprime fractions. two matrices (NR. Clearly. The set of such matrices is denoted P(s) e IIV x P(s ).e. of P R nonsingular The matrix R is called a common right divisor of NR 1 and DR 1 • It is called the greatest common right divisor (gcrd) of NR 1 and DR 1.f. A rational transfer function is expressed as the ratio of two polynomials in s. DR) is called a right fraction (r. of P ~ NR 1 = NRR r.1• Note that the gcrd's of (NR. NL) a left fraction of P(s) and a nonsingular matrix i.3. multiplying a polynomial matrix by a unimodular matrix does not affect its rank or the degree of its determinant. Similarly.3 Multivariable Adaptive Control 279 references just mentioned. that is.1 ) is also a right fraction so that P = NR 1 D£1 1 = NRDR. det(DR(s))#O for almost all s. all gcrd's are right equivalent. this topic is not well understood so far (even less than for SISO systems!). DR) is a right coprime fraction of P. the poles and zeros of P are defined as • pis a pole of P if det(DR(P)) = 0 • z is a zero of P ifdet(NR(z)) = 0 where (NR. whose elements are rational functions of s with real coefficients. DR e JRP xp[s] right coprime • DR nonsingular ~ = DL.e. with p rows and p columns. det (DL(s )) # 0 for almost all s • p ~ • NR. Similarly. DR).e. • p Similarly. We will say very little about the dynamic properties of MIMO adaptive control systems. iNL) is also a left fraction of Pes). A pair of polynomial matrices (NR.3. In fact. In analogy to the SISO case. (LDL. .NL ~ I ~ • P = NRDi ~ ~ I Given a right fraction (NR. a rational transfer function matrix may be represented as the ratio of two polynomial matrices..f. if any other common right divisor of NR 1 and DR 1 is also a common right divisor of R. DR = DR 1 R. In general. DR e IIV xp[s] • DR nonsingular. DR r. DR) and a gcrd R" the matrix R 2 = R 1R is also a gcrd if R is unimodular. It follows that ref and lcf of a plant P must satisfy det(DR(s)) = det(DL(s)) and det(NR(s)) = det(NL{s)). "the" gcrd is not unique. NL). In fact.
and the second step is to reduce the matrix to a column reduced form. DR) is a right coprime fraction of P(s). DR) is a right fraction of P(s) and DR is column reduced. DR) such that DR is column reduced. Properness.Lj) := J controllability index of P Polynomial Matrix Division ~ j =I J. If (DL.i(DL) = E where max(a(DL)u) are the row degrees of DL.i(D L) : = I If (NR.(DR) IR~. Note also that p Consider. (NR. defined as follows: correspond to the alternate definition for a minimal statespace realization. NL) is a left fraction of P(s ). J The converse is true if we introduce the concept of column reduced matrices. a.l n = system order max (J. DR}.lj} is invariant (see Kailath [ 1980]). Given two polynomials fi and standard division algorithm provides q and f such that d.j(DL) ~ coefficient of the term of degree acJ(DR) in (DR)iJ P(s) P(s) E IRgxp(s) IRg. we define observability indices of P v = max (vJ := observability index of P vi a.lJ = qd+r af <ad It may also be shown that the definition of controllability indices This procedure is equivalent to separating the strictly proper and not strictly proper parts of a transfer function: . Row Reduced Matrices Similar facts and definitions hold for left fractions and are briefly summarized.(DR) ~ fi(s) acj(NR) < acj(DR) ~ fi(s) E E ~ E IRVP(s) IRg.xt(s) A matrix is called column reduced (also column proper) if Dhc is non singular.i(NL) :=:.f. In other words. then a. The procedure is somewhat lengthy and is discussed in Kailath [ 1980]. first.lj It is a remarkable fact that { J. multiplying further on the right by some appropriately chosen unimodular matrix. First define the highest column degree coefficient matrix Dhc (Dhc)iJ = The matrix DL is called row reduced if Dhr is nonsingular.3 Multi variable Adaptive Control 281 Properness.(DR) coefficient of the term of degree a.x/(s) ~ ac/NR) < a. and DR is column reduced. the n = order of the system = ~ J~l fz J. If CNR.i(NL) < a. the controllability indices are the same (modulo permutations) for all r. ac. then a. with DL row reduced.x/(s). and DL is row reduced. it is always possible to find a ref (NR..x/(s) When (DL. we call The set of observability indices is invariant. The highest row degree coefficient matrix Dhr is defined as: (Dh.j(DL) a. and define the column degrees as ac/DR) = max[a(DR)u] I The following fact is easily established: fi{s) fi{s) E IRg XP(s) ~ acJ(NR) :=:. ac.)iJ = Pes) Pes) E IRg x P(s) if if lim P (s) exists lim hs) = E IRg. DR) with DR column reduced.x/(s) 0 Consider a right fraction (NR.i(DL) in (DdiJ a. Given a matrix fi{s) E IRgxp(s} or P{s) E IRg.j(NR) :=:.280 Advanced Topics in Identification and Control Chapter 6 Section 6. They are different from the controllability indices. Column Reduced Matrices We restrict our attention to proper and strictly proper transfer function matrices.c. although related by p controllability indices of P J. The first step is to obtain a right coprime fraction (using a Hermite row form decomposition}. NL) is a left coprime fraction of fi{s). scalar polynomials.
3. The concept of an interactor matrix follows in a natural way. Since SRDR = NR. (s) UJI(s) uJZ(s) 0 0 0 fact for left fractions. RR.. Therefore. The . rettueedo. .QRDR. 0 (s) is equivalent to the knowledge of a monic polynomial € such that (s) lim up 1(s) • any polynomial uiJ(s) is divisible by s (or is zero) The matrix Kp is called the high frequency gain of the plant Pes).1 The elements of NRDR. The integers r.3.(DL) is the Proof of Proposition 6.3 Multivariable Adaptive Control 283 d d d From this observation. The matrix i. The matrix €(s) is not unique. A AA Let RR = SRDR. E IIV x P[s] such that We must assume here that P (s) is itself nonsingular. where i.t There exists {2R.2 Interactor Matrix and Hermite Form Interactor Matrix In Chapter 3. Pes) is a matrix whose €(s) Pes) = kp "* 0 Then. Note that the knowledge of the relative degree of a scalar transfer function Pes) E IRp. AAI A A A €(s)P(s) Us) Kp nonsingular 1 S+00 i:(s)~(s) Therefore NRDR.1 Polynomial Matrix Division E q polynomial . Let ~ 9& ee rev. leading to an equivalent definition. where the gauge of an is its relative degree. are rational functions of s. we observed that SISO model reference adaptive control requires the knowledge of the relative degree of the transfer function Pes).P x '{§~such that lim IR~xP(s) 8c/RR) < 8cJ(DR) a. The highfrequency behavior of MIMO systems is similarly determined by a polynomial matrix such that lim S+00 €(s) P(s) = Kp nonsingular Let NR. DR.e. C strictly proper scalar kp was earlier called the highfrequency gain of the plant P.(RL) < a. unless its structure is somewhat restricted. SR RRDR. the column degrees of RR must be strictly less than those of DR.A Chapter 6 Section 6..AEicolpmn redutpad. QL. by taking the following approach.2...282 Advanced Topics in Identification and Control fi f . Ih ~~oti Then i. extend the notion of the relative degree r of an SISO transfer function.=q+. Proposition 6. Hermite Another [ 1976]. one notes that proper rational functions of s form a division algorithm may also be defined. the relative degree of Pes) is equal to the degree of €(s). DL nonsingular. SR being strictly proper a. RR is a polynomial matrix. NL. which becomes 1 in the SISO case. RL QRDR + RR DLQL + RL IIV xp[s] with DR.x/(s). is found in Morse For this. Divide each numerator by its denominator and call the matrix of quotients QR. A similar proof establishes the = • 6 (s ) diag (s'. A element Normal Form approach. The extension of the concept of relative degree to transfer function matrices is not trivial and must take into account the highfrequency interactions between different inputs and outputs. Further. the following proposition is obtained in the multivariable case. It is shown in Wolovich & Falb [ 1976] that there exists a unique matrix €(s) satisfying the following conditions. Definition Interactor Matrix The interactor matrix of a nonsingular plant P(s) E unique matrix € E IR. A A A I = QR +SR where SR E IR~.3. describes the highfrequency interconnections between different inputs and outputs. D 6. (s ). ring. ) 1 0 At 0 0 u21 (s) • i.
sii 32 (s) (s +a)'' hzl(s) + a )' 2 1 u32 (s) so that (s +a )'P (s (s + a )' 2 H i. decompose P(s) = if(s)·U(s) i where H is the Hermite column form of P. lim €P = Ku nonsingular.2 Equivalence Interactor Matrix and E Hermite Form and Model Reference Control The significance of the Hermite normal form in model reference adaptive control may be understood from the following discussion. Since r) is unimodular.3.. its inverse is a proper rational matrix. such that the elements below the diagonal are either zero. (s) also satisfies the required conditions.CFBP). The matrix U is unimodular in IR~ x P(s ). so defined satisfies the required properties.. H 0 0 0 0 0 0 0 s.3. Next.2 We let € = if· 1 and show that it satisfies the conditions in the definition of the interactor matrix. note that P = if. D where aiiu(s) ~ ri . 3. or have relative degree strictly less than the diagonal element on the same row.feedforward controller and CFB is a feedback controller. Proposition 6. so that the closedloop transfer function is given by (I.3.t ci. and may be factored as Proof of Proposition 6.h31Cs)h3 1(s)) . The Hermite column form of i/ is a lower triangular matrix. one could 3_li. First. The offdiagonal terms satisfy P(s) = if(s)·U(s) U unimodular in 0 IRg xP(s) u21 (s) u31 (s) Shzl(s) s(h31(s).'2 0 sii 21 (s) s rp shp 1(s) U(s) = Ku nonsingular uniquely de~~e the Hermite normal form as follows Definition Hermite Normal Form Morse [ 1971f showed that.PCFF (u) ~ ~ Hermite Normal Form 1 ~ ~ Let €Cs) be the interactor matrix of P(s) Hermite normal form of Pes) for a = 0.I and a is arbitrary. As shown in the following proposition. IR~ xP(s ).1P = r).)1 The Hermite norma/form of a nonsingular plant P(s) unique matrix if E E IRgxP(s) is the IRg xP(s) such that Clearly. and i.3 Multivariable Adaptive Control 285 elements belong to a principal ideal domain. An arbitrary linear timeinvariant (LTI) controller may be represented as u = CFF(r) + CFB(Yp) where CFF is a. is a lower triangular matrix with I 's on the diagonal. but fixed a priori. with some slight modifications. The unimodularity of U is equivalent to lim S+QO s.PCFB). Let if (s) be the P(I.CFF(u) ~ ~ ~ 1 ~ Then €Cs) = if· 1(s) . the interactor matrix and the Hermite normal form are completely equivalent.284 Advanced Topics in Identification and Control Chapter 6 Section 6. that is.
the assumptions required for multiinput multioutput (MIMO) model reference adaptive control will look fairly similar to the assumptions in the SISO case.6 It is defined by MIMO Controller Structure Mo(r) M (A3) = HM 0 e IR~.. consider a right coprime fraction (IVR.1(s) C(s)(u) + A. Then.3. Therefore (6. II C).x/(s) (A2) where ii is a stable Hermite normal form of P. the transfer function U(J. Combining with the expression of the controller Controller Structure First. the model reference adaptive control problem is replaced by a problem where the reference model M is equal to the Hermite normal form of the plant.2) The controller is represented in Figure 6. let ac : : .3. Let ~ (s) be an arbitrary. P = fi{. Reference Input Assumptions The reference input r( ·) e IRm is piecewise continuous and bounded on IR +· a.2. the reference model must be the product of the plant's Hermite form times an (arbitrary) proper transfer function. Reference Model Assumptions The reference model is described by Figure 6. E IR~. and restrict the controller to I be proper. and A. Assumptions (Al) Plant Assumptions The plant is a strictly proper MIMO LTI system.. nonsingular. ii is assumed known.6 (s) e JRP x P(s ). 6.1(s)D(s)(yp) (6. In other words. denotes the maximum degree of all elements of Now.CFF ~ ~ ~ ~ Assume now that the plant is strictly proper.3 Model Reference Adaptive ControlController Structure With the foregoing preliminaries.CFBP). stable transfer function matrix and ii is the Hermite normal form of the plant.I c = JRP x P.1.4) where x is a pseudostate of P. vp). transfer function matrix In this manner.3.D : : .CFF is proper.3 Multivariable Adaptive Control 287 = HU (I . The controller structure used for multivariable adaptive control is similar to the SISO structure.CFBP). and the observability index v is known (an upper bound on the order of the S) >tern is.286 Advanced Topics in Identification and Control ~~ ~ ~ I~ Chapter 6 Section 6. For SISO systems. . II. note that all the dynamics of M may be realized by prefiltering 0 the reference input r.6. Define A(s) e JRpxP(s) such that A(s) diag [A. The plant is minimum phase. (s )] (6. therefore.I D = . Therefore. we define . this is equivalent to saying that a proper compensator cannot reduce the relative degree of a strictly proper plant. A' A. c A I' C(s ).3) where M0 is a proper.?(s) u where C 0 e C 0r +A . provided that adequate transpositions are made.CFF (u) so that r Ym = Mo(r) if The transfer function is equal to the reference model transfer function if ~ ~~ ~ ~ I~ = H(r) (6. monic..3. described by a square. Hurwitz polynomial of degree vI (where vis the observability index of?). I (where ac By the foregoing choice of Further.1) M = HU (I.3. obtained by setting a > 0.CFBP).6 A. DR) of P.
Then  1 Consequently.5) Section 6.. [)• As in the SISO case.· '· ~ ..c. Let (DL.. NL) be a left coprime fraction of P.11) .Advanced 1 vp1cs 1n ldenll}Lcatwn and Control Chapter 6 (6.: A s + c 2. 7.W 12)T . Multiply both sides by and A 1 on the left. .. Co' . I (6.17) Di 1 on The regressor vector w is defined as the right WT ·.Wf . Since ac.3. r to Yp is H if and only if the following H StateSpace Representation A statespace representation is obtained by defining the matrices C 1 .3..\ . reminiscent of the matching equality of Chapter 3 (6. [)• such that the transfer function from r. v. the output Yp is given by A A A A DN R] x Ac 0 r lA  Yp = NR[(A C)DR..::.kn. . .2 (6.2 IA A (6. Divide AKplfj I on the right by DL. c·.14) Since DL NR = NL DR' it is easy to show that the given solve the matching equality. • 1c wr.+ .I wf2l := A(Yp) i = l .::..T (l)T (6.:.3 The transfer function from equality is satisfied c I I . . o ••• Co.is a polynomial matrix. . aiJ. so that The controller structure is represented in Figure 6.16) Since if. s". we determine C 0.9) +DvI_A_ A s". . On the other hand [( A .3. D.. c·.3. We define the parameter error toa.e. 7) (6. with DL iliiN U c·.8) Co. = 9T W E JRP (6. let ADRx A C0 r + CDRx A A + DNRx = (6.3. .._ 1 E JRP x P such that Proof of Proposition 6._ I ' Do' ._ I vary with time..c.I tli. ".vr\.3 Mu/tivariab/e Adaptive Control AAIAI 289 (6. &viOD~&.18) and the matrix of controller parameters is (6.~rt .3.3.13) AKp H = QLDL + RL A A A so that A where aci(RL) < aci(DL) . v.Yp is H. are obtained almost as in the SISO case..c )DR Therefore.ACo(r) AA (6. ..12) The polynomial matrices D. ' D.~ON "'~w~s tx\s~s. .. + A c..3. ' c. .3.. (I)T T Wvl>Yp. Do.3 There exist MIMO Matching Equality S+00 S+00 so that A 1 c· is strictly proper and ac· .w lv PvT c/rNfJ~•~e. ( cF &f&....3.3. C _1 .o ( ltl(" A 11 ~ ~ fL r.(r . = a(RL) . Then. the scheme will be made adaptive.DNR]. Proposition 6.3..r~. By letting the controller parameter e.: .3.19) so that the control input is given by Taking the limit as s ..3.oo (6.3.2..(DL) . v1 A 5 . i..6) (6. . this leads us to a proposition guaranteeing that there exists a nominal controller of prescribed degree such that the closedloop transfer function matches the reference model (F/) transfer function.3. the vectors wfll and wf2l are defined by wfll := ~(u) A i.~ •.20) . .::. the foregoing equality may be transformed into a polynomial matrix equality. D 1 . .._\.l.1 "'ovu.10) First. Wvl) (2)T E JRPX 2 •P(6.3..
7 <I> ( MIMO Controller StructureAdaptive Form t) 8 (f) .v1 caciii).3. .26) where f (s) is a monic.29) ~Tv where the last equality follows because we assumed that Kp is known. they commute with any matrix. Multiplying both sides 1 of (6.3.30) This equation may be obtained by considering each component of e 2 .~:tl) ((Hi ) I yJ' i.25) Now. · · ·. For example.3. Dv1) T (2)7 vT VT e2 • r WJ (1)7 . ~T E JRPx(lvllP.27) As in the SISO case. ( and EV := W Cvh Do. WvI•Yp'Wt (1)7 . the normalized gradient algorithm for (6.3. we make the following assumption (A4) High Frequency Gain Assumption The highfrequency gain matrix Kp is known.3 Multivariable Adaptive Control 291 (6. .. D1. standard update laws are easily extended to the multivariable case. we define (CI.1 (u)+D*A. Since Aand i are given by (6.and applying both transfer function matrices to u leads to i=1.27) e2 6. with e 2 E JRP. Hurwitz polynomial such that af = a(If .22) (6. .1 {yp) + (6.1( w) (6.3.23) and we look for an update law for 8 (t ). .1cc·A. .21) C~ (Hi).3.3. Consider the matching equality e =  g vef 1+"(VTv g. Consequently.24) and multiply both sides by A. with similar properties. .3.. we let Tv (6.. .3.3.1 {yp) + Figure 6.3.1 on the left and Dii 1 on the right so that (6..1(Yp> + L.3.3.4 Model Reference Adaptive ControlInput Error Scheme For simplicity in the following derivations.25) by i . .1( G* 7 w) e•T i . using (6.3.3.e· (6.2).1 (yp)) C~ (Hi)..3.26). . w. define i(s) diag [i(s)] E JRPXP(s) (6. (6.I ( W T)) Ll(wT) 8Tvi. However..1(u) <I> e.28) so that. 'Y > o (6.290 Advanced Topics in Identification and Control Chapter 6 Section 6..1) (the maximum degree of all elements of fi. that is. The error equation is a linear equation.8* E JRP x 2vp i . Co = C~.29) becomes  Co CO = = Kp 1 u Car+eTw (6..1 ). but a multivariable one.
the remaining poles are all at the origin (d step ahead control). 6.292 Advanced Topics in Identification and Control Chapter 6 Section 6. In fact. More research will be needed for the dynamics of MIMO adaptive control systems to be wellunderstood. we presented flexible indirect adaptive ~~ntrol schemes. The model reference adaptive control objective with M = if is somewhat restrictive: Indeed. we established a parameterization of ~he adaptive controller. The combmatwn of modern control theories based on factorization ~pp. Convergence properties have not been established. and may be calculated without matrix polynomial inversion (see Wolovich [1984]). a more adequate scheme was presented such that the internal dynamics of the closedloop system are actually those of the reference model. We then turned to the extensiOn of model reference adaptive control schemes to MIMO systems. In continuous time. 6. as discussed in Chapter 3. In the SISO case. [1984]).3. unless if is diagonal (cf. An mput error scheme was finally presented.3. In discrete time.3 Multivariable Adaptive Control 293 forming p scalar error equations. This is not very desirable for implementation. This is to be contrasted with the situation for pole placement. More research will be needed to better u~de:stand the dynamics of MIMO adaptive control schemes. and the remaining poles are defined by if. all zeros are cancelled by poles. the offdiagonal elements depend on the unknown plant parameters. Applications to pole placement and ~he factonzatwn approach were discussed. and proved their global stability under a richness cond1t10n on the ex~ge~eous reference input. the requirement of the knowledge of the Hermite form may be too much to ask for. The scheme may also be extended to the unknown highfrequency gain case. all desired dynamics are generated by prefiltering the input signal. When if is not diagonal. some procedure must be devised to prevent C 0 from becoming singular (cf. Singh & Nar~ndra [1982]. An output error version of the model reference adaptive control scheme is found in Singh & Narendra [1984]. following lines parallel to the SISO case. . Similarly. and collecting the standard SISO update laws.3. Stability proofs for MIMO adaptive control follow similar paths as for SISO (Goodwin & Long [ 1980]. Singh & Narendra [ 1982]).5 Alternate Schemes The multivariable input error scheme presented here is equivalent to the scheme presented by Elliott & Wolovich [ 1982].roaches an~ of MIMO recursive identification algorithms in a flexible mdtrect adaptive controller scheme is a promising area for further developments and applications. the matrix polynomial division itself is simpler than it looks at first. Then. we ~rst ~iscussed how prior information may be used in t?e context of tdentlficatwn.3. as it was in the SISO case. ~fter some preliminaries.3. a similar scheme is obtained by Goodwin & Long [ 1980]. a leastsquares algorithm may also be defined. Indeed. A possible advantage of the indirect approach is that the interactor matrix may be estimated online (Elliott & Wolovich [ 1984]). It is fairly straight forward to derive an indirect scheme. where the general Diophantine equation needs to be solved. However. the uniqueness of the controller parameter is not guaranteed by proposition 6.. Indeed.4 CONCLUSIONS In this chapter. based on the solution of the matching equality given in the proof of proposition 6. c0 # 0 in SISO). An interesting contribution of the proposition is to show that the solution of the matching equality requires only one polynomial matrix division. In discrete time.
A number of applications of these techniques have been made. In this chapter. Krener. if there are errors in the model of the nonlinear terms. g respectively.1 Basic Theory SISO Case A large class of nonlinear control systems can be made to have linear inputoutput behavior through a choice of nonlinear state feedback control law. one obtains Y1 (7. Meyer & Cicolani [ 1980]. Isidori. h all smooth nonlinear functions. 1986] and Byrnes & Isidori [ 1984].1) In (7.n. gi: IR.n.L1h + v) g 1 (7. Differentiating y with respect to time.2) Lfh(x) + Lgh(x)u where L 1h (x) : IR n .2.2.3 along with its applicatiOn to the adaptive control of rigid robot manipulators. we collect some suggestions for future work. based on Craig. The remainder of the chapter is organized as follows: we give a brief review of linearization theory along with the concept of a ~inimu~ phase ?onlinear system in Section 7.2. 7. the cancellation is no longer exact and the inputoutput behavior no longer linear.1) f(x) + ~ gi (x) ui i = I smooth IR n. at first. such as Singh and Rugh [ 1972].3) yields the linear system (linear from the new input v toy) y = v The control law (7.2 LINEARIZING CONTROL FOR A CLASS OF NONLINEAR SYSTEMSA REVIEW 7.IR and Lgh (x) : IR n .1 INTRODUCTION In recent years there has been a great deal of interest in the use of state feedback to exactly linearize the inputoutput behavior of nonlinear control systems. Also the h/s are scalar valued functions from IR. In this chapter.4. we suggest the y = ah a f(x) x ah + g(x)u ax (7.1 Introduction 295 CHAPTER 7 ADAPTIVE CONTROL OF A CLASS OF NONLINEAR SYSTEMS use of parameter adaptive control to help make more robust the cancellation of the nonlinear terms when the uncertainty in the nonlinear terms is parametric.3) has the effect of rendering (n. Isidori [ 1985. x E IRn. Their chief drawback however seems to be in the fact that they rely on exact cancellation of nonlinear terms in order to get linear inputoutput behavior. The discrete time case and sampled data cases are more involved and are developed in Monaco & NormandCyrot [ 1986]. Consider. u E IRP. Hsu & Sastry [ 1987]. f. g. The theory of linearization by exact feedback was developed through the efforts of several researchers.1).IR. Good surveys are available in Claude [ 1986]. the singleinput singleoutput system f(x) + g(x) u 7.I) of the states of 294 . GoriGiorgi & Monaco [ 1981] in the continuous time case. In Section 7.1.2. If Lgh (x) is bounded away from zero for all x.IR stand for the Lie derivatives of h with respect to f. y E IRP and f.B(x)v) u = L h (.Section 7. a function will mean an infinitely differentiable function.2. for example of the form p y with x E h(x) (7. Freund [ 1975]. the state feedback control law (of the form u = a(x)+. Consequently. We discuss the adapt~ve versiOn of thts control strategy in Section 7.2.n to IR.1. The results of this chapter are based on Sastry & Isidori [ 1987].
As before.2.5) linearizes the system (7 . differentiate the j th output y1 with respect to time to get p Ljh(x) + (LgL1 h)(x)u (7. then the inputs do not appear in (7. i = I (7. .2. In the instance that Lgh (x) = 0 meaning Lgh (x) = 0 for all x.10) The procedure described above terminates at some finite 'Y if the row vectors { ~~ (x).1h is bounded away from zero.2 and LgL. fx r L 1 L L Kp h (x)} are linearly indepenThen.4) In (7. MIMO CaseStatic State Feedback For the multiinput multioutput case.)(x) = 0. if 'Y is the smallest integer such that Lg L}h = 0 for Y/ 1 = L/ 1 h1 +~Lg.2. u E JRP. = Llhi + ~ (Lg. i =I 1 p (7. fx L1h (x).h.. one differentiates (7 .1) unobservable through appropriate choice of state feedback.2.2. h1. but is equal to zero for some values of x. we consider square systems (that is systems with as many inputs as outputs) of the form + A(x) L 'Yph I P (7. the control law Y.2. (L/ 1 . are assumed smooth. that is.h.)u.h. Define 'Y 1 to be the smallest integer such that at least one of the inputs appears in y/ 1.2) further to get ji = (7. g.2.2.6) yields y'Y = L Kp LI ')' 1 .296 Adaptive Control of Nonlinear Systems Chapter 7 Section 7. .7) Here x E IRn.1h is not identically zero.11) x Y1 j(x)+g1(x)u1 + · · · + gp(x)up h1(x) If A (x) e JRP xp is bounded away from singularity (meaning that A ..2. the state feedback control law .) # 0. 1 More generally.1(x) exists for all x and bas bounded norm).2.2. Ljh(x) stands for LI(L1 h)(x) and LgLih(x) = Lg(Lih(x)).8). Note that the theory is considerably more complicated and incomplete if Lg L .(L/J. if Lg L1 h (x) is bounded away from zero for all x. 'Y .4) to yield In (7.8).8) u = 1 (Ljh(x)+v) LgL1 h(x) ji = v (7. Now. for some x. .2 Linearizing ControlA Review 297 (7 .2... then the control law i = with at least one of the Lg.. . (7.4). . y E JRP and/..2. Define the p x p matrix A (x) as (7. note that if each of the (Lg.2.9) 0.1 h I A (x) v (7.h}) U.2.9) may be written as I 'Ypl h p dent for all x.
Now. we get LI '"~"hP yields the linear closed loop system [ (7.2. w 1 )C(x. Thus.1(x.2. A(x) T(x) = The two large blocks in the equation can be condensed to yield: [ Ylz . w 1 ) is bounded away from singularity. any further control objective such as model matching.2. l (7. pole placement or tracking can be easily met.7) now reads as .2.2.2.(x. decoupling is achieved as a byproduct of linearization.2. for example.11) reads al1(x)l [ a21 (x) WI + (7. If B (x.2.298 Adaptive Control of Nonlinear Systems Chapter 7 Section 7.15). Now. The procedure ends in finitely many steps if and only if the system is right invertible (for details. The feedback law (7. i.2.2. linearization may still be achieved by using dynamic state feedback. Thus.2. once linearization has been achieved. linearizing. we review the methods in the case when p = 2 (two inputs.13) y 1'"~ 1 + 1 Y2 'Yz +1 l = + a11 ~ a21 L 82 L/h1 + Litza 11 w 1 ] Liz L1· 2 + L gp21W1 '"~'h ~ [w 1] Wz Note that the system of (7.2 where Linearizing ControlA Review 299 u A (x) 1 + A (X). we may compress A (x) to one column. 7. w 1 ) 1 r~~] (7.3) when applied to the system (7.2. defining the new inputs w = T. MIMO CaseDynamic State Feedback If A (x) as defined in (7. and using (7. note that if B(x.17) yields the linearized system [~ 21(x) ll(x) a 00 ] [~:::::l = [ :. then the control law [:~] = B.1) of the original n states are rendered unobservable by the state feedback.16) Note the appearance of the control term w • Specifying w is 1 1 equivalent to the placement of an integrator before w I> that is to the addition of dynamics to the controller. w 1 ) is singular.14) The control law (7 .e:. (7. the first term on the righthand side) is not in the range of A (x ).14).2.12) [gl(x) gz(x)] = [gl(x) gAx)] T(x) Differentiating (7. (7.2.2. see Descusse & Moog [ 198_5]). w 1) + B(x.2 Minimum Phase Nonlinear Systems The linearizing control law (7. To keep the notation from proliferating. w 1) [w Wz 1 ] (7. (n. w 1 ).2.15) . two outputs).e.2.18) with T(x) E IR 2 x 2 a nonsingular matrix.17) is a dynamic state feedback. and decoupling control law.11) (i. A happy consequence of this is that a large number of SISO results are easily extended to this class of MIMO systems.12) is referred to as a static state feedback linearizing control law. Consequently.10) is singular.v? 1 1 + ] +1 = C(x. A(x) then has rank 1 for all x.1(x) u.1) results in a first order input output system. Using elementary column operations. w 1 )+B.2. the foregoing procedure may be repeated on B (x.1 v (7. and the drift term in (7.2.X = f(x) + g1(x) w1 + gz(X) Wz (7. To Also.13) is in addition decoupled.
i1 y Z2 /1 (Zt. The dynamics of the states rendered unobservable are indeed the socalled zerodynamics of the system (see (7. z2) + g1 (zt. Definition Strong Relative Degree The system (7 . i. consider the linear case. .e TAx + v) eTb (7. . z 1y)T and by z2 e IRn y the vector (Ty + 1.IRn . i.23)). the linearizing control laws may be thought of as being the nonlinear counterpart of this specific pole placement control law.1) is said to have strong relative degree 'Y if s (7. Thus.2.beT I e Tb) A are located at the zeros of the original system. . . Tn chosen such that a~ ( Ti (x) ] g(x) = 0 for i = 'Y + 1 .2..X = [/beT eTb lAx+ _jz_v eTb (7. . n is a diffeomorphism onto its image (see Isidori [ 1986]). 7. . b) In the instance that the system has strong relative degree 'Y. LgLj.2. Z2) . Then. the condition Lgh(x) # 0 and the control law of (7.2.( . f(x) = Ax.e.2. Clearly..2 Comments Linearizing ControlA Review 301 see this more clearly.1 h(x) is bounded away from zero. . there exists a neighborhood U o of x o such that the mapping u =  1. it is possible to verify that. the system be minimum phase. If we denote by z 1 e JRY the vector (z 11 .2.1 The SingleInput SingleOutput Case The first definition to be made is that of relative degree (or polezero excess). We start with the singleinput singleoutput case.20) defined as Z11 Z12 From the fact that the control law (7.21) and Ty + 1 . . .2.19) resulting in the closed loop system T: Uo.1) is said to have strong relative degree 'Y if the ?utput Y needs to be differentiated 'Y times before terms involving the mput appear.. it follows that the equations (7 . . for each x o e IR n. and h (x) = eT x.. it is important to have the closed loop polezero cancellation be stable. g(x) = b. This motivates the understanding and definitions of minimum phas. ..2. nonlinear systems. in order to have internal stability (and boundedness of the states). and the last at the origin. Tn)T.300 Adaptive Control of Nonlinear Systems Chapter 7 Section 7. .1) of the eigenvalues of the closed loop matrix (I .2.2.3) is *'> eTb # 0 a) The system (7.19) yields a transfer function of _!_ from v to y it follows that (n . z 2) u 'll(ZJ.1) may be replaced by z 11 Z12 and for all x.2.e.
Definition Minimum Phase The nonlinear system (7 .22) Section 7..e.2. have no finite escape time). in which case we call the system exponentially minimum phase. ..ax g(x) La = {x E uo I h(x) 1 g1(z1>z2) = · · · = L/. .e.302 Adaptive Control of Nonlinear Systems Chapter 7 (7.1h(x) x E L0 LFih Note that this is a vector field on La because ](x) is tangent to Lo.2..2.1h(x) = 0} and hereby defined the vector field ](x) = f(x) L/h(x) g(x) LgL/.23) are referred to as the zerodynamics. for example. If ..Y "(1) + · · · + a 1 ( Ym . z2) (7. .2. If x = 0 is an equilibrium point of the undriven system.2.1h(x) 0} can be made invariant by choosing u = [f 1(zl>z 2)+v) i~ the soc_all~d _Lie bracket of]. then the dynamics z2 = g(x).Y) (7. If the system IS globally minimum phase).2. 'l'.Ym (7. .1h(x) I ](x) ' as above '1'(0. n.23) only locally around any point X of IRn.2. ThiS IS Ill turn guaranteed by requiring that the vector fields in question be globally Lipschitz continuous.22) and the zerodynamics of (7.1g(x) are complete (i.1) through (7. f(O) = 0 and h (0) = 0 (without loss of generality). that is. . They have shown that a global version of the notion of zerodynamics is that of a dynamical system evolving on the smooth manifold of IRn 0 .2.22) are globally defined if and only if the vector fields represents LgLlin the new coordinates.2. b) The previous analysis identifies the normal form (7. Also note that the input does not directly influence the z 2 states. (7.(x) for i = 'Y + I . z 2) represents L 1 T.. Note that the subset adlg = = aga] ax f(x) . ad1 g iterated i times.2.2. Comments a) This definition may be strengthened to exponentially stable.24) The dynamics of(7. connected a~d the zerodynamics are globally asymptotically stable (I.23) are the dynamics on this subspace.2.25) is not implen:ented by differentiating y repeatedly but rather as a state feedback law smce La= (x E IRn:h(x) = Lfh(x) = ··· = Ll. g and ad} g = ad1 .26) It is important to note that the control law of (7. 1 h(x) f (zi> z 2) represents Ll h(x) and gl (zi> Zz) 1 ~o is.1) is said to be globally (locally) mrmmum phase if the zerodynamics are globally (locally) asymptotically stable. T?e utility of the definition of minimum phase zerodynamics anses In the context of tracking: if the control objective is for the output ~(t) to track a prespecified reference trajectory YmU ).2. The representation of the system (7. where g(x) while = g(x) LgL/. then the normal forms of (7. adf.22) is called a norma/form. Recent work of Byrnes & Isidori [ 1988] has identified necessary and sufficient conditions for the existence of a globally defined normal form.2 Linearizing ControlA Review 303 In the equations above. then the control mput v  y'Y + a 'Y ( Ym "(1 m . ad7g(x).(z 1 .25) results in the following equation for the tracking error e 0 := y.
.. () V1 t = dV1 .z2) ~ a3iz2l 2 .2.2.22) has continuous and bounded partial denvatlves m z 1 . To be concrete.2.2. there exists v 1 2) such that (z 2 a 1lz2l 2 ~ v1(z2) ~ a2lz2l for lz2i ~ k (7.30) is similar to (7.go to zero as t 1 tends to 00 .2.' z2) ~n (7 . it only remains to show that z2 is bounded. it is sufficient to ask only that the zerodynamics of (7.27) Z1> By assumption.25) results in bounded trackmg.5.1 ).2. it is easy to establish that z 2 is bounded. .24)(7. This remark is especially useful in the adaptive context where the assumption of minimum phase zerodynamics may be replaced by exponential attracti vi ty.. b) The assumptions of proposition 7 .22) yields for all t i.:S: 0 for I z2l ~ [ a.1· T?e followmg proposition guarantees bounded tracking.z2 dz2 "' Zl>Z2 dz2 (7. Proof of Proposition 7 . if some of the eigenvalues of d '11(0. .1 are bounded Then z2 is bounded for small enough z 1> that is.24)(7. In fact. the hypothesis of proposition 7.1 is a global propositiOn.z 2 andym. c) However. If the zerodynamics were only locally exponentially stable.Y.1 call for a strong form of stabilityexponential stability.+···+ a1 is a Hurwitz 1 · r 1 polynomial.1 are also bounded and ~o is z.1 can be repeated to yield bounded tracking.30) dv 1 2 d 'lt(O. . Proposition 7. .2.304 Adaptive Control of Nonlinear Systems 1 Chapter 7 Section 7.2 that Linearizing ControlA Review 305 Ifa . Fu~her.27) for the system (7. eo . the proposition would yield that Bounded Tracking in Minimum Phase Nonlinear Sys the zerodynamics of the nonlinear system (7 .2. the control law (7. counterexamples to the proposition exist if the zerodynamics are not exponentially stablefor example.1 ( 'It is globally Lipschitz since it has bounded partial derivatives). . It is then easy to verify that all trajectories of the undriven zerodynamics (7. .1 From the foregoing discussion.are bounded.27) for the Lyapunov function I z 2 12.2. y~.y..2.1 (as we did for theorem 5.2. except that it holds outside a ball of radius k. depending on which definition of the zeros of an L TI system one chooses to generalize a) the dynamics of the maximal controlled invariant manifold in the kernel of the output map. that is. (7. z 2 ) I dz 2 evaluated at z 2 = 0 lie on the jw axis.2. D Comments a) Proposition 7. 7. · · .2..23) are globally exponentially stable.2. yJ1.2. Y '. ~ (z 1.:l) Using this along with the bounds in (7.2. y.2.e.' then y.1 can be weakened considerably without affecting the conclusion. for small enough Ym. ·.25) yields bounded 1z 1 {t)l ~ k Using (7.23) eventually converge to a ball and that the proof of proposition 7. Ym. eo .1) as defined ~n (7 . Further.( dv1 )) ) < a 3 1z 2 12 + ('lt(zl>z 2 )'1t(O.2 The MultiInput MultiOutput Case Definitions of zerodynamics for the square multiinput multioutput case are more subtle..2. a'Y are chosen so that s"~+a'Ys'Y.2. the control law (7.2.2. .23) is (globally) exponentially stable and 'It has bounded derivatives. Condition (7. x E IR n is bounded and y(t) converges to Ym (t )..2. It iS now easy to see .23) converge asymptotically to a bounded set (a form of exponential attractivity). In particular.28) for some k (large).2.2.2. There are three different ways of defining them. Since (7.29) where 1 is the Lipschitz constant of 'lt(zl>z2) in z.Ym· . then it is easy to see t h at eo. if Ym• Ym. or .2.2. y"~. that iS trackmg with bounded states.3.1 terns If v1 . We accomplish this by using the converse theorem of Lyapunov of theorem 1. the exponential minimum phase hypothesis can be replaced by the condition zi'It(O..z 2) ~ a3lz2l Z2 I I dv1 dz2 . as pointed out in Isidori & Moog [ 1987].27).:S: a41 z21 (7.
It is also pointed out that the three different definitions co. hp.1) read as Then.2...1 h I> h 2· .33) IR n Z21 m by T1(x).1 0) is nonsingular. We do not discuss it here since we will not use it. . . / 1(zl> z2) represents L/'h 1(x) and g 1 (z 1. .2.1 hP) Also. Lfhl . LJ' 'Y . then we proceed as follows.2. ·+'Yp Yp (7.2.306 b) c) Adaptive Control of Nonlinear Systems Chapter 7 Section 7. specified by (7. define z 2 E (7. In these coordinates. if A (x) as defined in (7 ... . . Let u be a linearizing control law.2. . the equations (7. for example I and z 1 E IR.1) is not decouplable by static state feedback. LJ'Y P . .2.incid~ if the nonlinear system can be decoupled by static state feedback. that is /(0) = 0 and h 1 (0) = = hp(O) = 0.2. . Proposition 7 .35) Then. the .32) m Above. L / y. the definition of the zerodynamics is considerably more involved.34). if 0 E IR n is an equilibrium point of the undriven system.. More specifically. m wh1ch case the definition parallels the development of the SISO case above. . For this class of systems the extension to model reference adaptive control is easy: for the singleinput singleoutput case.. Z2(nm) = Tnm(X) = with 2 T = (z(.34) are independent of the choice of linearizing feedback law.2.2.34) It is verified in Isidori & Moog (1987] that the dynamics of (7.1 and the remarks following it can be verified to hold with the hypothesis being on the zerodynamics of (7. or the dynamics of the inverse system. z 2) the first row of A (x) in the (z 1> z 2 ) coordinates. In the instance that the system (7..m by zl T = (hi.xm (7.I h 2> ... the zero dynamics are the dynamics of (7. Define 'Y1+ .2. consider Ym( t) to be the output of a linear time invariant reference model with input r( t ).. zf) representing a diffeomorphism of the state variables x.2. +I) 307 the output constrained dynamics (with output constrained to zero).3 Model Reference Control for Nonlinear Systems The discussion thus far has been restricted to tracking control of linearizable nonlinear systems. provided that the relative degree of the reference model is greater than or equal to the relative degree 'Y of the nonlinear system..2.2 Linearizing ControlA Revie11• Y2 zl (y.31) Y1 Ym = cT. The zerodynamics are defined as follows. Z1(y 1 +1) Z1(y 1 +2) 7.
In the adaptive control sequel to this section. . Ymz has relative degree 'Yz and so on. Also. . our estimates of the functions f and g are nl fe(x) M(s) 1 diag [ s "~I + a I"'is "~I. we will assume that h is known exactly but we will discuss how to relax this assumption later. nl i =I ~ e}1>(t)L1.3) ge(x) = ~ (7 . quently.3.3) is replaced by U = This is not unlike the results of Section 6.3 ADAPTIVE CONTROL OF LINEARIZABLE MINIMUM PHASE SYSTEMS In practic~l implementations of exactly linearizing control laws. . As above.(Ljh )e + 1 V] (7.e.1) with Lgh(x) bounded away from zero.. (Lgh )e representing the estimates of L1 h..I h [Ljh + Y~ + "11·~=0I ai+l [y~. the dimensions of the reference model and its parameters do not play much of a role except to generate Conse (Lgh )e [. the model error eo:= y. where linear multivariable plants are found to match their diagonal Hermite forms.308 Adaptive Control of Nonlinear Systems Chapter 7 Section 7.1) Ym CmXm g(x) = ~ e)z>'gJ(x) j =I (7. . At the outset.3.4). Relative Degree One Case Consider a nonlinear system of the form (7.yi] 7. n 2 are unknown parame.h (7. Note that the dimensions of the model play no role.2. We discuss the use of parameter adaptive control to get asymptotically exact cancellation. e)Z>(t) stand for the estimates of the parameters of!>'.6) .4) s"~P + · · · + Here the subscript e stands for estimate· and e}l>(t). (7. + a II ~ ep> ( t )fi(x) i =I n2 (7. 1 ters and fi(x). 1· = 1 .I + .2.. i.3..3.3.Ym satisfies the equation (7..2. n 1 .J. and the model of the form 7. . . the chief drawback IS that they are based on exact cancellation of nonlinear terms. we will consider the tracking scenarios for compactness.1 • . Further. .3. If there is any uncertainty in the knowledge of the nonlinear functions f a.2.3.2. 1 .2. with (Lfh )e.1 SingleInput SingleOutput. as we have noted above. e)Z>' respectively at time t.· e(z>' . At time t. This also relates to the fact that the tracking error e0 : = y .nd ~· the c?ncellation i~ not exact and the resulting inputoutput equation IS not hnear.38) apl j =I e)2>(t) g1 (x) (7.24)(7.2 Linearizing ControlA Review 309 control law (7. y.3.25) is easily modified to U = LgLj. Lgh respectively based on (7. the linearizing control law (7.2.3).3.36) For the multiinput multioutput case. ym.3. gj(x) are known functions.5) Ym. let f(x) and g(x) have the form nl f(x) = ~ Ol 1>'Ji(x) i =I n2 Xm = Amxm +Em r = (7.37) · where e(l>' .Ym satisfies M(s)e 0 where = 0 (7 . .2) with Em E IR nm X p' Cm E IRP X nm' the class of models that can be matched is that for which Yml has relative degree 'YI. .
3.7) ' 8 ).(Ljh)e + Ym + k sgn (YmY)) (7. 7) is bounded away from zero and Ym is bounded.14) Unfortunately.9) The Lyapunov function v(e 0 . and the exponentially minimum phase assumption (proposition 7. asymptotically converging to y.3. a 2 . o > 0 such that to+ o s +a which is of the form of the SPR error equation encountered in Chapter 2. E L 2 .2. we need wa continuous function of x (since (Lgh )c is Jounded away from zero)to be bounded. (7.3.3. Theorem 7.oo.3. and rjJ = 8 .3. after some calculatwn Then the gradient type parameter update law rfJ = .1 Adaptive Tracking Consider an exponentially minimum phase.3. Now note that given a bounded eo.2) yields.11) may be written t .3.11) a) The preceding theorem guarantees that e0 converges to zero as where w E IR n 1 + n 2 is defined to be the concatenation of w t. Ym bounded implies y bounded.1 ). ¢ converge exponentially to zero if w is persistently exciting.2.e parameter error.10) The control law for tracking is v = Ym + a(ymY) Comments and yields the following error equation relating the tracking error eo : = y _ Ym to the parameter error rfJ T = ( rP (l)T rP (2)T)T (7. 0 L/. Nothing whatsoever is guaranteed about parameter convergence. the condition (7 .11). From this. 8 (t) E IR n I+ nl the parameter estimate.5) in (7. Define the control law eo = I. it follows that x is bounded.1) are bounded. Proof of Theorem 7.3. Hence w is bounded and eo is uniformly continuous.3.310 Adaptive Control nl vi Nonlinear Systems Chapter 7 Section 7. ¢) = V2 e6 + 1 T ¢ is decreasing along the 12¢ trajectories of(7.3. r/!) = ae6 .13). i.: : .3.I h and ((Ljh)e. easy to see that both e 0 . with v(e 0 .2. with the assumptions on f.1).8) yields bounded y(t ). and so e0 tends to zero as t . then using (7.3. no. It is.3. ).3.oo.3.2. The following theorem may now be stated. since w is a complicated nonlinear function of X. and e0 e w<2l E 1Rn2 (7. if there exist a 1 . Equa· tion (7.3.2.12) (Lgh)e = ~ 8)2l(t)Lg)h j =I (!)' (7.( t ).14) is usually impossible to verify explicitly ahead of time. Therefore.(r/J T w) a 2I ~ J w w T dt to ~ a 1I (7.3.e0 w (7. all state variables of (7.15) .3.12) by the "sliding mode" control law U = (Lg!h)e {. gas given m (7. Further. b) One other popular way of dealing with parametric uncertainty is to replace the control law in (7.v) (Lgh )e eo and rfJ are bounded.: 0.3 U Adaptive Linearizing Control = 311 (7.3).13) y with = v + r/J(I)r wOl + ¢<2lr w<2l (7.8* th.3.1 ). w2. (2)' (Lg!h)e [(Ljh)e+Ym+a(ymY)] If we define 8* E IR nl +nl to be the nominal parameter vector (8 If (Lgh )e as defined in (7.e.1 (7. however. To establish that e is uniformly 0 continuous (to use Barbalat's lemmalemma 1. or alternately that 0 is bounded. (7.nline~r system of t~e form (7.
.1).1 h (.1h)e] +"·+a1(YmY) (7.18) (7. (LJ. .3. In fact.1 h)e  (L/ h)e (L 8 LJ. I ' J ' control law to be Implemented is V where . 8~2{ (t) are kept in a certain parameter range which guarantees that (L8 h )e is bounded away from zero.1 h (7. J ' For the purpose of tracking. } = o(ll8(1J 8(ll 8(1J 8oJ 8<1J 8(2l (1l (1) r2) 1 . I Y = L h (x ). 1 be the k(large!) dimensional vector of parameters 8( 1l 8(2) 1 . 81 81 8k . Let 8 1 .. 1 1 . large control activity and other undesirable behavior. 2.3.1 is that (Lgh)e be bounded away from zero for all x. Lth and (L1 h )e. and so on.1 h by their estimates.3.21) losophy is not at odds with adaptation as discussed in theorem 7. • o(ll 8(2l 8<1J 8(1) Thus. the yZ. 1 .3. the error e 0 goes to zero.7) may indeed go through zero. We define these as follows LJh + LgLF Ih (L g LJ. in which case the 8}ll8)1l and 8} 2l8)ll of (7.20) and (7. the gain k needs to be large. . c) An hypothesis of theorem 7. 7.3.. .3. h )8} 1l 8)ll 1 (7. we need to replace LJ h and LgLJ. .2). gj and the ¢/s above.3. it could be used quite gainfully when the parameter error ¢ ( t) is small.11) is then replaced by e + k sgn e = d( t ) (7. The development of the preceding section can easily be repeated if we define each of the parameter products to be a new parameter. + ay(yZ.3. For example. say by E (by modifying the update law (7. The nonadaptive linearizing control law is then of the form = and so on..312 Adaptive Control of Nonlinear Systems Chapter 7 Section 7.2 Extensions to Higher Relative Degree SISO Systems We first consider the extensions of the results of the previous section to SISO systems with relative degree y.3. E IRk . for example If 'Y = 3 8 contains 8(ll 8(2) ' ' l ' . . ) .(Ljh)e) Note that for 'Y :=:: 2.23) u = 1 L 8 LJ. .3. .2 h = 0 with L 8 LJ.22) The adaptive control law then is u = (7. resulting in unacceptable chatter. L}h .3. k '.3. even if the 'true' Lgh is bounded away from zero.8 representing the parameter error) Iff and g are not completely known but of the form (7. .1 y'Y.(LJ h )e + Ve) .. This may be bounded using bounds on fi.. 2.+ay[y. that is. .3.19) are not linear in the unknown parameters 8i.21) are parameters.3. if¢ ( t) is large. In the absence of precise information about Lth.13) as discussed in Chapter 2 and Chapter 3). However.17) This yields the error equation (with <I> : = 8 (t) .LJ h + v) (7.::.18). auxiliary techniques need to be used to guarantee that (L8 h )e is bounded away from zero. (7.20) where d( t) is a mismatch term (depending on the difference between Lgh and (Lgh )e...3.1 h bounded away from zero.1.Y are obtained as state feedback terms using the tracking law to be implemented is 1 y= L h (x ). It is then possible to see that if k > sup I d ( t) I . Lgh = L 8 L1h = · · · LgLJ. in fact in finite time.3..V (Ljh . This phil n2 i = n1 2. (7.3.1) +···+ a1(YmY) Y.3 Adaptive Linearizing Control n1 n1 313 The error equation (7. 1j = 1 L8 i(L11 h)8)2l8)1l (7. . Since (Lgh)e as defined by (7.LJ. in which the parameters 8~2 l (t).19) (.1 h)e L 8.3. = 1j =1 L 1i (LJ. (7.16) and (Ljh )e = i 2. Ve = YZ.3. One popular technique is the projection technique..3.
31) Using the fact that constants commute with i. if such a signal e 1 were measurable. able since ~ := i .3. {3 1.3. For the purposes of adaptation. several parameter update laws are immediately suggested. and the mismatch between the ideal tracking control v and the actual tracking control ve· For definiteness. we get (7. oOlo< 2 l].24) Adaptive Control Using an Augmented Error Motivated by the adaptive schemes of Chapter 3. since the terms in the regressor multiplying it are zero.3. An exception is a large class of electromechanical systems of which the robot m~ nipulator equations are a special case.3.3.30) I Lfih l where we used the hybrid notation of previous chapters and dropped the exponentially decaying initial condition terms.34) e0.3.25) e0 = L. q.1(s)(w) (7.32) e1 =eo+ [<~>Ti.3. respectively.32) which is used for the analysis.3. with L}h not explicitly available since they may not be known functions of x.1 (s)(w)i. we get L(s) s'~ + a y s'~ 1 + ···+ aI (7.35) and so on.T Wz (7.32). .31) can be obtained from available signals. Define the augmented error Note that w 1 and w2 can be added to get a regressor w.1(s ). eo'~.3. + {3 I (7. Indeed. In such systems.30) in (7.27) Equation (7. with eT = [0(1). . Lf2h .I + . When the L}h 's are not available.28) (7. It is of interest to note that 0(2) cannot be explicitly identified in this case..3. For convenience.24) may be written as wf w2T . we will denote strictly positive real.3... the basic tracking theorem would follow easily.1 (s)(<I>Tw)) e1 with the transfer function = /3yer]. (b) 'Y is a generic L 2 n L 00 function.I are not measur From the error equation (7.3. from the mismatch between the ideal and actual linearizing law.33). 0(2).26) is that eo.3.33) is a linear error equation.33) s'~+aysyl + · · · + a1 (7. .. The difficulty with constructing the signal in (7 . see Section 7.3.3.1 +···+{3 1eo /3yS y.3. we will use the following notation {3 is a generic Lz n L 00 function which goes to zero as too. . For example. consider the case that 'Y = 2 and Ill = n2 = I..3.[o [ (X 0 L}1h 0 Lg 1L11 h 0 0 (. (a) .24) of the error equation.3. Consider now the form (7.3 Adaptive Linearizing Control 315 law of the previous section with e 1 as given by (7.314 Adaptive Control of Nonlinear Systems Chapter 7 Section 7. we could use an error of the form e1 = eo+[eTi.1 (s)(w)i. oOloOl. the normalized gradient type algorithm: 0 .27) is strictly positive real and the adaptive As in the stability proofs of Chapter 3.3 below.26) Note that e 1 in (7.29) so that equation (7.26) yields the desired conclusion.1 (s)(0Tw)] (7.Ym (7. Also note that w is a function of x and also 0(t ).3.3.T WI + q. . Then..1(s)(<I>Tw) (7. we define the polynomial The two terms on the righthand side arise. the following approach may be used.3.3.(Ljh)e + Ve) (LgLfh )e (7. /3y may be chosen so that the transfer function in (7. Using (7.3.3. unlike (7. we have that (7.
7. Although the scheme is based on the output error e0 . i 2 since the transfer functions M. From the results of Chapter 2. and noting that x is a diffeomorphism of zbz 2 we see that .46) t .3. <I> T w The differential equation for z 1 = (y. a number of properties of 4>.42) Proof of Proposition 7.. e1 follow immediately.1 is stable (since i .3.3 (a) Bounds on the Error Augmentation Using the swapping lemma (lemma 3.6. f. Y. that is the truncated L 00 norm.1 = M makes it identical to the input error scheme.3.23) applied to a? expon~ntlal~y minimum phase nonlinear system with parameter uncertamty as given Ill (7 .3 Basic Tracking Theorem for SISO Systems with Relative Degree Greater than 1 . 8) has bounded derivatives in x. L2 n L 00 • I<I>Tf.. we get (7.3.3..3.3.3.37) Remark: The proof is similar to the proof of theorem 3..36) with the update law (7 . If Ym. Theorem 7 .45) <I> with ~ = = l + ~T~ (7. and skM(s) are all proper stable transfer functions.3.39) we get i.1 are bounded. . Ym'~.3. we have that (7. s'~.. Ljh.44) Using (7. y.44).3.2).46).II z21!t :::. with no assumptions on the boundedness of~· Proposition 7. we have (with notation borrowed from the lemma) (7. (7.1(s )(w) yields bounded tracking.1M are strictly proper.. (7.3.oo and x is bounded.3. .1 is strictly proper and stable. 00 and for all t (7.3..3.22)(7.(w)i. M(s) i :_ s'~ 1 J (<I> T w) + I Ym .3.e.3 Adaptive Linearizing Control 317 zll 1 refers to the norm~~~ I z(r)l.2 Properties of the Identifier Consider the error equation (7.3. Proof of Theorem 7 . Ym.3. 8.40) Using the fact that 4> E L 2 and i 6.43) (LgLJ. Combining (7..41) Then <I> E L 00 .5).46) can also be obtained for i 1 . y.316 (c) 11 Adaptive Control of Nonlinear Systems Chapter 7 Section 7.2.44) and (7.6. h. the choice i.4. Equations similar to (7.1 are bounded by hypothesis. Then the parameter update law .1h)e is bounded away from zero. LgLjh are Lipschitz continuous functions. Consider the control law of (7. kll wilt+ k (7.3. .1 in the linear case.44) in the exponentially minimum phase zerodynamics (7. and w(x..3. .38) Using lemma 3. .<l> E L 2 nL :$.e1 ~ Since <I> is bounded and Ym.1 )(7 .1 is). We are now ready to state and prove the main theorem..3.Ym as . .J. y '~ 1 f is Ym '~I J (7. g..3. .3.4 and the fact that we get 1 I <I>T Ht)l for some 'Y E ')'(! + ll~llt) ic. 11llwllt+l3 (b) Regularity ofw.2: See theorem 2.3. i.1(<I>Tw)l:::.
57) in (7.43) and (7.49) For <I>T w. one can establish that <I> T ~ 11 + II~ !It has bounded derivative and so is uniformly continuous.3.3. 4> E L 00 we get Using the estimate (7.(s )(<I> w) .14)).1w is regular by corollary 3. we have only assumed parameter uncertainty in f and g. but not in h. Oj and so on.48) Section 7.3..3.3 above ifw is persistently exciting in the usual sense (cf(7.6.3.54) for e 1.3.6).45) we get that II xllt so that ~ +k (7 .> ~ note that i.2) to e0 = L.3.3.60) Using (7. and <I>. we get II <I> Tw!!t Using (7.4.3. leo! !It + k = ~ let!+ /311 + /3 (7.3. <I> T w.1.49).42). By theorem 2.53) with (7. however.3.3. leo! Using (7.3.50) II wilt+ k (7.3. x.3. Oj to be close to oio1 . (7. It is important.0 as t . 0 Comments (7. that the number of parameters increases very rapidly with 'Y· eo (7. c) Thus far.61) But from (7. o.57) are bounded and (7.7. it may also not be a good idea to constrain the estimate of o.6. It is not hard to see that if h depends linearly on unknown parameters. we see from (7.3.3.oo.3.3.3.3.3 Adaptive Linearizing Control 319 and Using the hypotheses (7.1h depend linearly on some unknown parameters.3.6.56) ~ k lleollt + k eo!!t (7.3. Oj.61) that e0 goes to zero as t+ oo (as in proof of theorem 3. followed by (7.3.56) b) In several problems. = I eo! Since ~ ~ /3 lleollt + /3 + /3 II~ lit (7.54) where {:J Now 0 as t 00.52) Since fJ. d) Parameter convergence can be guaranteed in theorem 7.53).47) (7 . we find Thus w is regular ==.3.57) in (7. then we can mimic the aforementioned procedure quite easily.3.59) is related to w by stable filtering ll~llt ~ k !{_(<I>Tw) dt 4Jw + <I>Tw (7. This in turn c:1n be easily verified to yield bounded w. 8i £Jj in the transient period may not be Oi{t) Oj(t ).35) appears not to take into account Combining (7. we have ~ let!+ /3llwllt + /3 <I>Tw!!t T leo! ~ let!+ !JII • l + /3 (7.59) yields (7. Applying the BOBI lemma (lemma 3.51) I !{_<I> Tw!!t dt ~ kll wilt + k kll <I>T w!!t I eo! ~ fJ lleollt + /3 (7.3.3.51) yields the regularity of<I>Tw.58) and using (7.3.3.3.3. prior parameter information such as the initial existence of o. It is then clear that the development of the previous theorem can be carried through. it turns out that L/h and LgL/.3.3.55) Using (7.6.53) a) The parameter update law (7. Note however. to note that the best estimate of (c) Stability Proof From the regularity of~.48) we get II xllt ~ kll w!!t + k II x !It ~ k II wilt + k that II aw 1ax I and II aw I ao11 II w!It ~ k II w along with the established regularity of <I>T w.3. Since parameter convergence is not guaranteed in the proof of theorem 7.53). using lemma 3.3.318 Adaptive Control of Nonlinear Systems Chapter 7 (7.
66) depend on a complicated fashiOn on the unknown parameters o< 1l'.(q) (7. For the sake of tracking.71) This may be abbreviated as M(q) = ~ i =I Ol 2l' M. we will illustrate our theory on an important class of such systems which partially motivated the present work (see Craig.3.320 Adaptive Control of Nonlinear Systems Chapter 7 Section 7. but fortunately they depend linearly on unknown parameters such as payloads. v is chosen to be.2 With 'Y 1 = · · · = 'Y n = 2. The parameter update Jaw P .. Thus. Hsu.1(q) ~ M.62) .(qmq) (7. if A(x) and the h. and ij.2 in other cases) can be readily adapted.3. its dynamics may be described by an equation of the form (7.1 if v 1 = v2 = · · · =Vp = l. .(q) ij ¢)2l i =I (7.3. phase.'(q) C(q.68).(ll' C. (7. (q.nxn is the positive definite inertia matrix.3. C(q. In applications.63).3..72) where We IR.3 Adaptive Linearizing Control n2 321 (7 .69) is referred to as the compute torque scheme.69) anddthe overall control law (7.3. we choose not to discuss them here. gravity and friction terms. q.3.3.3. We sketch only a few of the details of the application relevant to our present context.3. it is easy to see how the linearizing. the schemes of the previous sections (those of Section 7.66) u +v (7. and (7.3.eo = W<P (7.q so that ni In (7.3. q) + M(q)v (7. frictional coefficients.63) in state space form with xT = (qT · T) andY.3. 6 5) 3 7.3. decoupling static state feedba1 k control law for minimum.68) that the quantities in equation (7.2. q) are not known exactly. then the linearizing control law is also the decoupling control law. M(q) e IR. we see that the system is decouplable in the sense of Se~t~on 7. n represents the joint angles of a rigid link robot manipulator.M. q) = u (7.68) is replaced by U = M(q)ij + C(q. o(2)' while the equation (7 3 68) depends on them linearly. so that ni Me.12) by C(q. the interested reader is referred to the paper referenced previously. . q) represent the Coriolis.67) LT LJ"hn while the decoupling control law is given by u = C(q.. Adaptive Control of MIMO Systems Decouplable by Static State Feedback From the preceding discussion.q) + Me(q)v (7. the law (7.1(q) ~ Cj(q. 70) Let eo = qm. M(q) and C(q. and Sastry [ 1987])the adaptive control of rigid link robot manipulators.nx(n 1 +n2) · fi Is a unction of q.3.3.3. Instead.3.n represents the control input to the joint motors (torques).64) eo+azeo+a. If q e IR.'s depend linearly on certain unknown parameters. and those of Section 7. Not~ v = ijm + az(iJmiJ) + a. q) ¢Yl j =I nz + Me. Therefore.63) Ce(q. and u e IR. q') Writing the equation (7. To make the scheme adaptive. q) = ~ j =I O. = q. and <P is the arameter error vector.3. The details are more notationally cumbersome than insightful. q) Recall that if A (x) is invertible. square systems can be made adaptiveby replacing the control law of (7.3.
eo+f3teo is chosen SO that (s+(3.2 is not necessary in this application since both y.4. Further. wt). We considered the class of continuous time systems decouplable by static state feedback. Hagander & Sternby [1984]). b) The notions of zerodynamics are not as yet completely developed. The extension to continuous time systems not decouplable by static state feedback is not as obvious for two reasons a) The different matrices involved in the development of the control laws in this case. Consequently the parametric dependence of the control law is complex.1 cannot be folio wed.resent chapter is only a first step in the development of a comprehensive theory of adaptive control for linearizable systems. even in the linear case. . however. b) While the "true" A (x) may have rank less than p.2. The discrete time and sampled data case are also not obvious for similar reasons: a) The nonadaptive theory. w 1 ) depend in extremely complicated fashion on the unknown parameters.1) is not linear in uk in the discrete time case and a formal series for (7. the P. its estimate Ae(x) during the course of adaptation may well be full rank. NormandCyrot & Stornelli [ 1986] is fairly complicated since (7. This can be shown to yield bounded tracking.4 CONCLUSIONS We have presented some initial results on the use of parameter adaptive control for obtaining asymptotically exact cancellation in linearizing control laws. T(x).4 Concluding Remarks 323 where el = positive real. 7.y are available as states so that the L 1h. Note that the system is minimum phasethere are in fact no zero dynamics at all.4. using prior parameter bounds. the zeros of a sampled system can be outside the unit disc even when the continuous time system is minimum phase and the sampling is fast enough (Astrom. C(x. Hsu et a/ [ 1987]). The error augmentation of Section 7. It is.'s do not have to be estimated.3.)/(s 2 +a2S+at) is strictly Thus.1) in uk needs to be obtained (and inverted!) for the linearization.73) Section 7. B(x. namely. as discussed in Monaco.3. As in other examples.322 Adaptive Control of Nonlinear Systems Chapter 7 (7. in which case the procedure of Section 7. it is important to keep Me(q) from becoming singular. unfortunate that the signal W is a function of qbut this is caused by the form of the equations and may be avoided by modifying the scheme as in the input error approach (cf.
. This is due in great part to the dual control aspect of adaptive systems: the reference input plays a role in determining the convergence and robustness by providing excitation to the identification loop. and since the work of Rohrs et a!. For the parameter convergence proofs.1 General Conclusions 325 CHAPTERS CONCLUSIONS 8. The operator relationships that were used for continuous time systems (Lp spaces) also hold true for discrete time systems (/p spaces). This highlights the separation between adaptation and control. While all the schemes were shown to be globally stable. A major goal of this book has been the presentation of a number of recent techniques for analyzing the stability.r the reference . t~e assumptions that went into the derivation ~f the schemes were.Section 8. we presented a. although the adaptive systems discussed in this book have identical stability properties in the ideal case. but was also underlying in the proofs of exponential convergence (Chapter 2). Averaging techniques have proved extremely useful and it is likely that important developments will still follow from their use. and an indirect scheme were denved m a umfied framework. much remains to be done to precisely define design methodologies for robust adaptive systems and in particular a better understanding of which algorithms are more robust. Closer connections will probably emerge from the application of averaging methods. and global stability (Chapter 3). This is a crucial consequence of the nonlinearity of the adaptive systems. and extracted frequencydomam con~:htwns. An outp~t er~or sch~me. many derivations may be simplified in the discrete time case because continuity conditions (such as the regularity of signals) are then automatically satisfied. we have attempted to give. This also had implications for the transient behavwr of the adaptive systems. First. For the study of robustness. While simulations are extremely valuable to illustrate a point. we exploited Lyapunov and a~eragmg methods. 8. For instance.nerahzed harmonic analysis. Besides these fundamental questions of analysis. For the stability proofs.2 Future Research Adaptive control is a very active area of research.model. and from thefrequencydomain results that they lead to. we used results. ~equence of lemmas drawn from the literature on inputoutput Lp stability . and there is a great deal more to be done.. we presented several schemes for the adaptive identlficatwn and control of linear time invariant systems. The area of robustness is essential to successful applications. . a unified treatment of cu~rent methodologies for the design and analysis of adaptive control algonthms. Methods for the analysis of adaptive systems were a focal point of this book. pa~am~ter con~ergence and robustness of the complicated nonlinear dynamics mherent m the adaptive algorithms. an input error scheme. they are useless to prove any global behavior of the adaptive algorithm. A major problem remains to quantify robustness for adaptive systems.1 General Conclusions In this book. there is . our presentation of the continuous time results may be transcribed to the discrete time case with not much difficulty. qmte different. at a fairly advanced ~evel of rigor. While we did not deal explicitly with discrete time systems. Current theory does not allow for the comparison of the robustness of different adaptive systems. It is interesting to note that the twotime scale approximation was not only fundamental to the application of averaging methods to convergence (Chapter 4) and to robustness (Chapter 5). that makes rigorous analysis techniques essential to progress in the area. We feel that a complete mastery of these techniques will lay the groundwork for future studies of adaptive systems. from ~e. and makes the connections between direct and indirect adaptive control more obvious. the input error adaptive control scheme did not require a strictly positive real (SPR) c~ndition fo. Indeed. it has been understood that the questions of robustness for adaptive systems are very different from the same questions for linear timeinvariant systems. In fact. As was observed in Chapter 5. and the relation to nonadaptive robustness concepts. algorithms that are stable for some inputs may be unstable for others.
and adaptive control of nonlinear systems. There is a simple instance of these results which is in some sense representative of the whole family: consider the problem of adaptively stabilizing a first order linear plant of relative degree I with unknown gain kp.2 Future Research 327 evidence that their behavior is drastically different in the presence of unmodeled dynamics. but it is still in its early development. the control law (8. system.by state feedback. the relative degree of the plant. Chief under the assumptions that could be relaxed was the one on the sign of the highfrequency gain.2) and do =  y} (8. and the minimum phase property of the plantit is interesting to ask if these assumptions are a minimal set of assumptions. While past successes of adaptive control have been on systems of rather low bandwidth and benign dynamics. hght we1ght robot manipulators. Nussbaum [ 1983] proposed that if the sign of kp is unknown. ~he future years are going to be ones of experimentation on more challengmg systems. and of its uncertainty.2. He then showed that Yp  oo. Mudgett and Morse [1985].2.2.2.3) as before.1) with kP different from zero but otherwise unknown. explicitly those linearizable by state feedback. Yp = apYp+kpu (8.0 as t .e. The chief difficulty with implementmg the hneanzmg control law is the imprecise knowledge of the nonlinear functions in the dynamics. the order of the plant. Two other areas that promise explosive growth in the years to come are adaptive control of multiinput multioutput (MIMO) systems. most of which are logical extensions and outgrowths of the developments presented in the previous chapters. Adaptation then has a role in helping identify the nonlinear funct10~s online to obtain asymptotically the correct linearizing control law. which are largely dependent on the knowledge of a process model. There is a great deal of excitement in the control community at large over the emergence of such custom multiprocessor control architectures as CONDOR (Narasimhan et a/ [ 1988]) and NYMPH (Chen et a/ [ 1986]). and the like. complicated algorithms can now be implemented at very high sample rates.3) can be shown to yield Yp . such advances will make it possible to exploit adaptive techniques on high bandwidth. While we have extensively discussed the analysis of adaptive systems. In addition to all these exciting new directions of research in adaptive control. the feedback gain d6 cos(d ) of 0 (8. we also feel that great strides in this area will come from experiences in implementing the algorithms on several classes of systems. However. A "Universal" Theory of Adaptive Control While all the adaptive control algorithms developed in this book required assumptions on the plantin the singleinput singleoutput case.s such as flexible space structures. In turn. aircraft flight control systems. Adaptive control for MIMO systems is especially attractive because the traditional and heuristic t~ch niques for SISO systems quickly fall apart when strong crosscouphngs appear. The years to come will see a proliferation of techniques to effectively map these adaptive algorithms onto multiprocessor control architectures. Conversely. some of which are often specified in table lookup f~rm. and aP unknown. We presented in this book what we feel is the tip of the iceberg in these areas.2. Th1s approach was discussed in this book.. and of today's multiprocessor environments.4) alternates in sign ("searches for the correct sign") as do is . with do(t) remaining bounded. Indeed. i. 1987]. One may hope that the recently introduced averaging techniques will help to better connect the frequencydomain properties of adaptive and nonadaptive systems. Note also that research in the identification of MIMO systems is also relevant to nonadaptive algorithms. the sign of the highfrequency gain. More needs to be understood about the sort of prior information needed for MIMO adaptive systems. A better understanding of whic~ algorithms ~re more robust will also help in deriving guidelines for the Improved des1gn of robust algorithms. Nussbaum [1983]. we have found it valuable in the implementation of an adaptive controller for an industrial robot (the Adept1) and are currently working on a flight control system for a vertical takeoff and landing aircraft (the Harrier). and Martensson [ 1985] among others. Heuristically. A very large class of nonlinear systems is explic!tly linea~izab~e . that these assumptions can be relaxed was established by Morse [1985.326 Conclusions Chapter 8 Section 8.2.4) with (8.oo. the adaptive control law (8. If the sign of kp is known and assumed positive.2. we now present a few other new vistas which are not as obvious extensions.2) can be replaced by u = d6 (t) cos ( d0(t)) Yp 0 as t  (8. With the advent of microprocessors. the incorporation of various forms of prior knowledge into blackbox models of MIMO systems also needs to be studied.
a conjunction of several logic steps and control algorithms. there is a great deal of work needed to implement a given adaptive algorithm. the understanding. the parametenze. It also decides which of a family of control laws to use and communicates its inferencing procedures to the operator. A supervisor provides alarms and interrupts. and a finite dimensional statespace. either as a linear or nonlinear system because of the complexity of the physical pr~cesses. . Morse [ 1988]). While this framework is extremely attractive from a practical point of view. it is fair to say that no more than a few case studies of expert control have been implemented. adaptive ~ystems. The parameterized controller is assumed to control the process. involved. Further.3). a very interesting study is found in Astrom et a! [ 1986]. if necessary.t. In the context of adaptive control. a parameterized controller. Adaptation and learning in this framework consists in refining the rulebased model on the experience gained during the course of operation of the system. and state of the art in learning for rulebased models is rudimentary. (8.2. and expertise about the system bemg controlled (such as the amount of noise.2.e. ign process is called a rulebased expert controller design. modeling and control of a given process.s to give a minimal set of assumptions on the process. th~ ~umber of unknown parameters. the assumptwns are to contain as special cases the algorithms presented thus ~ar. The resulting composite algorithm is often referred to as a rulebased control law with the adaptation scheme being one of the rules. Adapted from their work is Figure 8. While the topics in the title have the same general philosophical goals.. and the tuner to guarantee global stability and ~symptot1c performance of the closed loop system. namely.2.. to inject new excitation. and adaptation law or tuner (cf. Tuning is said to have taken place when a suitable tuning error goes to zero. Such a composite de . around the adaptive control algorithm.. Adaptive systems may be seen as the interconnection of a plant.d controller. prior knowledge. Learning has. the scheme has stimulated a great deal of interes~ to derive adaptive control schemes requiring a minimal set of assumptwns on the plant (universal controllers). i. Learning. Connectionism and all those things . which of a library of identification algorithms to use and. The goal of a universal theory i. the bandwidth of the parameters vanat10n. based on the level of excitation.1. involving th~ use of heuristics. Expert and Learning Control Systems As the discussions in Chapter 5 indicated. however. A further objective is to develop a unified framework which would subsume all the algorithms presented thus far. although adaptive control algorithms form an especially attractive area of application. While the transient behavior of the algorithm (8. The design and ev~luation of such composite systen:s is still an open area of research for nonadaptive as well as. This may be coded as several logic steps or rules. RuleBased.328 Conclusions Chapter 8 Section 8.).4) is poor. The rules then serve as a model of the plant from which the controller is constructed as a rulebased system. illustrating the structure of an expert control system using an adaptive algorithm.3)) until it is large enough and of the "correct sign" to stabilize the equation (8. the fields of identification and adaptive control have made the largest strides in becoming a design methodology by limiting their universe of discourse to a small (but practically meaningful) class of systems with linear or linearizable dynamics. Figure 8. in which the plant to be controlled cannot be easily modeled. Such a theory would be extremely valuable from a conceptual and mtellectual standpoint. One can conceive of a more complex scenario..1: Expert Adaptive Control System The rulebased system decides.2 Future Research 329 decreased monotonically (by (8. Adaptation. and the tuner assumed to tune the controller. A controller then has to be built by codifying systematically mto rules the experience gained from operating the system (this is referred to as querying and representation of expert knowledge).:. been merely parameter updating. the order of the plan.2.1 ).wwledge. The sophistication and performance of t~e controller is dependent on the amount of detail in the model furmshed by the expert k.
12) IS obtamed by integrating ( 1.5) Inequality By assu~ption. image recognition. (1.330 Conclusions Chapter 8 The goals of connectionism and neural networks (see for example Denker [ 1986] and the parallel distributed processing models) have been far more lofty: the universe of discourse includes human systems and the learning mimics our own human development.11) by parts. We feel.4. that in the years to come. associative memory storage elements and the like.4.4. and it is interesting to note that the 'learning' algorithms implemented in the successful algorithms are remarkably reminiscent of the gradient type and leastsquares type of update laws studied in this book.2 Let r(t) Ja(r)x(r)dr 0 I (Al.$ r(t) + u(t) so that (1. by assumption r(t) = a(t)x(t) . there will be a confluence of the theories and techniques for learning. APPENDIX Proof of Lemma 1. consequently.s(r))dr (Al.$ a(t)r(t) + a(t)u(t) (Al.4.4.2) that is. for some positive s(t) r(t)a(t)r(t)a(t)u(t) + s(t) 0 (Al.4.) and s(.1) so that.4) Since exp(.$ Je' 0 a(r)u(r)dr (A1.4. x(t) . A few applications of this work have been made to problems of speech recognition. 0 331 .3) Solving the differential equation with r(O) I = 0 r(t) = Je' 0 I Ja(u)du (a(r) u(r).) are positive functions I  I Ja(u)du r(t) .11) follows.4.4.
4) to J II J I C(r)<l>(r. using the triangular inequality We use a similar procedure. w y. while <l>(r. 7) becomes t0 + o K(u)ll 2 t0 + o ::. /33'.cr::.5. cr)e(cr)l II K(cr)ll I C(cr)w(cr)l dcr Note that to+ 8.8). cr) e(cr)l 2 11 K(cr)ll 2 dcrdrl to i to It is sufficient to derive ec.:. (A2.5.5.1) starting at e(cr).3) Changing the order of integration.7) to ::. A E IR n xn. Therefore.5.5. and the condition on K(.: ~I to+o w(to)l  Consider the trajectories x(r) and w(r)..5) and the system under output injection w(t) = [A(t) + K(t)C(t)) w(t) z(t) = C(t)w(t) (A2.5. the integral in the last parenthesis becomes t0 +o t0 +o I C(r)(w(r). cr)e(cr) is the solution of system (A2. ~~~ to T C(r)<l>(r.5.. jiC(v)w(v)l 2dv jiC(r)<l>(r. using (A2.5.x(r))l 2 = IJ T C(r) <l>(r.u)e(cr)l 2dcrdr]I to to w(r). corresponding to systems (A2.. cr)e(cr)l 2drdcr u (A2. On the other hand.2 Appendix Appendix We consider the system x(t) A (t)x(t) C(t)x(t) y(t) (A2.5. o)K(cr) C(cr) w(cr)dcr to (A2. and the UCO of the original system I 0 to + fa IC(r)(w(r)x(r))l 2dr l 333 i (A2.6) · [toiofiiK(cr)II 2 1C(r)<l>(r. E IR n.5. cr)K(cr) C(cr) w(cr)dcr 2 I 2 J II K(u)ll J I C(r) <l>(r.8) Inequality (2.4) I C(r) w(r)l 2 ::.5. Then T [ l' C(v) w(v)j 2 dv li (A2. cr) e(cr)l 2 dr du to u 2 (A2.5. cr)e(u)l 2 11k(u)ll 2dcr to to (A2. Derivation of {3 2' using the definition of the induced norm and Schwartz inequality. I e(u)l = 1. C E IR m xn.6) and (A2.x(r) J<l>(r.5. IR n xm.).uations the inequalities giving f3t'• Derivation of {3 1' /32'. I C(r)x(r)l 2 .4).1) and (A2. Z E IRm.5.. using the UCO property on the original system.1) so that.5. 8. 7) follows directly from (A2.2) respectively. using (A2.332 Proof of Lemma 2. and I j to 0 fl C(v) w(v)l to 2 dv fl C( r) <l>(r.5.2) K E where X. with identical initial conditions x(t 0 ) = w(t 0 ). .. kof32 T ::.5.5.
334
T
Appendix
Appendix
335
+I IC(T)<I>(T,a)K(a)C(a)w(a)da 12
to
a2
~
to+ o
2
I
to
[<wT +eT)x
J
dT
~
a1
00
forallt 0 ~0
(A2.6.1)
:s; I C(T)X(T)I
2
By assumption, e
E
L 2, so that I (eT x) 2 dT :s; m for some m ~ 0.
0
+ [ [.1 C(a) w(•)l
:s; I C(T)X(T)I 2
T T
I C(
')~(',a) e(•)l
Since w is PE, there exist a, {JI> {3 2 > 0 such that
II K(•)II
dr
(A2.5.9)
to+ u
{3 2 ~
J (wTx) 2 dT
to
~ {3 1
for all t 0
~
0
(A2.6.2)
Leto~a(l+ ~
to+o
].a 1 ={JI>a 2 =m+{J2 [1+
2
~]
to+o
10
sothat
+II C(v) w(v)J 2 dv II C(T)<I>(T, a)e(a)J 2 IJK(a)JJ 2 da
t0 to
I
to
1 r
[ (wT + eT)x
J
dT
and, for all t 0 :s; t :s; to+
t
o
t0 +
~ J (wT x) 2 dT J (eT x) 2 dT
to
to+o
o
2
II C(T)W(T)j 2dT :s;
to
T
I
to
I C(T)X(T)j dT +I I I C(v)w(v)
to to
2
l
2
dv
~ {3
and (A2.5.10)
t 0 +o
1
(1 +
~]
2
m
=
a
1
(A2.6.3)
· IJC(T)<I>(T, a)e(a)j
to
jjK(a)jj 2dadT
J [(WT + e
to
2
t0 +o
T) X]
dT _:s;
J ( TX)
W
lo
to+o
dT +
J (eTX)2 dT
lo
and, using the BeilmanGronwall lemma (lemma 1.4.2), together with the UCO of the original system
2 IJC(T)W(T)j dT :s; fJ2I W(to)l
2 to
I
(A2.6.4)
0
2
exp
[!
fiC(T)<I>(T, a)e(a)J 2 11K(a)ii dadT]
Proof of Lemma 2.6. 7
(A2.5.11) We wish to prove that for some 10 + o
az
to to
o, a 1, a2 > 0, and for all x
2
with I xi = 1 (A2.6.5)
for all t and in particular for t = to+ o. The in;egral in the exponential can be tran~formed, by changi~g the order of integration, as in (A2.5.8). Inequality (2.5.8) follows dtrectly from (A2.5.8) and (A2.5.11 ). 0
~ J
to
[.H(wT)x] dT
~
a1
forallto~O
Proof of Lemma 2.6.6
We wish to prove that for some
Denote u = wT x and y = H(u) = H(wT x) = H(wT)x (where the last inequality is true bec.ause x does not depend on t ). We thus wish to show that
10 +
o, a I> a 2 > 0, and for all x
with Ixi = I
o
a2 ~
J y 2(T)dT
to
~ a1
for all t 0
~
0
(A2.6.6)
Since w is PE, there exists a, {JI> {3 2 > 0 such that
336
lo + u
Appendix
t0 + o
Appendix
337
fh ~
J u 2(r)dr
~ {3 1
for all to~ 0
(A2.6.7)
lo
J y 2(r)dr
lo
~
In this form, the problem appears on the relationship between truncated L 2 norms of the input and output of a stable, minimum phase ~TI system. Similar problems are addressed in Section 3.6, and we wlll therefore use results from lemmas in that section. Let o = mu, where m is an integer to be defined later. Since u is bounded, andy= H(u), it follows that y is bounded (lemma. 3.6.l).and the upper bound in (A2.6.6) is satisfied. The lower bound ts obtamed now, by inverting if in a similar way as is used in the proof of lemma 3.6.2. We let
z(s) =
~ ~1
[
m ({:JI  ;
k3 u) k2 k 4 ]
(A2.6.13)
Note that rIa is arbitrary, and although k 1 depends on r 1a, the constants f:J1> k3, and a do not. Consequently, we can let r 1 a sufficiently small that f:Jt(rla)k3u~{:J 1 12. We can also let m be sufficiently large that m{:J1 I 2 k2 k 4 ~ {3 1. Then the lower bound in (A2.6.6) is satisfied with
ll'J
a'
(s +a)'
u(s)
(A2.6.8)
0
(A2.6.14)
where a> 0 will be defined later, and r is the relative degree of H(s). Thus
y(s) =
(s +a)'
Proof of Lemma 3.6.2
H(s)z(s)
A
a'
(A2.6.9)
The proof of lemma 3.6.2 relies on the auxiliary lemma presented hereafter.
Auxiliary Lemma
The transfer function from z(s) to y(s) has relative degree 0. Being minimum phase, it has a proper and stable inverse. By lemma 3.6.1, there exist k t> k 2 ~ 0 such that
lo+o
Consider the transfer function
J
lo
10 +6
z 2(r)dr ::::;; k1
J
lo
K(s)
2 y (r)dr
=
a'
(s+a)'
a>O
+ k2
(A2.6.10)
(A3.6.l)
Since
u is bounded
lo + o
where r is a positive integer. Let k (t) be the corresponding impulse response and define
00
T
Ju
lo
2 (r)dr ::::;; k3o
(A2.6.ll)
Then
g(t  r) =
J k(u)du
a'
(r 1)!
IT
J k(t oo
u)du
t  r
~0
(A3.6.2)
for some k 3 ~ 0. Using the results in the proof of lemma 3.6.2 ((A3.6.14)), we can also show that, with the properties of the transfer function a' I (s +a)'
10 +o lo+o
k(t) =
t
~
0
(A3.6.3)
J u 2(r)dr
to
::::;;
J z 2(r)dr
to
+ ~ k3o + k4
(A2.6.12)
and k(t) = O_for t <0. It follows that k(t) ~ 0 for all t, and
00 I
where k 4 is another constant due to initial conditions. It follows that
II
k/11
=
Jk(u)du
0
=
J k(t u)du
00
= I
(A3.6.4)
Similarly
r.~
338
g(t)
Appendix
f'"
1;, 0
Appendix
339
=
e••
00
,±, ~a•'
I
(A3.6.5)
O
~~ L
k(t a)da] it(,jdT
(A3.6.12)
and g(t)
=
0 for t < 0. It follows that g(t) 2. 0 for all t, and Using the results of the auxiliary lemma
=
llgll1
D
Ig(a)da
0
I g(t a)da
00
r
a
(A3.6.6)
z(t)
=
I
t(f) + u(t)u(O)g(t) Ig(t T)U(T)dT
0
(A3.6.13)
We are now ready to prove lemma 3.6.2. Let r be the relative degree of
If, and
z(s) =
Since g(t) is exponentially decaying, u (O)g(t) can be included in t(l ). Also, using again the auxiliary lemma, together with lemma 3.6.1, and then the assumption on it follows that
u,
a'
(s +a)'
u(s)
(A3.6.7) (A3.6.14) Since a is arbitrary, let it be sufficiently large that !_ k 1< l. Consea quently,
kz + ____ a_
I
where a> 0 is an arbitrary constant to be defined later. Using (A3.6. 7)
y(s)
=
(s+a)' H(s)z(s)
a'
(A3.6.8)
Since the transfer function from z(s) to y(s) has relative degree 0 and is minimum phase, it has a proper and stable inverse. By lemma 3.6.1 (A3.6.9) We will prove that (A3.6.10)
=11
r I  kl a
Z1
Z1
liP
II tllp +
r
r k1 a
so that the lemma will be verified with a1 Derivation of (A3.6.10) We have that
I
=
c1b1>
az
= c1b2 + Cz.
D
'' ~
ii
:=
c 1 II
II P
+
Cz
(A3.6.15)
Proof of Corollary 3.6.3 (a) From lemma 3.6.2.
=
z{l)
t(f) + Ik(t T)U(T)dT
0
(A3.6.11)
(b) Since
is strictly proper, both y and
y are bounded.
where t(l) is an exponentially decaying term due to the initial conditions, and k (I) is the impulse response corresponding to the transfer function in (A3.6. 7) (derived in the auxiliary lemma). Integrate (A3.6.ll) by parts to obtain
I
(c) We have that y = H(u) and y = H(u). Using succesively lemma 3.6.1, the regularity of u, and lemma 3.6.2, it follows that for some constantsk1, ... , k 6
IPI
.:;;
k1lltitll 00 k3ll ulll 00 k5ll Y1ll 00
+kz
0
.:;;
+ k4 + k6
z{l) = t(f) + u(t) I k(t a)da u(O) I k(l a)da
oo oo
::;
(A3.6.16) D
»tE:)'I::,.J~) ~Q ~o.r (f\~.t.,o) 1..... • Fe)"\. r "'> o J .  .
~~
The proof can easily be extended to the vector case.
340 Proof of Lemma 3.6.4 Let
Appendix
Since
Appendix
341
(A3.6.17) where fi 1 is strictly proper (and stable). Let h 1 be the impulse response corresponding to H1• The output y(t) is given by
t
~
(W¢)
(A3.6.24)
it follows that
d dt (x  W ¢)
Y1 Y2
=
A (x  W ¢) 
w
¢
.
(A3.6.25)
T .
y(t) = t(t) + hou(t) + Jhl(t r)u(r)dr
0
(A3.6.18)
H(wT ¢) H(wT) ¢
The result then follows since
=
where t(t) is due to the initial conditions. Inequality (3.6.9) follows, if we define
y 1  Y2
A
A
Hc(Hb(w )¢)
(A3.6.26)
I h 0 I /31 (t) + J h 1(t I
0
t
0
 T) I /31 ( T) dr
(A3.6.19) Proof of Theorem 3.7.3 The proof follows the steps of the proof of theorem 3. 7.1 and is only sketched here. (a) Derive properties of the identifier that are independent of the boundedness of the regressor The properties of the identifier are the standard properties obtained in theorems 2.4.12.4.4
and
t
'Y2(t) :=
I t{t)l
+ I hoi (32(t) + Jl h1(t r)l f32(r)dr (A3.6.20)
0
Since t e L 2 and h 1 e L 1 n L 00 , we also have that I tl E L2, e L 1 n L 00 . Since (3 1 , (3 2 e L 2 , it follows that the last term of (A3.6.19) and similarly the last term of (A3.6.20) belong to Lz n L 00 ,
I hd
llf,r(t)w(t)l
=
l;ct)111
wtll 00 +(s(t)
t
(A3.7.1)
and go to zero as t  oo (see e.g., Desoer & Vidyasagar [ 1975], exercise 5, p. 242). The conclusions follow directly from this observation. 0 Proof of Lemma 3.6.5 Let [A, b, cT, d] be a minimal realization of H, with A E IR.mxm, b E IR.m, c E IR.m, and de IR. Let x: IR.+IR.m, and y 1 : IR+IR such that
f3 e L 2 n L 00
if; e L 00 if; e L 2 nL 00
forallt~O
am+l(t)~kmin>O
. fhe inequality for am+ 1(t) follows from the use of the projection in the update law. We also noted, in Section 3.3, that if 1r is bounded and am+ 1 is bounded away from zero, then 0 is also bounded, and the transformation has bounded derivatives. The vector q of coefficients of the polynomial q is also bounded. By definition of the transformation, O( 1r*) = o•. Therefore, if; E Loo, f E L2 n Loo implies that ¢ E Loo, ¢ E
.X
Yl
Ax+ b (wT ¢)
CT X
(A3.6.21) ; (A3.6.22)
and W: lR + IR m x n, y 2 : lR + IR such that
W AW+bwT
Y2
cTW ¢
Thus
L2 n L 00. Also, we have that (km I II am+ I II 00) ::5: Co(t) ::5: km I kmin· for all t ~ 0. (b) Express the system states and inputs in term of the control error As in theorem 3. 7.I.
fi (wT ¢)
Y2 + (d wT) ¢
(A3.6.23)
8). .5) Ao f M~o [ Co m km J. 7 .17)) coM.. (A3.c0) r + ¢r wJ qr ~' (\A.4) Xo where we divided by A Xo to obtain proper stable transfer functions.qkpfzp a 111 + 1 (C. with the error c0 . 7.. From proposition 3.J)t/1 = km Ao co [<coco)M.7. 7..7. using again the swapping lemma.b. we have that i = Ao (cf. 7.1 (A3. the lefthand side of (A3.5) becomes.5).am + 1 ) AA [ ~~ A Ao X dp Ao 0 co +A(¢ .a 111 + 1 d .(¢Tw)] Ao Ao The righthand side is very reminiscent of the signal z obtained in 1 the input error scheme..~ 1 (u)+.9) = .1. instead of r.9)) Yp . The output error is then equal to (using (3.q ~ + qdp am+I(J d*) + (am+IJ•i.1c:Prw) J] Ao Ao [ am+ I A• am+ I 'A Jm A 'A Jm .prJ(w)] Ao (A3.7. Using a similar approach as in the comments before the proof.5) becomes (A3. The polynomial ~ 0 is Hurwitz and q is bounded.(c r) + 1<:Prw)] 0 (A3. 7... km [ .qo [ a~a· ] + .q + . Applying both sides of (A3. A filtered version of the signal M.7.2) (u) = _!_ Ao co (J. (A3.d *) + and since U = Cor+ 0Tw.c0)r) + J."0 A kp kp dp dp J dAp (A3 •7 3) • Therefore i.7.7.2)..7) followed by the swapping lemma (and using the notation of the swapping lemma) km co [coco 1ccor) + 1c:Prw) + co Ao Ao co co [..342 (c) Appendix with Appendix 343 Relate the identifier error to the control error We first establish an equality of ratios of polynomials.7.6) a 111 + 1 (~C). it follows that : M. ob· OC> where the transfer functions A A s.4) to u clo M ((co.kpqfzp a 111 +1 (cc*) + (am+lq*kpq)np q ~ .7) Ao Ao The righthand side of (A3.+ A A P .3. / ~ 0 is a bounded operator. (A3.1 P1(u) Ao = 1<u)J.co.odm+qdp) am+ 1 (d .7.Aoc co A [ Ao co A 0 b(w )¢ A r _.a m+ 1 /o [E~ E• ] ka:p cc* dJ· A] Ao dmfzp_ .(o*'w) Ao Ao (A3.3.C• co ~~o J .~ od m .8) On the other hand. 7.+ C . using (A3.Ym = We now transform this polynomial equality into an operator equality as in the comments before the proof.T w).c*).7. (3. ."O A+ q.r Jew). so that the operator qT s.P(u) =rp appears._ ) (A3.J • PA + (c0 ~~o •) co MA  I P A 1 (A3.(kp .((c 0 . and Src result from the application of the swapping lemma. then we transform it to an operator equality.C*) + a 111 + 1 ( ~.
2.2. We showed that ~. an inequality such as (3.(t. x) jd(r.2.2.8) Since.J [:'jte·• j jrr'dr 1 + .' (A4.x) (d) = O.2.2. x)e<t + e Je•UTl(w 0(t. for some {3. x)dr 0 J I Using the fact that I f Je•(lr)wo(t' x)dr 0 wo(t. jew.2.1) by parts so that the first part of the lemma is verified. t 0 ~ (A4. {3.+'Y[~ je. Note that. H:'jte·• j .2. Similarly.(t.X) (A4.1. 0 q ~.'~f. x))dr (A4.x) bd(r.2.1 0) and part (a). ¢./ ~0 is a bounded operator.7. x) = wa(t. in turn. I d(t.344 I Appendix Appendix 345 (A4. e..wm.2. and therefore i .. (A4.1' :::. using (A4.He) From the assumptions From the assumptiOJ! on 'Y. M. x). [ :· w.1• :::. .3) aw. x) Wo(t.wo(to.X) = EW.(w 1/1) + q S.2.4.10) w.7) Establish the regularity of the signals As in theorem 3.x)j:::.3) and the fact that w0(0.(t. I•w. x)e<t (A4.19) can be obtained. t' e.1. x E Bh· Integrating (A4.x)j:::.d(t. .w 0(t. x).oo.(t.w is the same as the relative degree of P.b(w )1/1 km Ao ~ ~ ( T §. ~T T ~ ( ~ ~T . xll " .2. "((!) · t I wo(t for all t.(t.1.) can be rewritten as I (A3. t'.2. +"f[~ ](l+y.'~}.10) wo(t.w 0(r. from (A3. c0 E L 2 . qT s.11) 0.6) 0 and.Ym e L 2 and tends to zero as too. for all t' ~ 0.7.)e_y. (e) Stability Proof I jw..(t. ~~~lb 'Y ( :'] t'el' + 7"1 [ :' 1r'e"dr' (A4.(t.1) +to.2) .x)j :::. The boundedness of all signals in the adaptive system then follows as in theorem 3. x)j :::. x)dr (A4. 7.1) This. As before w regular implies that {3.2.2. so that w. 7.' [:· j (A4.9) Proof of Lemma 4.. the same result is true for Consequently.2. {3. we also have that 'Y(l):::.7.(t. x)j :::. and I {3y.M>. implies that jew.•e'd.5) (A4. so that. .x) 8t . ] w.1. x)e•(1r)dr 0 + J.(t.c S.1 Define I + [ . x). From (A4.0 as t .4) .+{3. x). 'Y(l)te<~+e[e•(IT)(tr)"((tr)dr M ~0 is a stable transfer function and since qT is bounded. and t'e. (A4. Since the relative degree of the transfer function from u . Yp. it follows that HE) e K.e e•(l r) w 0(r.
with convergence f~~ction 'Y(T)Ixl.17) Applying (A4. using (4. there exists a unique x e Bh such that (A4.X) ax = _i_ [!d(T.16) and (A4.6) of lemma 4.Ju.2. provided that Hf) < 1.2.15) If 'Y (T) = a 1 T'. the solution z of (A4. However. with r denoting the standard gamma function.2.2. f/av(z) + f f(t. From the smoothness of w.2.13) 0 Define h'(f) = Defining Hf) = 2 a l. It can be checked. Consequently.9) (although the function Hf) may t' be different from that obtained with ad (ax x) ' these functions can be replaced by a single Hf)).12) 0 may not belong to Bh.2. Hf)h.15) to the system (4.2.2. We now show that • for all z e Bh'. with an add1twnal factor I xl.) < 1.. and inequality (4. z. x) is zero mean. d(t.2. then the righthand side of (A4.8) and (4.2.2.2.2.2.2. for any x. if a solution z exists to (A4.8) and implies that. ax et(IT)dT (A4.10). x) . the proof of lemma 4.2. = aw..!6) and.3 The proof proceeds in two steps. for all x e Bh'• there exists a unique z e Bh such that (A4. and is bounded.15).2. The first part follows directly from (A4. I x. Given z e Bh.2.2.2.1 O) of lemma 4.15) is satisfied. that the mapping Fx(z) = x . b ded quality (4.1) 0 Proof of Lemma 4. x) is zero mean. x) . a l (A4.1 becomes in e . there is no guarantee that a solution z exists in Bh.2.. ad (t .1 can be applied to ad~. Proof of Lemma 4.2.15) is satisfied.!av(z).zl ::. and since it is unique in U.2. the map xz defined by (A4.2. 0).2.2. given x E Bh. z + fW.2.15). 1 +~(f)] (A4.2.zl ::. [ I +f az 1: . F has a unique fixed point z in U. Similarly.16). For x E Bh. Hf)h also follows from (A4.1 can be extended.2. Apply lemma 4. aw. inequality (4. it is also unique in Bh (and actually in R n). (t . Since d(t. the corresponding x such that aw.2.8) implies (4. This leads directly to (4. (a) For f sufficiently small.14) Since ad~. The fact that I x. 00 Ja l(T') 0 1 'eT'dT' = a lr(2r)::.(t. (b) The transformation of variable leads to the differential equation (4. Note that smce ax 1s oun ' and d(t.. z. x) is Lipschitz. the second part of the lemma is verified.1' = a f'(lr) 1 'e'I' ~ + f w.346 Appendix x Appendix = z 347 (A4.1. z) with respect to z.15) may not exist in Bh.2.X)e•(/T}dT] ax 0 • = J [_i_d(T. X)J 0 In both cases.(t. z satisfying (A4. but if it exists. Let f ::. lemma 4. f.2.2. but outside Bh'.f w. and let f 1 such that ~ (f.at ] [ +f [!u.2. min [h(IHf)). it follows that the map is a homeomorphism. Consequently.15).8) can be computed explicitly sup al(t') 1 re. (4. f). Consequently.2 Define w. O) = 0 for all t ~ 0.(t.15).15) is well defined.17) and note that h'(f)h as f0..2.. a l (A4. it is again unique in Bh. (A4. This solution is also a solution of (A4. (t. it must lie in the closed ball U of radius Hdh around x. z) is a contraction mapping in U.x) as in lemma 4..2.2. f)) . the transformation is a homeomorphism. and for t fixed. z) 1 ::.15).2.
2. z(O) = f)] (A4. Since this is not guaranteed a priori.. z.348 Appendix Appendix 349 + f [Ju.fav(z). f)l ~ 1. z. and by using (4. + 1 ] of length t:.1).19) where IPI(t.4. (4.4.XavCt )I can be calculated by integrating the difference (4.4.2. y (t ). z.Z.1 The proof is similar to the proof of lemma 4.f)l ~ Hf)lzl +~(f)lllzl +fl21zl Xo (A4.2) := f/avCz) + fp'(t. f)+ fP2(t. XavCt) E Bh'• with h'<h.f)l For f ~ fJ> (4. f) Xo IP2 (t ' z' 0 y' dl k2 IYI (A4..4. Proof of Lemma 4. (4. z E Bh. f ~ f 1. We consider the transformation of variable X = Z + fW.4.2) into the system (4. f) aw.f) [I +E ~:· r 1 := Hf)kt I zl and [ ffav(z) + fp'(t.4. Tlf]. provided that f and h0 are sufficiently small. and z (t) (to be defined) remain in Bh.a. By assumption. the steps of the proof are only valid for as long as the condition is verified. ] + f k2 Bound on 1y(t)l JIy(T)I dT 0 I (A4. We will show that by letting f and h 0 sufficiently small. z B h. Z. f). we transform the original system (4.2.1. 8 [I+ f :' ] has a bounded inverse for 8 Consequently.2.4. 0) .f(t' z' 0)) t To obtain a bound on I y(t )I . A bound on the error I z (t).Z.2. and AT will be defined later).5) [I + f az' aw ]1[ p'(t. from a contradiction argument. o)] (A4. The differential equation for y is . z..21) Proof of Theorem 4. It then follows. we can let x(t) be arbitrarily close to Xav(t) and y(t) arbitrarily small. z. and the results of lemma 4. (4.4. that x(t).4.3.4.4.Z.11 )( 4.1) becomes Z = [I + faa:' r 1 ~ davjlz(T)Xav(T)I dT + fHf)kljlz(T)I dT 0 0 f kav(Z) + [f(t. 0). z. y. using the assumptions.4. z.(f)lzl (A4.22) 0 for all t :2: 0.20) ffav(z) + fp(t. y(t) E Bh for all t E [0. we divide the interval [ 0. T If].4) z where p(t.4.2.18) or f favCz) + fPI (t.2 IP'(t.2 The proof assumes that for all t E [ 0.Jet.2. T If] in intervals [ l.3) (A4.(t . t. the solutions x(t ).4 ).4.12) I I and 1 P (t. f az favCz) ] (A4.4.2).1) lz(t)Xav(t)l with f ~ fJ> such that Hf 1)< I.11 ). Z.4.6) + [!(t. X.4.10) implies that E all t . z satisfies the differential equation ~ 1 Hfl) [Hf)lav+Hf)+~(f)!l]lzl (A4.::: 0. 7) and (4.4. f) z(O) = (A4.Hf1) [Hf) + Hf)ll + f/2 + Hf)lav] I zl 1/.T (the last interval may be of smaller length.2. z. Using lemma 4. f) where. z) (A4. z + fW.
..Y+Eg(t. for 1 Bound on z(t).AT/2 E (A4./ 2 + dav + k2(mkc + kb) !..) + tkb .r 12 + k2t(mkc + k6 )) te<i". y) dr I I jz(t)Xav(l)j::::. it follows that (A4. and Ax. T/t].AT/2)) = l _ e>. h (cf.. and (A4.4.T) i (Ax.15) l y. = A (x(t )). From the assumptions. A. davjJz(r)xavCr)Jdr+tHt)k 1 Jhdr 0 0 + Je I.XavCt )j and.11 ).17) where Yi = y(ti).Ti/2ho + 1.l + E ~ h [(1 3 +1 4 ) + (1 1 +1 2)kaAT) Ek6 ..4. ti + 1l We assumed in the proof that all signals remained in Bh. (4. using the BellmanGronwalllemma (lemma 1. is given by y(t) A (x(ti )). from a simple contradiction argument.AT)) ::::. (A4.. y) (A4.T/ 2 + t(mkc+kb)) dr 0 (A4. Assumptions i~l ~ )=0 [me>..11) : :.18) (A4. jy(t)j t kb XI /2 e>t. TIE].11) be denoted by recursion formula for Yi• so that and use (A4.9) I + tk2J [mhoe>. using the bound on j y (t )j e . and tT be sufficiently small so that.11) as a k2 mho lav 1 e 1a•·T ::::.2) J JJAx.. 0 ::::.13) for some bT.14) and using the assumption Yo Bho• (A4. mhoe>.. Ax (I .4.e>. It follows. where Ax.. By assumption.l!::::. using the uniform exponential stability assumption on A (x) Jy(t)J : : . E(/J+lz)hkaA z(t).12)(A4.4.8) y = Ax. 'f(t)aT and.) Yi + E Je I.4. XavCt) E Bh'• for some h' < h.T) ' g(r.4.4..·l : : .12) Choose AT sufficiently large that or It follows that AT~ ilnm 2 (A4.14) Combining (A4. (t + Ht)) [ k1 h + f.Xj(rti)::::. kaJ.4.~ ~ J=O [e>. it gives a bound on jy(t)j for all t E [0. l I [Ht)klh + k2mhoe>.4.(ll.4.x.(1l.4. I Ax (I . that the solutions x (t ). so that the solution Y (t )..4.. it follows that for all t 0 ti..4. Let h 0 .(IT)dr Let the last term in (A4.) + mtkce>. tJ.4.6). tT ::::. 1.4..s. ti + 1] as follows (A4.4.y)+(Ax.. (A4. m Jyd e>.Ax)Y = Since the last inequality does not depend on i. ti + 1 ].4.4.4.4. we have that mho+ dmkc + k 6 ) ::::. y(t) and z(t) remain in Bh for all 1 E [ 0.19) (A4. using (4. and that 'f(t)bT .27)).Ax.4. (I I..4.t/ 2 +t(mkc+kb) 351 (A4. . x .7) I 1 Appendix : : . me>.350 Appendix A (x) y + E g(t .(1l.1 2h e>.16)).10) so that..AT/2 . Ax)Y(r)dr (A4.. for all t ::::..Xav(t) We now return to (A4.4. h h' (cf.e ' ho + tkc ·E [ Using this result in (A4. t E [ ti. X....4.. so that all steps of the proof are valid.10) (A4. and to the approximation error.16) and is rewritten on the time interval [ ti .19) is in fact satisfied over the whole time interval.
4. (A4. Bound on II ax.2).4.4. Consider the following Lyapunov functiOn (A4. + aA (x) eA(x)l axi (A4.4.2. (4.20) (A4. using the foregoing inequalities v.2).4.4.4.y)::::.30) With (A4. ax. _where x(z) is defined in (4.!!. It Pz follows that a' 1 (1zl 2 + eA T(X)I Q [ a~i eA(x)l] 1dt (A4. Under the hypotheses. a4 such that. and strictly positive constants a.31) I if::::.27) The partial derivatives in parentheses solve the differential equation .27). Bound on I ap.29) +yl 2) 2 ::S: v 1(z. a3.4. form exponential stability of A(x).4.4. for all y.4.4.4.4. so that ' .4.4.9).4.24) aA(x) rom the boundedness of . V(Xav) ::::.26) a 1 I Xavl 2 ::::.4.y) ::s: w3lzl 2 (A4. using (4. (4.azll and I zl On the other hand. thUav + Ht)k 1 + k 2 ) z Using these results in (A4. Defining a'. = A (x) lr_j_ eA(x)t] ax.4.4) v(Xav) I : :. (4.4.28) with zero initial conditions.352 Appendix ""I 2 I Appendix 353 Proof of Theorem 4.4.4.9).21) (A4. .p.4.2) can be bounded. az..11). p 2 are defined in the comments after the definition of uniaz .8) and (4.. it follows that F The derivative of v 1 along the trajectories of (4.a eA(x)t = feA(x)(tr) aA(X) eA(x)rdr 0 ax.(z.3).4.).21) is to be taken along the trajectories of the averaged system (4.1).z is well defined and 1xl ::s: 21 zl ). II aaxz II and < I + ~(t) < 2 (A4.4.. E IR . (4. this implies that II aP(x)/ ax 11 is bounded by some kp.4)..4. there exists a function v(Xav): IRnIR+.23) where P(x ). We now study the stability of the original system (4.4. for all Xav E Bh.12) + tHE)k!a4izl 2 + Ekza41ziiYI (A4.= min(a.  a zl Xavl (A4._ [ _j_ eA(x)t] dt ax.25) for t ::s: EJ (so that the transformation x... ax I 00 Note that P(x) can be defined by P(x) = JeAr(x)t QeA(x)tdt 0 (A4.4.11)( 4. through the transformed system (4.:: 0.25). az(lzi 2 +YI ) (A4.3 The proof relies on the converse theorem of Lyapunov for expone~tially stable systems (theorem 1. and from the exponential stability of axi A (x ). and noting the fact that.25).4. We now calculate bounds on the terms in (A4.22) so that WJI Xav 12 The derivative in (A4.
..io . (6. then r/J (ti). all signals are actually guaranteed to remain in Bh so that all assumptions are valid.4) Note that fort #0.8).32) pI(tt) Appendix 355 ti k0 ! O.2.2. •.A.2.1) (A6. f dt 0 (A6..A(t))z + (b +t. Also. }.2 Lemma A6. with (A6.t 1 .b(t))r (A6..2. t 2 .2. with this definition.2.4.2.2.0 exponentially. 0 Auxiliary Lemmas for Section 6.2.4. . if the intervals ti + 1 .. b both bounded and converging to zero as .ao. o Lemma A6.1 (A4.5). t.: kph Then (A6. ~ a 1! for all ti (A6.• _E_(p1¢) Note that..33) as 0.4.2. 7) so that (A6.2 E /3a2 [ P2 2 + 2f . for E. The same conclusion holds for the x. Then (A4.2. If a 1 > 0 and the sequence ti is infinite.9) with the sequence of resetting times {0. the lemma shows that rjJ (ti) is bounded.34) so that the z. [k o+ ] ko a1 I r/J (ti. t 2 .2.2.8) and.1 Consider the least squares identification algorithm described by (6.9) yields (A6.ti are bounded. h0 sufficiently small. y system is exponentially stable with rate of convergence w (E) (v 1 being bounded above and below by the square of the norm of the state).9).2 Consider the following linear systems .2.t 2 . t 1 .2.2. . t 1 . r/J (tJ.2.5) [tav + HE)kl + k2) ]IYI 2 Proof of Lemma A6.qJ.1 0).4. (A6.1 >I (A6.0 as i .4. that is (A6.1i 354 Appendix (A4.2) I 'I' "'(t~·)l ::.4)  a2 2/3 k2a4 2/3 ql2d4a2f . with (4.. given the transformation (4.11) with A stable and t.3) it follows that If w satisfies (I+ l J WWT dt I. y system.2.1 0) (A6. Further. while Thus (A6.2.9) Recursion on (A6. Comments If a1 = 0.i Azo + br (A +t. P2 Let E:S:E 1 be sufficiently small that a(E)>O and 2w 2 a(E):S:q(E).6) a2 q(E).
2.18) too.17) to del (A6.17) 0 l . there exist unique polynomials fie and de of order at most n .2. and fip monic of order~ n. 0 0 0 0 anI an 2 an anI an a.1. there of degree at most n.16) Thus.20) For T sufficiently large. {3z 0 {3.15) Proof of Lemma A6. Then given an arbitrary polynomial dc~(s) of order 2n ..12) .l. we may choose q so that fie is of order ::::. m (exp(. we may modify (A6. dll : : .2. t3n t3n I t3n {3z {3. that is fie kp fip and dc1. + b. respectively.2. Let kp be a real number.12).2.2. 0 t3n I t3n2 az {3. Ek for all t :2: T (A6. {33 {32 /34 !33 0 0 a3 az a4 a3 l u.2.18). a>O and t :2:r> T 1 • Using this estimate.Vdc1 + q kp fip Proof of Lemma A6. anSn1+ . It is useful to note that if a II </>(t. Then.l. defining the error e(t) : = z(t).q dp (A6.I so that (A6.l.2. az 0 0 0 0 0 {3.. n. an Snbn SnI I e = Ae + ~ z + ~b r (A6.356 Appendix (a Appendix 357 (A6. + d.2. +a. de to those of dc1 is a. there exists k > 0 (independent of f) and a T( E) such that de~. n . + {3..13) for some m.. f/ the input r is bounded Then given f > 0. since the other two polynomials in (A6. n. ~ and ~b are arbitrarily small.l.3 Solution of the Pole Placement Equation Consider two coprime polynomials: dp monic of order n.. 2n .l.zo(t ). de is constrained to be of order ::::.2 From lemma 6. n . D Lemma A6. (A6.19) I z(t)z 0(t)l ::::.2.l. Now. Let fie de adel . as the quotient obtained by dividing dc1 by dp ). it follows that A +~A (t) is asymptotically stable and that there exists T 1 such that the state transition matrix of A + ~A (t) satisfies Since dp is of order n. it is easy to show that z(t) is bounded. we see that 0 0 0 0 0 an 0 0 0 0 0 0 l /3n udel kp fip + v del dp Further. so that e may be showed to satisfy (A6.qdp)kpnp + (vdc1 + qk"np)d" =de~ for an arbitrary polynomial q.2. respectively so that exist polynomials then the linear equation relating the coefficients of fie.l (for instance.3 Since kp fip and dp are coprime and of order n .a(t T))) (A6. Sn+t3nsnl + . a.2.2.2. v 0 0 l (A6.2. we have that del dp kp fip ne de dznS2nl + ..14) + + +a. 0 a.2. are of order::::.
8388. 5. Vol." IEEE Trans..2. Johnson. B. S.O. R. Riedle. 2.D. Moore. 303307. Kosut. 3.21) is of the form A (O")IJ.O. 1. no.. 1. Anderson. 1983. Dasgupta. 1985.O. Vol. 1986. R. & A.. "Adaptive Systems and Time Varying Plants. 1969." Automatica. 359 .V. "Adaptive Systems." IEEE Trans. no.D. J.D. = d where o• is the nominal plant D parameter. Vol. Englewood Cliffs. 1985. Johnstone. Control. B. B.. P. B. Control.O. pp. Kokotovic. Network Analysis and Synthesis.O. "Global Adaptive Pole Positioning. 3. "New Results in Linear System Stability. PrenticeHall. Johnstone. pp. pp. 1973. 1977.. Anderson. pp. Praly. and 0~ the nominal controller parameter.D. Lack of Persistency of Excitation and Bursting Phenomena. Stability of Adaptive Systems. & B.L. & S.O. no. Bitmead.. Anderson. 367377. 1985. Vongpanitlerd. L. 7.2.M. Vol.D. Vol. B. Anderson. "Exponential Stability of Linear Equations Arising in Adaptive Identification. I. Tsoi. 1122. Anderson. pp. B. pp. Mareels.O. & J.M.D. 398414. C. Anderson." SIAM J.O.B.D. 37. "On the Convergence of a Model Reference Adaptive Control Algorithm with Unknown High Frequency Gain..M. & R. 21. B. Anderson. on Automatic Control." Systems & Control Letters. on Automatic Control. New Jersey.R. Vol. B. AC22. no.D." Int. no. 247258. Anderson. & R.R.C.21) REFERENCES h I I (A6.·lf 358 Appendix aI a2 dl d2 i j i I l I an I an bl dn I dn dn +I dn +2 I (A6. Massachusetts.Y. MIT Press.D. Cambridge. AC30. Passivity and Averaging Analysis..
"On Parameter Convergence in Adaptive Control. Clary J. pp. no. Control. Vol. Applications to Stabilization and Adaptive Control. & Y. pp 494500. California. R.. Anderson. "Global Stability Proofs for ContinuousTime Indirect Adaptive Control Schemes. Vol.. 277286. Chen. pp. Texas. pp. Vol. K. M. no. pp. Asymptotic Methods in the Theory of Nonlinear Oscillators. 1987. Bai. 1984. 1986. of the 23rd IEEE Conf on Decision and Control. Astrom. 643647...=. 5. "Robustness of ModelReference Adaptive Control Systems with Unmodelled Dynamics. "Zeros of Sampled Systems. 224229. "Stability. 1984. E)(~h:J. Stanford University. 1986.J. & A. Bodson. Bai. of the 22nd IEEE Conf on Decision and Control." Proc.. M.. 1984... 3. Theory of Ordinary Differential Equations.. 2. 1987. San Antonio. Gordon & Breach New York 1961." Automatica. "Satyr and the Nymph: Software Archetype for Real Time Robotics. Boyd. 15. 1982. 39..illl Si8Rill 1/8 St1teility of J:ion 1ineu "gg. pp. S. 3138.J. pp. "Adaptive Feedback Control. Sastry. Minnesota. M. 75. McGrawHill. 3.. "Input Error versus Output Error Model Reference Adaptive Control. Dallas. 15691573... AC19. J. Balachandra. "Sm. 10. Astrom. J. New York. Mitropolskii. "SelfTuning Regulators for a Class of Multi variable Systems. M. New York.. Multivariable Feedback Systems. Dissertation. Isidori. "Dual Effect. Astrom. BarShalom. pp. 1988. J.. R. Mareels. AC32. Clarke. K. S. Sastry. 44. Burdick. "The Stability of Solutions of Linear Differential Equations. Texas. pp. Vol. Minneapolis. "Robust Algorithms in Adaptive Control. Tuffs.A. 1983. V. 982987. Vol. & J.S. pp. 7.." Systems & Control Letters." Ph. S. Lab. M. Linear Stochastic Systems. 2. 1988. 9. to appear. I. 1943. 137148. D. The Basic Algorithm. Armstrong. Riedel." Automatica. Bodson. Claude. pp. and Robustness of Adaptive Systems. Vol." Int. ( Bodson. Levinson. & S.. pp. 201214.J. SpringerVerlag. 185199. & N. Vol. E. "Nonlinear Averaging Theorems. and Separation in Stochastic Control. Geometric Methods in the Theory of Ordinary Differential Equations. of the IEEEACM Joint Conference. J. Hagander. 1955.." Proc. Wittenmark. 145157. pp. E. "Parameter Identification Using Prior Information. l>r~t\M . UCB/ERL M86/66.R.. 1984..R. E. Vol.J. Information Syst. & J. Bogoliuboff.. Astrom.B. C. & K. & P.. 261283. "A Generalization of the Method of Averaging for Systems with TwoTime Scales.. "Necessary and Sufficient Conditions for Parameter Convergence in Adaptive Control. 1987. Nevada. Byrnes. Stanford.A. K. Las Vegas." IEEE Trans. pp. Tse. J.360 Rejerences References 361 Arnold. 1984. D." Duke Math.. & C." IEEE Trans. Nevada. John Wiley. of Adaptive Control and Signal Processing. New York." Proc. SpringerVerlag. of the Automatic Control Conference pp. & P. pp. University of California." Automatica. 1986. 1979." Int.D.W." Proc. Sternby. no.. "Interactions Between Excitation and Unmodeled Dynamics in Adaptive Control.N. ~ol ~J~ f\~ . 23. 209215.. Astrom.. of the 23rd IEEE Conf on Decision and Control. ~~r_ct. "On Self Tuning Regulators.A. 1973. Vol. of the IEEE... B. Vol. Vol. Las Vegas. 6. & S. no." Automatica. N.S. Cook. C. Certainty Equivalence. Electronics Research Laboratory. & S.J. "Generalized Predictive ControlPart I. Vol.. no. 12821285. V.I.. Control. Z. F. no. Anton." Int. 6. I 984." Proc. I.." Archive for Rational Mechanics and Analysis. Mohtadi. & S." Proc. & S.1281. & R. on Automatic Control.O. 311319. Bodson. pp.D. Sastry.W. Berkeley. Las Vegas. Bitmead.. no.. Boyd. 3. Nevada. Arzen. 1986. Caines. 1987. no. 1986. Desoer. pp." Systems & Control Letters. M. 22. Vol. ttol Systems• * t~t!llieatioa lO the R:ubasmcss of a ftlttAe ~drcaw.E. Sastry. & B. 197 5. & P. K.E...A. Dordrecht. on Automatic Control. Chen. K. 2. Vol. P." Algebraic and Geometric Methods in Nonlinear Control Theory. Convergence. Vol. 12 76. pp. B. Callier.J. Y. J. 1982. & S. 629639. K. Fearing.D. "Effect of the Choice of Error Equation on the Robustness Properties of Adaptive Control Systems. 537543." Automatica. Memorandum No. & E. Dissertati~n. 185217. and the Determination of Parameter Convergence Rates in Adaptive Control. pp. "Analysis of Rohrs Counterexamples to Adaptive Control. Coddington. no.J.P. 22. Bodson.M. ' ' Borison. 1974. Sastry. 455473. Bellman. Sastry. Sastry. 1983. "Expert Control.W. Astrom. 1988. New York. P." Automatica. 20. 58. of the 23rd IEEE Conf on Decision and Control. "Everything You Always Wanted to Know About Linearization. 1986.J." Ph. Sethna." Proc.
pp. 5770. "A Review of JLFor Case Studies in Robust Control. & S.. J. Vol. 1984. and Bifurcations of Vector Fields. no.416.C.. & S." IEEE Trans. pp... 6. on Automatic Control.H. "Decoupling with Dynamic Compensation for Strongly Invertible Affine Nonlinear Systems. Mayne.. J. Vol. . Das. J. & S. Goodwin. Vol. 1984. Hill. Vol." IEEE Trans. 13871398. "Connections Between FiniteGain and Asymptotic Stabthty. 4. 6.D. 23.C. no.). Vol. P.M. Freund. "A Parameter Estimation Perspective of Continuous Time Model Reference Adaptive Control. 1985.l. "Direct Adaptive Control for Linear Systems. Wolovich.C. Peter Peregrinus.. Minneapolis. Kokotovic. 1984. Dynamic System Ident(!ication. G. pp. & S. J. Australian National University. G. & M.C. & C. 4. SpringerVerlag. pp.C.S. Elliott. no. Das. Vol. on Automatic Control. G. Vol. Meredith. & W. Control. Blankenship (Eds.T. 1963a. London. "Adaptive Identification and Control. "DiscreteTime Multivariable Adaptive Control. 533545." Int. no. 1963b. & D. 340352. pp.25. J." Proc of the lOth IFAC Congress. "Advances in Multivariable Control. J.. Fu.. Vol.J. no. 1.E. & R. 221229. 12511281. SpringerVerlag. H. Donalson. 1967. no. 1983. A. Feedback Systems: lnputOuput Properties.S. Self Tuning and Adaptive Control: Theory and Applications. 1980. Elliott. & M. Wolovich. New York. L.. Cristi. Vol. pp. 1980. 1985. 2. PrenticeHall. & M. Vidyasagar. New Jersey. Stein. H. Goodwin. J. pp. 42. 793795. Sastry. no. Vol.. J. 1974..: & P. 1982. S.. Vol. of Robotics Research. on Automatic Control. R. J. J. Doyle.no. J. J. Wolovich. New York.J.. System Ident(/ication. 1987. of the 1986 Conference of the American Institute of Physics. 241251.T. on Automatic Control." IEEE Trans. lEE." IEEE Trans. Goodwin. A.C. Elliott. no. pp." Proc. Academic Press. Dissertation. "The Structure of Decoupled Nonlinear Systems. AC30.. Hale. Sastry.J. 6. 251262." Int.. 20. 348356. 1987. Berlin. & J. IEEE Trans. Bodson." Automatica. Huntington. H.L.362 J<ejerences Ke}erences 363 Craig. Krieger. Egardt. Bensoussan. "A Model Referenced Parameter Tracking Technique for Adaptive Control SystemsThe Principle of Adaptation. Harris." Notes for the Honeywell Workshop in Multivariable Control. 2. "Slow Drift Instability in Model Reference Adaptive SystemsAn Averaging Analysis.AC26.. "Global Stability of Adaptive Pole Placement Algorithms. Munich. & W. 1984. on Automatic Control. on Automatic Control. AC29. New York. & K. Dasgupta.C. 743753. 1986. 1241 1245. Payne.. 1981. Control. Desoer.pp. Vol. M.C.O. no. "A Model Referenced Parameter Tracking Technique for Adaptive Control SystemsStability Analysis by the Second Method of Lyapunov. Nonlinear Oscillations. Dugard. Sin. of Control.. 6. 21. Vol. of Control. "Adaptive Control of Mechanical Manipulators. 651659. New York. AC25. pp. Vol." Ph. on Automatic Control. Englewood Cliffs. D. 82. 45. Hsu. & P." Automatica. on Applications and Industry. pp. pp.A. 82. Doyle. 1987. 1975.C." IEEE Trans. Ordinary Differential Equations." Singular Perturbations and Asymptotic Analysis in Control Systems.. New York.. 112. Descusse. 1975. Goodwin. New York. 1984.. P. 1986." Automatica. D. 395402. White. no. "New Stability Theorems for Averaging and Their Application to the Convergence Analysis of Adaptive Identification and Control Schemes. "Analysis of a Type of Model ReferenceAdaptive Control System. Guckenheimer. 2.. 1981. 1965. Adaptive Filtering Prediction and Control. 1977. 1628. Academic Press. P.K. Holmes. L. "Arbitrary Adaptive Pole Placement for Linear Multivariable Systems.F. 1979. pp. Snowbird. New York. Donalson. Stability of Motion.). AC. Stability of Adaptive Controllers. Leondes. & R.M. on Applications and Industry. L." Int." IEEE Trans. no. Doyle. Dymock. "A Parameter Adaptive Control Structure for Linear Multivariable Systems. Billings.. Dion. pp. .Q. 1984. W. " Neural Networks for Computing. 20. & K. C. Vol. Vol. D.. & G.. Long. pp. "Parameterization Issues in Multivariable Adaptive Control. "On the Stabilizability Condition in Indirect Adaptive Control. 42. W. and C. 1985. SpringerVerlag. Hahn. 3. Goodwin.O." Int. De Larminat. Vol. 449456. G. 8. P. J. Springer Verlag. Sastry. pp.." IEEE Trans.. (Ed. & C. 503527. Denker. Australia. Elliott. G. Eykhoff. H. Leondes. & P. AC27." Int. P. Moog.. no. Fu. "Generalization of Results on Multivariable Adaptive Control. Dynamical Systems. 1987. pp. Caines. AC25..A. Ramadge. Utah. J. no. 3. Vol.C. B." IEEE Trans. & G. Lecture Notes in Control and Information Sciences. 5. E. Wiley & Sons. H." Proc. ~oy~~n. Canberra. C. 1980. pp. A. Hall. "Multivariable Feedback Design: Concepts for a Classical/Modern Synthesis..
364
References
References
365
no. 5, pp. 931936, 1980. Horrocks, T., "Investigations into ModelReference Adaptive Control Systems," Proc. lEE, Vol. Ill, no. 11, pp. 18941906,1964. Hsu, P., M. Bodson, S. Sastry, & B. Paden, "Adaptive Identification and Control of Manipulators without Using Joint Accelerations," Proc. of the IEEE Conf on Robotics and Automation, Raleigh, North Carolina, pp. 12101215, 1987. Hunt, L.R., R. Su, & G. Meyer, "Global Transformations of Nonlinear Systems," IEEE Trans. on Automatic Control, Vol. AC28, no. I, pp. 2431, 1983. Ioannou, P.A., "Robust Adaptive Controller with Zero Residual Tracking Errors," IEEE Trans. on Automatic Control, Vol. AC31, no. 8, pp. 773776, 198~. Ioannou, P.A., & P.V. Kokotovic, Adaptive Systems with Reduced Models, Springer Verlag, New York, 1983. Ioannou, P.A., & P.V. Kokotovic, "Robust Redesign of Adaptive Control," IEEE Trans. on Automatic Control, Vol. AC29, no. 3, pp. 202211, 1984. Ioannou, P.A., & G. Tao, "Frequency Domain Conditions for Strictly Positive Real Functions," IEEE Trans. on Automatic Control, Vol. AC32, no. l, pp. 5354, 1987. Ioannou, P.A., & K. Tsakalis, "A Robust Direct Adaptive Controller," IEEE Trans. on Automatic Control. Vol. AC31, No. 11, pp. 10331043,1986. Isidori, A., "Control of Nonlinear Systems via Dynamic State Feedback," Algebraic and Geometric Methods in Nonlinear Control Theory, M. Fliess & M. Hazewinkel (Eds.), Riedel, Dordrecht, 1986. Isidori, A., & C.l. Byrnes, "The Analysis and Design of Nonlinear Feedback Systems: Zero Dynamics and Global Normal Forms, " Preprint, Universita di Roma, "La Sapienza", Rome, 1988. Isidori, A., A.J. Krener, C. GoriGiorgi, & S. Monaco, "Nonlinear Decoupling via Feedback: a Differential Geometric Approach," IEEE Trans. on Automatic Control. Vol. AC26, pp. 331345, 1981. Isidori, A., & C.H. Moog, "On the Nonlinear Equivalent of the Notion of Transmission Zeros," Modeling and Adaptive Control, C. Byrnes and A. Kurszanski (Eds.), Lecture Notes in Information and Control, SpringerVerlag, 1987. James, D.J., "Stability of a Model Reference Control System," AIAA Journal, Vol. 9, no. 5, 1971.
Johansson, R., "Parametric Models of Linear Multivariable Systems for Adaptive Control," IEEE Trans. on Automatic Control, Vol. AC32, no. 4, pp. 303313, 1987. Johnson, C.R., Lectures on Adaptive Parameter Estimation, PrenticeHall, Englewood Cliffs, New Jersey, 1988. Johnson,. C.R., B.D.O. Anderson, & R.R. Bitmead, "A Robust Locally Optlmal Model Reference Adaptive Controller," Proc. of the 23rd IEEE Conf on Decision and Control, Las Vegas, Nevada, pp. 993998, 1984. Kailath, T., Linear Systems, PrenticeHall, Englewood Cliffs, New Jersey, 1980. Kalman, R.E., "Design of SelfOptimizing Control Systems," Trans. ASME, Vol. 80, pp. 468478, 1958. Kalman, R.E., & J. Bertram, "Control System Analysis and Design Via the 'Second Method' of Lyapunov," J. of Basic Engineering, Trans. ASME, Vol. 82, pp. 371400, 1960. Kalman, R.E., & R.S. Bucy, "New Results in Linear Filtering and Prediction Theory," J. of Basic Engineering, Trans. ASME, Ser. D, Vol. 83, pp. 95108, 1961. Koivo, H., "Multi variable SelfTuning Controller," Automatica, Vol. 16, pp. 351366, 1980. Kokotovic, P.V., "Recent Trends in Feedback Design: An Overview," Automatica. Vol. 21, no.3, pp. 225236, 1985. Kokotovic, P., B. Riedle, & L. Praly, "On a Stability Criterion for Continuous Slow Adaptation," Systems & Control Letters, Vol. 6, no. 1, pp. 714, 1985. Kosut, R.L., & C.R. Johnson, "An InputOutput View of Robustness in Adaptive Control," Automatica, Vol. 20, no. 5, pp. 569581, 1984. Krasovskii, N., Stability of Motion, Stanford University Press, Stanford 1963. ' Kreisselmeier, G., "Adaptive Observers with Exponential Rate of Convergence," IEEE Trans. on Automatic Control, Vol. AC22, no. 1, pp. 28, 1977. Kreisselmeier, G., "An Approach to Stable Indirect Adaptive Control," Automatica, Vol. 21, pp. 425431, 1985. Kreisselmeier, G., _"A Robust Indirect AdaptiveControl Approach," Int. J. Control, Vol. 43, no. 1, pp. 161175, 1986. Kreiss~lmeier, G., & B.D.O. Anderson, "Robust Model Reference Adaptive Control," IEEE Trans. on Automatic Control, Vol. AC31, no. 2, pp. 127133, 1986.
366
References
References
367
Kreisselmeier, G., & K.S. Narendra, "Stable Model Reference Adaptive Control in the Presence of Bounded Disturbances," IEEE Trans. on Automatic Control, Vol. AC27, no. 6, pp. 11691175, 1982. Kumar, P.R., & P.P. Varaiya, Stochastic Systems: Estimation, Identification and Adaptive Control, Prentice Hall, Englewood Cliffs, New Jersey, 1986. Landau, Y.D., "Unbiased Recursive Identification Using Model Reference Adaptive Techniques," IEEE Trans. on Automatic Control, Vol. AC21, no. 2, pp. 194202, 1976. Landau, Y.D., Adaptive ControlThe Model Reference Approach, Marcel Dekker, New York, 1979. Lion, P.M., "Rapid Identification of Linear and Nonlinear Systems," AIAA Journal, Vol. 5, no. 10, pp. 18351842, 1967. Ljung, L., System Identification: Theory for the User, PrenticeHall, Englewood Cliffs, New Jersey, 1987. Ljung, L., & T. Soderstrom, Theory and Practice of Recursive Identification, MIT Press, Cambridge, Massachusetts, 1983. LozanoLeal, R., & G.C. Goodwin, "A Globally Convergent Adaptive Pole Placement Algorithm without a Persistency of Excitation ial~ t '1 Requirement," IEEE Trans. on Automatic Control, Vol. AC 30, ) no. 8, pp. 795798, 1985. _, Mareels, I.M.Y., B.D.O. Anderson, R.R. Bitmead, M. Bodson, & S.S. Sastry, "Revisiting the MIT Rule for Adaptive Control," Proc. of the 2nd IFAC Workshop on Adaptive Systems in Control and Signal Processing, Lund, Sweden, 1986. Martensson, B., " The Order of Any Stabilizing Regulator Is Sufficient A Priori Information for Adaptive Stabilization," Systems and Control Letters. Vol. 6, pp. 8791, 1985. Mason, J.E., E.W. Bai, L.C. Fu, M. Bodson, & S. Sastry, "Analysis of Adaptive Identifiers in the Presence of Unmodeled Dynamics: Averaging and Tuned Parameters," IEEE Trans. on Automatic Control, Vol. AC 33, 1988. Meyer, G., & L. Cicolani, "Applications of Nonlinear System Inverses to Automatic Flight Control Design  System Concepts and Flight Evaluations," AGARDograph 251 on Theory and Applications of Optimal Control in Aerospace Systems, P. Kent (Ed.), NATO, 1980. Monaco, S., & D. NormandCyrot, "Nonlinear Systems in Discrete Time," Algebraic and Geometric Methods in Nonlinear Control Theory, M. Fliess & M. Hazewinkel (Eds.), Riedel, Dordrecht, 1986.
Monaco, S., D. NormandCyrot, & S. Stornelli, "On the Linearizing Feedback in Nonlinear Sampled Data Control Schemes," Proc. of the 25th IEEE Conf on Decision and Control, Athens, Greece, pp. 20562060, 1986. Monopoli, R.V., "Model Reference Adaptive Control with an Augmented Error Signal," IEEE Trans. on Automatic Control, Vol. AC19, no. 5, pp. 474484, 1974. Morgan, A.P., & K.S. Narendra, "On the Uniform Asymptotic Stability of Certain Linear Nonautonomous Differential Equations," SIAM J. Control and Optimization, Vol. 15, no. I, pp. 524, 1977a. Morgan, A.P., & K.S. Narendra, "On the Stability of Nonautonomous Differential Equations .X= [A + B (t )]x, with Skew Symmetric Matrix B (t )," SIAM J. Control and Optimization, Vol. 15, no. 1, pp. 163176, 1977b. Morse, A. S., "Global Invariants Under Feedback and Cascade Control," Proc. of the International Symposium on Mathematical System, Lecture Notes in Economics and Mathematical Systems, Vol. 131, Springer Verlag, New York, 1976. Morse, A.S., "Global Stability of ParameterAdaptive Control Systems," IEEE Trans. on Automatic Control, Vol. AC25, no. 3, pp. 433439, 1980. Morse, A.S., " A ThreeDimensional Universal Controller for the Adaptive Stabilization of Any Strictly Proper Minimumphase System with Relative Degree Not Exceeding Two," IEEE Trans. on Automatic Control," Vol. 30, pp. 11881191, 1985. Morse, A.S., "A 4(n+ I) Dimensional Model Reference Adaptive Stabilizer for 'Any Relative Degree One or Two, Minimum Phase System of Dimension nor Less," Automatica, Vol. 23, pp. 123125, !987. Morse, A.S., " Towards a Unified Theory of Parameter Adaptive Control," Report no. 8807, Dept. of Electrical Engineering, Yale University, 1988. Mudgett, D.R., & A.S. Morse, "Adaptive Stabilization of Linear Systems with Unknown High Frequency Gains," IEEE Trans. on Automatic Control, Vol. 30, no. 6, pp. 549554, 1985. Nam, K., & A. Arapostathis, "A Model Reference Adaptive Control Scheme for Pure Feedback Nonlinear Systems," Preprint, University of Texas, Austin, 1986. Narasimhan, S., D.M. Siegel, & J.M. Hollerbach, "CONDOR: A Revised Architecture for Controlling the UtahMIT Hand," Proc. of the IEEE Conference on Robotics and Automation, Philadelphia, Pennsylvania, pp. 446449, 1988.
~VO ~
lt S 1v tJ AP...t tl 0 p.._ k lq=t'~ i E C& "flL feNS. o t.J AU r\) "'"""'""' C
(.o~V 1"0 " 1 v ol · A,C It J
~.1
JH) A
'T'"''
06Sf ~
368
References
References
369
Narendra, K.S., "Correction to 'Stable Adaptive Controller DesignPart II: Proof of Stability'," IEEE Trans. on Automatic Control, Vol. AC29, no. 7, pp. 640641, 1984. Narendra, K.S., & A.M. Annaswamy, "Robust Adaptive Control in the Presence of Bounded Disturbances," IEEE Trans. on Automatic Control, Vol. AC31, no. 4, pp. 306315, 1986. Narendra, K.S., & A.M. Annaswamy, "A New Adaptive Law for Robust Adaptation without Persistent Excitation, " IEEE Trans. on Automatic Control, Vol. AC32, no. 2, pp. 134145, 1987. Narendra, K.S., A.M. Annaswamy, & R.P. Singh, "A General Approach to the Stability Analysis of Adaptive Systems," Int. J. Control, Vol. 41, no. l, pp. 193216, 1985. Narendra, K.S., I.H. Khalifa, & A.M. Annaswamy, "Error Models for Stable Hybrid Adaptive Systems," IEEE Trans. on Automatic Control, Vol. AC30, No.5, pp. 339347, 1985. Narendra, K.S., Y.H. Lin, & L.S. Valavani, "Stable Adaptive Controller Design, Part II: Proof of Stability," IEEE Trans. on Automatic Control, Vol. AC25, no. 3, pp. 440448, 1980. Narendra, K.S., & R.V. Monopoli (Eds.), Applications of Adaptive Control, Academic Press, New York, 1980. Narendra, K.S., & L.S. Valavani, "Stable Adaptive Controller DesignDirect Control," IEEE Trans. on Automatic Control, Vol. AC23, no. 4, pp. 570583, 1978. Nussbaum, R.D., "Some Remarks on a Conjecture in Parameter Adaptive Control," Systems and Control Letters, Vol. 3, pp. 243246, 1983. Ortega, R., L. Praly, & J.D. Landau, "Robustness of DiscreteTime Direct Adaptive Controllers," IEEE Trans. on Automatic Control, Vol. AC30, no. 12, pp. 11791187, 1985. Ortega, R., & T. Yu, "Theoretical Results on Robustness of Direct Adaptive Controllers : A Survey," Proceedings of the 5th Yale Workshop on Applications of Adaptive Systems Theory, pp. 115, 1987. Osburn, P.V., H.P. Whitaker, & A. Kezer, "New Developments in the Design of Model Reference Adaptive Control Systems," Paper no. 6139, Institute of the Aerospace Sciences, 1961. Ossman, K.A., & E.W. Kamen, "Adaptive Regulation of MIMO Linear Discretetime Systems without Requiring a Persistent Excitation," IEEE Trans. on Automatic Control, Vol. AC32, no. 5, pp. 397404, 1987.
Parks, P.C., "Liapunov Redesign of Model Reference Adaptive Control Systems," IEEE Trans. on Automatic Control, Vol. AC11, no. 3, pp. 362367, 1966. Peterson, B.B., & K.S. Narendra, "Bounded Error Adaptive Control," IEEE Trans. on Automatic Control, Vol. AC27, no. 6, pp. 11611168, 1982. Polderman, J.W., " Adaptive Control & Identification: Conflict or Conflux ?," Ph.D. dissertation, University of Groningen, Groningen, The Netherlands, 1988. Popov, V.M., Hyperstability of Control Systems, SpringerVerlag, New York, 1973. Prager, D., & P. Wellstead, "Multivariable PoleAssignment and SelfTuning Regulators," Proc. lEE, Part D, Vol. 128, pp. 918, 1981. Praly, L., "Robustness of Model Reference Adaptive Control," Proc. of the Jrd Yale Workshop on Applications of Adaptive Systems Theory, pp. 224226, 1983. Riedle, B.D., B. Cyr, & P.V. Kokotovic, "Disturbance Instabilities in an Adaptive System," IEEE Trans. on Automatic Control, Vol. AC29, no. 9, pp. 822824, 1984. Riedle, B.D., & P.V. Kokotovic, "A StabilityInstability Boundary for DisturbanceFree Slow Adaptation with Unmodeled Dynamics," IEEE Trans. on Automatic Control, Vol. AC30, no. 10, pp. 10271030, 1985. Riedle, B.D., & P.V. Kokotovic, "Integral Manifolds of Slow Adaptation," IEEE Trans. on Automatic Control, Vol. AC31, no. 4, pp. 316324, 1986. Rohrs, C.E., "How the Paper 'Robustness of ModelReference Adaptive Control Systems with Unmodelled Dynamics' Misrepresents the Results of Rohrs and his Coworkers," Int. J. Control, Vol. 41, no. 2, pp. 575580, 1985. Rohrs, C.E., L. Valavani, M. Athans, & G. Stein, "Robustness of Adaptive Control Algorithms in the Presence of Unmodeled Dynamics," Proc. of the 21st IEEE Conference on Decision and Control, Florida, pp. 311,1982. Rohrs, C.E., L. Valavani, M. Athans, & G. Stein, "Robustness of ContinuousTime Adaptive Control Algorithms in the Presence of Unmodeled .E>ynamics," IEEE Trans. on Automatic Control, Vol. AC..:.3o, no. 9, pp. 881889, 1985. Samson, C., "Stability Analysis of an Adaptively Controlled System subject to Bounded Disturbances," Automatica, Vol. 19, pp. 8186, 1983.
370
References
References
371
Sastry, S., "ModelReference Adaptive ControlStability, Parameter Convergence, and Robustness," IMA Journal of Mathematical Control & Information, Vol. I, pp. 2766, 1984. Sastry, S., & A. Isidori, "Adaptive Control of Linearizable Systems," Electronics Research Laboratory Memo No. M87/53, University of California, Berkeley, California, 1987. Sethna, P.R., "Method of Averaging for Systems Bounded for Positive Time," Journal of Math Anal and Applications, Vol. 41, pp. 621631,1973. Singh, R.P., & K.S. Narendra, "A Globally Stable Adaptive Controller for Multivariable Systems," Tech. Rept. 8111, Center for Syst. Sci., Yale University, New Haven, Connecticut, 1982. Singh, R.P., & K.S. Narendra, "Prior Information in the Design of Multivariable Adaptive Controllers," IEEE Trans. on Automatic Control, Vol. AC29, no. 12, pp. 11081111, 1984. Singh, S. N., & W.J. Rugh, "Decoupling in a Class of Nonlinear Systems by State Variable Feedback," J. Dyn. Syst. Measur. and Control, Trans. ASME, Vol. 94, pp. 323329, 1972. Sondhi, M.M., & D. Mitra, "New Results on the Performance of a WellKnown Class of Adaptive Filters," Proc. of the IEEE, Vol. 64, no. 11, pp. 15831597, 1976. Staff of the Flight Research Center, "Experience with the X15 Adaptive Flight Control System," Nasa Technical Note NASA TN D6208, Washington D.C., 1971. Tomizuka, M., "Parallel MRAS Without Compensation Block," IEEE Trans. on Automatic Control, Vol. AC27, no. 2, pp. 505506, 1982. Tsypkin, Ya.Z., Adaptation and Learning in Automatic Systems, Academic Press, New York, 1971. Tsypkin, Ya.Z., Foundatior > of the Theory of Learning Systems, Academic Press, New York, 19 7 3. Unbehauen, H. (Ed.), Methods and Applications in Adaptive Control, Springer Verlag, Berlin, 1980. Vidyasagar, M., Nonlinear Systems Analysis, PrenticeHall, Englewood Cliffs, New Jersey, 1978. Vidyasagar, M., Control System Synthesis: A Factorization Approach, MIT Press, Cambridge, Massachusetts, 1985. Vidyasagar, M., & A. Vannelli, "New Relationships Between lnputOuput and Lyapunov Stability," IEEE Trans. on Automatic Control, Vol. AC27, no. 2, pp. 481483, 1982.
Volosov, V.M., "Averaging in Systems of Ordinary Differential Equations," Russian Mathematical Surveys, Vol. I 7, no. 6, pp. 1126, 1962. Whitaker, H.P., "An Adaptive System for Control of the Dynamic Performance of Aircraft and Spacecraft," Paper no. 59100, Institute of the Aeronautical Sciences, 1959. White, A.J., "Analysis and Design of ModelReference Adaptive Control Systems," Proc. lEE, Vol. 113, no. 1, pp. 175184,1966. Widrow, B., & S. Stearns, Adaptive Signal Processing, PrenticeHall, Englewood Cliffs, New Jersey, 1985. Widder, D. V., An Introduction to Transform Theory, Academic Press, New York, 1971. Wiener, N., "Generalized Harmonic Analysis," Acta Mathematica, Vol. 55, pp. 117258,1930. Wittenmark, B., & K.J. Astrom, "Practical Issues in the Implementation of SelfTuning Control," Automatica, Vol. 20, pp. 595605, 1984. Wolovich, W.A., "A Division Algorithm for Polynomial Matrices," IEEE Trans. on Automatic Control, Vol. 29, no. 7, pp. 656658, 1984. Wolovich, W.A., & P.L. Falb, "Invariants and Canonical Forms Under Dynamic Compensation," SIAM J. Contr. Opt., Vol. 14, no. 6, pp. 9961008, 1976.
267. 119 output error direct adaptive control. 274 multivariable.I._290 model reference identifier. 164. 125 input error direct adaptive · control. high gain. 327 using M. 103. 166 two time scales. 71. 75. 320 parallel. 233 Certainty equivalence principle. of gradient algorithms with SPR error equation. 9. 277. 118 indirect. 158 Averaging. Ill output error. I model reference. 124. 279 . 6 parametric. 115. I series. 129. 5. 155 of the output error direct adaptive control. 59 identifier with normalized least squares and covariance resetting. rule. 183. 8. 5 universal theory of. 12 Adaptive pole placement. 305 of nonlinear systems.INDEX Adaptive control. instability analysis. 330 Controllable canonical form. 99. 16 7 Convergence. 186 nonlinear. 122 Autocovariance. 272 Algorithms identifier with normalized gradient. 280 Computed torque. 130 Bursting phenomenon. 162. 85 of the identifier. left. 9. 80 output error direct adaptive control. 154 of the indirect adaptive control. 268 Class K. 63 indirect adaptive control. 294. 14. 286 direct. 270 of the input error direct adaptive control. 236 mixed time scales. 20 Averaging. 321 Connectionism. 25 Column reduced. 123. 155 Coprime. 40 Autonomous. 329. 307 of robot manipulators. Ill input error. 55 Convergence function. I 79 BIBS stability. 192 one time scale. 8.T. !59. I 03. relative degree I.
39 Gradient algorithm.2. 311 partial. 15. 279 Ga111.1. 100 positive real.6. 44 Ocadzone. 216 MIMO systems. 19 Corollary 3. 299. for nonlinear systems.1. left. 142 Lemma 4. 233 high gain identifier.6. 269 Lemma 6. (PE).2.2.5 Swapping Lemma. 62 propagation equation. 52 Hyperstate.252 Diophantine equation. 39 lemma. I 8 extended. 48. 62 with covariance resetting.6 Small Gain Theorem.6. 10. 58 with projection. 24 exponentially stable. 242 Expert control. 320. 12 stability. 315 control.2 Output/Input L 1. 325 Equation error identifier. 42 Proposition 1. 23 uniformly asymptotically stable. augmented. 82 identification.1 Input/Output L. transfer function. 19. 57 input. 59 Prior information. 236 Interactor matrix. 121. 129. 58 Index high frequency. 65. 159. 305 stable.4. Ill. 248. 57.2. 126 Hermite normal form.1. convergence. 27!\ right coprime. 177. 43. 262 Identification.1 Matching Equality.5.Cross Correlation. 62 reselling. 277 Persistency of excitation.4. 279 Minimum phase. 25 Positive real (PR). 154155. I 0. 19 definite functions. 214 375 Nonlinear. 86 Lemma 2.5. 21 Lp spaces. 50.6. right. 123. 292 Projection. strictly positive real. 303 derivative. 305. 45 model reference. 253 Learning. 24 uniformly stable. 236 due to bursting. 71. 251 Parameter. 10 Identifiability condition. 100. 228. 76. 139 Lemma 3. 302 Model signals. 52 Propositions Proposition 1. II 7 Parameter. 56. 330 Nominal plant. 119. 23 Lemma 1. 38 redesign. 73. 81 Proposition 3.6 PE and L 2 Signal.5. 141 Lemma 3. 7.2 Convergence of the Identifier. 18 Observer. 77 Proposition 2. 24. adaptation. 109 tuned. I.1. 20 Polynomial matrix division. 272 Divisor. 47. 7.3 Uniform Asymptotic Stability of LTV Systems. 257. modified strictly positive real. 249.2. I 00 frequency domain approach. 302 Norms.3 Perturbation Formulation of Averaging. 295 Linear.2. 7. 270 Lie.2 KalmanYacuuovitchPopov Lemma. 127. 20 Normal form. 140 Covariance. 72. 282 Kalman filter. 265 Piecewise continuous. 21 Proposition 1. 63 Lemmas Lemma 1. 95 Parameterization. 233. 7 . 18 Lyapunov equation. 15. 58.2. 2 78 Dual control. 280 observability. 258 Index. 53. 100.261 normalized. 33 Proposition 1.6. 251. 61. Ill. 90.1 Approximate Integral of a Zero Mean Function. 82 Prefiltering. common right. 328 Factorization approach to controller design.2 Bellman Gronwall Lemma.3 Properties of Regular Signals. 62 with forgetting factor. 83 Lemma 2.4 Exponential Stability of LTV Systems. 283 scheduling. 2 78 left coprime.6. 295 decoupling.4. 58 equation. Stability. 189 Cross spectral measure. 139 In de. slow drift. 242 variation. 82 tuned. 221.5.6. 23 7 Error.\' Lemma 3. I 04. 19 definite matrices. 23 unstable. 190. 326 poles and zeros. scmidctinite matrices. 168 Lemma 4. 106 . I. 274 Fraction. 269 Initial conditions. 52nonlinear system. 42 Proposition 2.3 Linear Filter Lemma. 251 relative. 48. 134 Monic. 279 right. 59. 167 Measurement noise. 53 Equilibrium point.2.5 Exponential and Uniform Asymptotic Stability.374 Coprime.6. 34 Proposition 1. 169 Lemma 4. 12 Generalized harmonic analysis.I PE and Autocovariance. 73 Lemma 2. 257 Proper. 278 greatest common right. 51. 12.6. 102.6.2 Uniform Complete Observability under Output Injection. 241. 265 Lemma 6. 121. 51. error. II. 63 windup. 48. 57. linear.2. 21 asymptotically stable.175. controllability.4. 133 equation. 48.3.6. I 06.2 Persistency of Excitation under Perturbation. 20 Linearizing control. 61 Leakage term. 141 Lemma 3. I 09 nominal. Stability.2 Smooth Approximate Integral of a Zero Mean Function. 118 parameter. 170 Lemma 4. 113 Input saturation. 114 output.6. 76 structural. 329 Least Squares Algorithm. 266 Lemma 6. 45. 48. 279 Corollaries Corollary 1. 284 Hurwitz. 84 equation. 293 Lipschitz. 7 PE through LTI Systems. bracket. 34 Proposition 1.4. 2. 281 Positive. 52 Neural networks.1 Barbalat's Lemma. 117 Instability. 23 Lemma 2. 281 Information matrix. 2 3 theory.1 Perturbation Formulation of AveragingTwo Time Scales. 184 Lemma 6. 70 Proposition 2. 250. 261 normalized.1. 4.2 Linear Filter Lemma.6. 62 Cross correlation. 288 Mean value. 86 Lemma 3. 277. 25 Matching equality.
6 7 Theorem 2.2 Basic Averaging Theorem. 254 (see Averaging) Slow drift instability. 84 Theorem 2. 73 Theorem 2. 14 Sensitivity. 316 Uncertainty.5. 328 Selftuning. 316 adaptive.1. 143 Spectral line.2 Instability Theorem for Two Time Scale Systems. 254 amodification. 209. 221 Theorem .3 Exponential Stability Theorem.2 Exponential Stability Theorem.3 Basic Theorem of Lyapunov. 248 Robustness.2 Linear Error Equation with Normalized Gradient Algorithm. 26 Theorem I.7. 5 Exponential Stability Theorem.3 Basic Tracking Theorem for SISO Systems with Relative Degree Greater than I.5.4 Basic Averaging Theorem. 254 modifications.. 304 Proposition 7.I Stability Proof Input Error Direct Adaptive Control. 7 Stability of the IdentifierUnstable Plant.2. 185 Theorem 4. transfer functions.2.1.1 Fundamental Identity. I I 7 slow.3 Stability ProofIndirect Adaptive Control.3 Effect of Initial Conditions and Projection. 33 Stationary. 236 Solutions. l 7 3 Theorem 4. 58 hybrid update. 175.4. 209 additive.3. 41 Spectral measure. 230 Theorem 5. 211 Robust identification.1. 177. 251 regressor vector filtering. 46.1 Small Signal l/0 Stability. 23 7 Theorem 5. 216 Unimodular.3 Exponential Convergcnc~ of the Identifier. 376 Index control. 52. 155 Theorem 4. 3 Exponential Parameter Convergence and Sufficient Richness. 205. 70 Relative degree. 307 Regressor. 221 to output disturbances. 52 Sufficient richness.3. 239 Theorem 6. 7 Lyapunov Lemma. 310 Theorem 7. fast. 253 X15 aircraft.2 PE and Sufficient Richness. 93 Theorem 3. 82 Strictly proper. l 72 Theorem 4. 225.4. 310. 211 structured.8. 316 Rate of convergence. 275 State.3 Robustness to Unmodeled Dynamics. 35 Universal controller.3 SPR Error Equation with Gradient Algorithm. 214 nonadaptive. 277 of adaptive pole placement.4. 75 Index Theorem 2. 328 expert controller design.1 PE & Exponential Stability.3. 49 Reference model. 209.2. 266 input error direct adaptive Algorithm. 143 Theorem 3. 215 Row reduced.3 MIMO Matching Equality. 213. 273 Theorem 6. 64 Theorem 2.4 Linear Error Equation with Normalized LS Algorithm and Covariance Resetting. 21. 69 output error direct adaptive control.3. 228 Theorem 5.2 Interactor Matrix and Hermite Normal Form Equivalence. 58 Strictly positive real (SPR).1 Bounded Tracking in Minimum Phase Nonlinear Systems.. 36 Theorem 1. 328 Unmodeled dynamics. 288 Proposition 7.1 Instability Theorem 377 for One Time Scale Systems. 229 Rohrs examples. 40 Steepest descent. 253 Regular signals. 212 Slow adaptation. I 55 Theorem 3. 7.5.3. 75.4.1 Polynomial Matrix Division. 183 indirect adaptive control.5.7.2.8. 198. 300 Robust control. 233.5 Stability of the Identifier. 211 multiplicative. 304. IS l Theorem 3. 261 Recursive.2 Properties of the Identifier.2. 7 . 162 generalized. 58 filtering.5.1 Linear Error Equation with Gradient Algorithm. nonadaptive.5 Asymptotic Stability of the Adaptive Identifier using the Factorization Approach.4. 7. 85 Theorem 2.3 Asymptotic Stability of an Indirect Adaptive Scheme.6 Exponential Stability of LTV Systems. 277 Theorem 7.3. 71 Theorem 2. 279 Update law. II 7 State transition matrix. 65 Theorem 2.1 Adaptive Tracking.4. 52 strong for nonlinear systems. 213 Uniformly completely observable (UCO).4 Asymptotic Stability of the Adaptive Pole Placement Proposition 3. 263 unstructured. 262 Proposition 6.5 Exponential Convergence of Gradient Algorithms and SPR Error Equations.3. 92 Theorem 2. 25.6.2. 6 .5. 69 Theorem 2. 226 to unmodeled dynamics. existence. 114 Proposition 6.6. 64 Theorem 2. 28 Theorem 1.4 Modified SPR Error Equation with Gradient Algorithm. 41 Stability (see Equilibrium) family of matrices.2 Stability Proof Output Error Direct Adaptive Control. 143 Stabilizing compensator. 149 Theorem 3.4.3. 310.5. 316 Tracking. 271 Theorem 6. 50. 38 Theorem 2. 280 Rule based control. 185 Theorem 5. 209. 99. 149 of adaptive identifier using factorization approach.6.2 Robustness to Disturbances. 5. 286.2. 56. 7. 284 Proposition 6. 273 of the identifier. 282 Proposition 6.4. 31 Theorem 1. 92 Theorems Theorem 1.3. I. 85 Theorem 2. 253 slow adaptation.4. 162 Statevariable filters.1 Converse Theorem of Lyapunov.5. 151.
48. 45.9). "0". line 16: replace "[0.3). line 9: in eqn (2. line 15: replace "− g " by "− " 2 2g p. 50.29).0. A ]" p. C ]" by "[C .0. line 3: replace "Re(M (j ω)" by "Re(M (j ω))" p.53). 37. A + KC ]" p. line 18: in eqn (2.January 30. line 16: replace "(1. 83.4.19). w T (t )]" by "[w T (t ). line 27: in eqn (2.28). replace "y (t )" by "yp (t )" 2 p.0. line 19: delete "that is the projection can only improve the convergence of the algorithm" p. C ]" by "[C .3. line 21: in eqn (2. replace "φ(t + δ.5. replace "am e 1" by "am e 1 " 1 1 p. 84. 51. line 19: in eqn (2.0.0.30). replace "≤ 0" by "≥ 0" p. line 21: in eqn (2. replace "w (t ) w T (τ)" by "w (τ) w T (τ)" − p.5).5.4. 51. replace "e 1 (τ)" by "(θT (t ) w (τ) − yp (τ) )2" 2 p. 12.0. line 11: in eqn (0. line 17: replace "for all t ≥ 0" by "for all t ≥ t 0 ≥ 0" p.13). 66. 67. − g w (t ) w T (t )]" ˆ ˆ p.41).56). line 3: in eqn (2. line 5: in eqn (2. 50. 39. replace "β(t )" by " β(t ) " at both places p. replace "P (0) = P 0" by "P (0) = P (0)T = P 0" p. 48. 74.0. line 7: replace "[A + KC . line 7: in eqn (2. replace "am + km b 0(t )" by "am − km b 0(t )" * p. 64.3. 0]" p.Corrections . 50. 68.4. 74. line 16: in eqn (2.7)" by "(1.9). t )" by "Φ(t + δ. line 4: replace "[A . 83. line 7: in eqn (2.30). replace " d (P −1 ) d (P −1 ) " by " " dt dt p. replace "dt " by "d τ" p. Convergence. replace "e 1 (τ)" by "(θT (t ) w (τ) − yp (τ) )2" p. t )" p.38). 2000 Adaptive Control: Stability.0. replace "d τ" by "d σ" p. line 14: in eqn (1. 37. and Robustness p. w T (t )]" by "[w T (t ). line 2: replace "Re(M (j ω)" by "Re(M (j ω))" ˆ ˆ p. line 1: in eqn (1. 24. 74.15)" by "Note that (0.14)" p.7)" p.38). replace "(θ r )" by "(r )" p.0. 50. line 1: replace "β" by " β " p. replace "(b 0(t ) − b 0 )" by "km (b 0(t ) − b * )" 0 2 p. line 2: replace "c T " by "c T (t )" . 13. 50.0. line 18: in eqn (2. 74.3. line 9: in eqn (2. 52.0.18).29). 22. line 1718: replace "[− g w (t ) w T (t ). line 5: in eqn (1. 69. line 18: in eqn (1.3. replace "P 0" by "P 0 1 " at both places −1 p. replace the lower limit of the integral.61). by "t 0" p. 22.5.4. replace "P 0" by "P 0 " at both places p. line 17: replace "Note that (0.
88. line 10: replace "Morse [1979]" by "Morse [1976]" ˆ ˆ p. replace "Sale" by "Scale" ˆ ˆ p. line 3: adjust eqn (2. line 10: replace "lemma 2. line 5: in eqn (3. replace "n α1" by " n α1" 2 p. line 11: replace "ξ ∈ R pxp (s )" by "ξ ∈ R pxp [s ]" p. replace "β(t )" by " β(t ) " at both places p. last 3 lines: insert factors of 1/2 as follows t0 + δ t0 + m σ 2 e 1 (τ ) d τ t0 + m σ 2 x 1 (τ ) d τ t0 + δ 1 2 a − b 2" 2 ∫ t 0 1 ≥ 2 ∫ t 1 + 2 − 0 ∫ t 2 x 2 (τ ) d τ t0 + δ 0 2 x 2 (τ ) d τ t0 + m σ 2 ∫ − t0 ∫ x 1 (τ ) d τ +m σ 2 2 ≥ 1 γ3(m σ) em (t 0) 2 − m α2 φ(t 0 ) 1 n α 1" 2 p. line 1: in eqn (2.3. replace "n α1" by " p.3.1. line 21: replace "φ(t ) → 0 as t → 0" by "φ(t ) → 0 as t → ∞" ˆ ˆ p. line 5: in proposition 6. replace "e 1" by "e 0" p. 89.6. Control.6" by "lemma 2.6.5" by "lemma 2. replace "P (z )" by "P (r )" p. 282. vol. 284.17). 148.6" p. line 8: replace "lemma 2. delete "Let DR be column reduced and DL be row reduced" ˆ p.6. line 13: in the caption for Fig. Beghelli S.41).38).1. line 27: replace "DL row reduced" by "DL column reduced (such a matrix fraction description always exists (cf.5.6. 247. 89. 271. 247. 138. 148. line 4: replace " s + am s + am p. line 1: in eqn (5. replace "γ3 (m σ)/2" by "γ3 (m σ)/4" kp kp ˆ M (φr r + φy yp )" by " (φr r + φy yp )" p. & R. 143.6" p.30). line 16: in the proof of proposition 6. c T ] by [c T .6" by "proposition 2. 692696. 88.2p.6. line 13: replace [A . 283.3.27). 282. replace "− ε r (t )" by "− g r (t )" p. line 31: in the equation giving φT (t ) v (t ) . 5. 114.’’ IEEE Trans. on Autom.6. 89. line 22: in eqn (4. delete "and DR column reduced" ˆ ˆ p. line 10: in eqn (3. line 14: replace "the same conditions as β" by "the same conditions as β(t ) " p. 21. 163.4. 155.49). 161. line 7: in eqn (2. ‘‘A New InputOutput Canonical Form for Multivariable Systems. line 5: in eqn (2.40).15.6.5. 288. 155. pp. line 12: in the title. Guidorzi.39) to read " −γ m σ 1 γ (m σ) − γ1 e 2 ≥ γ3(m σ) / 4" 2 3 1 p.7" p. line 9: replace "theorem 2. 1976)" . 89.1. 101. replace "r 1(t )" by "r 1(t ) = sin(5t )" p. replace "+ ε(t )" by "− ε(t )" p.6. A ] p. 88.4. line 18: replace "Using the triangle inequality" by "Using the fact that (a − b )2 ≥ p.
6.6.. 5.3p. 356. Narendra.S. insert "When r = 0. 341. pp.. line 10: in eqn (A6. 360. replace " −∞ ∫ " (the ﬁrst integral) by "∫" 0 p.6. & K. AC18. u (t ) = z (t ). 339. line 1: in eqn (A3. "An Adaptive Observer and Identiﬁer for a Linear System.10)". line 15: under "Derivation of (A3." (we have that .. When r > 0. line 3738: the title of the paper should be "Exponential Convergence and Robustness Margins in Adaptive Control" instead of "Small Signal. line 1: in eqn (A3. line 22: insert Luders.12).6. so that (A3. replace "t r − k " by "t r − k " p. on Automatic Control.15).. replace "m ( exp − α −(t − τ ))" by " m e − α (t − τ) " p." IEEE Trans. line 14: replace "β(t )" by " β(t ) " at both places p..10) is trivially true. replace ":=" by "≤" p. line 12: in eqn (A3.3. 366.6. 496499." p. 338. .5).) t t p. G.22). no. 1973. 338.2. line 5: in eqn (6.13). 339. Vol. 290. replace "w " by "w T " p.
and Robustness.1991 © Marc Bodson." S. 1989. Homework Exercises Dec. PrenticeHall. . Sastry & M. Bodson. Convergence. 1991.Complement to: "Adaptive Control: Stability.
950952. el 180 tOO 140 120 100 s / // I! I. Show that the system with state variables (e0 . Given that the equation for y. . check the responses for m= 1 and g = 10. .!!o. the xaxis of the fignre becomes.2 Problem 0.2 Give an example of a control system where parametric adaptation is required or desirable. with interleaving of the stable and unstable regions.3 Consider the MRAC algorithm of chapter 0. Let r = r 0 he constant.. =m and theyaxis becomes trz=g. The algorithm is the same as the one given by eqn.. James in "Stability of a Model Reference Control System. can be solved independently of the others. an external factor.4: Stability Regions for the MIT Rule Assuming that the plant is f(s)= l/(s+ 1). an initial uncertainty or an inherent change in the system under control. the dynamic behavior of the algorithm is quite complex. b) Simulate the responses of the adaptive system for a few values of the parameters to check the validity of the fignre. 1971. rule. If ~~ /. 9. show that the remaining equations constitute a linear timevarying system x = A(t) x. a slow state variable.) is linear timeinvariant Calculate the poles of the system as functions of a.T.6).3. Problem 0. Let all initial conditions be zero and plot the error e0 for 50 seconds. while for small enough g.G. no. a) Plot the responses over 20 seconds of 8(1). (0. it is stable for all"'· Otherwise.4. 5.. a) Write the differential equations of the overall adaptive system and show that it is a nonlinear timevarying system. Problem 0. and r0 • Plot the locus of the poles as r0 varies from 0 to""· Discuss what happens when ro=O. Figure P0. and the reference input r=sin(m t). 2 (1) for the M. In particular. 5 and 2. the reference model is ~(s)= l/(s+ 1). k. then for g = 10 and "'= 1.50 and 80.4 The stability properties of the MIT rule were investigated by DJ. 1.. Could such . the Lyapunov redesign and the indirect scheme of chapter 0. Problem 0.<O." AIAA Journal. for small enough"'· the MIT rule is stable for all g. vol. 35. Identify the source of variation and indicate whether it is due to a nonlinearity. which are shown on Fignre P0. pp. b) Repeat part a) with r(l) = e•.. 5. and when k. Let r(t) =sin(I). e 0 (1) and v(l)=e~(l)+k. using the Lyapunov redesign. 8(0) =I and all other initial conditions be zero.I.1 Consider the plant /l(s)= 2/(s+ I) and the reference model ~(s)= I I (s+ 1). I. Comment on the responses in view of the analytical results. The author calculated the stable and unstable regions of the algorithm.. Note that. I 25 .
In each case.geoJp .(A)?:.< 1(A).. give A.(AB)e R and .( (A B)+ (A Bl)...4 Prove the following facts.mn(A)Ixl2 s xr AxSA.(A) f) det(S).. 2 (A) d) A?:. u and y P are the input and output of the plant.(AB) and A. 0 if and only if . 1(A). b) Find examples of 2x2 symmetric (but not diagonal) matrices such that A?:..(A.~... I fz(t)= t+ I .tO implies A.:.<.t 2(A) implies xi x 2 =0 wbere x 1.S Consider the plant F(s) = 1 /(s+ a.O ifandonlyif A.t~. L_...(AB)?:.(A)=A. l 1(t)=ea' for a>O and a.. Give A.(A)e R c)A=AT and . L 2 . and d is an adaptive gain with update law d=. a22 > 0..(A) and UT U=l..<.1 a) Determine which functions belong to L 1 .(As) 2:0 c) For A=AT?:.1 )=(<.) and the reference model M(s)= I /(s+ a.(A)Ixl 2 . All matrices are real and square (except in a)).(A). CTC2:0 and CTC>O b) A=AT implies .<..O u=dyp+r where r is the reference input (bouoded).2 a) Show that A e R 2' 2 is positive definite if and only if au > 0.O b) A?:. a)ForallCeR""". 0 implies Re A. All matrices are real and square. x 2 are the eigenvectors associated with . A.mn(A)r' IIAII=A.(s1 AS) g) A=Ar>O. c) Find an example of two symmetric positive definite matrices A and B such that the product is neither symmetric nor positive definite. Use only the definitions of eigenvalues and eigenvectors. B=Br?:.o IJ(t) = t'. You may use the fact that A= AT implies that A=UT AU withA=diag A.1 .)..3behavior be observed with a secondorder linear timeinvariant system ? Problem 1. and au a22 > (a 12 + a21 ) 2 I 4. Let the controller be given by A. A< 0 and A is neither positive nor negative semidefinite.(A)?:.O d) For A= AT >0 e) For A=AT?:.3 Prove the following facts. L_.1. a) ForA=Ar A?:....O implies Problem l. I I J.. Problem 1.<. L 1 . 0.. L 2 .O A.(t)= 'it+ I b) Show that I e L 1 () L_ implies I e L 2 • Problem 1. 0 ifrank(C)=n Problem 1.
indicate whether the system is stable.8 Consider the linear timevarymg syswm (from Vidyasagar's Nonli11ear Systems Analysis. or uniformly asymptotically stable. 1978) Jc.x)= l+xz f(t. 0) bounded (with bound b)..x)=e' x f(t.x)=Ax b) e' f(t.A(t)x where A(t). What conclusions ean you draw from this example ? Problem 1. ( (a!)t . PrenticeHall.x)= e'+e' . x) globally Lipschitz in x (with constant/) and f(t. asymptotically stable. Discuss the stability properties of the equilibrium point at x = 0. with f(t. and give the Lipschitz constant Let x and f be scalars.O Find m . ( l+acos t 2 1asintcost Show that the transition matrix ct>(t. r) satisfies e(at)rcost 1asintcost ) l+asin2 t cl>(I.) Find expressions for k1 .4 a) Is the overall system described by a linear or nonlinear differential equation ? Is it timeinvariant or timevarying ? b) For constant g > 0. +k 2 el(tt. k 2 . x 0 . f(t.) lx0 1 Problem 1.Ax x(to).) s.li Let Jc. a) X f(t. I x(l) I S:k3 + k 4 el(tt. x(t0 ) .x)=. e srnt and that the matrix A(t) bas eigenvalues that are independent of 1 and located in the lefthalf plane for I< a< 2. pT > 0 and Q. x).7 Consider the linear timeinvariant system Jc. What stability properties does the system have if g(t)>O for all t? Problem l. Show that k.x 1 +e' f(t. Problem 1. k3 and k4 as functions of I x 0 I. band I. uniformly stable. 0 and a > 0 as functions of the matrices P and Q such that lx(t)l$mea(tt.9 Determine which of the following functions are globally Lipschitz. Does e 0 > 0 as t> ? c) Let g(t) be a continuously differentiable function of time.O) ..xo Assume that there exist matrices P.x)={i c) e) d) !) exex f(t. QT > 0 such that AT P+PA+Q.
=(~ =!) As=(=~ ~) Aa=(=~\) A·=(~ ~) ~=(=!!) 117=(_1212) A= ( au Ozt an give the conditions on a 11 . L c Figure PI.ll. recall the RouthHurwitz test. or unstable. asymptotically stable. Problem 1. and so that it is stable in the sense of Lyapunov.x2 dt RC c =Xt dx2 dt I L b) Consider the "natural" Lyapunov function consisting of the total energy stored in the system v=zXt+zX2 C 2 L 2 ..11 Consider the circuit depicted in Fig. a 21 and a22 so that the equilibrium point is asymptotically stahle. Problem 1. a 12 ..5Hint: use the meanvalue theorem of elementary analysis to relate the Lipschitz constant to fJj /fJx. Pl. =Ax..= . Show that the system is described by the differential equations dxt I 1 .X t .t determine whether the following matrices correspond to an equilibrium point x = 0 that is stable in the sense of Lyapunov.10 a) Given the linear timeinvariant system . A lit=( 01 02) Az=( 21 43) i\·=U b) For a general matrix n Utz ) A.ll: Circuit a) Let x 1 be the voltage on the capacitor C and x 2 be the current in the inductor L. Him: to find simple conditions on the elements of the matrix.
a 2 . a 2• What stability properties of the equilib Problem 1. .12 a) Consider the system dx = x' dt x(O)=x0 Determine whether the function on the rightband side is locally Lipschitz.. d) Repeat parts a) to c) for the system dx _. and such that P :<> lxjZ For R = L = C = 1. What is the minimum nunber of vectors needed so tbat the matrix x. b) Show that tbe differential equation can be solved exactly and deduce what the stability properties of the equilibrium point are. c) Given an arbitrary vector x E R". for zero initial conditions . give the specific values of the constants rium can you deduce from this Lyapunov function ? a. or globally Lipschitz..!!. can be positive definite ? Problem 1. c) Find a Lyapunov function to confirm the conclusions of part b)._ u(s) s+a with a> 0.=e x dt x(O)=x0 e) Repeat parts a) to c) for the system dx 1 =X dt t+ 1 Problem 1..14 Consider the linear timemvanant system with transfer function H(s) = y(s) = .. but never positive definite.6What stability properties of the equilibrium can you deduce from this Lyapunov function ? c) Use the Lyapunov lemma to find a Lyapunov function such that a 1 1xl2 :<> v :<> a 2 1xl2 for some strictly positive constants a 1. show that the matrix A=xxT is always symmetric positive semiOOfinite.13 x(O)=xo a) Calculate the eigenvalues and eigenvectors of A= ( 5 3 3 ) 5 b) Find the decomposition A=UT AU and find a symmetric positive definite matrix A 112 such that A=A 112 A 112 .. Show that..
(R2) w=A.7llyll.=( b. Show that the plant transfer function may be realized by two statespace representations (Rl) and (R2): (Rl) x=A. Show that (Rl) is observable if 11.b. w+b. d) Using the PopovBelevitchHautus test. w 2(r)d.=a' and A.9) and (2. x+b. b'are as defined in equations (2. Him: The PopovBelevitchHautus test indicates that a statespace realization is controllable if and only if rank (sfA B)=dim (A) n for all s and observable if and only if rank ( C ) = dim (A) slA for all s Note that it is only necessary to check the rank for values of s for which det (sl. show that any unobservable mode must (indeed) be a mode of A. Problem 2.. r where A.. Find and plot .t(t)= Jw 0 .al .. S llhll1 Hull_ where II h 11 1 is the L 1 norm of the impulse response h(t) corresponding to H(s). and II y II_. from the eigenvalues of A0 . d. J ... In the same manner.1 a) Consider the identifier of section 2. deduce that the extra modes are those of A. so that (R2) is detectable.2 a) Let w(l)= sin (I). II u II_ are the L_ norms of y(l) and u(t) respectively. r y.2. 2 (r)dr b) Show that there exist a~o a 2 and 8 > 0 such that t 0 +6 a 2 .2.:. Problem 2. oT c. show that (R2) is controllable if P' JP are coprime. b'T J b) Show that (Rl) is controllable. are coprime. c) Show that (R2) is nonminimal and. a'.2.=c~ w where A.T A+~.=A+b...A)= 0.. b b1 =b.10). Give the value of II h 11 1 as a function of a and b and show that there exists a nonzero u{t) such that the inequality above becomes an equality.
gwwT . Is w persistently exciting? Why ? b) Consider the following proof of exponential convergence for=. What conditions must a. _61 I w2('<) . 6 >0. with w PE: Let v=.>O. d) Repeat part c) with w(t)=( asm (r+..P(t) w(l) wT(t) P(t) A>O .(r)) 8(0)=0(0) P(O)=PT(0)=P0 >0. Why is this proof incorrect? (4) Problem 2.3 a) Consider the function w(t) such that w(t)= 1 w(t)=O te [n.n+(l/2)"+ 1 ) otherwise. Show that. is positive definite? Problem 2. for all t0 .(to)~ .('<) ~ .(r)dr r0+d (2) '• ~2g. do not converge to zero exponentially..T (to +6) .2.....Q.(O)=.>O .)d. w(r)wT('<)d'< .(t0 +6) (3) where the last inequality follows from (I). satisfy so that the matrix R.. v(to+6) and exponential convergence follows as in the proof of theorem 1.) ~n(t) and replacing w2('<) by w('<) wT('<).5.T . /'(t)= A P(t).r(r)w('r)wr(r). ) d'l" whereR...4 This problem investigates the parameter convetgence properties of the leastsquares algorithm with forgetting factor B(t) =.. If w is PE. t 0+6] .y.(to+6) and ro+• (I) v(t0 )v(r0 +6)=2g J ..T ('<) ..T (to) .=gw2 .P(t) w(l) (BT (I) w(t). the solutions of . = lim to+l . this implies that v(to)v(to+6)<!:2ga. independently of t0 R. so that~=.8for all to:<!: 0.2 g (. and 'I"E [t0 . however... forallpositiveintegersn Plot w(t) and show that there exists 6 > 0 such that 'o+6 f w2(.o '• for all to:<!: 0. c) Show that the following limit exists. Therefore.T(t 0 +6) I ...T w) 2 SO.
.=a.(k+l)T] b) Extend the results of part a) to a general nth order system . However. and T (this is the soealled ste[rresponse discretetime equivalent of the continuoustime system). Problem 2.C. kD.9' I converges to zero as 1/1. a) Consider the first order system y.:Cx c) Adapt the results of section 2.< if P(l) is bounded. c) What optimization criterion does the leastsquares algorithm with forgetting factor solve ? d) Show that Ill( I) 11'1 converges to zero exponentially with rate .+k.tO. a. show that YD[k]:8 •• w[k] where (/ is a vector of unknown parameters and w is an observer state vector. subject to Yn[k]=8T[k] w[k] Using Lagrange multipliers.9' 1> 0 as 1> ~ if Jw(r)wr(r)dr:<:a(t)/ 0 t and a(l)> ~ as 1> ~. find expressions similar to (2. e) Show that P(l) is bounded if w is PE.0. What is the discretetime transfer function P(z): Yn(z)lfD(z)? forte [kT. Him: First derive d(r')l d1 and d(r' 8)/ d1.9a) Let . show that the solution of this optimization cnterion is . Show that if a(l) :2: k 1 for some k > 0. then 18(1). i.30).29) and (2.2 to the discretetime by describing an identifier structure for an nth order plant with transfer function P(z). The derivations follow lines parallel to the continuoustime case.r The continuoustime system is interfaced to a digital compoter through discretetime signals rv[k] and YD[k] such that r(l):rD[k] YD[k] = y.<: 0. as functions of k. show that YD and rD satisfy exactly a first{)rder difference equation YD[k] =aD YD[k 1] + kD rD[k 1] and find expressions for aD.e. see G. What is the effect of initial conditions in that case? d) The socalled projection algorithm is the estimate 8[k] such that l8[k]8[k1]12 is minimized. In particular. Goodwin & K. Sin ( 1984) for more details.t:Ax+br y.< .0.. Use eqn (2. when all the observer poles are at the origin.S.5 This problem constructs discretetime algorithms for the identification of sampled&ita systems.y.0. What if a :<:e'? b) For. Give conditions that the polynomial i(z) must satisfy and indicate the simplifications that arise when i(z) = z•.(k T) By solving the differential equation.30) to show that 18(1)..
Compare the convergence propenies.1 Br1 C A. Problem 2. R 2. g) Repeat part t) for the leastsquares algorithm. T=O.7 Let M(s) be a rational. • (or[k] w[j].10 IJ[k] = IJ[kl]. Problem 2. which minimizes J(IJ[k])= I.6 a) Find conditions on a 1 . Replace the denominator wr[k] w[k] by 1 + w 7 [k] w[k] to avoid possible division by zero or by a very small number.1 t) Simulate the responses of the identifier with the projection algorithm.0 1[k].Yo[k] wT[k] w[k] Show that IJ[k] is also the projection of IJ[k1] on the set of IJ[k] 's such that or[k] w[k]. P 22 [k] and Pdk]. YD[k]. Show that if M(s) is SPR.YD[k] =0.l s. a 2 .6) is SPR. C~o C2 (all> 0) such that the impedance Z(s)= V(s) I f(s) (see fignre P2. e) Find the (batch) /eastsquares estimate. b) Show that M(s) is SPR if and only if there exists a minimal statespace representation t=Ax+bu y=cT X such that c=b and A+ Ar =Q for some Q= Qr >0. show that the recursive formula for the leastsquares estimate is P[k] = P[k l] P[k1] w[k] wr[k] P[k1] 1 + wT[k] P[k1] w[k] k _ k _ P[kl]w[k](Or[kl]w[k]yD[k]) IJ[ ]IJ[ l] l+wT[k]P[kl]w[k] Hint: To find the difference equation for the covariance matrix. Let kP = 1. a) Let s(t) be the step response associated with M(s). p2 such that the transfer function isSPR.w[k] IJT[k1] w[k]. and all initial conditions be zero. b) Find conditions on R1. Plot rD[k]. the following identity is useful (A+B crt =A1 A1 B(l +C A. in addition to the other variables. and IJ2 [k] for rD[k]=sin (kT) and for rD[k] = 1. p1.Yo[j]) 2 }=1 Then. ap= 1. . strictly proper transfer function. then (ds) dt  >0 and t•O (dzs) dt2 <0 t•O Hint: Use the KalmanYacubovitchPopov lemma. Let P[O] =I and plot P 11 [k]..
Q=Q1 >0 such that and with [A.II  + "r( b) o' Figure P26: Circuit Hint: Use again the KalmanYacubovitchPopov lemma. A e R"'"'. w(x) is a known. be R". that is... globally Lipschitz function of x and the vector x(t) itself is available.=Ax. we Rm. Use the KalmanYacubovitchPopov lemma to find an adaptive control law such that x.+br . Assume that A is asymptotically stable and [A. show that If there exist matrices P=P 1 >0.j wl A1 r 1 P=(j ml A1 r 1 Q(j ml Ar' b) Consider the system t=Axb9 oT w(x)+bu where x..8 a) Prove one direction of the KalmanYacubovitchPopov lemma.. The vector 9" is completely unknown._. b] is controllable._ (JJ 2 Re M(j(JJ)>O Hint: Show first that P (j ml Ar' +(. e". Problem 2. and u e R. A and b are known exactly.Xm ~ 0 as t ~ ~ where t.b] controllable Then Pb=c M(s)=c 1 (sf Ar' b is stable Re M(jm) > 0 for all m and lim .
y 0 [k] Show that so that B(k] = 6l[k .BT[kl]w[k]yo[k] w[k] I+ wT[k] w[k] Define the a priori error e[k)=BT[kl) w[k). ue R"'.t is not available and one does not wish to differentiate x explicitly for noise considerations. including the boundedness of all the signals. linear error equation can be obtained.I:]=B[k]6' Using the results of part a). and that a gradient identification algorilhm can be derived.t and u are available. . Assume that x.l:] = 6l[k.1 + wT[k] w[k] w[.t=A* x+B* u where xe R". b) Let oT Yolk]=B w[k] and .Yolk) and the a posteriori error e•[k] = eT[k] w[k].9 This problem examines the extension of the singleinput singleoutput identification results to multivari· able systems where all the states are available for measuremenL a) Consider the statespace representation .e•[t] w[. A'e R""" and B'e R"""'. A' and B' are completely unknown. Problem 2. the modified projection algorithm is a gradient algorithm for the a posteriori error e•[k].(. What stability properties can you establish? b) Propose an algorilhm for the case when .1].1]. if .10 This problem investigates lhe stability properties of the projection algorithm. Show that a multivariable. Problem 2. show that c) A sequence x[k] is said to be in 1.12 where r e L_.l:] e[k] In other words. which is lhe discretetime equivalent of the gradient algorithm. a) Consider the modified projection algorithm B[k]=B[k1]. Prove the result.
and p converge and describe Ute type of convergence. Problem 2. . converges and describe the type of convergence (exponential.g pw~. 1 c) Consider Ute leastsquares algorithm withforgeuingfactor (i=.[O]I forallk e[k] (iv) (I+ wT[k] w[k])l/2 e 12 (v) (I+ wT[k] w[k])l/2 + 0 as t+ ~ t+~ . ·+ a 1 s• + P. Find the solutions of the differential equations. Indicate to what values.[kl]l+0 as Does (vi) imply that 9[k] tends to some limit ask+~ and.I. . Find the solution of the differential equation. w where e 1 .I¢[kl]ls. and p(O) = p 0 > 0.2 can be extended to proper transfer functions of the form Yp s f(s) • ( ) = P(s) = a•+ Is +a. P=gp 2 w~ with g > 0.1 + .13 ~ 1.. b) Consider the leastsquares algorithm (i=gpw~. '=I lx[k]JP <~ and x e 1~ if x[k] is bowtded for all k. Discuss the case when w0 =0. 0 arbitrary. wiUt g > 0 and ¢(0) = . a) Consider the gradienr algorithm (i=gwij. w are all scalars. Hinr: note that : 1 (¢ p.12 Consider Ute linear error equation e 1 =. let w(t) = w 0 be constant.)=0. ¢(0) = .s + .11 Show bow Ute identification meUtod described in section 2. if so.. Show that (i) l¢[k]I:S... . 0 .ntJ (vi) l¢[k].. .). Indicate to what value . asymptotic. is Ute limit necessarily e" ? Problem 2. ·+ P1 • •I Specialize Ute results to a fustorder system (n = I) to check the method. Discuss the case when w0 =0. Funner. s•.
Indicate ro what values . a) Show how x(B) can be obtained from /{8 1. . A> 0. .. 306310. ¢1(0) = ¢10 .).)SO and that outside S. The stabilizing term a p 2 is an interesting alternative to covariance resetting.)>0.8. the number being known. 1990.r x > 0 for 8 on the boundary ? . with g > 0. . Problem 2. Discuss the case when w0 =0.g(. d) Consider the stabilized /eastsquares algorithm with forgetting factor P=g pw~. 61 that the filters (sfb1 can be replaced by (at least) two other vector transfer functions. Problem 2. a) Show that the problem can be put in the form of a linear error equation of the type described in chapter 2. introduced ro keep the adaptive parameters within a specified region in which the nominal parameters are known to lie.b 0.)=0. See figureP2. P=. pointing outwards. ••• . and p{O) =Po> 0.A p+ p2w~+a p2) with g > 0. a> 0. no.(s). differentiable boundary given by f(B 1. Let x(B) be a vector perpendicular to the plane tangent to the boundary of Sat 8. Discuss the case when w 0 = 0. Him: it may be useful ro remember partial fraction expansions. vol.B2. What condition mustS satisfy so that .. and p converge and describe the type of convergence. and p converge and describe the type of convergence. The stabilized algorithm was proposed in the discretetime framework by G. Indicate to what values. .•.14P=g(Ap+p2 w~) Him: start again from~(. o. where S is a closed set with a smooth. Him: show that pis bounded away from 0 ro establish the convergence properties of (J. 35. on Automatic Control. Give the solution of the differential equation for p.e. ••• . indicate how the vecror e• is related ro the polynomials k. pp. Give explicit formulas relating the a's and {J's to the components of e• for the transfer function Ar' b) Repeat part a) for the other transfer function.(s) and J. Assume that Be S if and only if f(B" . Give the solutions of the differential equations.15 This problem investigates the properties of the projection modification.13 A periodic signal y(t) is measured.8. . It is assumed that a finite number of terms is sufficient ro represent y(t) exactly. f(8 1. 0 . in "Stabilized LeastSquares Type Adaptive Identifiers.15 for n=Z. Assume that 8 • e S. and p(O) = p 0 > 0. Kreisselmeier.(0) = . b) Determine whether the estimates are guaranteed to converge to the values of the coefficients of the Fourier series. and justify your answer. Its period T is known and one wishes ro design a recursive scheme ro estimate the coefficients of its Fourier series. p1).. Problem 2. It is sbown that a different projection modification is necessary for the leastsquares algorithm in order to preserve the properties of the Lyapunov function. n." IEEE Trans. 3. Give a recursive algorithm to estimate the coefficients of the Fourier series. For the first transfer function.14 a) Consider the assertion made in the book on p.
model reference adaptive control algorithm. for the gmdient algorithm (z =.g e 1 w). The scheme is an indirect. letting z=. Problem 3. Pu. v.Tw)2 d) Repeat part c) for the leastsquares algorithm.g P e 1 w.... the following projection can be used if/(11)=0 and(zT x)>O Optional: Show that this modification consists in projecting in the transformed space 8 space. However. note here that the properties of the Lyapunov function are not preserved by the projection modification. assuming that 11(0) E S.. P 22 . et. b) Show that the modification B. c) Show that. Prove also that. 9 = Z . as functions of z. X xT X if/(11)=0 and (zT x)>O =z otherwise consists in projecting the update vector z on the tangent plane when the boundary of the set S is reached and the vector is pointing outwards. P 12 . a) Consider the first order system . r 112 11 instead of the f) Consider the case when 11 e R2 and the set S is the set of 11's such that 111 :5 with 112 arbitrary.1S: Projection modification. 2g(.(zT X). z2 . and the Lyapunov function v = . Write explicitly what the update laws for 111 and 112 are.T r' .IS Figure P2. the projection modification guarantees that 11(t) E S for all t. and 11t (compare the gradient and the leastsquares algorithms).T. for v = .1 This problem considers the sampled{)ata control of a first{)rder system.g P w wT P.. 1'=. e) Show that to preserve both properties for the leastsquares algorithm.
aM are arbitrary.l)::(. a. I< 1..10) and a control law derived from part a).16The continuous time system is controlled by a digital computer through discrete time signals uv[k] and Yv[k] such that forte [kT. Let the reference model be given by • s+3 M= 7 (s+. Let k. b) Consider lhe polynomiallj in section 3. combining a discrete time identifier using the modified projection algorithm (cf problem 2.1 such that by proving that there exist unique polynomials sP=y+ j'.= I..s+5"") Simulate the response of the adaptive scheme for r(t) = 2 sin (1) and P= s+ I s2 + I ..2. T=O.2).=l. Problem 3. g~ E R and t•. wis the state of JS in a nonminimal statespace repre Him: First show that there exist uniqiU! f.2. but I a. Include a standard modification to the identification algorithm so that the estimate of the gain is bounded away from 0 (assume that the sign and a lower bound of the gain is known)..(k+l)T] Let the control law be given by Show that there exist nominal values of the controller parameters c• and d• such that the discrete time transfer function from r D to y D matches the reference model transfer function Ai(z)= 2H_ za11 where kM..I such that . b) Simulate the responses of an indirect adaptive control algorithm. in section 3. j' 7 t r" Indicate how J' and g' can be obtained from fl and £. Show that sentation (give the A.2 a) Consider the vector w.. and c matrices). is this polynomial guaranteed to be Hurwitz ? What can you conclude about the stability of the controller itself (as represented in Figure 3. b. Problem 3.. Let r 0 [k]=sin [kT].. Under the assumptions... g• E R. t of degree at most n.ls and M(z) be the discrete time equivalent of 3/(s+3).3 a) Consider the output error MRAC scheme of chapter 3.
Note: This problem illustrates the fact that two to!ally different unstable systems can be stabilized by a single adaptive controller. as the simulations show. and the interested student is invited to investigate other algorithms based on the leastsquares estimates.. the transformation from x to z defined by x(t)= z(t)+ e W(t) z(t) is onetoone and is differentiable with bounded derivatives.z(l) 1<2£ TIIA(. J 0 and W(l)= J(A(~)A. bounded. Further I x(l).. b) Repeat part a) with the plant JS=~ (s1)2 and the same adaptive controller.)z d) Show tnat if is exponentially stable.)IIM I z(l) I c) Show that z satisfies a differential equation t=e A. X( I)= e A(t) x(t) a) Let I T A. then X=£ A(t) x(t) is exponentially stable fore sufficiently small. z+e 2 (l +eW)1 (A W W A.e.. Again.=T A("r)d~ x(O)=x 0 where A(t) E R.)dT 0 I Show that W(t) is periodic and bounded b) Show that. and periodic function oft with period T. comment on the responses.. tne averaging results will be proved in the (simpler) linear periodic case. Poor convergence results are often observed with gradient algorithms. Plot e 0 and the adaptive parameters for 1000 secoods. However.17Let g = I and all initial conditions be zero.is a piecewise continuous.1 In this problem. Calculate the nominal values of the adaptive parameters and comment on the responses. t. Problem 4. for e sufficiently small. Hint: Use the Lyapunov lemma . some surprising responses can be observed within the coofines of the Lyapunov stability results...
B(t) are piecewise continuous and bounded. Find an expression for r.. by showing that If: A(t).. for x 0 = 0 and for llull. < .!:=f(t.5 Plot 1n ( x ) and In ( x~).. Problem 5. the following identifier is created w=apw+r .P+PwwT P) Problem 5.18Problem 4. show the asymptotic stability of the parameter vector and characterize the type of convergence. Find an expression for the averaged system..1 a) Consider the system t(l) = A(t) x(t) + B(t) u(t) A>0 Ptove a special case of theorem 5.l:=g(cos(r)+sin(2r)) 2 x g=0. Problem 4.x) such that x = 0 is a stable equilibrium point. b) Find an example of a system . and discuss the convergence properties of. and . and . llull. but . if x 0 "# 0. 1n particular. P=gPwwTP IP(O)=IPo P(O)=P(Ol =Po>O where a gain g > 0 has been included..2 The following problem illustrates the tradeoff between speed of convergence and robustness.. r.1. b) Repeat part a) with the leastsquares with forgetting factor P=g(A. when w is PE.25 b) . as a function of the constants m and a of the exponential stability definition. . One wishes to identify k r Exploiting the fact that a P is known. Assume that w is stationary.1: = f(t.... (ii) lx(t)l converges to a ball of radius r .2 Calculate the averaged system and simulate the responses of the original and averaged system for a) . llull.l:(t) = A(r) x(t) is exponentially stable Then: (i) llxll..3..3 a) Consider the least·squares algorithm p=gPwwT. and of the bounds on A(t) and B(t). a) Let the plant where k P is unknown and a P > 0 is known. x) + e has unbounded trajectories for any e > 0.. Solve for P..l:=gsin2(r)x 2 g=landg=0.
5. 9.o. 1. J I. k. Let r..w+r Simulate again the responses. P(O) I.PP 2 w 2 ) g>O J. .sin (I) and experiment with the values g . 25. 2.>O Explain how this identifier is re1nied to the algorithms presented in the book and what the nominal value of the parameter 8 is. and the other inital conditions be all zero. Simulate the responses for I.I9iJgPw(Bwyp) Pg(J.0. b) Replace the equation for w by wap. a.. I. with a.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.