ADAPTIVE CONTROL OF ILL-DEFINED SYSTEMS NATO CONFERENCE SERIES I Ecology II Systems Science III Human Factors IV Marine Sciences V Air-Sea Interactions VI Materials Science II SYSTEMS SCIENCE Recent volumes in this series Volume 5 Applied General Systems Research: Recent Developments and Trends Edited by George J. Klir Volume 6 Evaluating New Telecommunications Services Edited by Martin C. J. Elton, William A. Lucas, and David W. Conrath Volume 7 Manpower Planning and Organization Design Edited by Donald T. Bryant and Richard J. Niehaus Volume 8 Search Theory and Applications Edited by K. Brian Haley and Lawrence D. Stone Volume 9 Energy Policy Planning Edited by B. A. Bayraktar, E. A. Cherniavsky, M. A. Laughton, and L. E. Ruff Volume' 10 Applied Operations Research in Fishing Edited by K. Brian Haley Volume 11 Work, Organizations, and Techonological Change Edited by Gerhard Mensch and Richard J. Niehaus Volume 12 Systems Analysis in Urban Policy-Making and Planning Edited by Michael Batty and Bruce Hutchinson Volume 13 Homotopy Methods and Global Convergence Edited by B. Curtis Eaves, Floyd J. Gould, Heinz-Otto Peitgen, and Michael J. Todd Volume 14 Efficiency of Manufacturing Systems Edited by B. Wilson, C. C. Berg, and D. French Volume 15 Reorienting Health Services: Application of a Systems Approach Edited by Charles O. Pannenborg, Albert van der Werff, Gary B. Hirsch, and Keith Barnard Volume 16 Adaptive Control of III-Defined Systems Edited by Oliver G. Selfridge, Edwina L. Rissland, and Michael A. Arbib ADAPTIVE CONTROL OF ILL-DEFINED SYSTEMS Edited by Oliver G. Selfridge Edwina L. Rissland and Michael A. Arbib University of Massachusetts Amherst, Massachusetts Published in cooperation with NATO Scientific Affairs Division PLENUM PRESS· NEW YORK AND LONDON Library of Congress Cataloging in Publication Data NATO Advanced Research Institute on Adaptive Control of /II-defined Systems (1981 : Moretonhampstead, Devon) Adaptive control of ill-defined systems. (NATO conference series. II, Systems science; v. 16) "Proceedings of the NATO Advanced Research Institute on Adaptive Control of /11- defined Systems, held June 21-26, 1981 in Moretonhampstead, Devon, England"- T.p. verso. Bibliography: p. Includes index. 1. System theory-Congresses. 2. Adaptive control systems-Congresses. I. Selfridge, Oliver G. II. Rissland, Edwina L. III. Arbib, Michael A. IV. Title. V. Title: /11- defined systems. VI. Series. Q295.N37 1981 003 83-17699 ISBN 978-1-4684-8943-9 ISBN 978-1-4684-8941-5 (eBook) 00110.1007/978-1-4684-8941-5 Proceedings of the NATO Advanced Research Institute on Adaptive Control of III Defined Systems, held June 21-26, 1981, in Moretonhampstead, Devon, England © 1984 Plenum Press, New York Softcover reprint of the hardcover 1s t edition 1984 A Division of Plenum Publishing Corporation 233 Spring Street, New York, N.Y. 10013 All rights reserved No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording, or otherwise, without written permission from the Publisher PREFACE There are some types of complex systems that are built like clockwork, with well-defined parts that interact in well-defined ways, so that the action of the whole can be precisely analyzed and anticipated with accuracy and precision. Some systems are not themselves so well-defined, but they can be modeled in ways that are like trained pilots in well-built planes, or electrolyte balance in healthy humans. But there are many systems for which that is not true; and among them are many whose understanding and control we would value. For example, the model for the trained pilot above fails exactly where the pilot is being most human; that is, where he is exercising the highest levels of judgment, or where he is learning and adapting to new conditions. Again, sometimes the kinds of complexity do not lead to easily analyzable models at all; here we might include most economic systems, in all forms of societies. There are several factors that seem to contribute to systems being hard to model, understand, or control. The human participants may act in ways that are so variable or so rich or so interactive that the only adequate model of the system would be the entire system itself, so to speak. This is probably the case in true long term systems involving people learning and growing up in a changing society. Some kinds of true complexity arise when simple systems are put together into numerous and complex assemblies; one might hope that thermodynamics showed the way. But, alas!, the simple truths that thermodynamcis seizes from simply behaving molecules may well be unique in science; most other examples -- from schooling of fishes to economic models of simple consumers -- are just too intractable when they are built on any basis that corresponds roughly to reality. One kind of difficulty seems to arise inevitably from considering purposive behavior, whether directly by people or imputedly by complex mechanisms: that is the interaction of different purposes and goals. Sometimes the goals are in immediate v vi PREFACE conflict, so to speak, sometimes they are cooperative, but in real systems they are always embedded in hierarchical structures of goals subsumed under the values of the designers and operators. Nevertheless, although we cannot attain with some systems a degree of understanding and control that we might with others, that does not mean that we cannot strive to find out what controls are possible. The notion of control here does not necessarily mean just the dimensions of decision and manipulation; rather it also includes the concepts of improving understanding, gathering data, and making inferences about the feasible structures of affectors and effectors. That is, we are interested also in the science, as well as in the technology. The techniques of adaptive control are well enough in hand for the large class of dynamic systems where the plant and its dynamics can be analyzed a priori using the mathematical techniques of modern control theory. The purpose of this volume is to evaluate the control techniques that may be applicable when the system is not well defined, or may even be frankly unknown; such systems would include, inter alia: human control systems where the person is learning complex communication networks, such as those involving people biological systems economic systems man-machine systems No general rules are offered that will enable control in all, or even most, such systems. Rather, we analyze the state of the art, propose new approaches, and discuss possible applications and impl ications. Most of the papers were presented at a NATO Advanced Research Institute held in Moretonhampstead, Devon, England, from June 21 to June 26 of 1981 to analyze the current technology and the current problems, and to assess the possibli ties and to formulate a program. For the purpose of this volume, the papers have been grouped into six sections. Section 1 offers a variety of "Perspectives": a dialogue constructed by the Editors to dramatize a number of the issues addressed throughout the book; perspectives from the point of view of man-machine system, adaptive algorithms, and economics; as well as a historical view of adaptation in biology. Section 2 looks at adaptation and control using the well-defined mathematical techniques of control theory and nonlinear analysis, in contrast to Section 3 in which the planning techniques of Artificial Intelligence and the power of the computer take over from the closed forms of the mathematician. The final 3 sections look at different application areas in which issues of adaptation are raised with respect to ill-defined systems; "Motor skills" in Section 4, PREFACE vii "Language Acquisition and Adaptation" in Section 5, and "Development and Evolution" in Section 6. We thank Dr. Bayraktar of the NATO committee for making the Advanced Research Institute possible, Devon for being so beautiful in June of 1981, and our authors for all they contributed to the Institute and this volume. We also thank Barbara Nestingen for her invaluable editorial assistance. Massachusetts, March 1983 Oliver G. Selfridge Edwina L. Rissland Michael A. Arbib CONTENTS PERSPECTIVES A Dialogue on Ill-Defined Control. .. •..• 1 M.A. Arbib, O.G. Selfridge and E.L. Rissland Humans and Their Relation to Ill-Defined Systems 11 N. Moray Some Themes and Primitives in Ill-Defined Systems. . • • • . . 21 O.G. Selfridge The Use of Optimal Control in Economics. • • . . • • • • • • • 27 P. Rowlatt Biological Views of Adaptation - Some Historical Accounts. •• 39 F.E. Yates CONTROL THEORY AND NON-LINEAR ANALYSIS Adaptive Behavior in Manual Control and the Optimal Control Model. • . • • • • • • • • . • • • • 5l S. Baron Regulation, Feedback and Internal Models • • • • • • • • • • • 75 W.M. Wonham The Dynamics of Adaptation in Living Systems • • • • • • • • • 89 F.E. Yates ARTIFICIAL INTELLIGENCE Adaptive Control: From Feedback to Debugging •..•••••• 115 M.L. Minsky The Role of the Critic in Learning Systems ....•••••• 127 T. Dietterich and B.G. Buchanan ix x CONTENTS Examples and Learning Systems . . • • • . . . . . • . . • 149 E.L. Rissland Conceptual Models o~ III-De~ined Systems. . . . • . . • . .. 165 R.M. Young IDTOR SKILLS Creativity in Skilled Per~ormance 177 L.H. Sha~~er The Concepts o~ 'Adaptation' and 'Attunement' in Skill Learning. • . . . . • . . . . . 187 H.T.A. Whiting Visuomotor Coordination: From Neural Nets to Schema Theo~ ••••••••••••••••••• 207 M.A. Arbib LANGUAGE ACQUISITION AND ADAPTATION Richly Speci~ied Input to Language Learning . . . • 227 L. Gleitman and E. Wanner The Sensible Brain: Adaptation as a Positive and as a Negative Factor in the Reorganization o~ Neuropsychological Systems A~er Brain Damage 251 P. Lavorel DEVELOPMENT AND EVOLUTION Piaget and Education. 277 C.J. Brainerd Implications and Applications o~ Piaget's Sensorimotor Concepts ................... . 289 J. Churcher Failure is Not the Spur 305 M.A. Boden Genetic Algorithms and Adaptation • • . • • • • • . • . • •• 317 J.H. Holland CONTRIBUTORS. • 335 AUTHOR INDEX. • 337 SUBJECT INDEX • 343 A DIALOGUE ON ILL-DEFINED CONTROL Michael A. Arbib, Oliver G. Selfridge and Edwina L. Rissland Department of Computer and Information Science University of Massachusetts Amherst, MA 01003 USA To give an introduction to the concepts addressed in the present volume, the Editors have contrived a dialogue between two scholars, imaginatively naned A and B, which reflects the current state of our inter-discipline. It captures many of the views expressed in general discussions throughout the meeting. A. When we speak as technicians seeking to understand control strategies, a system is a "them," as distinct from the controllers who are an "us"-dividing the universe into "us" and "them." The task of control theory is to provide an algorithm for how "us" is to control "them ," and such an algorithm must be based on criteria of optimality or satisfiability, but the appropriate criteria are often unclear in an ill-defined situation. For example, in writing a technical paper we have the conflicting criteria of "keep it short" and "provide the reader with background material." Very often we do not have a single criterion for our control strategy. The different criteria may even be mutually inconsistent. Of course, if we had a well-defined criterion for trading off amongst these criteria, then we could optimize on that weighted criterion function. But, in general, we do not. Thus, one aspect of our concern with ill-definedness is not adaptive control of ill-defined systems, but ill-defined criteria for control of systems. B. My most horrifying example is based on the fantasy that finally we have incredible computers which control the economy and production and the weapons system and everything else, and we