Connectionist Learning for Control: An Overview

|This report is an introductory overview of learning by connectionist networks, alsocalled arti cial neural networks, with a focus on the ideas and methods most relevant to the controlof dynamical systems. It is intended both to provide an overview of connectionist ideas for controltheorists and to provide connectionist researchers with an introduction to certain issues in control.The perspective taken emphasizes the continuity of the current connectionist research with moretraditional research in control, signal processing, and pattern classi cation. Control theory is awell{developed eld with a large literature, and many of the learning methods being describedby connectionists are closely related to methods that already have been intensively studied byadaptive control theorists. On the other hand, the directions that connectionists are takingthese methods have characteristics that are absent in the traditional engineering approaches.This report describes these characteristics and discusses their positive and negative aspects. It isargued that connectionist approaches to control are special cases of memory{intensive approaches,provided a su ciently generalized view of memory is adopted. Because adaptive connectionistnetworks can cover the range between structureless lookup tables and highly constrained model{based parameter estimation, they seem well{suited for the acquisition and storage of controlinformation. Adaptive networks can strike a balance between the tradeo s associated with theextremes of the memory/model continuum. yThe author acknowledges the support of the Air Force O ce of Scienti c Research, Bolling AFB, through grantAFOSR{87{0030, which made this chapter possible, and the support of the King's College Research Centre, King'sCollege Cambridge, England, where much of it was written. Special appreciation is expressed to Chuck Anderson,Judy Franklin, Mike Jordan, Mitsuo Kawato, Rich Sutton, and Paul Werbos for useful discussions of the materialpresented in this chapter and many helpful suggestions on improving its presentation. This report is based on atalk given at the NSF Workshop on Neurocontrol, University of New Hampshire, October 1988. A version of thisreport will appear in Neural Networks for Control, T. Miller, R. S. Sutton, and P. J. Werbos (Eds.), The MITPress, Cambridge, Massachusetts.