Neural Networks for Adaptive Control and Recursive Identification: A Theoretical Framework

Abstract

Massively parallel arrays of simple processing elements, the so-called “neural” network models, can be used to greatly enhance and extend techniques for identification and control of complex, nonlinear dynamic systems. However, the design of practical algorithms capable of ensuring prespecified performance levels requires a comprehensive, cross-disciplinary treatment, drawing techniques and insights from such diverse fields as machine learning, constructive approximation, nonlinear dynamic stability, and robust systems analysis. The goal of this paper is to review techniques for assembling all of these elements into an integrated and systematic framework, highlighting both constructive neural network analysis and synthesis methods, as well as algorithms for adaptive prediction and control with guaranteed levels of performance. These designs also contain supervised network training methods as a special case, and hence provide tools for evaluating the quality of a training set, avoiding overtraining, and maximizing the speed of learning.KeywordsNeural NetworkAdaptive ControlTracking ErrorOutput WeightNetwork InputThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.