Minimum prediction error neural controller

Abstract

Three approaches to adaptive control of nonlinear processes using artificial neural networks are presented. They are all based on minimum d-step-ahead prediction error control. All of them are capable of starting at random network weights, and the startup behavior was acceptable. The combination of classical stochastic approximation and the network as an associative memory was superior to more 'neural' approaches. The indirect approach is shown to be a suitable method when used with a recirculation network, making it possible to solve the predictive control with the network itself. The direct approach has a very strong tendency to converge to a local (sometimes unacceptable) minimum.<>