Community Research and Development Information Service - CORDIS

Abstract

In recent years, multilayer and recurrent neural networks have emerged as important and powerful tools for adaptive control of complex dynamical systems. A number of different neural network models and neural learning schemes have been applied to system controller design with varying degrees of success. Most of these are based on the back-propagation learning algorithm. The derivation of such a learning rule, which is a steepest descent optimisation technique, should be modified so that the adjustment rule minimises the error observed at the plant output and the network's error indirectly. However, systematic design methods, which guarantee the stability of the adaptation process and the success of the overall design, are still missing. This paper investigates a new approach in the design of learning algorithms for multilayer neural controllers; viewing neural controllers as nonlinear adaptive systems, Lyapunov's stability theory is used to derive adaptation laws which guarantee the stability of the training process under weak conditions, while requiring minimal information about the plant's dynamics. The concepts are illustrated by an application to robots' trajectory control.

Additional information

Authors: RENDERS J-M, ARC-Robotique, FSA, Université Libre de Bruxelles, Bruxelles (BE)
Bibliographic Reference: Paper presented: IMACS-IFAC International Symposium on Mathematics and Intelligent Models in Sys. Simul., Bruxelles (BE), Sept. 3-7, 1990
Availability: Available from (1) as Paper EN 35699 ORA
Record Number: 199011486 / Last updated on: 1994-12-02
Category: PUBLICATION
Original language: en
Available languages: en