Adaptive neural controller design for robots

Abstract

This paper investigates the feasibility of utilizing neural networks for the design of controllers for robots. It is argued that the usual error backpropagation (EBP) learning algorithm cannot be readily used for the training of neural controllers. Instead, in order to ensure the convergence of the training process and the stability of the closed-loop system, a stability approach must be taken to derive a learning algorithm. Viewing neural controllers as nonlinear adaptive controllers allows one to utilize the techniques developed in adaptive control theory. With this in mind, we use Liapunov's stability approach, in the same way as it is used for adaptive control systems, to develop a learning rule for neural network controllers that would guarantee the stability of the training process under mild conditions. These controllers do not require a priori information about the plant dynamics.