Chapter 7 - Neural Network Control of Robot Arms and Nonlinear Systems

Abstract

This chapter presents new controller structures to initialize the neural network (NN) weights. NN controllers are designed that give guaranteed closed-loop performance in terms of small tracking errors and bounded controls. Applications are given to rigid-link robot arms and a class of nonlinear systems. Both continuous-time and discrete-time NN tuning algorithms are given. New NN properties – such as strict passivity – avoid the need for persistence of excitation. New NN controller structures avoid the need for preliminary off-line learning, so that the NN weights are easily initialized and the NN learns on-line in real-time. No regression matrix need be found, in contrast to adaptive control. No certainty equivalence assumption is needed, as Lyapunov proofs guarantee simultaneously that both tracking errors and weight estimation errors are bounded. To guarantee performance and stability in closed-loop control applications using multilayer (nonlinear) NNs, it was found that the standard delta rule does not suffice, but that the NN tuning rules must be modified with extra terms.