A fast new algorithm for a robot neurocontroller using inverse QR decomposition

Abstract

A novel system identification scheme and adaptive control algorithm for a class of nonlinear systems is described, which is based on the computational properties of artificial neural network (ANN) models. The application of this to the direct control of robot manipulators using ANNs is then presented. An inverse QR decomposition (INVQR) and a weighted recursive least squares method for network weight estimation is derived using Cholesky factorisation of the data matrix. The use of higher derivatives means that the system can be linearised so that the resulting equations are linear, and can be solved for many of the weights simultaneously using some RLS algorithm. The purpose of the research described in this paper is to introduce a new class of linearised training algorithms for feedforward neural networks. The implementation of these algorithms with INVQR-WRLS is described.