RBF network pruning techniques for adaptive learning controllers

Abstract

This paper presents two simple and efficient methods for pruning a Radial Basis Network (RBF) used in an adaptive controller architecture for a robotic manipulator. The methods presented in this paper are Weight Magnitude Pruning (WMP) and Node Output Pruning (NOP). The above pruning methods are simulated on a trajectory tracking task of a three degree of freedom robotic manipulator arm. The RBF based inverse dynamics controller is presented with a task of learning the inverse dynamics of the plant in a closed loop control. Simulation study shows that implementation of an inverse dynamics control law in such manner makes the controller more robust towards uncertainties and disturbances. Pruning RBF network improves controller performance in the case of modelling errors and reduces computational costs, thus making such controller more suitable for implementation.