Pruning of RBF Networks in Robot Manipulator Learning Control

Abstract

Radial Basis Function Neural Networks are well suited for learning the systemdynamics of a robot manipulator and implementation of these networks in thecontrol scheme for a manipulator is a good way to deal with the system uncertaintiesand modeling errors which often occur. The problem with RBF networkshowever is to nd a network with suitable size, not too computational demandingand able to give accurate approximations. In general two methods for creating anappropriate RBF network has been developed, 1) Growing and 2) Pruning.In this report two dierent pruning methods which are suitable for use in alearning controller for robot manipulators are proposed, Weight Magnitude Prun-ing and Neuron Output Pruning. Weight Magnitude Pruning is based on a pruningscheme in [8] while Neuron Output Pruning is based on [2]. Both pruning methodsare simple, have low computational cost and are able to remove several unitsin one pruning period. The thresholds used to nd which neurons to remove arespecied as a percent and hence less problem dependent to nd.Simulations with the two proposed pruning methods in a learning inverse kinematiccontroller for tracking a trajectory by using the three rst joints of the ABBIRB140 manipulator are conducted. The result was that implementing prunedRBF networks in the controller made it more robust towards system uncertaintiesdue to increased generalization ability. These pruned networks were found togive better tracking in the case of unmodeled dynamics compared to the incorrectsystem model, not pruning the RBFNNs and a type of growing network calledRANEKFs. Computational costs were also reduced when the pruning schemeswere implemented.NTNU has a manipulator of the type ABB IRB140 and the learning inversekinematic controller with pruning of RBF networks should be implemented andtested on this in real-life simulations.