Pruning of RBF Networks in Robot Manipulator Learning Control

Radial Basis Function Neural Networks are well suited for learning the system dynamics of a robot manipulator and implementation of these networks in the control scheme for a manipulator is a good way to deal with the system uncertainties and modeling errors which often occur. The problem with RBF networks however is to find a network with suitable size, not too computational demanding and able to give accurate approximations. In general two methods for creating an appropriate RBF network has been developed, 1) Growing and 2) Pruning. In this report two different pruning methods which are suitable for use in a learning controller for robot manipulators are proposed, Weight Magnitude Pruning and Neuron Output Pruning. Weight Magnitude Pruning is based on a pruning scheme in [8] while Neuron Output Pruning is based on [2]. Both pruning methods are simple, have low computational cost and are able to remove several units in one pruning period. The thresholds used to find which neurons to remove are specified as a percent and hence less problem dependent to find. Simulations with the two proposed pruning methods in a learning inverse kinematic controller for tracking a trajectory by using the three first joints of the ABB IRB140 manipulator are conducted. The result was that implementing pruned RBF networks in the controller made it more robust towards system uncertainties due to increased generalization ability. These pruned networks were found to give better tracking in the case of unmodeled dynamics compared to the incorrect system model, not pruning the RBFNNs and a type of growing network called RANEKFs. Computational costs were also reduced when the pruning schemes were implemented. NTNU has a manipulator of the type ABB IRB140 and the learning inverse kinematic controller with pruning of RBF networks should be implemented and tested on this in real-life simulations.