Adaptive learning in incremental learning RBF networks

Abstract

This paper presents a modification to the incremental learning algorithm originally proposed by Bernd Fritzke.This algorithm is a single stage approach to build a RBF neural network in the training phase. It combines the unsupervised and supervised learning stages to incrementally generate the RBF architecture. The algorithm uses accumulated error information to insert new neurons dynamically in the hidden space. The algorithm maintains constant parameters to control the growth of the network. This may result in non-optimal architectures. Modifications that provide adaptive capabilities to the learning parameters have been proposed in this work. Two benchmark data sets have been used for training. Results show an improved performance over the original incremental learning algorithm. © Springer-Verlag Berlin Heidelberg 2004.