An Enhanced Structure Learning Mechanism for Determining the Structure of Radial Basis Neural Networks

Abstract

In traditional approach it takes signi¯cant time and e®ort to determine the struc- ture of a neural network because it is determined by carrying out a trial and error approach. The proposed method is capable of ¯nding neural network parameters, especially the number of hidden nodes with a minimal time and e®ort instead of us- ing the traditional trial and error approach. We used cluster validity mechanisms to identify the number of hidden nodes and fuzzy c-means clustering algorithm to iden- tify the hidden node parameters (center and width).To evaluate the cluster validity we slightly modify existing cluster validity index called Xie-Beni's separation index and determine the best number of clusters in dataset using that approach. In ad- dition to these parameters we adaptively change the learning rate parameter during the training process. Results show that the proposed approach gives better classi¯- cation rates compared to the traditional approach for the datasets we used. Also it gives similar classi¯cation rates when compared with the best results reported in the literature. Therefore we can consider this algorithm as an alternative approach to set the number of hidden nodes and other learning parameters automatically while the training is carried on.