Radial Basis Function Network Training Using a Fuzzy Clustering Scheme.

Training algorithms for radial basis function (RBF) networks usually consist of an unsupervised procedure for finding the centres and a supervised learning algorithm for udating the connection weights. Good netwiork performance will often be dependant on the RBF centre locations but the K-means clustering or related methods which are often used, can be sensitive to the initial conditions and this can result in local minima and a deterioration in overall network performance. In the present study, a fuzzy clustering scheme is implemented to locate the radial basis function centres in a manner which overcomes the sensitivity to initial conditions and improves overall network performance. Artificial and practical data sets are used to demonstrate the properties of the fuzzy clustering scheme.