Fast and efficient training of RBF networks

Radial basis function (RBF) networks are used in many applications, e.g. for pattern classification or nonlinear regression. Typically, either stochastic, iterative training algorithms (e.g. gradient-based or second-order techniques) or clustering methods in combination with a linear optimisation technique (e.g. c-means and singular value decomposition for a linear least-squares problem) are applied to find the parameters (centres, radii and weights) of an RBF network. This article points out the advantages of a combination of the two approaches and describes a modification of the standard c-means algorithm that leads to a linear least-squares problem for which solvability can be guaranteed. The first idea may lead to significant improvements concerning the training time as well as the approximation and generalisation properties of the networks. In the particular application problem investigated here (intrusion detection in computer networks), the overall training time could be reduced by about 29% and the error rate could be reduced by about 74%. The second idea rises the reliability of the training procedure at no additional costs (regarding both, run time and quality of results).