Robust centre location in radial basis function networks

One of the most important and broadly accepted solutions for the centre selection problem in radial basis function networks is due to Moody and Darken (1989), who propose to cluster the input vectors via the k-means clustering method. In this paper, an alternative robust solution, based on the spatial median, is proposed and compared also to the k-medoids method (Kaufman and Rousseeuw, 1990). Some simulation studies in the context of classification problems show that, when outlying data are present, the solution we propose improves the network performance and allows to define a more parsimonious network.