On normalising radial basis function networks

摘要: Normalisation of the basis function activations in a radial (RBF) network is common way achieving partition unity often desired for modelling applications. It results functions covering whole input space to same degree. However, normalisation can lead other effects which are sometimes less desireable This paper describes some side fundamentally alter properties functions, e.g. shape no longer uniform, maxima be shifted from their centres, and guaranteed decrease monotonically as distance centre increases -- many cases re-appear far centre. examines how these phenomena occur, analyses theoretically experimentally effect on least squares solution weights problem.