A better selection of patterns in lazy learning radial basis neural networks

Abstract

Lazy learning methods have been proved useful when dealing with problems in which the learning examples have multiple local functions. These methods are related with the selection, for training purposes, of a subset of examples, and making some linear combination to generate the output. On the other hand, neural network are eager learning methods that have a high nonlinear behavior. In this work, a lazy method is proposed for Radial Basis Neural Networks in order to improve both, the generalization capability of those networks for some specific domains, and the performance of classical lazy learning methods. A comparison with some lazy methods, and RBNN trained as usual is made, and the new approach shows good results in two test domains, a real life problem and an artificial domain.