Generalization of Adaptive Incremental Learning RBF Network Trained with Significant Patterns

摘要

In this paper we present a new idea of training the incremental learning RBF network with a set of patterns which are closer to decision boundaries amongst the different classes of training samples which constitute the input space. We call them significant patterns. We discuss the selection of significant samples from the data set and present learning and generalization characteristics of adaptive learning RBF network. The resutls are compared with the best known published recent works. It is found that incremental RBF network trained with significant samples can achieve very good generalization capabilities besides developing small optimal architecures compared to other methods.