An adaptive learning algorithm aimed at improving RBF network generalization ability

This paper proposes a new adaptive learning algorithm of network structure aimed at improving RBF network generalization ability. The algorithm determines the initial number and center vectors of network hidden units by using forward selective clustering algorithm with decaying radius, and then adjusts them by using cluster sample transform algorithm based on impurity and variance and gets the final center vectors. The determination of widths of hidden units considers both the dispersivity of inner samples and the distance between clusters. Thus we get the final hidden structure. After determining the hidden structure, the back-propagation algorithm is used to train the weights between the hidden layer and output layer. The experiment of two spirals problem proves that our algorithm has higher generalization ability indeed.