A new adaptive RBF network structure learning algorithm

The network structure learning algorithm is an important aspect of network research. This paper proposes a new adaptive PBF network structure learning algorithm. The initial hidden network structure is determined by using forward selective clustering algorithm, and then a cluster sample transform algorithm based on impurity is proposed to adjust the hidden structure and we get the final hidden structure. After that we use the classical back-propagation algorithm to train the weights between the hidden layer and output layer. The experiment of two spirals problem proves that our algorithm can achieve higher training accuracy and testing accuracy in both the presence of noise and absence from noise.