A three-phase RBFNN learning algorithm for complex classification

To improve radial basis function neural network (RBFNN) classification ability, this paper proposes a three-phase RBFNN learning algorithm. Firstly, the initial hidden structure of the network is determined by dynamic decay radius clustering algorithm. Then the hidden centers are modified by the sum squared error (SSE) rule, and the radius widths are calculated with the within-cluster and between-cluster distances. Finally the pseudo-inverse algorithm is utilized to train the weights between the hidden layer and the output layer. The experiments are implemented on Iris and Wines datasets, which shows that the proposed RBFNN training algorithm has a better classification ability compared with the conventional methods.