Abstract: Aiming at designing the simplest RBF neural network architecture,a RBF neural network structure design algorithm based on mutual information was proposed in this paper．The relevance measure between each hidden and output units can be acquired by estimating the mutual information between the output matrix of the hidden unit and output unit,using k-nearest-neighbor statistics．And the simplest RBF neural network architecture can be achieved by removing the least related hidden units from the trained neural network one after another according to the relevance measure．This algorithm has the self-recovery mechanism,and the information processing capacity of the neural network can be ensured in the process of the simplification of the network’s architecture．The simulation results on the artificial datasets and the real-world benchmark datasets show the effectiveness and stability of the algorithm．

Key words: RBF neural network,Structural optimization,k-nearest-neighbor statistics,Mutual information