A PSO-based subtractive clustering technique for designing RBF neural networks

When designing radial basis function neural networks, the central task is to set parameters of radial basis functions. In this paper, subtractive clustering is improved by particle swarm optimization (PSO) to automatically select the number and locations of radial basis functions. Subtractive clustering is used to find center prototypes and then PSO Flues their locations iteratively. Comparative experiments were executed between subtractive clustering and PSO-based subtractive clustering proposed in this paper for designing RBF neural networks on several datasets. The experimental results suggest that the PSO-based subtractive clustering algorithm can be successfully applied to design RBF neural networks with competitive classification accuracy and small number of radial basis functions. The RBF neural networks evolved by PSO-based subtractive clustering have stronger generalization ability than the ones evolved by subtractive clustering.