Abstract:

This paper investigates a hybrid self-generated training algorithm of RBF neural networks based on combining multi-encoding particle swarm optimization and back propagation (MPSO-BP).In the algorithm， each particle consists of binary and real parts which correspond to code the number of hidden units and parameters (i.e.， widths， centers and weights)， respectively.  Furthmore， a special fitness function is introduced to ensure accuracy with a few centers and to improve the self-adaptive and general capability of RBF neural networks. Simulation experiment results show that the proposed method has fewer hidden nods， higher accuracy， and better generality compared with other methods and the superiority is clearly.