Swarm Intelligence Algorithm For Rbf Neural Network

RBF neural network has been successfully used in many fields,because of its capability of simple structure,fast training speed and good generalization ability.The choices of positions of hidden layer and the corresponding widths are very important and directly affect the goodness of fit of overall network approximation capabilities.This paper studies several swarm intelligence optimization algorithms and applies them in the training of RBFNN based on some typical structure optimization algorithms.The advantage of solving complex optimization problems makes swarm intelligence optimization algorithms show broad applications and strong vitality.This paper studies several swarm intelligence optimization algorithms including Bacterial Colony Chemotaxis(BCC),particle swarm optimization(PSO),Micro-artificial fish-swarm algorithm.These algorithms are improved and adapted separately then they can be suited to be used in the training of RBFNN.In the process of seeking the minimum points,this paper aims for the minimum of the generation error,and the positions of hidden layer and the corresponding widths are optimized at the same time.This paper proposes a new coding style and combines steepest descent algorithms in the BCC algorithms in order to increase the speed of calculation.The extended RPCL algorithms is used to confirm the number of particle and the re-initialization mechanism based in the global information feedback is introduced to the PSO algorithms to remain the activity of particle during training the RBFNN.And the experimental result is ideal.This paper proposes Micro-artificial fish-swarm algorithms based on artificial fish-swarm algorithms.The algorithms decrease the number of fish-swarm and increase the calculation speed and the activity of the fish-swarm is improved at the same time,so the generalization ability of the RBFNN is also improved.We apply the RBFNN optimized by these swarm intelligence optimization algorithms in experiments of pattern recognition.The results show that the generation ability is better than the RBFNN optimized by SGA and k-means algorithms.