The radial basis function (RBF) is well known dynamic recursion neural network. However, RBF weights and thresholds, which are trained by back propagation algorithm, the gradient descent method and genetic algorithm, will be fixed after the training completing. The adaptive ability is bad. To improve RBF identification performance, particle swarm optimization (PSO), which is a stochastic search algorithm, is employed to train and adjust RBF structure parameter online. The simulation experiments show that PSO-NN has less adjustable parameters, faster convergence speed and higher precision in multimodal functions identification.