Abstract: Due to the unchangable structure of conventional RBF(radial basis function) neural networks in the learning process,a new structure of dynamic RBF neural network is designed.The presented algorithm is based on the sensitivity analysis(SA) on the model output of neural network.The redundant nodes in the hidden layer is pruned by judging the effect of the output of the nodes in the hidden layer on the whole network output,and thus the dynamic pruning of the neural network is accomplished.The results of approximating non-linear functions and modeling the non-linear systems show the perfect performance of this algorithm,and comparing with the Minimal RBF(MRBF) neural network,the proposed SARBF can achieve the smaller testing error,the shorter training time and a compact neural network structure.