Abstract: Suitable architecture is critical for a neural network to model time series and it determines forecast performance.So far,most methods assume that the architecture will keep fixed once it has been trained.However,many time series are nonlinear and their structures often change,which requires that the architecture of network should vary with time.Therefore,a variable structure radial basis function (RBF)neural network is proposed in this paper.Furthermore,sequential Monte Carlo(SMC)method is applied for time series on line prediction in the variable structure RBF network model.At last,the data of CRU global steel price index are analyzed,and experimental results indicate that the variable structure RBFnet work proposed is effective.

Key words: prediction, time series, radial basis function neural networks, sequential Monte Carlo method