An improved method of wavelet neural network optimization based on filled function method

BP algorithm of neural network don´t obtain global minimum sometimes, furthermore, it is possible to create many local minimum so that the optimum solution can´t be found. In order to solve this question, one parameter filled function method is presented which can calculate value fast. We combine it with modified BFGS (Broyden-Davidon-Fletcher- Powell) to get a new algorithm for global optimization of wavelet neural network. The algorithm obtain the first local minimum by BFGS, then filled function method is used to obtain another smaller local minimum, this process is repeated for some times so that the network structure and weight value are optimized till global minimum is found. This method is used to train Shanghai stock index, then better network performance is obtained.