A NOVEL LEARNING METHOD FOR WAVELET NETWORKS

Wavelet networks(WN) has many merits and has been applied to system identification and pattern recognition. But it is not convenient to design the structure and initial parameters of WN. This paper presents a systematic method for WN design, named a two-step learning method. At first, wavelet networks will be constructed using the modified Gram-Schmit algorithm by combining with AIC. A parsimonious networks structure and initial parameters are obtained. Then, two interior parameters for wavelet networks, the dilation and translation factors, will be optimized by means of genetic algorithm(GA) and-hierarchical optimization algorithm. In this way, more accurate approximation can be realized without any increase of wavelet cells. Finally, this novel wavelet network is used to identify a nonlinear dynamic system, and the simulation results show that this method is effective and feasible.