Abstract: Wavelet neural network is a new type of neural network which combines wavelet theory and BP network and has become a popular tool in approximation and forecasting fields. To improve the performance of the wavelet network, the strategy of evolving wavelet neural networks is proposed, which uses hierarchical chromosome to encode the structure and parameters of the wavelet network, and combines genetic algorithm and evolutionary programming to design and train the network simultaneously through evolution. By means of this strtegy, not only the local minimum problem in training can be overcome and the convergence performance can be improved, but also the structure of the network can be optimized. With respect to function approximation, sunspots time series forecasting and condition forecasting for hydroturbine machine, experimental results are presented to show the superior performance and potential of the proposed hierarchical evolving algorithm based wavelet network.

Key words: Forecasting, Evolutionary computation, Evolutionary programming, Genetic algorithm, Wavelet neural network, Dual algorithm, Multiple load cases, Multiple objective optimization, Structural compliance, Structural topology optimization