Improving Wavelet-Networks Performance with a New Correlation-based Initialisation Method and Training Algorithm

Wavelet-networks are inspired by both the feedforward neural networks and the theory underlying wavelet decompositions. This special kind of networks has proved its advantages over other networks schemes, particularly in approximation and prediction problems. However, the training procedure used for wavelet networks is based on the idea of continuous differentiable wavelets, but unfortunately, most of powerful and used wavelets do not satisfy this property. This paper presents a new initialisation procedure and a new training algorithm for wavelet neural-networks that improve its performance allowing the use of different kind of wavelets. To show this, comparisons are made for chaotic time series approximation between the proposed approach and the typical wavelet-network