A basis selection algorithm for wavelet neural networks

In this paper, a basis selection algorithm for wavelet neural networks is proposed. The algorithm repeatedly selects a wavelet basis to reduce residual error between network output and desired output. The newly selected wavelet is orthogonalized relative to all previously selected wavelets, and the optimal output weights are finally determined for all selected bases. The proposed approach has the advantage of being computationally less expensive than standard orthogonal least square algorithms. It also has better accuracy results than residual based selection algorithms. (C) 2002 Elsevier Science B.V. All rights reserved.