Previous application of wavelet theory to nonlinear function and dynamic system approximation has produced networks that lack parameter interpretability. Although wavelet neural networks can model any system, with suitable training, they do not contribute to an explanation of the underlying system dynamics. Orthogonal wavelets, however, may offer a useful route to transparent models. This paper introduces a new technique for constructing orthogonal wavelet networks based on orthogonal least squares. Problems with conventional regularization are highlighted and a heuristic solution is proposed.
展开▼