Fast convergence algorithm for wavelet neural network used for signal or function approximation

Abstract

A new way to set the initial values of the wavelet neural network's parameters is proposed in order to improve the convergence speed. Experiments on linear polynomials, exponent functions, sin & cos functions and a certain multistage simulation function show the neural network has a much faster convergence speed and can be widely used for approximating many kinds of signals and functions. A discussion on the merit of this method is given. The experiment results are satisfactory.