Adaptive multiresolution filtering to forecast nonlinear time series

There are two ways to improve the identification process of a dynamic system using an artificial neural network: 1) preprocessing the training values to extract characteristics of the data; and 2) adapting the architecture of the network. In this paper we used an adaptive scheme of multiresolution filtering to decompose the series into other series for an easier analysis. The scheme proposed uses genetic algorithm to find the optimal bank of filters without previous knowledge of the behavior of the system to be identified. A new variation of the algorithm using random individuals is proposed to avoid local minima. The objective function proposed is the estimation quadratic error of a multilayer perceptron using the Levenberg-Maquardt learning.