Non-linear adaptive filters based on genetic algorithms with applications to digital signal processing

This paper presents the application of genetic algorithms to the on-line adaptation of non-linear adaptive filters-adaptive systems applicable to, for example, stochastic signal estimation, system identification and the optimization of electronic or optoelectronic signal processors. Given the filter topology, the corresponding filter parameters are estimated using a time-dependent moving error criterion. The genetic algorithm's ability to track temporal changes in the signal statistics is achieved by the use of partial hypermutation. The proposed methodology is applied to the problem of non-linear and non-stationary signal estimation by using the adaptive filter as part of a non-linear prediction error filter. Simulation results for the estimation of autoregressive and bilinear stochastic signal models and a comparison to the least mean squares algorithm are presented demonstrating the suitability of the approach.