Genetic algorithms for adaptive nonlinear predictors

In time series analysis adaptive predictors are used for the modeling of stochastic signals. Compared to linear filters, nonlinear predictors allow one to model a greater variety of signal characteristics, e.g., chaotic behavior and limit cycles. Nonlinear predictors require a suitable adaptation algorithm that is able to work in nonstationary environments. To this end genetic algorithms are proposed for the on-line adaptation of non-linear predictors. The methodology is applied to the problem of nonlinear and nonstationary signal estimation by using the adaptive predictor as part of a nonlinear prediction error filter. Experimental results for measured fire signals demonstrate the excellent adaptation properties of the genetic algorithm.