Combining evolutionary and stochastic gradient techniques for system identification

In the present contribution, a novel method combining evolutionary and stochastic gradient techniques for system identification is presented. The method attempts to solve the AutoRegressive Moving Average (ARMA) system identification problem using a hybrid evolutionary algorithm which combines Genetic Algorithms (GAs) and the Least Mean Squares LMS algorithm. More precisely, LMS is used in the step of the evaluation of the fitness function in order to enhance the chromosomes produced by the GA. Experimental results demonstrate that the proposed method manages to identify unknown systems, even in cases with high additive noise. Furthermore, it is observed that, in most cases, the proposed method finds the correct order of the unknown system without using a lot of a priori information, compared to other system identification methods presented in the literature. So, the proposed hybrid evolutionary algorithm builds models that not only have small MSE, but also are very similar to the real systems. Except for that, all models derived from the proposed algorithm are stable.