Evolutionary multimodel partitioning filters for multivariable systems

It is known that for the adaptive filtering problem, the Multi Model Adaptive Filter (MMAF) based to the Partitioning Theorem is the best solution. It is also known that Genetic Algorithms (GAs) are one of the best methods for searching and optimization. In this work a new method, concerning multivariable systems, which combines the effectiveness of MMAF and GAs' robustness has been developed. Specifically, the a-posteriori probability that a specific model, of the bank of the conditional models, is the true model can be used as fitness function for the GA. Although the parameters' coding is more complicated, simulation results show that the proposed algorithm succeeds better estimation of the unknown parameters compared to the conventional MMAF, even in the case where it is not included in the filters bank. Finally, a variety of defined crossover and mutation operators is investigated in order to accelerate algorithm's convergence.