ADAPTIVE FILTERING WITH ADAPTIVE p-POWER ERROR CRITERION

In a previous paper, we have proposed a novel approach to choose an optimal p-power error criterion for adaptive filtering, by minimizing a Kullback-Leibler information divergence (KL-divergence). However, the method requires exact knowledge of the noise PDF, which reduces its practicality since in real life scenario the noise distribution is usually unknown. In the present paper, we propose a more practicable method, in which the optimal p value is approximately determined by using the error samples on-line, without resorting to a priori knowledge of the noise. The mean-square convergence of the proposed algorithm is studied, and an upper bound of the step-size is derived. Monte Carlo simulation results confirm the effectiveness and superiority of the new method.