A fractional filter based on reinforcement learning for effective tracking under impulsive noise

Abstract

It is valuable and meaningful to suppress impulsive noise in system identification. Existing algorithms usually only consider impulsive noise with small frequency and amplitude. Furthermore, few researchers pay attention to the tracking performance of these algorithms. This paper builds a framework based on the deep deterministic policy gradient (DDPG) algorithm with the ability to explore and correct. The enhanced fractional derivative is introduced to further improve the performance of this reinforcement learning-based framework. Thus a fractional least mean square filter algorithm based on reinforcement learning (FrlMS) is proposed. The stability of the FrlMS algorithm is analyzed. Compared with the competing algorithms, the simulation experiments in system identification show that the FrlMS algorithm has satisfactory tracking performance in the face of impulsive noise, especially when the frequency and (or) amplitude are (is) large.