Robust Adaptive System Identification and Noise Filtering Using Bregman Divergence Based Algorithms

This article proffers a new set of least mean square (LMS) algorithms based on the Bregman divergence class of information-theoretic divergence, which are named as Kullback-Leibler LMS (KLLMS), Itakura-Saito LMS (ISLMS), and ßLMS. To achieve a tractable analysis, a bound on the learning rate of proffered algorithms is derived. Steady-state mean-square deviation (MSD) and computational cost provide insight into the performance and efficiency of the proffered algorithms in the presence of Gaussian and impulsive noise. Simulation results also show the effectiveness of the proffered algorithms in adaptive noise filtering.