Robust incremental normalized least mean square algorithm with variable step sizes over distributed networks

Abstract

An improved incremental normalized least mean square (INLMS) algorithm is developed by minimizing the Huber cost function, which is robust against impulsive noises, over distributed networks. To significantly suppress impulsive noises, a recursive scheme based on the incremental cooperation strategy is designed for updating the cutoff parameter in the Huber function. Since the proposed algorithm can be interpreted as a variable step size INLMS algorithm, it has faster convergence rate and lower steady-state error than some existing incremental distributed algorithms in both impulsive and non-impulsive noise environments. In addition, to track a sudden change of the unknown system, a modified method of resetting the cutoff parameter is developed.