Improved Non-negative Least Mean Square Algorithm using Gaussian CDF as Reweighting Function

Abstract

Adaptive filters have played a major role in signal processing these years. In systems identification problem, adaptive filters such as least mean square (LMS) are considered as beneficial approach because of their effortlessness and sturdiness. Due to physical restrictions in a realistic unknown target system, non-negativity constraint is usually trivial. Non-negative restraint is appropriate for unknown systems due to physical limitations: For more feasible solutions, it is better to impose non-negative limits into optimization problem. So, Non-negative least mean square (NNLMS) and its variations were put forward to solve problems of Wiener filtering with non-negative weights. In this paper we propose a new version of non-negative least mean square with considering CDF of Gaussian distribution as a useful reweighting function. The presented algorithm is analyzed in sparse system identification problem by Mont Carlo simulation in order to show that the simulation results follows the theory models properly with better performance. Furthermore, e compare our approach with other variations of NNLMS to confirm the advantage of our proposed GR-NNLMS algorithm.