Combination of momentum term based natural gradient algorithm with variable smoothing factor

The convex combination of two momentum term based algorithms with different momentum factors is an effective solution to highlight the tradeoff between convergence rate and steady-state error of a blind source separation system. As the smoothing factor is chosen in the range from 0 to 1, however, the performance of this convex combination is restricted. In this paper, a novel variable smoothing factor approach is proposed to optimize the proportions of the two separating subsystems. The value span of the smoothing factor is effectively expanded by modifying the updating rule for each BSS subsystem, which leads to a better tradeoff between convergence rate and steady-state error. The simulation experiments showed the good performance of the proposed variable smoothing factor method for separating the instantaneous mixed signals.