Variable Momentum Factor Algorithm for Nonlinear Principle Component Analysis

In this paper, a variable momentum factor algorithm is presented for improving the performance of the momentum term based nonlinear principle component analysis (PCA). Firstly, a smoothed error function is defined to describe the estimation error between the estimated separating matrix and its optimal value. Then, using a nonlinear function, the variable momentum factor is obtained according to the smoothed error function. Computer simulation results of adaptive blind source separation demonstrate that the proposed approach leads to faster convergence rate and lower misadjustment error than the momentum nonlinear PCA just with small increase in computational complexity.