Natural Gradient Improvement Methods in Blind Source Separation

This paper studies the characteristic of NGA (Natural Gradient Algorithm), and propose a set of improved natural gradient blind separation algorithm by applying data preprocessing and constructing learning factor and nonlinear function. For data preprocessing we use de-mean and whitening method to preprocess original data to reduce the amount of computation during iteration in BSS (blind source separation) greatly. The main work we do are study various learning factors and nonlinear functions for the natural gradient algorithm and propose a learning factors in iteration and two kinds of nonlinear functions for adaptive convergence. By the way the nonlinear functions can be used to separate both real and complex signals. The simulation results show that the paper constructed learning factor and nonlinear function are suitable for the convergence speed and precision, and can make the kernel function have adaptive convergence capacity and good stability also.