Blind Signal Processing based on Information Theoretic Learning with Kernel-size Modification for Impulsive Noise Channel Equalization

This paper presents a new performance enhancement method of information-theoretic learning (ITL) based blind equalizer algorithms for ISI communication channel environments with a mixture of AWGN and impulsive noise. The Gaussian kernel of Euclidian distance (ED) minimizing blind algorithm using a set of evenly generated symbols has the net effect of reducing the contribution of samples that are far away from the mean value of the error distribution. The process of ED minimization between desired probability density function (PDF) and output PDF is considered as a harmonious force interaction on PDF shaping between concentrating force and spreading force. The spreading force is composed of the difference between output sample values themselves, and is directly related with the output information potential and the output entropy that leads to the output distribution spreading out. The proposed kernel-size modification scheme is to impose loose discipline on the spreading force by employing larger kernel-size so that the long distance between two outputs which are correct symbol-related and far-located is less likely to be treated as impulsive noise. From the simulation results, the proposed kernel-size modified blind algorithm not only outperforms correntropy blind algorithm, but also significantly enhance robustness against impulsive noise. Key-Words: Impulsive noise, blind equalization, kernel-size modification, source symbol assignment, information-theoretic learning, PDF, correntropy, CMA.