A New Kernel-Width Adaptation Method for Minimum Error Entropy

초록

Minimum error entropy (MEE) as a method of information theoretic learning has been effectively used in many applications such as channel equalization, machine learning and automatic control in Gaussian or non-Gaussian noise environments. The choice of kernel width is very sensitive and has important effects on system performance. The conventional kernel width adaptation methods based on optimization of error probability density estimation lead the kernel width to converge to the square root of 2 times of error variance (a very small value in the steady state), therefore, the error power increases in white noise situations and the width can have sudden rises in impulsive noise environments. In this paper, by applying average rate of change of error power over the small width interval to kernel width adjustment, a new kernel width adjustment method directly controlling system error is proposed. In the experiment of channel equalization, the proposed method shows the same stable learning convergences in both noise environments and yields significantly widened ranges of kernel width selection for the applications of MEE. © 2021, Korean Institute of Communications and Information Sciences. All rights reserved.

키워드