Adaptive Kernel Adjustment for Unsupervised Learning for Equalization

초록

Information-theoretic learning (ITL) has been successfully applied to machine learning, state estimation, and signal processing under non-Gaussian noise environments. The kernel width for ITL algorithms based on the linear combiner structure is very sensitive to signal processing conditions, and even a carefully chosen and fixed width cannot be readjusted in accordance with changing stochastic situations. In order to solve this critical problem, adaptive kernel adjustment methods for unsupervised information-theoretic learning algorithms under impulsive noise are presented in this paper. The conventional kernel adjustment method designed to maximize similarity between the actual distribution and the estimated one is revealed in this paper to be unacceptable under impulsive noise. But the proposed approach, which minimizes the averaged power of decision-error and employs its average rate of change as a gradient for kernel width update, showed fast and stable convergence and produced output samples converging rapidly to each symbol point.

키워드