Performance Analysis of Learning Algorithms Based on Simplified Maximum Zero-Error Probability

초록

The common Mean Squared Error (MSE)-based learning algorithms are known to yield insufficient performance in non-Gaussian noise environments. In contrast, learning algorithms developed from the Minimum Error Entropy (MEE) criterion can overcome these obstacles. One of the MEE drawbacks is known for requiring sufficient error samples to correctly calculate error entropy, which in turn makes the system complicated. A recently proposed learning method for simple and efficient calculation of error entropy utilizes the difference between two consecutive error samples and in an iteration times k. Inspired by the fact that the variance of could be larger than that of or , this study proposed a new simple learning algorithm based on the Maximum Zero Error Probability (MZEP) criterion and its learning performance was analyzed through adaptive equalization experiment in a communication system model. The proposed simplified MZEP (SMZEP) shows convergence faster by about two times and lower steady-state MSE by about 1 dB than the simplified MEE (SMEE), indicating that the proposed SMZEP can be more appropriate for efficiency-requiring learning systems than the existing SMEE.

키워드