Performance analysis of Wilcoxon-based machine learning nonlinear equalizers for coherent optical OFDM

Abstract

In the recent research on the mitigation of nonlinearities in CO-OFDM systems, it has been seen that various types of non-robust algorithms (based on minimization of least square error principle) are used for learning of nonlinear equalizer. Moreover, it is well known that performance of nonlinear equalizer learned by robust algorithms is not easily affected by the outliers. In this paper, some robust algorithms such as Wilcoxon Multilayer Perceptron (WMLP), Wilcoxon Generalized Radial Basis function (WGRBF) and Wilcoxon Robust Extreme Learning Machine (WRELM) for the performance enhancement of CO-OFDM system have been analyzed. Subsequently, the performance enhancement capability of both the algorithms i.e., robust and non-robust has been compared in this study. It has been observed that the nonlinear equalizers trained with Wilcoxon approach based learning algorithm offer improved performance in terms of Q-Factor as compared to non-robust algorithms. In this study K-means machine learning based training algorithm is used to cluster the points at their desired locations. From obtained numerical results, it has been observed that the improvement in Q-Factor with Wilcoxon multilayer perceptron algorithm w.r.t its non-robust solution is ~ 0.65 dB which is significantly higher than the value ~ 0.2 dB with both the other mentioned robust algorithms w.r.t their non-robust counterparts. From the comparison of robust algorithms performance on the basis of convergence rate, it has been professed that the WRELM converges 100 and 7 times faster than WMLP and WGRBF respectively.