Abstract

With the rapid growth of machine learning, neural network (NN)-based nonlinear compensation (NLC) for coherent optical communication systems have become widespread due to its robust nonlinear fitting capability, reduced complexity and real-time operation. NNs can adaptively evaluate nonlinear impairment and reduce computational complexity while improving performance as compared to digital signal processing (DSP)-based NLC algorithms such as digital back propagation (DBP) [1,2]. In this work, an NLC scheme based on Bernoulli-Restricted Boltzmann Machine (RBM) is proposed. This method factors the present symbol's nonlinear degradation to recover the transmitted symbols and aims to combine RBM's effective information extraction capabilities with a logistic regressor. The proposed semi-supervised model is based on a two-layered RBM which works in an unsupervised manner by learning the probability distribution over its sample training data inputs. The embeddings of RBM are fed to a logistic regressor (LR) based on LBFGS solver [3], for classification along with a target vector which consists of the true class labels. RBM-LR operates to discover latent factors in the data that can further be used to create a nonlinear decision boundary [4].

© 2023 IEEE