A Research On Decision Rules Of Fiber Coherent Optical Communication System Based On Machine Learning

As the development of emerging technologies such as Internet of things and 5thGeneration,the requirement of communication system capacity is increasing.The urgent need of higher capacity drives the revolution on modulation format.Compared with conventional square QAM constellation,the probabilistic shaping constellation has capacity which is closer to Shannon capacity and become a feasible solution of higher capacity.However,the decision algorithms in receiver have some deficiencies while the transmission technology is developing.The maximum a posteriori probability decision algorithm does not take the fiber-optic nonlinearity into account,and the compensation algorithm for fiber-optic nonlinearity has high complexity.Thus the current decision rules cannot meet the requirement of developing fiber-optic communication technology in future.Recent years,the machine learning(ML)method shows its powerful ability of learning and fitting data in many fields,and it is also applied to fiber-optic communication.In this paper,we take advantages of ML method which is good at fitting nonlinear function to construct classifier that can work properly with the influence of fiber-optic nonlinearity.Our work is present as followed:1.We analyze the common ML methods and select support vector machine(SVM)to construct the classifier.We construct several sub-classifiers to decrease the complexity according to the constellation.The SVM decision methods is more tolerant than MAP decision.With the simulation and experiment data,SVMbased decision method outperforms the MAP decision.2.We analyze the performance of ML under PS constellation and fiber-optic nonlinearity,and utilize the Markov Chain Monte Carlo(MCMC)to make up for deficiencies of SVM.The samples generated by MCMC have better performance than common methods as shown in Quantile-Quantile plot.The simulation and experiment data shows the MCMC-based SVM can work properly under PS constellation,and it has a Q value gain of up to 0.5d B compared with MAP decision.