Abstract: Nonlinear effect has been becoming a major bottleneck in high speed visible light communication (VLC) system. In this paper, we propose a supervised learning algorithm, support vector machine (SVM) for improving constellation classification in geometrically shaping (GS) VLC system. By taking the in-phase and quadrature components of the signal as feature vectors, an optimal classification plane can be built, and the symbol error introduced by nonlinearity could be therefore reduced. The performances of several GS designs are conducted and compared. Simulation and experimental results show that SVM could significantly reduce the error rate, compared with conventional classification scheme based on Euclidean distance. Among all simulations with SVM, the system with the data rate of 1.2 Gbit/s, circle169 GS-16QAM performs the lowest symbol error rate.

Key words: visible light communication(VLC), support vector machine(SVM), geometrically shaping(GS), anti-nonlinear