Geometric Constellation Shaping for Fiber-Optic Channels via End-to-End Learning

• DOI: 10.1109/jlt.2023.3276300
• Publication Date: 2023-05-15T18:35:24Z
• Authors: Ognjen Jovanovic; Francesco Da Ros; Darko Zibar; Metodi P. Yankov
• Keywords: Backpropagation

End-to-end learning has become a popular method to optimize constellation shape of communication system. When the channel model is differentiable, end-to-end can be applied with conventional backpropagation algorithm for optimization shape. A variety algorithms have also been developed over non-differentiable model. In this paper, we compare gradient-free based on cubature Kalman filter, model-free and fiber-optic modeled by split-step Fourier method. The results indicate that provide decent replacement in terms performance at expense computational complexity. Furthermore, quantization problem finite bit resolution digital-to-analog analog-to-digital converters addressed its impact geometrically shaped constellations analysed. Here, show when optimizing respect mutual information, minimum number levels required achieve shaping gain. For generalized gain maintained throughout all considered levels. Also, imply autoencoder adapt size given conditions.