Generative Adversarial Network and End-to-End Learning for Optical Fiber Communication Systems Limited by the Nonlinear Phase Noise

There is an exponentially growing demand for communicating information. Optical fiber represents the most wideband communication known to date but its communication performance is limited by the effects of fiber nonlinearity. In this paper, we present an end-to-end model for geometric constellation shaping that can be used for any nonlinearity-limited optical telecommunication channel. Our approach is based on state-of-the-art methods in deep learning: generative adversarial network and end-to-end system learning via an autoencoder. We argue that our proposed implementation is capable of determining the optimal geometric constellation shaping for any channel but for specific evaluation we used a well known Nonlinear Phase Noise Channel. Our model outperformed the conventional QAM constellation in terms of symbol error rate and resulted in high performance while being robust to variations of the input power and not requiring additional retraining in the large range of the power levels. We have simultaneously trained a generator capable of emulating the genuine channel and an autoencoder designed to find the optimal constellation shaping for the imitated channel. These results show that our system can determine an optimal constellation shaping even in the channel-agnostic environment.