Bi-GRU Enhanced Cost-Effective Memory-Aware End-to-End Learning for Geometric Constellation Shaping in Optical Coherent Communications

Abstract

We propose a cost-effective and memory-aware end-to-end learning scheme utilizing bi-directional gated recurrent unit (bi-GRU) for geometric constellation shaping (GCS) under the first-order regular perturbation (FRP) auxiliary channel. The performance of the proposed system has been numerically verified at a 32 GBd 5-channel wavelength division multiplexing (WDM) 64 quadrature amplitude modulation (QAM) 800 km optical coherent communication system. Results show that the proposed bi-GRU based GCS scheme can achieve a performance gain over square 64QAM in mutual information (MI) with 0.12 bits/symbol and a Q-factor gain of 0.4 dB at optimal launched optical power. When transmission distance is extended to 1280 km, a generalized mutual information (GMI) gain of 0.136 bits/symbol is observed. Additionally, compared with the bi-directional long short-term memory (bi-LSTM) based GCS, the proposed bi-GRU scheme has lower computation complexity with similar system performance.

Keywords

• Bi-directional gated recurrent unit
• end-to-end learning
• geometric constellation shaping