Constellation Design with Metaheuristic Deep Learning for Future Wireless Networks

Geometric constellation shaping is a promising technique for improving the transmission capacity of future communication systems. In this paper, we propose a hybrid automatic repeat request scheme with chase combining (CC-HARQ) that utilizes a metaheuristic learning-based neural network for geometric shaping. Based on a fixed mapper for the first transmission, we employ an autoencoder to train the geometric symmetry constellation used for retransmission, with metaheuristic initialization and data preprocessing. Simulation results show that the proposed CC-HARQ scheme significantly outperforms the conventional square constellation rearrangement scheme in terms of bit error rate performance and is comparable to the incremental redundancy HARQ scheme specified by the 5G standard.