MINE-based Geometric Constellation Shaping in AWGN Channel

The use of high-order constellation modulations is imperative to improve the spectral efficiency, for both radio frequency/laser-based satellite systems and optical wireless communications. The geometric shaping (GS) optimization as one typical constellation shaping method drives the improvement of communication capacity and system performance. This paper presents a novel mutual information neural estimation (MINE)based GS method to optimize the high-order constellations in pure additive white Gaussian noise (AWGN) channel, which uses the deep neural network (DNN) to estimate the mutual information (MI) value and maximize the MI to approach the AWGN capacity asymptotically. The proposed system trains both the encoder and MINE networks by back propagation, and does not need to train a decoder for optimization and thus can avoid the loss caused by the decoder. Simulation results show that the MINE-based shaping design outperforms the unshaped M-ary quadrature amplitude modulation (QAM) in terms of MI values. Note that the capacity gain increases slightly as the order M increases. Furthermore, the proposed scheme is promising for constellation design in various channel models, such as the phase noise and the fading channels, once the channel model used in MINE is matched, which can be a future research topic.