Constellation Shaping for Phase Noise Channels with Deep Learning Approach
Amir Hossein Omidi, Ming Zeng, Leslie Ann Rusch · 2022
Constellation shaping can significantly enhance the capacity of communication systems without incurring extra bandwidth. The statistics of the noise in the system dictate the appropriate constellation geometry. While additive white Gaussian noise (AWGN) channels have been widely studied, the more challenging phase noise channels are rarely investigated. To fill in this gap, in this paper we optimize the constellation shaping for phase noise channels using a deep neural network (DNN) end-to-end learning approach called autoencoder. The proposed DNN method is compared with state-of-the-art baseline algorithms in terms of symbol error rate (SER). It is shown via simulation that our method provides improved SER for all considered phase noise values, with significant improvement at the high phase noise regime.