Deep Learning-Based Transceiver Design for Additive Non-Gaussian Impulsive Noise Channels

This paper presents a deep-learning (DL)-based transceiver design on additive non-Gaussian impulsive noise (IN) channels. At first, an impulsive generative adversarial network (IGAN) with specific regularization terms is proposed to capture the IN behaviours and used as the channel noise simulator (CNS). Subsequently, the transmitter and receiver are jointly optimized to develop both optimal transmit signal and detection. Furthermore, to enhance the detection performance, the multi-level wavelet signal recovery network (MWSRN) is applied to construct a preprocessor to combat both IN and channel fading without pilots. Illustrative simulation results show that the proposed scheme can achieve better convergence and bit-error rate (BER) performance under different IN settings than various existing approaches.