Deep Learning-based Multiuser Physical Layer Communication Without Known Channel

With the recent development of deep learning (DL), DL-based autoencoder techniques provide a novel paradigm for end-to-end physical layer optimization. In this paper, we address the dynamic interference in an end-to-end communication system with a multiuser Gaussian interference channel. In this context, the standard constellation is not optimal under high interference conditions. To address this issue, we propose an adaptive learning algorithm for learning and predicting dynamic interference. Note that existing DL-based autoencoders are unable to train end-to-end learning systems by deep learning without a known channel. Thus, we propose a generative adversarial network (GAN)-based training scheme to imitate the real channel. Simulation results show that compared with traditional PSK and QAM modulation schemes, our proposed adaptive learning-based auto encoder can achieve significantly lower block error rate (BLER) in presence of interference. Besides, the BLER performance of our proposed GAN-based training scheme is close to that of the optimal training scheme with known channel on different channel models.