A Concurrent Training Method of Deep-Learning Autoencoders in a Multi-user Interference Channel

Abstract

Autoencoder (AE) has been proposed recently as a promising, and potentially disruptive approach to design the physical layer (PHY) of future beyond-5G networks. In this paper, we propose a novel concurrent and interactive training method for autoencoder-based physical layer for the multiuser interference channel scenario, which is highly relevant to beyond-5G dense systems. In our approach, AEs use a concurrent training approach in order to minimise the interference in the presence of other AEs. Deep learning-based AE design for a multi-user interference channel is proposed in a dense urban environment where the AE is adapted to the interference from nearby intelligent base stations (BSs) by considering other BSs randomly distributed over 200 meters. Interference is modelled based on a 3GPP channel model in an urban microcell environment. We simulate and test the system, and compare the performance with a conventional QPSK modulation in an additive white Gaussian noise (AWGN) channel. Results show that at low signal-to-noise ratios, the AE in an interference channel delivers similar performance to that of QPSK in a noisy channel. However, at high signal-to-noise ratios, every trained AE outperforms the theoretical limit for QPSK systems in a noisy channel, in an interference-limited environment. The learned constellations is studied. We show that AEs can augment their own transmitted symbols based on the interference from nearby BSs via a learning process, which can counteract the effects of channel noise.