Signal Restoration and Prediction for End-to-End Learning of Practical Wireless Communication System

Abstract

Autoencoders (AEs) have been proposed to learn the physical layer of wireless communication systems in an end-to-end fashion. However, it is challenging to jointly train the transmitter part and the receiver part of an end-to-end AE, since closed-form transfer functions of wireless channels are unknown in practice. In this paper, two solutions, namely signal restoration and signal prediction, are proposed to address this challenge. Deep neural networks (DNNs) along with deep AEs are employed to implement these solutions. And these solutions are further evaluated in a practical wireless communication system with superheterodyne architecture, bandpass channel noises, and quantization noises. Evaluation results show that with the aid of the proposed solutions, end-to-end AEs can be jointly trained and perform better. The proposed signal restoration solution also helps improve the performance of conventional modulation schemes, e.g., quadrature amplitude modulation (QAM).