End-to-End Learning for OFDM: From Neural Receivers to Pilotless Communication

The benefits of end-to-end learning has been demonstrated over AWGN channels but has not yet been quantified over realistic wireless channel models. This work aims to fill this gap by exploring the gains of end-to-end learning over a frequency- and time-selective fading channel using OFDM. With imperfect channel knowledge at the receiver, the shaping gains observed on AWGN channels vanish. Nonetheless, we identify two other sources of performance improvements. The first comes from a neural network-based receiver operating over a large number of subcarriers and OFDM symbols which allows to reduce the number of orthogonal pilots without loss of BER. The second comes from entirely eliminating orthogonal pilots by jointly learning a neural receiver together with either superimposed pilots (SIPs), combined with conventional QAM, or an optimized constellation. The learned constellation works for a wide range of signal-to-noise ratios, Doppler and delay spreads, has zero mean and does hence not contain any form of SIP. Both schemes achieve the same BER as the pilot-based baseline with 7% higher throughput. Thus, we believe that a jointly learned transmitter and receiver are a very interesting component for beyond-5G communication systems which could remove the need and associated overhead for demodulation reference signals.

I. Introduction

End-to-end learning has attracted a lot of attention in recent years and it is considered to be a promising technology for future wireless communication systems [1], [2]. Its key idea is to implement the transmitter, channel, and receiver as a single neural network (NN), referred to as an autoencoder, that is trained to achieve the highest possible information rate [3], [4]. Since its first application to wireless communications [5], end-to-end learning has been extended to other fields including optical wireless [6] and optical fiber [7]. However, most of the literature is either simulation-based on simple channel models, such as additive white Gaussian noise (AWGN) or Rayleigh block fading (RBF), or experimental, but performed in static environments [3], [8]. Such setups do not account for the Doppler and delay spread encountered in practical wireless systems that lead to variations of the channel response in both time and frequency. The evaluation of end-to-end learning on more realistic channel models is overlooked in the existing literature but critical to bring the technology from theory to practice.