End-to-end Physical Layer Communication using Bi-directional GRUs for ISI Channels

Abstract

This paper presents a deep learning autoencoder for end-to-end physical-layer communications, in the presence of intersymbol interference (ISI) and additive white Gaussian noise (AWGN). Both the proposed transmitter and receiver employ the bi-directional gated recurrent unit (Bi-GRU) layers, and they are trained jointly with binary cross-entropy loss. By doing so, the transmitter learns tailored modulation constellations considering the channel impulse response while the receiver performs equalization and demodulation simultaneously. Experiments conducted over representative ISI channels reveal that the proposed autoencoder outperforms the Viterbi-based maximum-likelihood sequence estimation algorithm with perfect channel state information, when $E_{b}/N_{0}$ takes low to medium values, i.e., less than 13 dB.