RECURRENT NEURAL NETWORKS FOR PILOT-AIDED WIRELESS COMMUNICATIONS

Recently, deep learning (DL) has been successfully applied in physical-layer communications and shown great success and competitive results to conventional systems. In this paper, we propose a novel recurrent neural network (RNN)-based communication system, based on the autoencoder concept. We develop a structure to mimic the working principle of a pilot-aided equalizer and integrate it as a learnable part of the system to support the task of channel estimation and equalization. The system shows competitive results under flat and frequency selective fading channels. The model can be trained to deal with any predefined number of channel taps (multipath components) of specific strengths. The system can also be generalized to deal with arbitrary strengths of the taps, which was infeasible in previous deep learning-based communication systems due to the absence of a guiding pilot. We assess the system performance for various alphabet and encoding sizes showing the BLER vs EBNO and the learned constellations.