Advanced Learning Architectures and Spatial Statistics for Beam Selection with Multi-Path

In this paper, we investigate the applicability of machine and deep learning (ML/DL) techniques to uplink beam selection problems with hybrid beamforming architectures. The goal is to select the element in the codebook of analog beamformers (ABF) yielding the highest sum-rate. The multi-antenna system operates in 5GNR's Frequency Range 2 (mmWave band). In this context, the presence of multi-path propagation along with the use of multi-carrier signals precludes the use of (single) angle-of-arrival information as an input to the learning system. Therefore, we investigate alternative statistics such as the singular vector/values of the (multi-carrier) channel matrix, the average covariance matrix, or the covariance matrix at a given subcarrier. Besides, we propose a novel ML/DL architecture enabling a continuous operation of the system which avoids the spectral efficiency losses associated to periodically switching to a dedicated ABF for the estimation of such statistics.