Learning and Data-Driven Beam Selection for mmWave Communications: An Angle of Arrival-Based Approach

Abstract

This paper investigates how angle-of-arrival (AoA) information can be exploited by deep-/machine-learning approaches to perform beam selection in the uplink of a mmWave communication system. Specifically, we consider a hybrid beamforming setup comprising an analog beamforming (ABF) network with adjustable beamwidth followed by a zero-forcing baseband processing block. The goal is to select the optimal configuration for the ABF network based on the estimated AoAs of the various user equipments. To that aim, we consider 1) two supervised machine-learning approaches: k -nearest neighbors (kNN) and support vector classifiers (SVC); and 2) a feed-forward deep neural network: the multilayer perceptron. We conduct an extensive performance evaluation to investigate the impact of the quality of CSI estimates (AoAs and powers) obtained via the Capon or MUSIC methods, fluctuations in the received power, the size of the training dataset, the total number of analog beamformers in the codebook, their beamwidth, or the number of active users. The computer simulations reveal that performance, in terms of classification accuracy and sum-rate is very close to that achievable via exhaustive search.