Deep Learning based Beam Selection for Wide Beam Sweeping System

In massive MIMO systems, a large number of beam pairs should be tested during a beam sweeping process, which causes large overhead and latency. We study how to reduce the number of the beam pairs without the performance degradation of deep neural network based scheme and the additional feedback of hierarchical beam sweeping process. We perform the wide beam sweeping process using the smaller number of beams with the smaller number of antennas. The DNN using the received signals from the wide beam sweeping process selects a beam among the original beam candidates. The simulation results show that the DNN using the wide beam sweeping process with the reduced overhead and latency has the better performances than the conventional DNN based beam selection methods and the similar performances to the exhaustive search method in terms of the beam accuracy and throughput.