Performance Analysis and Improvement of Deep Reinforcement Learning based Hybrid Beamforming Design for Multi-User Systems

For the future millimeter wave (mmWave) systems, communication using massive antennas has been considered for exploiting the large scale antenna array gain enabled by the decrease in wavelength. Hybrid beamforming (HBF) is a promising technology to balance between the hardware constraints and achievable data rate for the mmWave systems. However, precise beamforming is required since the beam width is very narrow as the number of antennas increases in such mmWave systems. This requires a high computational complexity and makes to apply HBF even harder as the number of users increases. To address this problem, we propose a deep reinforcement learning (DRL) based HBF design for multi-user multiple-input single-output (MU-MISO) systems. Simulation results show that the proposed DRL adopting a dueling architecture based HBF enhances the average throughput compared to the existing supervised learning (SL) based algorithms. In addition, the average throughput maintains competitive performance compared to unconstrained conditions without quantization loss and unsupervised learning (USL) based schemes.