DRL based Beam Selection and Hybrid Beamforming for Intelligent Reflective Surface assisted Massive MIMO System

This paper presents a deep reinforcement learning (DRL)-based hybrid beamforming design and beam-user selection for the intelligent reflective surface (IRS)-assisted multiuser massive multiple-input multiple-output (MIMO) downlink system. The conventional hybrid beamforming in massive MIMO provides multiple directional beams, but due to high absorption loss and low penetration power at millimeter wave (mmWave) frequencies, the distant or shadowed users suffer from low signal-to-noise ratio (SNR). The integration of IRS with the massive MIMO enables extended communication within a beam, hence increasing the system capacity. In this paper, we train a deep Q-network (DQN) agent with Adam optimizer. The DQN agent selects the beam-user pair from the cascade channel state information which is used for beamforming (BF) design. Simulation results show that the proposed DRL-based BF design outperforms the Greedy-based BF scheme by 7.52% and 16.47% with 128 and 256 IRS elements, respectively, at 20 dB SNR.