Integrated SDR-Based Passive Beamforming and DRL-Driven Analog Beam Selection in IRS-Assisted Massive MIMO System

Abstract

This paper presents a deep reinforcement learning (DRL)-based hybrid beamforming (DHB) and semidefinite relaxation (SDR)-driven passive beamforming (SPB) 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 massive MIMO enables extended communication within a beam, hence increasing the system capacity and coverage. The optimization problem tackled in this paper is inherently complex and nonconvex, primarily due to the coupling of multiple variables and strict unit-modulus constraints. The discrete nature of beam selection further complicates the problem, rendering it NP-hard, as it involves combinatorial optimization over binary variables representing beam-user associations. To address this, we reformulate the problem into a quadratically constrained quadratic program (QCQP) followed by semidefinite relaxation (SDR) for IRS phase shift optimization. Then, based on the knowledge of the optimized IRS phase shifters, DRL is used to obtain the analog beamformer at the base-station. The DQN agent selects the beam-user pair with the help of the optimized cascaded channel state information to form the analog beamformer. Simulation results show that the proposed knowledge-based DRL beamforming design outperforms the state-of-the-art beamforming schemes in terms of transmit power, energy efficiency, and spectral efficiency. Specifically, the proposed design reduces the transmit power 20.37% for the target SNR of 10 dB.