Joint Beamforming and Metasurface Reflection: A Lightweight Design for Energy Efficiency via Deep Reinforcement Learning

Intelligent reflecting surfaces (IRSs) continue to gain a growing research interest for their potential to support next-generation wireless communications without incurring additional power consumption. In this work, we propose a deep reinforcement learning (DRL)-driven and IRS-aided active/passive beamforming solution for multi-user multiple-input single-output (MISO) settings in beyond 5G networks, which is both lightweight and energy-efficient. The proposed solution is based on a hybrid finely-engineered design that leverages two Twin-Delayed DDPG (TD3) agents. Compared to classical optimization techniques, our numerical evaluation shows that the proposed DRL approach achieves 60% reduction in online computation complexity at the expense of only 1 dB higher power consumption.