Adaptive Intelligent Reflecting Surfaces for Enhanced Wireless Communication via Multi-Agent Deep Reinforcement Learning

This research introduces an innovative model-free framework for Intelligent Reflecting Surface (IRS) control in multi-user wireless communication systems using Multi-Agent Reinforcement Learning (MARL). Departing from traditional methods that rely heavily on detailed channel state information (CSI), this approach emphasizes the dynamic adaptation of IRS phase shifts to optimize overall system performance, measured in bits per second per Hertz (bps/Hz). The MARL framework enables decentralized decision-making among multiple agents, offering superior adaptability to complex and rapidly evolving channel conditions. Comparative analyses with established methods, including Multi-Armed Bandit (MAB), Random Selection, Deep Q-Network (DQN), and Hierarchical Reinforcement Learning (HRL), reveal substantial improvements in system performance. In particular, for a base station (BS) configuration with two antennas (<inline-formula> <tex-math notation="LaTeX">$N_{B} = 2$ </tex-math></inline-formula>), MARL achieves superior performance compared to baseline methods, demonstrating its robustness in limited-antenna settings. Similarly, an IRS deployment with 32 elements enhances MARL&#x2019;s efficiency by leveraging improved signal reflections, leading to significant performance gains. Furthermore, the impact of batch size and K-factor on MARL&#x2019;s adaptability is evaluated, confirming its ability to optimize phase shift adjustments across diverse network conditions. Simulation results confirm MARL&#x2019;s efficacy in optimizing resource allocation and phase shift adjustments, demonstrating its scalability and potential for real-time deployment in large-scale wireless networks.