Collaborative Multi-Agent Deep Reinforcement Learning for STAR-RIS Assisted SWIPT: A Rate-Energy Pareto Frontier Analysis

Simultaneous transmission and reflection reconfigurable intelligent surfaces (STAR-RIS) have been proposed as a promising technique to improve the spectral and energy efficiency of future wireless communication networks. In this paper, we examine the joint rate and energy optimization problem in STAR-RIS assisted simultaneous wireless information and power transfer (SWIPT) systems using a collaborative Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MA-TD3) framework. Differing from existing single-agent solutions, the proposed solution allows multiple agents to work together to learn beamforming, power control, and mode splitting decisions in a manner that takes into account realistic constraints such as imperfect channel state information, nonlinear energy harvesting, and discrete phase shifts. The problem is formulated as a Pareto frontier analysis, which captures the rate-energy tradeoff for multiple users. Simulation results show that the proposed MA-TD3 solution can significantly outperform existing benchmark methods, achieving a near-optimal rate–energy trade-off while ensuring user fairness. Specifically, numerical results show that the MA-TD3 framework outperforms the conventional MADDPG method by 18.5% in sum rate convergence, and the STAR-RIS configuration offers a 30% larger coverage area for energy harvesting than the conventional reflecting surface. In addition, the work sheds light on the effect of STAR-RIS configurations, collaboration strategies among agents, and environmental uncertainties on the overall system efficiency. These results indicate the potential of collaborative learning to facilitate intelligent and energy-efficient wireless communication networks and facilitate the large-scale deployment of STAR-RIS assisted SWIPT in future 6G IoT and UAV-enabled communications.