Genetic Multi-Agent Reinforcement Learning for Multiple Double-Sided STAR-RISs in Full-Duplex MIMO Networks

Simultaneously transmitting and reflecting reconfig-urable intelligent surface (STAR-RIS) capable of manifesting the wireless channel provides the capability of signal reflection and refraction. However, conventional STAR-RIS has its limitation owing to signals impinging from one side of the surface, sup-porting either uplink (UL) or downlink (DL) users. Therefore, a novel concept of double-sided STAR-RIS (DS-STAR) becomes a promising solution, enabling signals impinging from both sides of the surface. In this paper, we consider multiple DS-STARs in a full-duplex (FD) enabled multi-input-multi-output (MIMO) system. We aim for maximizing joint UL/DL data rate by configuring transmit beamforming of the base station (BS) and UL users as well as configuration of DS-STARs, while ensuring quality-of-service (QoS) for both the UL/DL users. To tackle the complex problem, a genetic algorithm (GA) enhanced multi-agent Q-learning (G-MAQ) scheme is designed. MAQ considers a QoS-aware reward with each parameters as a sub-agent, whereas GA is applied to automatically optimize the hyperparameters of MAQ. In numerical results, we observe the significant im-provement of G-MAQ compared to that without hyperparameter optimization. Moreover, the proposed architecture of DS-STARs in FD networks achieves the highest rate compared to single-sided STAR-RIS, RIS and deployment without RIS/STAR-RIS. Additionally, the proposed G-MAQ scheme of DS-STAR FD sys-tems outperforms the other existing methods in open literature.