AI-enabled Resource Allocation for BD-RIS Empowered ISAC Systems

Reconfigurable intelligent surfaces (RIS) have emerged as pivotal elements in enhancing integrated sensing and communication (ISAC) systems. Among RIS architectures, the beyond-diagonal (BD-RIS) design stands out for its advanced beamforming capabilities compared to traditional diagonal RIS. This paper investigates the deployment of BD-RIS to optimize energy efficiency through passive and active beamforming strategies while ensuring robust communication and sensing quality. The task is particularly challenging due to constraints inherent in BD-RIS configurations, including orthogonality, quartic inequalities, and the fractional nature of the objective function. To address these complexities, we employ twin delayed deep deterministic policy gradient (TD3) models, a state-of-the-art deep reinforcement learning (DRL) approach. Numerical validations confirm the effectiveness of our proposed algorithm and highlight the benefits of integrating BD-RIS into ISAC systems. Simulations demonstrate that incorporating BD-RIS leads to significant energy efficiency gains compared to benchmarks using diagonal RIS and RIS with random phase shifts.