Attention-Based Spatial-Temporal GCN for Sensing-Aided Beam Prediction in RIS-Assisted ISAC Systems

Integrated sensing and communications (ISAC), recognized as a key technology of the sixth-generation (6G) communication system, simultaneously attends to dual functionalities of communication and sensing. This paper introduces a reconfigurable intelligent surface (RIS)-assisted two-stage ISAC system. The system utilizes uplink pilots to achieve sensing-assisted communication and implements predictive optimal beamforming to maximize the multi-slot system sum-rate. Conventional beamforming algorithms predominantly rely on perfect or estimated channel state information (CSI), which is idealistic or requires a large pilot overhead, making it unaffordable in mobile user scenarios. To address this challenge, this paper proposes a fusion framework named TD3-SATGCN, which integrates deep reinforcement learning (DRL) with an attention-based spatial-temporal graph convolution network (SATGCN) for non-convex joint beamforming. The proposed framework implicitly captures the spatial features from sensed user trajectories and pilots, which are further mapped into beamforming solutions in multi-slots without explicit CSI estimation. Furthermore, the poor generalization of artificial intelligence (AI)-based algorithms has hindered their deployment in communication systems. This article converts communication systems into graph topologies, harnessing the permutation/equivariance properties of GCNs to enhance generalizability. Simulation results under various scenarios indicate that the TD3-SATGCN reduces pilot overhead by 25% and achieves up to a 13.52% improvement in the system sum-rate compared to benchmarks without CSI.