Beamforming and Load-Balanced User Association in RIS-Aided Systems via Graph Neural Networks

This paper considers the joint optimization of user association (UA), base station (BS) transmit beamforming, and reconfigurable intelligent surface (RIS) phase adjustment in a RIS-aided multi-BS multi-user-equipment (UE) system with load balancing. The problem is challenging due to variable coupling and nonconvexity. We propose a novel graph neural network (GNN) approach, leveraging permutation equivariance and invariance properties of GNNs for enhanced generalization compared to conventional deep neural network (DNN) methods. By using uplink pilots as input, our proposed GNN method eliminates the need for explicit channel state information (CSI), which can be challenging to acquire in RIS-aided systems. Simulation results demonstrate the superior performance of our GNN-based method over traditional UA strategies in terms of sum rate and load-balancing violation penalty. Moreover, the benefits of deploying RIS are illustrated from a new perspective of facilitating load balancing without substantial compromises in sum-rate performance.