Investigating Large-Scale RIS-Assisted Wireless Communications Using GNN.

Channel estimation (CE) in reconfigurable intelligent surfaces (RIS)-assisted wireless communication systems is challenging when using traditional CE methods due to their computational intensity and inaccuracies, especially in large-scale RIS environments. These limitations directly impact the achievable data rate, which relies heavily on accurate channel state information (CSI) obtained from CE. To overcome these challenges, we propose a novel approach that utilizes graph neural networks (GNN) with region-specific training models. The GNN is employed to obtain CSI for carefully selected regions in a given large-scale area of interest (AOI) using a trial-based method, where different system configurations and parameters are tried, and the achieved performance for different assessing region sizes is evaluated. This ensures that the chosen regions effectively act as representative samples for the entire AOI. By leveraging the GNN-based CEs for these selected regions, we can accurately predict the performance for users in any AOI region. Additionally, we optimize the placement of double RISs to further enhance system performance. Extensive simulations are conducted to validate our approach and demonstrate its effectiveness in achieving accurate system performance with reduced complexity in large-scale communication systems.