Enhanced-GNN With Angular CSI for Beamforming Design in IRS-Assisted mmWave Communication Systems

Intelligent reflecting surface (IRS) is regarded as an efficient technology which reflects incident signals towards desired directions using reconfigurable reflective elements. The conventional approach to design active and passive beamforming involves measuring the channels and then optimizing the system parameters. However, this two-step procedure does not directly maximize the sum-rate due to imperfect channel estimation. To address this, recent research has proposed end-to-end learning using graph neural networks (GNNs). However, prior works exhibit instability in low SNR and suffer from degraded performance in multi-path scenarios. To overcome these challenges, we propose an Enhanced-GNN (E-GNN) approach. Our method enhances feature preprocessing from the angular channel and incorporates a graph attention mechanism (GAT) to improve sum-rate. Additionally, we leverage transfer learning for training in multi-user systems. Simulation results show that E-GNN effectively addresses the issues in low SNR scenarios and achieves performance close to the upper bound in multi-path scenarios.