Abstract
AbstractIntelligent Reflecting Surfaces (IRS) enhance wireless communication by optimising signal reflection from the base station (BS) towards users. The passive nature of IRS components makes tuning phase shifters difficult and direct channel measurement problematic. This study presents a machine learning framework that directly maximises the beamformers at the BS and the reflective coefficients at the IRS, bypassing conventional methods that estimate channels before optimising system parameters. This is achieved by mapping incoming pilot signals and data, including user positions, with a deep neural network (DNN), guiding an optimal setup. User interactions are captured using a permutation‐invariant graph neural network (GNN) architecture. Simulation results show that implicit channel estimation method requires fewer pilots than standard approaches, effectively learns to optimise sum rate or minimum‐rate targets, and generalises well. Specifically, the sum rate for GDNNet (GNN + DNN) improves by over linear minimum mean square error (LMMSE) and by over perfect CSI concerning the number of users, and by over LMMSE and by over perfect CSI concerning pilot length. Offering a feasible solution with reduced computing complexity for real‐world applications, the proposed GNN + DNN method outperforms conventional model‐based techniques such as LMMSE and approaches the performance of perfect CSI, demonstrating its high effectiveness in various scenarios.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call