RIS Phase Optimization via Generative Flow Networks

This letter introduces a new Machine Learning (ML) technique to learn phase shifting patterns for Reconfigurable Intelligent Surfaces (RISs). We leverage the Generative Flow Network (GFlowNet) paradigm and adapt it so as to compose a RIS phase control resulting in high communication rate. To generalize our approach for different physical layer scenarios, we use a channel chart as a latent representation of the wireless spatial environment to condition the GFlowNet. As such, the GFlowNet learns a scalable policy over RIS configurations that tailors the propagation environment in real-time. We evaluate our solution by means of simulations on a synthetic dataset, and the results corroborate its superiority compared to benchmarks, achieving more than 15% higher communication rates.