A Hypernetwork Framework for Learning Adaptive Beamforming Schemes in RIS Systems

This work develops a learning-based framework that directly exploits noisy pilots to optimize reconfigurable intelligent surface (RIS) systems while accommodating different service priorities and fairness via user weights. First, an adaptive beamforming configuration problem is formulated to generate the base station active beamforming vectors and RIS passive beamforming reflection coefficients that optimize the weighted sum-rate. Under mild regularity conditions, this problem is shown to attain a maximum. To learn approximate solutions, a novel hypernetwork-based beamforming (HNB) framework is proposed. Particularly, a beamforming network (BFN) exploits available information, including noisy pilots, to generate optimized beamforming configurations. Rather than learning one BFN, a hypernetwork is trained to dynamically generate BFN learning parameters from an input conditioning vector. When the conditioning vector is chosen as the user weights, the trained HNB can tune the BFN to the user weights without the need for retraining. Numerical experiments demonstrate that tuning allows the proposed HNB to perform close to an optimistic block-coordinate descent with perfect CSI benchmark and significantly outperform static learning where a BFN is directly trained to optimize beamforming configurations. Additionally, employing the HNB to also tune the BFN to location information considerably reduces the pilots needed to generate optimized beamforming configurations.