Learn to Optimize RIS Aided Hybrid Beamforming With Out-of-Distribution Generalization

Owing to the manually fixed step size, the conventional gradient projection (GP) method requires relatively long time to solve the reconfigurable intelligent surface (RIS) aided hybrid beamforming problem. In order to speed up the GP method, we propose to learn the step sizes by using deep learning. Since the proposed deep learning architecture has a coordinate ascent structure, every step in the deep learning is explainable. Due to the simple multi-layer architecture, the proposed unrolled GP method has a strong out-of-distribution generalization capability. Under a single training setting, the unrolled GP approach is tested under thirty nine different out-of-distribution settings. The extensive simulation results show that the unrolled GP method has larger achievable rate than the GP method under middle-to-high signal-to-noise ratio (SNR) settings, and the proposed method is ten times faster than the GP method for all settings. Codes are available url https://github.com/hexingit/RIS-aided-HBF.