Learning-based Beamforming and Phase Shift Design for RIS-aided mmWave Networks Under Sparse Channel Estimation

Reconfigurable intelligent surface (RIS)-aided mmWave communication offers a promising approach to overcome the issue of severe signal blockage that limits the data rate of mmWave networks. However, joint base station beamforming and RIS phase shift (BF-PS) design presents great challenges not only because of the intricate relationship between the beamforming and phase shift matrices in determining the network sum-rate, but also due to the lack of accurate channel estimation when the number of RIS elements gets large. In this work, we study learning-based BF-PS design for RIS-aided mmWave networks under sparse channel estimation. Our main finding is that, under the practical constraint of pilot overhead for RIS channel estimation, learning-based BF-PS design can yield superior performance to that achieved by the conventional block-coordinate-descent (BCD) algorithm. This work provides an initial study of learning-based BF-PS design under the compressed-sensing framework for mmWave channel estimation. The proposed BF-PS design adopts an enhanced four-stage neural-network architecture that improves upon the existing two-stage design by utilizing outputs from earlier stages to form new features for later stages. The new four-stage design using hybrid beamforming (HBF) outperforms those of the two-stage design using either HBF or fully digital beamforming (DBF).