Dual-Driven Learning for RIS-Assisted Multi-User MISO Beamforming and Reflection With Uplink Channel Information

Reconfigurable intelligent surface (RIS) can improve wireless transmission performance by passive reflective elements to reconfigure the wireless propagation environment. However, the traditional optimization approach has a high complexity when jointly optimizing phase shifts at the RIS and the beamforming at the base station (BS). This paper focuses on the sum rate maximization problem with the transmit power constraint and the phase constraint, for the RIS-assisted multiuser multiple-input-single-output (MISO) downlink transmission where the downlink and uplink channel reciprocity does not exist. To solve this problem, we propose a data and model-driven learning approach via a hybrid learning manner. Specifically, we adopt an optimal beamforming structure to effectively reduce the output dimension and to improve the training and testing efficiency of the neural network. Then, we provide a learning framework including several fully-connected neural networks to learn the mapping between uplink and downlink channels, power features of the beamforming, and phase-shift matrices, after which the optimal beamforming can be recovered by the proposed beamforming structure. Simulation results show that our proposed method achieves better rate performance than state-of-the-art data-driven learning approaches.