AI-Empowered Mode Selection and Beamforming for STAR-RIS-Assisted Communications

Simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) can enable full spatial coverage. In this paper, we investigate an artificial intelligence (AI)-empowered STAR-RIS-assisted multi-user communication system. We aim to maximize the system sum-rate by jointly optimizing the mode selection matrix of the STAR-RIS elements, the passive beamforming matrix of the STAR-RIS, and the beamforming matrix of the base station (BS). Due to the mixed-timescale structure, we propose a joint neural network (NN) design consisting of an advantage pointer-critic (APC) NN to optimize the discrete variables in the mode selection matrix, a fully connected NN, and a deep-unfolding NN to optimize the beamforming matrices. Specifically, the element mode selection problem is formulated as a Markov decision process (MDP), and an APC NN is carefully designed to solve it. The passive beamforming matrix of the STAR-RIS is optimized by employing a fully connected NN. Then, we apply an iterative weighted minimum mean-square error (WMMSE)-based deep-unfolding NN for the BS beamforming design. Simulation results verify that our jointly trained NN can outperform the conventional algorithms with reduced overhead.