An SST-IE Framework for Beamforming and Phase Shift Design in RIS-Aided Multi-User Networks

This paper proposes a novel semi-supervised, separate training, and iterative execution (SST-IE) framework for joint beamforming and phase shift optimization in RIS-aided multiuser networks. SST-IE employs three independently trained neural networks with carefully designed input features and loss functions, effectively leveraging the strengths of both supervised and unsupervised learning. Guided by the block coordinate descent (BCD) algorithm, SST-IE achieves superior sum-rate performance compared to the BCD algorithm while requiring significantly fewer iterations. Simulation results show that SST-IE outperforms representative deep learning and reinforcement learning (RL) approaches—including a two-stage method and a Deep Deterministic Policy Gradient (DDPG)-based RL algorithm—particularly in high-SNR regimes, while ensuring fair rate allocation among users.