Sensing-Assisted End-to-End Beamforming Based on Reinforcement Learning in LEO Satellite Communications

This paper proposes a sensing-aided end-to-end channel prediction and beamforming scheme based on reinforcement learning (RL) for multi-user communication systems in low earth orbit (LEO) satellite scenario. In case of channel non-reciprocity, the base station (BS) can directly use the uplink pilot information to complete the downlink beam design and data transmission. To improve the efficiency of channel prediction, by exploiting the features of integrated Sensing and Communications (ISAC) system, this paper uses the estimated angle-of-arrival (AoA) information as prior knowledge to enhance the self-learning efficiency of the channel prediction network. Simulation results show that the end-to-end channel prediction and beamforming scheme effectively enhances the system's transmission efficiency while reducing time overhead.