Channel Characteristics-Based Domain Adaptation for Multiuser MISO Beamforming Design

Deep learning-based beamforming solutions face the issue of channel mismatch when the communication environment changes. Although existing adaptive solutions, such as meta-learning, can solve the channel mismatch, there are still have two disadvantages. They rely on labelled data in the fine-tuning stage, and the characteristics of channel state information (CSI) has not been fully explored and transferred, which are important for improving adaptation performance. To address these challenges, the paper first proposes a novel domain adaptation framework based on channel correlation characteristics for sum rate maximization (SRM), named DA-SRM, which does not require labelled data during the transfer process. Then, two transformer units with different attention mechanism are applied to explore the long-range dependence characteristics of CSI. Simulation results show that the proposed DA-SRM and transformer-based DA-SRM achieve better adaptation performance than existing beamforming adaptive solutions.