AI-Enabled Near-Field Communications: User Movement Prediction and Beam Tracking

A novel near-field beam tracking framework is proposed based on the prediction of the movement of mobile users (MUs). A two-step approach is proposed to successively align the near-field beams in the polar domain and allocate the data streams to adapt to the variations of the channels’ degrees of freedom (DoFs). First, a Transformer-based algorithm is proposed to predict the locations of MUs, where digit-aware embedding and weighted mean square error (WMSE) are incorporated. The multi-head attention mechanism can attend to individual decimal digits of the MUs’ locations to reduce the prediction error from the digit perspective. Then, a hierarchical proximal policy optimization algorithm with a dual-tiered architecture is proposed to learn the successive hybrid beamfocusing and data stream allocation according to the predicted MUs’ locations. Dealing with the high-dimensional action space introduced by the near-field extremely large-scale antenna arrays, the dual-tiered sub-policies are hierarchically designed to reduce dimensionality and facilitate learning. The numerical results demonstrate that 1) the proposed algorithms outperform the baselines in terms of throughput due to the high predictive accuracy and beamfocusing gain; 2) the proposed two-step beam tracking scheme can achieve a similar throughput to the upper-bound scheme with perfect channel state information, while the performance gap of the non-tracking scheme is 64.8%; 3) compared to a fixed scheme, the proposed data stream allocation scheme benefits a performance gain which escalates with an increasing number of allocable data streams