A learning-based beam training method for multi-user milli-meter wave systems

In this paper, we propose a hierarchical learning-based beam selection algorithm for multi-users (MU) in vehicular-to-infrastructure (V2I) milli-metre wave communication systems, which attempts to maximize the sum-rate and reduce the beam training overhead. A graph learning model is developed to achieve the initial beam pair selection, exploring the potential correlation between users. The chosen beam pairs are then fine-tuned through a modified sequential downlink uplink sweeping method (SDU). Compared with the conventional beam training method, simulation results based on ray-tracing channel data confirm that our proposed method remains good sum-rate performance with much reduced training overhead. In the case of 8 users, the proposed method achieves the same sum-rate with the conventional SDU scheme, while the overhead of beam pair training is reduced by a factor of 1.26.