On the Generalization of Machine Learning for mmWave Beam Prediction

Owing to the crucial importance of the beam management (BM) procedure in millimeter-wave (mmWave) communication systems, machine learning (ML)-based beam prediction is currently being studied in the 3rd Generation Partnership Project (3GPP). The targets of these studies, in addition to reduction in reference signal (RS) overhead, latency, and power consumption, are to investigate low-complexity ML solutions that can cope up with wireless channel dynamism. Therefore, in this article, we investigate the aspect of ML model generalization and propose a low-complexity meta ensemble learning (MEL) design that facilitates efficient spatial domain beam prediction over different wireless channel conditions. Evaluation results over 3GPP specified channel models demonstrate that the proposed design can generalize well over different channel conditions and achieves a beam prediction accuracy of 93% and 81% over line-of-sight (LOS) and non-line-of-sight (NLOS) scenarios, respectively, with significantly lower computational complexity as compared to the state of the art.