Machine Learning-Assisted Beam Alignment for mmWave Systems

Abstract

Beam alignment is a challenging and time-consuming process for millimeter wave (mmWave) systems, particularly as they trend towards higher carrier frequencies which require ever narrower beams. We propose a beam alignment method that is assisted by machine learning (ML), where we train ML models to predict the optimal access point (AP) and beam – or the best few candidates – for a user equipment (UE) given just its GPS coordinates, which can be fed back by the UE or estimated by the network using emerging localization techniques. We train and evaluate the models with data generated by a state-of-the-art commercial ray-tracing software in a realistic urban topology. Even with dynamic scatterers and imperfect UE coordinates, our proposed method can greatly reduce the search space (number of candidates) for finding the optimal AP and beam. For example, in a 28 GHz scenario with 64 beams, our method reduces the search space by approximately 4x for AP selection and over 10x for beam selection while achieving over 95% accuracy. We provide our dataset and models for ease of reproducing and extending our results, which suggest that UE localization coupled with suitably trained ML models can significantly speed up current beam alignment procedures.