Beam Prediction for mmWave V2I Communication aided by Geolocation and Machine Learning

Millimetre wave (mmWave) systems require high beamforming gains to overcome the unfavourable impacts of high path losses at mmWave frequencies. Large antenna arrays enable such gains through highly directive narrow beams which then require multiple beams to cover the spatial directions of interest. The required beam management for such systems, particularly for mobile use cases such as the vehicle-to-infrastructure (V2I) scenarios, is challenging. Real-time optimal beam selection from codebooks consumes radio resources and incurs large training overheads. As a result, geolocation side information and machine learning (ML) algorithms are being explored to address beam management challenges. However, prior works have mostly applied their solutions using simulations that are based on synthetic datasets. Recently, real-world datasets based on extensive mmWave measurements have become available. Leveraging the real-world datasets, in this work, we evaluate and compare the performance of three ML (i.e., k-nearest neighbours, support vector machine and decision tree) algorithms on mmWave V2I beam selection aided by global positioning system latitude and longitude coordinates as the only two features for the ML. The results show the impact of codebook sizes on the accuracies of the ML algorithms under ten different scenarios. The results also reveal the limitations of the geolocation-aided beam prediction as average accuracy could go below 30% in some scenarios, and higher than 90% in other scenarios. These performance results point to the need for multi-modal approaches (involving a combination of different sensors' data) for efficient mmWave V2I beam prediction.