Impacts of Dataset and Codebook Sizes on ML-Driven Beam Prediction for mmWave V2I Communication

Machine learning (ML) can aid the challenging beam management operations in millimeter-wave (mmWave) communication systems. In this paper, we investigated the performance of four ML algorithms (i.e., K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Decision Tree (DT) and Naïve Bayes (NB)) on position-aided beam prediction, using beam prediction accuracy as the metric. We used the KNN, SVM, DT, and NB algorithms to investigate the effects of beam codebook sizes and dataset sample sizes on beam prediction accuracy performance in various vehicle-to-infrastructure (V2I) scenarios using real-world datasets from extensive mmWave V2I measurements. We also illustrated the results using confusion matrices to reveal the misclassification statistics across the different beams. For the four algorithms, the results show that the larger the beam codebook size, the lower the beam prediction accuracy. The results also show that the dataset split ratio does not significantly impact the beam prediction accuracy for the four algorithms. The results point to the need for multimodal approaches that employ a combination of sensor and communication data to improve the beam prediction performance.