Comparison of Machine Learning Techniques for Beam Management in 5G New Radio (NR)

Abstract

Abstract Aligning beams in the initial access of beam management is a challenging and timeconsuming process. Especially, when the number of antenna elements grow large to compensate for high path loss of millimeter waves. Machine learning methods have successfully been applied to the problem of beam selection and perform much better than traditional methods like exhaustive search. In this thesis, some different machine learning approaches are investigated: decision tree, random forest, Adaptive Boosting (AdaBoost), Support Vector Machine (SVM), Multi-level Perceptron (MLP), Q-learning, Deep Q-Network (DQN) and Double Deep Q-Network (DDQN). Each model is adapted to specific scenarios with different preprocessing steps. A total of three scenarios are explored which have been defined by the 3rd Generation Partnership Project (3GPP): Urban Micro (UMi), Urban Macro (UMa) and Rural Macro (RMa). The UMi and UMa scenarios are both implemented with an explicit city layout containing static receivers. The RMa scenario is uniformly distributed and divided into two datasets: one for static receivers and one for dynamic receivers along tracks. Each scenario has been generated by the stochastic channel model called QuaDRiGa. The aim of the thesis is to provide a fair comparison of machine learning models by testing them on data from one simulator. Results show that random forest and AdaBoost perform best overall on all datasets with up to 90% accuracy when predicting the optimal beam pair, which suggests that the search space can be significantly reduced.