Comparison of Centralized and Federated Machine Learning Techniques for Beamtracking in 5G/6G Systems

Beamtracking is an essential feature of 5G/6G systems allowing to maintain the active connection between user equipment (UE) and base station (BS). However, due to the use of antenna arrays with extremely narrow directional radiation patterns, this procedure consumes a significant amount of time and frequency resources. Thus, reducing the scanning state space is vital for improving 5G/6G systems performance. In this paper, we explore and compare the centralized and federated machine learning (ML) approaches for scanning state space reduction. Both options are based on long short-term memory (LSTM) neural networks but the latter utilizes federated learning techniques for beam direction prediction. Our numerical results indicate that the federated learning approach allows us to achieve the same accuracy when the overall training sets for centralized and federated learning are comparable even when users have heterogeneous characteristics in terms of their mobility patterns. This allows for a decrease in the training time utilized for ML-aided beamtracking.