Supervised learning and graph signal processing strategies for beam tracking in highly directional mobile communications

• Published in: Transactions on emerging telecommunications technologies Vol. 30; no. 9
• Main Authors: Ortega, Yosbel R., Guerreiro, Igor M., Hui, Dennis, Cavalcante, Charles C., Cavalcanti, F. Rodrigo P.
• Format: Journal Article
• Language: English
• Published: 01-09-2019
• ISSN: 2161-3915; 2161-3915
• DOI: 10.1002/ett.3687

The use of efficient beam tracking mechanisms becomes necessary in highly directional communication of the fifth‐generation systems. In this context, this paper analyzes the tracking problem and proposes a framework that exploits the samples in the user equipment (UE) dataset (historical UE dataset) to efficiently estimate and predict the channel state at the base station. The framework is composed of two steps. First, a supervised learning algorithm, namely, K‐nearest neighbors (K‐NN), is evaluated as a mean of (i) finding the most similar historical samples to some beam measurements reported by the UE and (ii) predicting the corresponding channel state. As a second step, a sampling and reconstruction strategy based on graph signal processing (GSP) called K‐nearest neighbors with reconstruction (K‐NN‐R) is introduced in order to reduce the beam search space during the beam measurement stage, which allows a more efficient usage of the feedback channel. Simulation results illustrate the performance of the proposal in terms of normalized mean square error in comparison with three traditional/baseline prediction techniques. The K‐NN technique provides a better performance than the baseline approaches for any length of the observed temporal window and with full beam sweep. Meanwhile, the K‐NN‐R framework outperforms the baseline approaches with only half of the beam pairs and throughout a significantly low length of observed temporal window.
This paper analyzes the tracking poblem and proposes a KNN technique and a KNN‐R based on GSP framework that exploits the samples in the historical UE dataset to efficiently estimate and predict the channel state at the BS. The KNN technique provides a better performance than the baseline approaches for any length of the observed temporal window. Meanwhile, the combined KNN‐R framework outperforms the three baseline methods with only half of the beam pairs and throughout an significantly low length of observed temporal window.