Rapid Beam Training at Terahertz Frequency with Contextual Multi-Armed Bandit Learning

Terahertz (THz) frequency technology holds great promise for enabling high data rates and low latency, essential for manufacturing applications within Industry 4.0. To achieve these, beam training is necessary to enable MIMO communications without the need for explicit channel state information (CSI). In this context, the Multi-Armed Bandit (MAB) algorithms are able to facilitate online learning and decision-making in beam training, eliminating the necessity for extensive offline training and data collection. In this paper, we introduce three algorithms to investigate the applications of MAB in beam training at Terahertz frequency: UCB, Loc-LinUCB, and Probing-LinUCB. While UCB builds upon the well-established Upper Confidence Bound algorithm, Loc-LinUCB and Probing-LinUCB utilize the location of the user equipment (UE) and probing information to enhance decision-making, respectively. The beam training protocols for each algorithm are also detailed. We evaluate the performance of these algorithms using data generated by the DeepMIMO framework, which simulates abrupt changes and various challenging characteristics of wireless channels encountered in realistic scenarios as UEs move. The results illustrate that Loc-LinUCB and Probing-LinUCB outperform UCB, showing the potential of leveraging contextual MAB for beam training in Terahertz communications.