Machine Learning for Robust Beam Tracking in Mobile Millimeter-Wave Systems

Narrow beams in millimeter-wave (mmWave) communication introduce significant beam misalignment challenges. In this paper, we introduce MAMBA-X, an enhanced version of the MAMBA beam tracking scheme. Basically, MAMBA uses a restless multi-armed bandit framework to capture the dynamics of mmWave links by discounting the relevance of past observations using a 'forgetting factor' (_{1}) and increases the weight of recent observations via a 'boost factor' (_{2}). Because the original MAMBA uses fixed values for _{1} and _{2}, it cannot quickly adapt to variations in user mobility. Moreover, if the time between consecutive beam selection instances is large compared to channel dynamics, past observations become obsolete. To tackle these issues, we first use the concept of beam coherence time to establish a bound on the beam selection intervals. Secondly, we show that the performance of MAMBA depends primarily on the value of _{1} which, in turn, depends on UE mobility. We develop a Long Short-Term Memory (LSTM) model to dynamically predict and update the optimal value of _{1}. Through extensive simulations at 28 GHz and using publicly available 5G NR experimental dataset, we evaluate MAMBA-X. Our results indicate that the total delivered traffic is improved by up to 46.8% relative to the original MAMBA and 142% compared to the default beam management scheme in 5G NR.