Physics-Informed Kernelized Thompson Sampling for Beam Management in Millimeter-Wave Systems

Abstract

Maintaining connectivity in directional millimeter-wave communications requires frequent beam probing and realignment, incurring substantial overhead. We address this challenge by formulating beam tracking as a kernelized multi-armed bandit over the two-dimensional beam-angle domain. We propose GP-TSAF, a Gaussian process Thompson sampling algorithm with a physics-informed array-factor (AF) kernel. Derived directly from the known array geometry and beam codebook, the AF kernel captures beam-pattern overlap, including sidelobe interactions and null-induced decorrelation, without learning the beam-correlation structure from data. This allows each received signal strength (RSS) observation to update the posterior over the full codebook, enabling data-efficient tracking under limited measurements. To improve robustness under mobility, we further introduce adaptive correlation scaling, which weakens cross-beam smoothing as the prediction error increases without requiring an explicit mobility model. For a stationary fixed-kernel variant, we establish a cumulative-regret bound whose kernel-dependent complexity is characterized by the AF-kernel spectrum. Evaluation on 30 GHz simulation datasets with 32- and 64-beam codebooks and a 28 GHz over-the-air dataset with a 28-beam codebook shows that GP-TSAF achieves lower mean oracle gap and cumulative regret than several baselines.