Beam alignment in mmWave phased arrays and RIS-assisted links is a stochastic bandit under both short TTI budgets and Doppler-induced non-stationarity. The arm space is a Riemannian manifold: for steering, for phase combining, for panel orientation, or the discrete torus with up to configurations for -level RIS (, bits/element); the intrinsic Matérn kernel of Borovitskiy et al.\ provides the base GP. We contribute two algorithmic pieces. A Kronecker-factorised intrinsic-product Matérn kernel on evaluating in table lookups, making GP-UCB tractable at where the extrinsic alternative is infeasible. AdaptiveGP-v2, an online sliding-window controller that selects by per-sample marginal likelihood, with predictive-variance and drift -score reset triggers and a post-reset -boost. On a four-speed (~km/h), -seed paired campaign at , AdaptiveGP-v2 is statistically indistinguishable from the hand-tuned fixed-window oracle at every speed (Holm–Bonferroni-corrected paired differences cross zero); the operational benefit is the absence of a deployment-time per-speed calibration step, not a mean-regret improvement. On four static 3GPP-style mmWave benchmarks, intrinsic-kernel GP-UCB reduces cumulative regret by – vs.\ codebook UCB1/Thompson and by – vs.\ Euclidean-ambient GP-UCB on the toroidal arm spaces; a wideband OFDM ablation on a ~MHz channel confirms the advantage persists under frequency-selective fading (~Mbps/UE at initial access vs.\ UCB1). A third-party-simulator sanity check on Sionna CDL is reported in Section~V.

生成讲解博客的中文版本

生成AI智能校验的中文版本