Combinatorial Unimodal Bandit Algorithms for Joint Path and Beam Selection in Ka‐Band Mobile Direct‐To‐ LEO Networks

Low‐Earth‐Orbit (LEO) satellite networks with Ka‐band technologies are critical for 6G global high‐capacity connectivity. However, high‐directional links, blockage sensitivity, and multi‐hop routing pose challenges for online path planning and beam selection. This article investigates the joint online path and beam selection (JPBS) problem in Ka‐band mobile direct‐to‐LEO networks, exploiting the Unimodal relationship between beam alignment angle and link quality to reduce exploration complexity. We propose a novel combinatorial multi‐armed bandit (MAB) framework based on the Unimodal Lower Confidence Bound (ULCB) mechanism, named UCombLCB‐JPBS, which restricts beam exploration to local neighborhoods of the optimal beam leader to avoid exponential arm growth. Unlike conventional combinatorial bandits, our algorithm breaks the linear regret dependence on beam count and accelerates convergence to the optimal path‐beam configuration. Simulations show that UCombLCB‐JPBS achieves significantly lower cumulative regret and higher throughput than LLR and CombUCB baselines. This work validates the potential of Unimodal bandit learning for high‐performance LEO‐6G communication networks.