Unimodal Bandit-Based Beam Selection for Ka-Band Mobile Direct-to-LEO Links

摘要

Directional mobile links to low-Earth-orbit satellites require beam selection from limited feedback while the preferred direction changes. This paper studies terminal receive-beam selection on a 20-GHz downlink using a graph of 19 beams arranged on a two-dimensional hexagonal lattice. Transmission success is determined by a fixed SNR threshold, giving all algorithms the same binary ACK/NACK reward and an exact Beta–Bernoulli update for Thompson sampling. unimodal UCB and unimodal Thompson sampling restrict exploration to the empirical leader and its spatial neighbors. Their stationary evaluation is complemented by periodic-restart variants for continuously drifting beam quality. Across 50 independent runs of 4,000 decisions, stationary unimodal Thompson sampling reduces expected-reward loss from 67.5 to 33.7 relative to unstructured Thompson sampling. For a direction drifting at 0.36 degrees per second, restarting every five seconds reduces dynamic loss from 679.7 to 249.7 relative to unrestarted unimodal Thompson sampling. The corresponding delivered goodput increases from 103.1 to 120.4 Mbit/s under the specified transmission duty factor. Sensitivity experiments examine three drift rates and three restart intervals. The results separate the benefit of spatially restricted exploration from the benefit of discarding outdated observations.