Optimizing the Sky: Machine Learning-Based Aerial Network Planning for UAM

초록

Urban Air Mobility (UAM) demands a reliable low-altitude communication fabric, yet planning aerial networks is intrinsically combinatorial because base-station (BS) activation and 3D beam orientation (swing/tilt/HPBW) jointly determine coverage and interference. We formulate this design as a high-dimensional optimization over real corridor data and present a two-stage BO-MARL framework with sequential multi-agent learning (MA-Sequential). First, Bayesian optimization (BO) explores BS on/off states and initializes beam angles to prune the search space. Then, a sequential MARL phase treats each active beam as an agent that updates its policy in turn, improving stability and sample efficiency relative to parallel multi-agent updates. A coveragecentric objective, augmented with distance and angular-overlap penalties, encourages wide coverage while curbing redundant infrastructure. Evaluations on an urban UAM corridor corroborate that the proposed pipeline delivers effective and efficient network plans, yielding robust convergence and improved coverage under diverse path conditions. © 2026 IEEE.

키워드