Skip to Main Content
Networks consisting of stationary data sources and sinks linked by mobile ferries allow high-latency communication over large distances. Network performance depends on the trajectory of the ferry, which usually should aim to take the shortest path that allows it to exchange all waiting data with each node. An accurate model of the system would permit optimal trajectory planning, but such a model is hard to discover. Consequently, extant ferry flight planners make simplistic assumptions about data rates, ignoring signal reflections and occlusions from aircraft and ground, noise sources, and aircraft dynamics. We present an approach based on reinforcement learning, which allows a ferry to learn to improve its performance while in service, obviating considerable effort in system identification. We compare hand-coded and learned trajectories through radio fields with a complex data rate model, and show that learning trajectory planners can quickly discover good flight paths, outperforming a standard handcoded policy.