This work explores the use of reinforcement learning (RL) for the multi-satellite, multi-target Earth-observing satellite (EOS) scheduling problem. In this work, a decision-making agent is trained independently in a stationary, single satellite environment on a fixed number of imaging targets that are updated as they are imaged and downlinked by the spacecraft. The trained agent is then deployed in a nonstationary, multi-satellite scenario where different spacecraft can share imaging targets. The distributed nature of the architecture simplifies the training process and required training time but requires cross-communication between satellites and is inherently suboptimal in terms of global reward due to the competition between agents. The prospect of this method is that the constellation design and size can scale up and down without requiring the decision-making agent be retrained. This method is evaluated and benchmarked for variousWalker-delta constellations and target numbers. Furthermore, extrapolation of the trained agents to orbits outside of the training distribution is explored. An increase in spacecraft leads to an increase in global reward, but a decrease in reward per spacecraft due to the competition among agents. Furthermore, local reward is highest for orbits with a longitude of the ascending node close to the training distribution.
REINFORCEMENT LEARNING FOR THE MULTI-SATELLITE EARTH-OBSERVING SCHEDULING PROBLEM
Proceedings of the 44th Annual American Astronautical Society Guidance, Navigation, and Control Conference, 2022 ; Kapitel : 74 ; 1351-1368
2024-01-01
18 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch