A Learning-Based Approach for Agile Satellite Onboard Scheduling

Autonomy increases the ability of earth observing satellites by allowing them to acquire more images. This is enabled by an efficient planning and scheduling algorithm which is able to make quick decisions onboard. Due to the NP-hardness of the agile earth observing satellite (AEOS) onboard scheduling problem, heuristic and metaheuristic algorithms seem to be appropriate to cope with increasingly enlarged problems. Also, the algorithms need to be intelligent enough to deal with dynamically changing situations onboard. Such algorithms are missing in the literature and we make the first attempt to propose a learning-based approach (LBA) for the AEOS onboard scheduling problem. LBA adopts an offline training - onboard scheduling paradigm where it trains a classifier using massive historical data offline on the ground and embeds this classifier to an onboard greedy construction algorithm. At each construction step, the greedy algorithm uses the classifier to test the potential of a task and arranges its observation time if it is accepted by the classifier. Extensive experimental results show that the proposed LBA is highly suitable for onboard use in terms of both solution quality and response time. In particular, LBA easily dominates state-of-the-art algorithms by producing very high quality solutions for large-size problems (with over 100 tasks) in seconds.