Deep Reinforcement Learning for Multi-satellite Collection Scheduling

Abstract

Multi-satellite scheduling often involves generating a fixed number of potential task schedules, evaluating them all, and selecting the path that yields the highest expected rewards. Unfortunately, this approach, however accurate, is nearly impossible to scale up and be applied to large realistic problems due to combinatorial explosion. Furthermore, re-generating solutions each time the tasks change is costly, inefficient and slow. To address these issues, we adapt a deep reinforcement learning solution that automatically learns a policy for multi-satellite scheduling, as well as a representation for the problems. The algorithm learns a heuristic that selects the next best task given the current problem and partial solution, avoiding any search in the creation of the schedule. Although preliminary results in learning a collection satellite scheduling heuristic still fail to outperform baseline domain specific methods, the trained system might be fast enough to potentially generate decisions in near real-time.