Deep Reinforcement Learning for Agile Satellite Scheduling Problem

Agile Earth Observation Satellite Scheduling Problem (AEOSSP) consists in selecting a subset of tasks from a given task set, which is then scheduled on agile satellite, to maximize the total reward of scheduled tasks. AEOSSP is a NP- hard problem and the existing solving methods mainly focus on the field of heuristic and meta-heuristic method, Theoretically, it is impossible to find a single heuristic method that works well on any problem instance. In this paper, inspired by RNN and the attention mechanism, we abstracted the problem from fixed scenarios and proposed an end-to-end framework based on deep reinforcement learning. This model treats neural network as a complex heuristic method constructed by observing reward signals and following feasible rules. The trained model can directly obtain a scheduling sequence without retraining each new problem instance. Compared with the general heuristic rules, experiments prove that this method is more effective and more robust.