A large-scale mission planning method for agile earth observation satellite

Due to the large search space, the large-scale agile earth observation satellite mission planning problem (AEOSMPP) is a challenge to existing algorithms. This paper proposes an end-to-end framework for solving the large-scale AEOSMPP using reinforcement learning. The framework combines graph pointer networks (GPNs) and independently recurrent neural networks (IndRNNs). We use mask vector to consider various constraints and train the model with REINFORCE. In this approach, once a single model is trained to find near-optimal solutions from a given distribution, it can be applied to different scenarios from the new same distribution without retraining. Experimental results show that for large-scale AEOSMPP, the proposed algorithm has a stronger generalization ability and can obtain a higher observation reward.