Attention-Mechanism-Based Method for Many-to-Many Orbital Reconnaissance with Limited Payloads

The many-to-many orbital reconnaissance (MMOR) scheduling problem focuses on optimizing the sequences of target inspections performed by reconnaissance satellites to minimize the costs of orbit transfers. As satellite constellations continue to expand, the complexity of planning MMOR missions has increased significantly, rendering traditional optimization methods inadequate for meeting real-time requirements. To tackle these challenges, we propose an innovative attention-based learning method aimed at enhancing the decision-making capabilities of reconnaissance satellites. We begin by formulating the MMOR problem in the form of a capacitated vehicle routing problem, a formulation that simplifies the problem by normalizing the payload capacities of reconnaissance satellites and the resource demands of the targets. We then design a neural network with an attention mechanism, which is trained using the REINFORCE algorithm. This model allows satellites to dynamically determine the inspection sequences. Through extensive simulations, we evaluate the impact of different hyperparameters and identify optimal configurations that maximize performance. Our results demonstrate that the proposed model consistently achieves near-optimal performance compared to traditional metaheuristic methods. Moreover, it exhibits strong generalization capabilities, performing exceptionally well across various scenarios with different target quantities. These findings suggest that the method could be useful for improving the efficiency of satellite mission planning.