Optimizing Earth Observation Satellite Schedules under Unknown Operational Constraints: An Active Constraint Acquisition Approach
Summary: arXiv:2604.13283v1 Announce Type: new
Abstract
Earth Observation (EO) satellite scheduling, which involves deciding which imaging tasks to perform and when, is a well-studied combinatorial optimization problem. Traditional methods typically operate under the assumption that the operational constraint model is fully specified beforehand. However, in real-world scenarios, constraints related to observation separation, power budgets, and thermal limits are often embedded within engineering artifacts or high-fidelity simulators, rather than being explicitly defined in mathematical models.
Introduction
This article discusses a novel approach to EO satellite scheduling that addresses the challenge of unknown operational constraints. Our focus is on scheduling under these unknown constraints, where the objective is clear, but feasibility must be learned interactively through a binary oracle.
Methodology
We introduce the Conservative Constraint Acquisition (CCA) method, a domain-specific procedure designed to efficiently identify justified constraints in practice while minimizing unnecessary tightening of the learned model. This method operates within the Learn&Optimize framework, which supports an interactive search process. The search process alternates between optimization under a learned constraint model and targeted oracle queries.
Results
Our experiments were conducted on synthetic instances featuring up to 50 tasks and dense constraint networks. The findings indicate that the Learn&Optimize approach significantly improves upon a no-knowledge greedy baseline. Moreover, it utilizes far fewer main oracle queries compared to a two-phase acquire-then-solve baseline known as FAO.
Performance Metrics
- For tasks with n ≤ 30, the average performance gap decreased from 65-68% (using Priority Greedy) to 17.7-35.8% with Learn&Optimize.
- At n = 50, where the CP-SAT reference is the best feasible solution found within 120 seconds, Learn&Optimize outperformed FAO on average, achieving a gap of 17.9% compared to FAO’s 20.3%.
- Learn&Optimize accomplished this while requiring only 21.3 main queries, a significant reduction from the 100 queries needed by FAO, and it executed approximately 5 times faster.
Conclusion
The results of this study demonstrate the effectiveness of the Conservative Constraint Acquisition method within the Learn&Optimize framework for EO satellite scheduling. By addressing unknown constraints interactively, this approach not only enhances the efficiency of scheduling tasks but also reduces the computational resources needed for optimization.
This research opens avenues for future work in the field of EO satellite scheduling, particularly in expanding the methodologies to accommodate a broader range of operational constraints and refining the interaction with oracles for even greater efficiency.