Optimizing Earth Observation Satellite Schedules under Unknown Operational Constraints: An Active Constraint Acquisition Approach

A new AI research paper tackles a critical bottleneck in space operations: scheduling Earth Observation satellites when the full rulebook is unknown. Authored by Mohamed-Bachir Belaid, the work addresses that real-world constraints—governing factors like safe distances between imaging tasks, power budgets, and thermal limits—are often buried in engineering simulators or legacy code, not in clean mathematical models. This forces schedulers to work with incomplete information, leading to inefficient plans.

The proposed solution is the Learn & Optimize (L&O) framework, which embeds a domain-specific procedure called Conservative Constraint Acquisition (CCA). Instead of a costly two-phase process of learning all constraints first and then optimizing, L&O performs an interactive search. It alternates between proposing an optimized schedule under its current best guess of the rules and asking a binary 'oracle' (like a simulator) targeted yes/no questions to check feasibility and learn. This active learning approach is far more query-efficient.

On synthetic but challenging test instances with up to 50 imaging tasks, the results are significant. Compared to a simple greedy baseline with no constraint knowledge (which had a 65-68% performance gap from optimal), L&O slashed the average gap to between 17.7% and 35.8% for smaller problems. For the largest 50-task scenarios, L&O achieved a 17.9% gap, outperforming a traditional 'acquire-then-solve' baseline (20.3% gap) while using 21.3 main queries instead of 100 and running about five times faster.