aRieL: reinforcement learning to plan exoplanet observations by the Ariel space telescope

An agent trained using simulation mediates between the competing objectives of the European mission and adapts to an expanded catalogue or a new scenario without the need for further training.

The Ariel mission of the European Space Agency must study the atmosphere of hundreds of exoplanets, which requires a constant choice of targets to be observed in a wide catalogue. Each decision consumes mission time and alters the remaining scientific opportunities: a long-term optimization problem that researchers propose to entrust to learning by strengthening.

Their environment, called aRieL, simulates observations and leads a PPO agent based on a Transformer Set to produce a planning policy. Compared to several reference heuristics, he found a good balance between the objectives of the mission: large samples at the most general level of observation (Tier 1), while completing the in-depth observations (Tier 3), covering well the planet population and wasting little time observing. This behaviour occurs when the target catalogue is greatly expanded, and changing the reward function changes the strategy learned consistently.

A trained policy on the basic scenario also adapts, without retraining, to a variant requiring repeated Tier 3 observations, showing the price to be paid for the rest of the study. The authors see this as an approach that can be applied to other observatories whose planning depends on the state of the mission.