A challenging, week-long antenna scheduling problem (12 antennas, 30 missions to track) was solved using three techniques: formulated as a QUBO problem and solved on the D-Wave hybrid (quantum-classical) solver, formulated as a MILP problem and conventionally solved, and formulated as a DeepRL problem in which an agent learns a policy to deconflict the schedule. All three techniques were able to solve overscheduled problems for multiple weeks with competitive results for the satisfied time fraction for each mission (actual / requested). The quantum hybrid approach showed promise for scaling to larger problems with shorter running times.

This paper presents a method to effectively adapt the baseline of a synthetic aperture radar based on Deep Reinforcement Learning in distributed Earth observation missions. We describe the approach, which uses the Proximal Policy Optimization algorithm and provides initial results for a toy example built around a hypothetical mission to measure the vertical structure of forests using a formation of 7 satellites carrying L-band synthetic aperture radars. We demonstrate that using a reward function based on expected science return over time and fuel usage; our Deep Reinforcement Learning planner can create plans with positive scientific returns while minimizing fuel usage.