Designing Low-Thrust Transfers near Earth–Moon via Multi-Objective Reinforcement Learning

Abstract

Multi-objective reinforcement learning is used to uncover a subset of the multi-objective solution space for low-thrust transfers between two southern halo orbits in the Earth–Moon circular restricted three-body problem. Multiple policies are trained in this scenario to recover transfers while maximizing their reward functions that reflect distinct relative weightings of two competing objectives: minimizing both the time of flight and propellant mass usage. These policies are simultaneously trained using an algorithm that is designated as multi-reward proximal policy optimization. Evaluating these policies successfully produces transfers between the two orbits with various geometries, propellant mass requirements, and flight times that span a subset of the trade space. These results are also compared to nearby solutions to a constrained optimization problem. A hyperparameter exploration is also performed to determine their influence on the behavior of the trained policies.