Low Thrust Minimum Time Transfer: Reinforcement Learning of Optimal Co-State Time Dependance

This paper proposes a deep reinforcement learning framework for computing minimum-time low-thrust orbit-raising trajectories. Spacecraft dynamics is modeled using h-e elements, and includes J2 perturbations, acceleration due to the Sun and the Moon, as well as the effect of Earth shadow. Machine learning was used to adapt a parameterization of the optimal control co-state. To reduce a tendency to wander and oscillate during co-state learning and exploration, we replace our previous co-state dependency and learning mechanism, which we used in a prior paper, with a method here to adapt a co-state function that depends on time. Our methodology is demonstrated through numerical simulations of orbit-raising scenarios to the geosynchronous equatorial orbit. The simulations indicate better convergence when compared to parameterization of co-state dependency on the current state.