Гремячих Леонид Игоревич
Automation of Satellite Collision Avoidance Maneuvers with Deep Reinforcement Learning
Науки о данных
Space debris presents a danger to artificial Earth satellites. The number of space objects will grow several times in a few years, due to the planned launches of constellations of thousand microsatellites. This increase in population will lead to a significant increase in conjunctions with space debris. To avoid collisions, spacecraft must undertake collision avoidance maneuvers. In this thesis we propose a novel solution to finding the optimal collision avoidance maneuvers based on state-of-the-art Reinforcement Learning methods: Monte Carlo tree search and Cross-Entropy method. It makes incorporating additional considerations (such as the need to fulfill the mission objective) accomplishable by modification of the reward function, without the need to change the algorithms. The experimental results show their performance to be on par with the analytical methods.
Текст работы (работа добавлена 27 мая 2018г.)