Spacecraft Orbiting and Uncertainty - Planning Surveillance

Abstract

Scheduling of the Space Surveillance Network (SSN) is a crucial operation for the maintenance of safety and operations in Earth’s orbit. However, the capabilities of the SSN are limited and the number of objects that are being tracked is increasing with every year. This work proposes harnessing Imitation learning (IL) to develop explainable schedules without the development of subjective functions, but instead learning from approved schedules. To that end is proposed a graph structuring of the situation that allows learning from expert solutions. Importantly, this proposed framework also removes fragmentation and discretisation requirements within the time and space domains, requirements that are present in other solutions and lower the asymptotic efficiency that can be achieved. However, the models that were trained in this work did not achieve these goals and showed a very strong competition between the capability to choose the correct pass to observe an object and choosing the correct time within the pass. The trained models also showed a significant maintenance of performance of a trained model on data inputs outside of distribution. Overall, this thesis provides the necessary background to understand the principles of decision making for developing an SSN schedule, shows the set up of a graph structure for the basis of an IL algorithm for scheduling, and presents the results that have been obtained to this point.