Sequential Assignment Problems and Their Application to Task Allocation in Satellite Constellations

Abstract

As satellite constellations grow in size, there is an increasing need for methods of autonomous, scalable, and real-time dynamic task assignment. The satellite constellation setting is unique among task allocation problems in that orbital mechanics dictates that assignments between agents and tasks must change frequently. This gives rise to the Sequential Assignment Problem (SAP), in which agents must be assigned to tasks at several consecutive time steps, while accounting for assignment values which are constantly evolving and the potential penalties associated with changing assignments. This thesis formalizes the Sequential Assignment Problem, and proposes two classes of algorithm to generate solutions to it. The first, Handover Aware Assignment with Lookahead, is an optimization-based algorithm which can generate assignments which are computable in polynomial time, fully distributed, and provably optimal within a bound. The second algorithm is Reinforcement Learning Enabled Distributed Assignment, which utilizes reinforcement learning to find performant solutions to a larger class of Sequential Assignment Problems. Experimental results show that both approaches vastly outperform the state-of-the-art across a variety of challenging, realistic domains with hundreds or even thousands of agents and tasks, with both approaches having clear strengths and weaknesses.