A decompose-and-learn multi-objective algorithm for scheduling large-scale earth observation satellites

Abstract

In recent studies, task scheduling problems of earth observation satellites (EOSs) still encounter large difficulties when they meet the increasing number of satellites. Moreover, multiple objectives and increasing number of tasks must be considered in business affairs. To effectively address these issues, considering satellite orbits as independent resources, a multi-objective algorithm is tailored for earth observation satellites scheduling problems (EOSSPs) in this paper. The algorithm includes a novel dynamic learning task allocation mechanism and a bidirectional sorting strategy with global patching method. Specifically, the mechanism works as an evolution operator to allocate tasks to appropriate orbits thus reducing problem complexity and the strategy is tailored to schedule observation windows to corresponding tasks. With the mechanism, original EOSSP can be decomposed into subproblems through the task allocation mechanism, with each task assigned to an corresponding orbit using adaptively updating guideline values. The decision space of a subproblem is limited within one single orbit, which will vastly decrease the complexity of the original large-scale EOSSP. Then, the bidirectional sorting strategy will schedule specific observation windows of each orbit to the candidate tasks. Since the algorithm is proposed to solve large-scale EOSSPs with multi-orbit and multi-objective, it is referred here as LS2MO-SS. Finally, extensive experiments are conducted on ten large-scale problems with different tasks to evaluate the performance of the multi-objective algorithm, the proposed evolution operator, and the bidirectional sorting strategy. The comparative results indeed validate the effectiveness of the proposed algorithms.