Dynamically Choosing the Candidate Algorithm with Ostasos in Online Optimization

Abstract

The increasing challenge in designing online algorithms lies in the distribution uncertainty. To cope with the distribution variations in online optimization, an intuitive idea is to reselect an algorithm from the candidate set that will be more suitable to future distributions. In this paper, we propose Ostasos, an automatic algorithm selection framework that can choose the most suitable algorithm on the fly with provable guarantees. Rigorous theoretical analysis demonstrates that the performance of Ostasos is no worse than that of any candidate algorithms in terms of competitive ratio. Finally, we apply Ostasos to the online car-hailing problem and trace-driven experiments verify the effectiveness of Ostasos.