Machine Learning for Online Algorithm Selection under Censored Feedback

Tornede, Alexander ; Bengs, Viktor und Hüllermeier, Eyke (Juni 2022): Machine Learning for Online Algorithm Selection under Censored Feedback. Thirty-Sixth AAAI Conference on Artificial Intelligence, Virtual, February 22–March 1, 2022. Proceedings of the AAAI Conference on Artificial Intelligence. Bd. 36, Nr. 9 S. 10370-10380

Abstract

In online algorithm selection (OAS), instances of an algorithmic problem class are presented to an agent one after another, and the agent has to quickly select a presumably best algorithm from a fixed set of candidate algorithms. For decision problems such as satisfiability (SAT), quality typically refers to the algorithm’s runtime. As the latter is known to exhibit a heavy-tail distribution, an algorithm is normally stopped when exceeding a predefined upper time limit. As a consequence, machine learning methods used to optimize an algorithm selection strategy in a data-driven manner need to deal with right-censored samples, a problem that has received little attention in the literature so far. In this work, we revisit multi-armed bandit algorithms for OAS and discuss their capability of dealing with the problem. Moreover, we adapt them towards runtime-oriented losses, allowing for partially censored data while keeping a space- and time-complexity independent of the time horizon. In an extensive experimental evaluation on an adapted version of the ASlib benchmark, we demonstrate that theoretically well-founded methods based on Thompson sampling perform specifically strong and improve in comparison to existing methods.

• Dokumententyp: Konferenzbeitrag (Paper)
• Publikationsform: Publisher's Version
• Fakultät: Mathematik, Informatik und Statistik > Informatik > Künstliche Intelligenz und Maschinelles Lernen
• Themengebiete: 000 Informatik, Informationswissenschaft, allgemeine Werke > 000 Informatik, Wissen, Systeme
• ISSN: 2159-5399
• Sprache: Englisch