Machine Learning for Online Algorithm Selection under Censored Feedback

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