Algorithm Selection as Superset Learning: Constructing Algorithm Selectors from Imprecise Performance Data

TL;DR: In this paper, a simple regression method based on so-called superset learning is proposed, in which right-censored runtime data are explicitly incorporated in terms of interval-valued observations, offering an intuitive and efficient approach to handling censored data.

Abstract: Algorithm selection refers to the task of automatically selecting the most suitable algorithm for solving an instance of a computational problem from a set of candidate algorithms Here, suitability is typically measured in terms of the algorithms’ runtimes To allow the selection of algorithms on new problem instances, machine learning models are trained on previously observed performance data and then used to predict the algorithms’ performances Due to the computational effort, the execution of such algorithms is often prematurely terminated, which leads to right-censored observations representing a lower bound on the actual runtime While simply neglecting these censored samples leads to overly optimistic models, imputing them with precise though hypothetical values, such as the commonly used penalized average runtime, is a rather arbitrary and biased approach In this paper, we propose a simple regression method based on so-called superset learning, in which right-censored runtime data are explicitly incorporated in terms of interval-valued observations, offering an intuitive and efficient approach to handling censored data Benchmarking on publicly available algorithm performance data, we demonstrate that it outperforms the aforementioned naive ways of dealing with censored samples and is competitive to established methods for censored regression in the field of algorithm selection