EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

2644. Anytime algorithm configuration

Invited abstract in session TC-3: Optimization and Machine Learning: Methodological Advances, stream Data Science Meets Optimization.

Tuesday, 12:30-14:00
Room: 1005 (building: 202)

Authors (first author is the speaker)

1. Elias Schede
Decision and Operation Technologies, Bielefeld University
2. Kevin Tierney
Decision and Operation Technologies, Bielefeld University

Abstract

Optimization algorithms contain parameters that greatly influence their behavior, such that finding good parameters with automated algorithm configuration (AC) tools has become a critical component in the algorithm design process. Optimization algorithms often possess the anytime property, meaning they can be stopped at any time during their execution and provide a feasible solution. However such algorithms are not specifically targeted by current AC techniques. Setting the parameters of anytime algorithms is especially difficult, as the parameters ought to provide robust performance across varying execution times. Traditional AC methods address this challenge by finding a one-size-fits-all parameter configuration, however finding a schedule of configurations, each targeted to a different runtime, can lead to better overall performance. We introduce a novel AC method for configuring anytime algorithms that produces viable configuration schedules that assign different configurations to different runtimes. Our proposed method harnesses an early termination mechanism for unpromising configurations using a machine learning model, and uses two novel MIP formulations to discard configurations and to create the configuration schedule, respectively. The resulting schedules can be used by decision makers to choose a suitable configuration given a specific runtime for an anytime target algorithm.

Keywords

Status: accepted

Back to the list of papers