Automated Algorithm Configuration

It is common that solution algorithms are parameterized, and their configuration influences the quality of results or the algorithms’ performance. A high quality configuration of an algorithm can make the difference between a viable use in production or problem instances being unsolvable in practical settings, and have an impact on costs and profit. The manual configuration of algorithms, however, requires domain knowledge and considerable time investment, while it still does not guarantee to result in a high quality configuration. This challenge can be addressed by automating approaches that efficiently search the parameter space and identify high-quality configurations. Automated algorithm configuration (AC) is an established approach used to tackle this challenge. Recent improvements in AC involve the introduction of machine learning methods and concepts into AC methods that replace or assist AC mechanisms. The machine learning methods improve performance in AC and increase adaptability to diverse AC scenarios. However, existing AC methods do not yet utilize intermediate runtime information, omitting to increase resource efficiency. Similarly, the potential of ensemble methods to increase robustness and performance in AC was not yet investigated. To address these limitations, this thesis presents several contributions to the field of automated algorithm configuration. AC is further extended by machine learning methods and concepts, as well as evaluated and applied to unconventional AC scenarios. In particular, an RAC method is extended by a surrogate model that includes problem instance features, and a gray-box method, utilizing intermediate algorithm output for identification of unpromising evaluations to increase resource efficiency. Further, an ensemble-based approach is developed in the offline AC setting, including three state-of-the-art surrogate models, to achieve higher adaptability to diverse AC scenarios and performance improvement over the individual components. Both methods show improved performance in comparative experiments.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bielefeld

Year

2025

Author dc:creator

• Weiß, Dimitri

Repository record source_url

https://pub.uni-bielefeld.de/record/3006393

OAI identifier oai:identifier

oai:pub.uni-bielefeld.de:3006393