On 16 May, our PhD student Jeroen Rook successfully defended his qualifier assessment as part of his PhD graduate program. His research focuses on automated algorithm selection and configuration frameworks that simultaneously find good algorithmic solutions for multiple performance objectives. Additionally, he aims to improve the statistical robustness of these meta-algorithmic frameworks.
Automated algorithm configuration and selection aim to improve the state of the art in solving challenging problems in artificial intelligence (AI) by optimizing an algorithm for a specific set of problem instances, and selecting an algorithm from a predefined collection of algorithms that, most likely, solves a given problem instance best, respectively. Both these meta-algorithmic frameworks successfully demonstrated to improve the state of the art in many AI domains, such as the Boolean satisfiability problem (SAT), traveling salesperson problem (TSP), machine learning (ML), and mixed-integer programming (MIP). One limitation is that most existing meta-algorithmic frameworks only focus on improving one performance objective, e.g., CPU runtime, model accuracy, or solution quality. In reality, performance usually consists of many (conflicting) measures. Additionally, secondary performance objectives are often also relevant to consider, such as the energy consumption of running the algorithm. With multi-objective automated algorithm configuration and selection frameworks, Jeroen aims to obtain even better-performing, versatile, and more robust algorithmic solutions across various AI domains.
To make the adoption of existing and new meta-algorithmic frameworks more widespread, Jeroen is also involved in developing the Sparkle platform. Sparkle integrates multiple meta-algorithmic paradigms into one familiar interface and makes them easily accessible for expert and non-expert users.
