Multi-Objective Approaches for Automated Algorithm Configuration and Selection
Jeroen Rook is a PhD student in the department Datamanagement & Biometrics. (Co)Promotors are prof.dr. H. Trautmann and prof.dr.ir. M. van Keulen from the faculty of Electrical Engineering, Mathematics and Computer Science, University of Twente and prof.dr. H.H. Hoos from the Radboud University Leiden.
This thesis presents significant progress in automating algorithm configuration and selection for scenarios involving multiple performance objectives -- an important step toward building more trustworthy, adaptable, transparent, and resource-efficient artificial intelligence systems. By advancing meta-algorithmic methods, this thesis makes a substantial contribution to the field of artificial intelligence.
The central focus is on managing conflicting objectives, such as computational efficiency, accuracy, and explainability, which frequently arise in real-world applications. Multi-objective scenarios require finding trade-off between various goals simultaneously -- improving on one often comes at the cost of the other goals. While existing automated frameworks excel in optimising for single objectives, they struggle in these multi-objective contexts. The thesis proposes extensions to current paradigms by incorporating multi-objective optimisation techniques into automated algorithm design.
A further challenge addressed is the adaptability of these methods to dynamic environments, where problem characteristics may shift over time and static models become inadequate. Ensuring algorithmic robustness under changing conditions is essential for real-world deployment.
Beyond the development of meta-algorithmic frameworks, the thesis introduces new approaches for evaluating performance in continuous multi-objective optimisation. It presents a novel evaluation metric that balances convergence in the objective space with diversity in the decision space, which are two aspects often overlooked by traditional measures. In addition, a robust statistical ranking method is proposed, offering more reliable algorithm comparisons in the face of instance variability and stochastic noise.
The effectiveness of the proposed methods is demonstrated through extensive experiments across various problems from artificial intelligence. Case studies include multi-modal multi-objective optimisation and sparse neural network training, highlighting practical trade-offs between competing goals, such as accuracy and efficiency.
Finally, to support broader impact and adoption, the developed approaches are made accessible through open-source software tools, enabling their integration in both academic research and applied industry settings.
