Benchmarking Beyond Statistics: Data-Driven Footprints for Explainable Black-Box Optimization скачать в хорошем качестве

Benchmarking Beyond Statistics: Data-Driven Footprints for Explainable Black-Box Optimization

Title: Benchmarking Beyond Statistics: Data-Driven Footprints for Explainable Black-Box Optimization Speaker: Tome Eftimov (https://cs.ijs.si/eftimov/) Abstract: This talk explores how emerging benchmarking and meta-learning methodologies are redefining the way we evaluate and select optimization algorithms, moving toward a trustworthy and explainable paradigm. Two promising directions will be highlighted. The first focuses on representative instance selection, ensuring that benchmarking data are diverse and generalizable rather than tailored to narrow or convenient test sets. The second introduces the concept of algorithmic footprints—digital signatures that capture how algorithms interact with problem landscapes, revealing which landscape features influence their success or failure. Together, these developments are paving the way for a new generation of explainable and automated optimization. By replacing simple statistics with interpretable, data-grounded insights, the field is advancing toward a future where black-box optimization becomes as transparent, reproducible, and knowledge-transferable systems.