We are happy to share that the paper “Unsupervised Multi-kernel Learning for Automated Algorithm Selection”, authored by Yihang Lu, Tome Eftimov, and Carola Doerr, has been accepted at the Parallel Problem Solving from Nature conference.

The paper investigates an unsupervised approach to automated algorithm selection in black-box optimization using multi-kernel clustering over heterogeneous landscape representations, including ELA, DeepELA, DoE2Vec, and TransOptAS. Instead of relying on expensive performance labels, the proposed framework groups problem instances directly from structural information and maps them to solver recommendations through a strictly separated evaluation protocol.

Experimental results on affine BBOB-derived selector tasks for Differential Evolution and Particle Swarm Optimization show that the proposed approach consistently outperforms naive feature concatenation and strong single-view baselines. The learned kernel weights further reveal that ELA and TransOptAS contribute most strongly to selector-oriented grouping, providing additional interpretability into the automated selection process.

Congratulations to all authors, especially to Yihang Lu for leading this work!