Study finds no single landscape representation dominates in black-box optimization

A new study systematically evaluates how four state-of-the-art landscape feature representations (ELA, DeepELA, TransOptAS, and DoE2Vec) agree or disagree in organizing optimization problem spaces. Using a diverse set of affine combinations of BBOB functions (MA-BBOB), the researchers applied extensive clustering analyses, coverage-based stability measures, and cross-representation similarity assessments. The results show that each representation structures the same problems in markedly different ways. ELA and TransOptAS form compact geometric structures, DeepELA offers a balanced intermediate view, and DoE2Vec achieves strong semantic alignment but with substantial fragmentation. No single representation dominates; instead, they capture complementary aspects of underlying landscapes.

The findings carry significant implications for automated algorithm selection and meta-learning. When tested across two algorithm families—Differential Evolution and Particle Swarm Optimization—the study reveals an inherent trade-off between how well a representation aligns structural landscape descriptions with observed algorithm performance. This indicates that no single representation can fully capture performance, suggesting that practitioners should adopt multi-view analyses or combine representations for downstream tasks. The paper, accepted at IEEE CEC 2026, provides guidance on selecting or merging representations based on specific optimization goals.