Reproduce “Landscape Analysis for Surrogate Models in the Evolutionary Black-Box Context” on OpenResearch

About this paper

Zbyněk Pitra, Jan Koza, Jiří Tumpach, Martin Holeňa

This research from Czech Technical University, Charles University, and the Czech Academy of Sciences rigorously analyzes how fitness landscape features dynamically correlate with the predictive accuracy of surrogate models within an evolutionary black-box optimization context. The study identifies robust and discriminative landscape features, including novel CMA-ES-specific metrics, that can inform adaptive selection of surrogate models, with Gaussian Process models generally exhibiting superior performance.

Key insights

• Many generic landscape features, particularly those related to clustering and y-distribution, were not robust or computable in the dynamic, non-uniformly sampled context of evolutionary optimization.
• Newly proposed CMA-ES-specific features and certain transformed dispersion and meta-model features exhibited high robustness and proved valuable for characterizing the fitness landscape during optimization.
• Landscape features, especially those related to dispersion and CMA-ES step-size, possess significant discriminative power to identify conditions where specific surrogate models achieve lower errors, enabling potential adaptive model selection.

Summary by alphaXiv.