Reproduce “Neural Exploratory Landscape Analysis for Meta-Black-Box-Optimization” on OpenResearch

About this paper

Zeyuan Ma, Jiacheng Chen, Hongshu Guo, Yue-Jiao Gong

Researchers from South China University of Technology developed Neural Exploratory Landscape Analysis (NeurELA), a neural network-based framework for dynamically profiling optimization status within Meta-Black-Box Optimization. NeurELA improved generalization performance on unseen algorithms and problem sets compared to existing methods, and exhibited efficient adaptation through fine-tuning, reducing reliance on manual feature engineering.

Key insights

• NeurELA's attention-based architecture effectively learns features that are more sensitive to dynamic optimization progress, providing clearer distinctions between exploration and exploitation phases than traditional static ELA features.
• The framework enables robust zero-shot generalization, indicating it extracts fundamental and transferable landscape characteristics applicable to unseen MetaBBO algorithms and diverse problem sets.
• NeurELA offers computational efficiency comparable to or better than hand-crafted baselines, particularly as candidate samples increase, demonstrating its scalability where traditional ELA methods often become impractical for high-dimensional problems.

Summary by alphaXiv.