Research map / Probabilistic and scientific modeling / Bayesian optimization
Neural Exploratory Landscape Analysis for Meta-Black-Box-Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeurELA, a two-stage attention neural landscape analyser that replaces hand-crafted ELA features in MetaBBO methods, trained end-to-end via Fast CMA-ES neuroevolution over multiple MetaBBO tasks.
Problem
MetaBBO methods rely on human-crafted exploratory landscape analysis (ELA) features to profile low-level optimization status; these features are correlated, sampling-sensitive, offline, and costly in function evaluations.
Approach
A learnable attention-based landscape analyser (population information embedding plus cross-solution and cross-dimension attention) that dynamically profiles optimization status and is pre-trained by neuroevolution to maximize relative meta-level performance gain across several MetaBBO tasks, so it can be used zero-shot or fine-tuned.
Key results
- NeurELA gives generally superior zero-shot performance over Original, ELA and DeepELA across unseen MetaBBO algorithms and problem sets, with the largest gain in the unseen algorithm + unseen problem case.
- Fine-tuning NeurELA needs less than 20% of meta-training epochs to reach similar zero-shot performance.
- Fast CMA-ES was selected as the neuroevolution optimizer over SEP-CMAES, R1ES, RMES and CMA-ES; the analyser has 3296 learnable parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the original landscape analyser with pre-trained NeurELA | Unseen MetaBBO algorithms · Unseen MetaBBO algorithms | relative performance | Helped |
| Replace the original landscape analyser with pre-trained NeurELA | MetaBBO tasks · BBOB-Noisy and Protein-Docking | relative performance | Helped |
| Replace the original landscape analyser with pre-trained NeurELA | MetaBBO tasks · Unseen algorithms and unseen problem sets | relative performance | Helped |
| Use ELA features instead of the original analyser | Evaluated MetaBBO tasks · Zero-shot tasks | relative performance | Mixed |
| Fine-tune pre-trained NeurELA jointly with the meta-policy | LDE, RLPSO, and GLEET · BBOB or Protein-Docking | performance gain: less than 20% meta-training epochs | Helped |
| Train NeurELA with Fast CMA-ES rather than another tested evolution strategy (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Helped |
| Replace Fast CMA-ES with SEP-CMAES for training (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Hurt |
| Replace Fast CMA-ES with R1ES for training (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Hurt |
| Replace Fast CMA-ES with RMES for training (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Hurt |
| Replace Fast CMA-ES with CMAES for training (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Hurt |
| Replace the two-stage attention module with a linear feed-forward layer (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Hurt |
| Increase Ts-Attn model complexity (ablation) | NeurELA · Unseen algorithms and unseen problem sets | relative performance | Hurt |
Surprising findings
- Traditional ELA features yield noisy, sometimes inferior results when integrated into MetaBBO tasks, attributed to correlations among feature groups.
- Increasing model complexity (more Ts-Attn layers or larger hidden size) causes zero-shot performance to drop rapidly, attributed to the limits of evolution strategies at large scale.
Limitations
- Training efficiency is low; zero-shot performance drops as model complexity grows because the meta-objective is non-differentiable and ES struggles at large scale.
- Evaluated only on three MetaBBO training tasks and a limited set of BBOB/noisy/protein-docking problems.
Experimental setup
Datasets: BBOB (COCO) BBOB-Noisy Protein-Docking
Benchmarks: BBOB BBOB-Noisy Protein-Docking
Metrics: Z-score normalized relative performance wall-time for feature computation
Baselines: Original hand-crafted landscape analyser ELA DeepELA DEDDQN RLPSO GLEET MelBA GLHF
Builds on: MetaBBO Exploratory Landscape Analysis DeepELA Fast CMA-ES MetaBox
Closest papers by meaning
- $\texttt{MetaDistill}$: Unlocking the Performance Ceiling for Pretrained Optimizers (ICML 2026 · similarity 0.87)
- Meta-Black-Box Optimization with Bi-Space Landscape Analysis and Dual-Control Mechanism for SAEA (AAAI 2026 · similarity 0.87)
- MetaBox-v2: A Unified Benchmark Platform for Meta-Black-Box Optimization (NeurIPS 2025 · similarity 0.86)
- Neuro-Visualizer: A Novel Auto-Encoder-Based Loss Landscape Visualization Method With an Application in Knowledge-Guided Machine Learning (ICML 2024 · similarity 0.83)
- Meta-Black-Box-Optimization through Offline Q-function Learning (ICML 2025 · similarity 0.83)
- SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.