Random subspaces beat full-space sampling for landscape analysis under tight budgets
New paper proposes sampling in random low-dimensional subspaces to stabilize ELA descriptors.
The article proposes an alternative sampling strategy for Exploratory Landscape Analysis (ELA) using random linear embeddings. Under limited evaluation budgets, classical space-filling designs often fail to provide reliable statistical results, resulting in noisy and unstable landscape descriptors. The proposed approach allocates the budget to randomly oriented low-dimensional subspaces. Tests on 20-dimensional BBOB benchmark problems from the COCO environment suggest that random linear embeddings are a promising alternative for budget-constrained ELA, though effectiveness depends on feature class and underlying problem. The paper was accepted at PPSN 2026.
- Proposes sampling in random low-dimensional subspaces (random linear embeddings) instead of full-space for ELA
- Tested on 20-dimensional BBOB problems from the COCO environment across multiple feature sets
- Shows improved robustness of landscape descriptors under limited evaluation budgets, but depends on feature and problem type
Why It Matters
Enables reliable landscape analysis for high-dimensional optimization with very few evaluations, crucial for expensive real-world problems.