Discovery–Closure Adaptive Search: Cost-Normalized Structural Resource Allocation for Constrained Black-Box Optimization

Abstract

Constrained black-box optimization increasingly combines feasibility management, adaptive search, and learned operator scheduling, making simple hybridization an insufficient novelty claim. This paper develops Discovery–Closure Adaptive Search (DCAS), which treats optimization as sequential resource allocation over executable structural transformations. Actions are credited by structural burden removed per unit declared cost, while availability gating prevents nonexistent repair actions from corrupting learning. Closure Complexity supplies the target-relative repair interpretation, and Discovery Plane Theory separates action-generation cost from realized compression. We establish cost-efficiency dominance, positive-scale ordering invariance, archive-closure consistency, and a conditional connection to nonstationary bandit guarantees. A corrected 3,600-run campaign gives DCAS a mean rank of 3.672 versus 6.792 for JADE-NCV, with A₁₂ = 0.914 and Holm-adjusted p = 9.86 × 10⁻⁴⁷. A 1,440-run cross-host study strongly favors DCAS-DE over JADE, while transfer to CMA and PSO remains mixed. A 240-run, 100–1,000-dimensional stress study favors the DCAS-DE and DCAS-PSO embeddings on the declared scalable families. Across 5,388 preserved records, the results provide reproducible mechanism-level evidence while retaining negative findings. They establish strong internal evidence and a frozen path to official CEC and released-author-code validation without claiming unexecuted global state-of-the-art superiority.