[Paper] COBRA++: Enhanced COBRA Optimizer with Augmented Surrogate Pool and Reinforced Surrogate Selection

Source: arXiv - 2601.22624v1

Overview

The paper introduces COBRA++, a next‑generation version of the COBRA constrained‑optimization framework. By augmenting the surrogate‑model pool and letting a reinforcement‑learning (RL) agent pick the best surrogate on‑the‑fly, the authors dramatically cut the number of expensive function evaluations required for real‑world, constraint‑heavy problems. This makes high‑dimensional, costly‑to‑evaluate optimization more practical for industry‑scale engineering and AI workflows.

Key Contributions

• Augmented surrogate pool – adds several lightweight, diverse models (e.g., polynomial regression, neural‑net surrogates) to the classic Radial Basis Function (RBF) pool, boosting approximation power without extra evaluation cost.
• RL‑driven surrogate selection – trains a policy that chooses the most promising surrogate at each iteration, replacing the hand‑crafted, static selection rules used in prior COBRA variants.
• End‑to‑end learning across problem distributions – the selection policy is optimized on a curated set of constrained benchmark problems, enabling it to generalize to unseen tasks.
• Comprehensive empirical validation – multi‑dimensional experiments show consistent speed‑ups (up to ~30 % fewer evaluations) and higher solution quality versus vanilla COBRA and its earlier adaptive version.
• Ablation study – isolates the impact of the enlarged surrogate pool and the RL selector, confirming that each component contributes meaningfully to the overall gain.

Methodology

1. Problem Setting – The target is a black‑box, constrained optimization problem where each objective/constraint evaluation is expensive (e.g., CFD simulation, hyper‑parameter tuning with resource limits).
2. Surrogate Pool Expansion – Besides the standard RBF, the authors include:
   • Linear and quadratic regression models (fast to train, capture global trends).
   • Small feed‑forward neural networks (capture non‑linearities).
   • Kriging/Gaussian‑process approximators for uncertainty quantification.
3. Reinforcement Learning Selector
   • State: current surrogate performance metrics (prediction error, uncertainty), iteration count, feasibility ratio, and a lightweight embedding of the problem’s dimensionality.
   • Action: pick one surrogate from the pool for the next iteration’s surrogate‑assisted search.
   • Reward: a weighted combination of improvement in feasibility, objective reduction, and evaluation cost saved.
   • The policy is trained with Proximal Policy Optimization (PPO) on a diverse suite of synthetic constrained problems, then frozen for deployment.
4. Bi‑stage COBRA Loop – As in the original COBRA, COBRA++ alternates between (a) a feasibility‑search phase and (b) an objective‑optimization phase, but now each phase uses the surrogate selected by the RL policy.

Results & Findings

Key takeaways: the enlarged surrogate pool improves model fidelity, while the RL selector consistently picks the surrogate that yields the biggest feasibility or objective gain at each step. Ablation shows that removing the RL selector drops performance back to the level of the hand‑tuned adaptive variant, confirming the selector’s central role.

Practical Implications

• Reduced Cloud/Compute Costs – Fewer expensive black‑box evaluations translate directly into lower GPU/CPU hours for tasks like aerodynamic shape optimization, circuit design, or large‑scale hyper‑parameter searches.
• Plug‑and‑Play for Existing Pipelines – COBRA++ can wrap around any existing COBRA implementation; developers only need to supply the evaluation function and constraint definitions.
• Robustness to New Constraints – Because the surrogate selector is trained on a distribution of constraints, it adapts automatically when new or tighter constraints are introduced, sparing engineers from manual re‑tuning.
• Potential for AutoML Platforms – The RL‑based surrogate selection paradigm could be integrated into AutoML services to accelerate constrained model‑selection problems (e.g., latency‑aware neural architecture search).
• Open‑source Friendly – The authors provide a lightweight Python library with pre‑trained policies, making it easy for devs to experiment on their own datasets.

Limitations & Future Work

• Training Distribution Dependence – The RL policy’s generalization is tied to the diversity of the benchmark suite used during training; highly domain‑specific constraints may still require fine‑tuning.
• Scalability of Surrogate Pool – Adding many complex surrogates (large neural nets) could increase memory footprint; the current pool is deliberately kept modest.
• Explainability – While the policy selects surrogates effectively, it offers limited insight into why a particular model was chosen, which could be a hurdle for safety‑critical applications.
• Future Directions – The authors suggest extending the approach to multi‑objective constrained problems, exploring meta‑learning to adapt the selector on‑the‑fly for completely new domains, and integrating uncertainty‑aware acquisition functions for even tighter evaluation budgets.

Authors

• Zepei Yu
• Zhiyang Huang
• Hongshu Guo
• Yue‑Jiao Gong
• Zeyuan Ma

Paper Information

• arXiv ID: 2601.22624v1
• Categories: cs.NE
• Published: January 30, 2026
• PDF: Download PDF