Heuristic Search as Language-Guided Program Optimization
New structured approach decomposes heuristic discovery into three stages, achieving up to 0.17 QYI improvement.
Researchers Mingxin Yu, Ruixiao Yang, and Chuchu Fan propose a structured framework for LLM-driven Automated Heuristic Design (AHD) in combinatorial optimization. Their method explicitly decomposes the discovery process into three modular stages: forward pass evaluation, backward pass feedback, and program refinement. Validated across four real-world domains, it outperforms baselines and shows existing AHD methods are restricted versions of this framework, enabling systematic component upgrades.
Why It Matters
Provides a principled, modular approach to automate complex optimization tasks, reducing reliance on manual trial-and-error and domain expertise.