文章背景与核心概要
在约束满足问题中，算法选择（Algorithm Selection）是一个至关重要的长期挑战。传统的特征提取器设计需要耗费大量的人力并依赖深厚的领域专业知识。为此，本文提出了一种自动化的智能体框架，利用大语言模型（LLM）在“检查-修复-验证”循环中自主合成可解释的、特定于问题的 Python 脚本。
该方法通过利用高层 MiniZinc 模型，生成能够计算如图密度、变量聚类和约束紧密性等结构属性的代码。在三种经典的组合优化问题（车辆路径问题、汽车排序问题以及固定长度纠错码）上的实验结果表明，这些合成的特征提取器在性能上超越了专家人工设计的特征（如 mzn2feat）以及基于 Transformer 的替代方案（如 trans2feat），同时具备完全的透明度和可检查性。
Synthesizing Feature Extractors: An Agentic Approach for Algorithm Selection
arXiv: 2608.17170
Authors: Hai Xia, Carlos Ansótegui, Stefan Szeider
Date: August 17, 2026
Subject: Artificial Intelligence (cs.AI)
arXiv: 2608.17170
Authors: Hai Xia, Carlos Ansótegui, Stefan Szeider
Date: August 17, 2026
Subject: Artificial Intelligence (cs.AI)
Summary
本文介绍了一种用于解决约束满足问题中算法选择问题的自动化智能体框架。传统上，设计用于捕捉问题结构的特征提取器需要大量的体力劳动和领域专业知识。作者提出了一种基于 LLM 的智能体，它在“检查-修复-验证”循环中运行，以合成可解释的、特定于问题的 Python 脚本。
通过利用高层 MiniZinc 模型，该智能体生成代码来计算图密度、变量聚类和约束紧密度等结构属性。在三个组合问题（车辆路径、汽车排序和固定长度纠错码）上的实证结果表明，这些合成的提取器优于专家人工策划的特征（例如 mzn2feat）和基于 Transformer 的替代方案（trans2feat），同时保持了完全的透明度和可检查性。
Summary
This paper introduces an automated, agentic framework for solving the algorithm selection problem in constraint satisfaction. Traditionally, designing feature extractors—which are essential for capturing problem structure—requires significant manual effort and domain expertise. The authors propose an LLM-based agent that operates within a "check-fix-verify" loop to synthesize interpretable, problem-specific Python scripts.
By utilizing high-level MiniZinc models, the agent generates code to compute structural properties such as graph density, variable clustering, and constraint tightness. Empirical results across three combinatorial problems (vehicle routing, car sequencing, and fixed-length error-correcting codes) demonstrate that these synthesized extractors outperform both expert-curated features (e.g., mzn2feat) and transformer-based alternatives (trans2feat), while maintaining full transparency and inspectability.
Key Contributions
- 智能体合成： 用自动化的 LLM 智能体取代了人工特征工程，该智能体构建可执行的、特定于问题的特征提取器。
- 可解释的方法论： 与“黑盒”神经方法不同，合成的 Python 脚本保持了人类可读性和可检查性。
- 卓越性能： 与现有的最先进特征提取方法相比，测试集准确率提高了高达 8.3 个百分点。
- 通用性： 在各种组合问题类别（包括车辆路径和纠错码）中得到了验证。
Key Contributions
- Agentic Synthesis: Replaces manual feature engineering with an automated LLM agent that constructs executable, problem-specific feature extractors.
- Interpretable Methodology: Unlike "black-box" neural approaches, the synthesized Python scripts remain human-readable and inspectable.
- Superior Performance: Achieved up to an 8.3 percentage point improvement in test-set accuracy compared to existing state-of-the-art feature extraction methods.
- Versatility: Validated across diverse combinatorial problem classes, including vehicle routing and error-correcting codes.
Access & Resources
Access & Resources
Citation
如果您觉得这项研究有用，可以通过以下方式引用： * NASA ADS * Google Scholar * Semantic Scholar
Citation
If you find this research useful, you can cite it via: * NASA ADS * Google Scholar * Semantic Scholar