Synthesizing Feature Extractors: An Agentic Approach for Algorithm Selection

Hai Xia, Stefan Szeider, Carlos Ansótegui

Keywords: constraint solving, algorithm selection, LLM, combinatorial optimization, feature extraction

Abstract: Algorithm selection for constraint satisfaction problems requires extracting features that capture problem structure. Manually designing feature extractors demands deep domain expertise and becomes a bottleneck when facing new problem classes. We present an automated approach using Large Language Models to synthesize executable Python scripts that function as interpretable feature extractors. Given a high-level constraint model in MiniZinc, an LLM agent generates code that constructs a typed graph representation and computes structural properties, such as graph density, variable clustering, and constraint tightness. We validate our approach on algorithm selection across 227 combinatorial problem classes from MiniZinc Challenges (2008--2025). Our synthesized extractors achieve 58.8\% accuracy versus 48.6\% for human-engineered extractors (mzn2feat), and outperform neural baselines by 6.8 percentage points on FLECC and 4.3 points on Car Sequencing while maintaining full interpretability. This demonstrates that program synthesis can automate feature extraction for constraint optimization without sacrificing transparency.

Primary Area: optimization

Submission Number: 24962