DGA$_2$D: Directed Graph-Guided Automated Algorithm Design with LLMs
The proposed framework, DGA$_2$D, aims to enhance automated heuristic design (AHD) for NP-hard combinatorial optimization problems (COPs) through the utilization of large language models (LLMs). Current AHD approaches driven by LLMs are constrained by inflexible solver templates and separate module adjustments, which impede the development of fully autonomous system-level algorithms. DGA$_2$D organizes the expansive programming landscape as a directed graph, with nodes symbolizing functional operators represented by potential code implementations, while directed paths create comprehensive algorithmic workflows. This framework tackles issues related to low reliability, extensive search spaces, and poor credit assignment. The research can be found on arXiv (arXiv:2608.00700).
Key facts
- DGA$_2$D is a directed graph-guided automated algorithm design framework.
- It uses large language models for automated heuristic design.
- It targets NP-hard combinatorial optimization problems.
- Existing methods are confined to rigid solver templates.
- DGA$_2$D structures program space as a directed graph.
- Nodes represent functional operators with multiple code implementations.
- Directed walks constitute complete algorithmic pipelines.
- The framework addresses reliability, search space, and credit assignment issues.
Entities
—