ReVEL: LLM-Guided Heuristic Evolution for Combinatorial Optimization
A novel framework named ReVEL (Multi-Turn Reflective LLM-Guided Heuristic Evolution via Structured Performance Feedback) has been unveiled to facilitate the automation of heuristic generation for NP-hard combinatorial optimization issues. This method categorizes heuristics into reflective groups that are aware of their behavior, comprising similarity-driven clusters for localized enhancement and diversity-driven clusters for exploratory searches. The LLM engages in iterative multi-turn refinement within each cluster, utilizing gathered performance feedback to analyze and enhance related heuristics across evolutionary cycles. Tests conducted on standard benchmarks for combinatorial optimization highlight ReVEL's superiority over current techniques. This framework tackles the challenge of creating effective heuristics, which often necessitates extensive domain knowledge, by employing LLMs for automated generation and refinement.
Key facts
- ReVEL stands for Multi-Turn Reflective LLM-Guided Heuristic Evolution via Structured Performance Feedback.
- It is designed for NP-hard combinatorial optimization problems.
- Heuristics are organized into behavior-aware reflective groups.
- Groups include similarity-driven for localized refinement and diversity-driven for exploratory search.
- LLM performs iterative multi-turn refinement using accumulated performance feedback.
- Related heuristics are jointly analyzed and progressively improved across evolutionary iterations.
- Experiments were conducted on standard combinatorial optimization benchmarks.
- The framework shows promise for automated heuristic generation.
Entities
—