ELEVATION: Exploiting Large Language Models to Improve Efficiency, Scalability, and Flexibility of Heuristics in Solving Optimization Problems

Heuristics are widely used to address optimization problems. However, manual heuristic design remains a labor-intensive process that requires considerable expertise. ELEVATION introduces an innovative multi-agent framework, inspired by Google’s FunSearch, which integrates Large Language Models (LLMs) with automatically generated heuristics to facilitate the automatic design of optimization solutions. Starting from basic heuristics or from scratch, the LLMs generate and refine executable code through a robust self-assessment mechanism that ensures solution quality. Although the Vehicle Routing Problem (VRP) constitutes the primary use case, the framework is inherently adaptable to a broad spectrum of optimization challenges. Empirical evaluations on a real-world case study demonstrate the potential of ELEVATION to automatically discover effective heuristics for complex optimization problems.