An LLM-Based Multi-Agent Framework for Evolutionary Blackbox Optimization

Optimizing blackbox functions requires a structured approach to ensure an effective and efficient search for the optimal solution, supported by a deep understanding of the problem domain, algorithmic behavior and programming abilities. Large Language Models (LLMs) offer a powerful tool to enhance this understanding by excel at solving diverse tasks such as text summarization, code generation, and question answering. This paper presents a novel method using a structured LLM-based multi-agent collaboration approach, to perform numerical blackbox optimization for complex or highly non-linear functions. The tasks of algorithm selection, constraint analysis, code generation, execution, hyperparameter tuning, and solution evaluation are handled by separate agents within the framework. Experiments on standard optimization benchmarks demonstrate the feasibility of mapping the entire optimization process from problem definition to evaluation, achieving competitive results.