LLM actor-critic based dispatching rule generation for dynamic job shop scheduling

Job shop scheduling is a fundamental production problem characterized by NP-hard complexity. Traditional
approaches rely on numerical data processing and require significant human expertise for heuristic design.
Although Large Language Models (LLMs) have demonstrated strong capabilities in processing unstructured
information, their application in scheduling research has remained limited. This paper proposes an actor
critic LLM-based framework for automated dispatching rule generation. The framework combines simula-
tion-based evaluation with two LLM components: an actor that iteratively generates dispatching rules and a
critic that provides performance- and structure-based feedback. Results indicate that the proposed approach
can generate competitive dispatching rules that statistically dominate state-of-the-art rules.