MASC: Large language model-based multi-agent scheduling chain for flexible job shop scheduling problem

The flexible job shop scheduling problem (FJSP) presents significant challenges in intelligent manufacturing due to the need for efficient resource allocation and dynamic rescheduling. Traditional solutions often require manual intervention and lack real-time adaptability. This paper introduces the Multi-Agent Scheduling Chain (MASC) framework, which leverages large language models (LLMs) for enhanced decision-making and automated equipment control. MASC integrates four agents to manage dynamic scheduling and real-time rescheduling effectively. SchedAgent (Scheduling Agent) uses an improved ReAct method, combining FJSP-specific scheduling indicators and fine-tuned decision algorithms to optimize results. The DialBag (Dialogue Bagging) method is also introduced, constructing a specialized dataset to prevent knowledge loss and enhance decision-making. This method allows the agent to retain knowledge across diverse scheduling contexts while improving performance in specific tasks. MASC was validated through simulations and real-world robotic experiments, handling both machine malfunctions and urgent job additions. These tests demonstrated MASC’s robust performance, with significant improvements in scheduling efficiency and rescheduling accuracy. Quantitative results showed that SchedAgent consistently achieved high ranking percentages, with an average ranking percentage of 84% at a 10-second solving time and 90% at a 30-second solving time. MASC provides a scalable and adaptable solution for intelligent manufacturing, demonstrating the potential of LLMs to automate and optimize production workflows in both static and dynamic environments.