MASC: Large Language Model-Based Multi-Agent Scheduling Chain for Flexible Job Shop Scheduling Problem

Introduction

This repository contains the implementation of the Multi-Agent Scheduling Chain (MASC), a framework leveraging Large Language Models (LLMs) and multi-agent systems to solve the Flexible Job Shop Scheduling Problem (FJSP). MASC integrates LLM-driven decision-making with a structured multi-agent approach to enhance scheduling efficiency, adaptability, and automation in dynamic manufacturing environments.

This repository includes:

• The SchedAgent implementation based on LLMs fine-tuned with QLoRA.
• The DialBag method, which enhances knowledge retention in LLM-based scheduling.
• Simulation and real-world robotic arm experiments validating MASC’s performance.
• Preprocessed datasets and experimental results for reproducibility.

---

Features

• Multi-Agent Coordination: Agents for observation, scheduling, planning, and control.
• LLM-Driven Decision Making: Enhanced scheduling through LLM-powered SchedAgent.
• Flexible Scheduling Capabilities: Supports both simulated and real-world execution.
• Knowledge Retention via DialBag: Improves scheduling efficiency without forgetting previous knowledge.

---

Requirements

• Python >= 3.8
• PyTorch >= 1.12.0 (2.0.0 and above are recommended)
• Transformers >= 4.38

---

Datasets and Benchmarks

ARC (AI2 Reasoning Challenge) [1]: A dataset for testing machine reasoning abilities.

HellaSwag [2]: A dataset designed to evaluate commonsense reasoning in language models.

MMLU (Massive Multitask Language Understanding) [3]: A dataset covering various knowledge domains for evaluating LLM generalization.

SchedQA: A dataset specifically generated using the gen_ReAct_data script to evaluate scheduling-related question-answering performance.

[1] P. Clark, I. Cowhey, et al., “Think you have solved question answering? Try ARC, the AI2 reasoning challenge,” 2018, arXiv:1803.05457.

[2] R. Zellers, A. Holtzman, et al., “HellaSwag: Can a Machine Really Finish Your Sentence?,” in Proc. 57th Annu. Meet. Assoc. Comput. Linguistics, pp. 4791-4800, July 2019, Florence, Italy.

[3] D. Hendrycks, C. Burns, et al., “Measuring Massive Multitask Language Understanding,” in Proc. Int. Conf. Learn. Representations (ICLR), 2021.

---

Experiment & Results

Our experimental results demonstrate:

Method	ARC (%)	HellaSwag (%)	MMLU (%)	SchedQA (%)
internlm2-7b	83.73	85.64	62.19	31.36
internlm2-7b + Single Round	72.2	81.01	61.01	31.47
internlm2-7b + DialBag	84.75	86.14	58.93	32.11

These results validate that DialBag effectively reduces knowledge forgetting while improving domain-specific scheduling accuracy.

---

Citation

If you use our work, please cite:

@article{MASC2025,
  title={MASC: Large Language Model-Based Multi-Agent Scheduling Chain for Flexible Job Shop Scheduling Problem},
  author={Zelong Wang,Chenhui Wan,Jie Liu,Xi Zhang,Haifeng Wang,Youmin Hu,Zhongxu Hu.},
  journal={Advanced Engineering Informatics},
  year={2025}
}

---

Acknowledgments

This work was supported in part by the National Key Research and Development Program of China (No. 2023YFD2100905), National Natural Science Foundation of China (No. 52205104).