LLM system interacts with simulation models in digital twins (work-in-progress, under construction)
A demo video with higher resolution: mix_real_experiment.mov
A demo video with higher resolution: mix_simulation.mov
The folder source_code contains the source code for reproducibility.
Follow the source_code/README.md for the source code to run the prototyp locally.
The LLM interprets the simulation steps in a cyclic manner, interacting with the data and control interface in a digital environment.
The system is designed to be independent from a specific LLM, meaning that any proprietary LLM or open-source LLM can be used to power the system.
The reasoning capability is the most essential, and GPT-4 performs significantly better than GPT-3.5 and other open-source models.
The user provides an objective to the multi-agent system, which then experiments with the simulation to heuristically explore solutions. Finally, the LLM agent provides a summarized solution to parameterize the simulation model.
- Design: introduces a framework that integrates a multi-agent system with LLMs to interact with a simulation model and find parametrization solutions for a process.
- Project Status: it is currently a work-in-progress research project and the paper is under peer-review.
- Application Area: we are investigating the LLMs' interaction with more sophisticated simulation models for industrial automation systems.
Details of this work has been documented in a pre-print paper, which is currently under review.
Xia, Y., Dittler, D., Jazdi, N., Chen, H., & Weyrich, M. (2024). LLM experiments with simulation: Large Language Model Multi-Agent System for Process Simulation Parametrization in Digital Twins. https://arxiv.org/abs/2405.18092v1
@misc{xia2024llm,
title={LLM experiments with simulation: Large Language Model Multi-Agent System for Process Simulation Parametrization in Digital Twins},
author={Yuchen Xia and Daniel Dittler and Nasser Jazdi and Haonan Chen and Michael Weyrich},
year={2024},
eprint={2405.18092},
archivePrefix={arXiv},
primaryClass={cs.AI}
}