Automating control system design: Using language models for expert knowledge in decentralized controller auto-tuning
Loading...
Date
10/11/2025
Authors
Ares-Milian, Marlon J.
Provan, Gregory
Quinones-Grueiro, Marcos
Journal Title
Journal ISSN
Volume Title
Publisher
Dagstuhl Publishing
Published Version
Abstract
Fully-automated optimal controller design for engineering systems is a challenging task. While, optimization-based, automated control parameter tuning techniques have been widely discussed in the literature, most works do not discuss expert knowledge requirements for system design, which result in significant human intervention. In this work, we discuss a multistage controller tuning framework for decentralized control that highlights expert knowledge requirements in automated controller design. We propose a methodology to automate the input-output pairing and stage definition steps in the framework using Large Language Models (LLMs) for a family of multi-tank benchmarks. We achieve this by proposing a mathematical language to describe the system and design an algorithm to bind this mathematical representation to the input prompt space of an LLM. We demonstrate that our methodology can produce consistent expert knowledge outputs from the LLM with over 97% accuracy for the multi-tank benchmarks. We also empirically show that, correct stage definition by the LLM can improve tuned controller performance by up to 52%.
Description
Keywords
Automated system design , Controller auto-tuning , Large language models , [FoodNutritionalSciences] , Geography, Planning and Development , Modeling and Simulation
Citation
Ares-Milian, M. J., Provan, G. and Quinones-Grueiro, M. (2025) 'Automating control system design: Using language models for expert knowledge in decentralized controller auto-tuning', 36th International Conference on Principles of Diagnosis and Resilient Systems, DX (2025) Nashville, United States, 22-24 September 2025, pp. 1-20. https://doi.org/10.4230/OASIcs.DX.2025.10