In this presentation, we will provide an extensive review of the practical challenges in ML-based MPC, with a particular focus on neural network (NN)-based MPC [11], as well as the state-of-the-art solutions to these challenges. The development of ML-based MPC can be categorized into three stages: data collection, modeling, and execution. Each stage presents a unique set of challenges that need to be addressed to ensure the feasibility of ML-based MPC. Specifically, we will discuss practical challenges such as data scarcity, the curse of dimensionality, and computational efficiency, as well as the current methods developed to resolve these issues.
Data scarcity and data quality are common challenges faced in the data collection process in industrial settings. The accuracy and reliability of ML models are highly dependent on the quantity and quality of the training data used. We will highlight novel techniques to address data scarcity, such as physics-informed machine learning (PIML) [12], transfer learning [13, 14]. PIML integrates fundamental physical laws, typically expressed as differential equations or conservation principles, into the learning process of ML models. By incorporating physics-based constraints, PIML improves model generalization and reduces the dependency on large datasets. On the other hand, transfer learning leverages knowledge gained from one domain or task and applies it to a different but related problem. In industrial applications, this often involves pretraining ML models on data-rich processes or simulated environments and fine-tuning them using limited real-world data from the target system.
However, while PIML and transfer learning provide promising solutions to mitigate data scarcity, both approaches present specific challenges when applied to the modeling of nonlinear dynamic systems and NN-based MPC. For example, PIML requires the careful selection of collocation points and the tuning of weight coefficients for different loss terms—tasks that are often performed through trial-and-error, which can be time-consuming and may not guarantee optimal performance. Similarly, transfer learning relies on the availability of a sufficiently similar source dataset, which is often difficult to obtain in industrial applications. Moreover, quantitatively assessing the similarity between source and target domains remains an open challenge, complicating the transferability and reliability of the learned models. In this talk, we will introduce novel methodologies designed to address these practical issues in both PIML and transfer learning, enhancing their applicability and effectiveness in complex, data-scarce industrial settings.
In addition to data scarcity, we will also briefly address the curse of dimensionality, a major challenge in modeling large-scale systems. This issue arises from the exponential increase in data requirements and computational complexity as the number of input features grows, which can significantly reduce the efficiency and effectiveness of machine learning algorithms. To overcome this challenge, reduced-order machine learning models and distributed ML-based MPC frameworks have emerged as promising solutions, tackling the problem from both modeling and control perspectives. Furthermore, during the execution phase of ML-based MPC, ensuring sufficiently fast computation times remains a critical challenge for real-time applications. To address this, we have developed input-convex neural networks and explicit MPC solutions to improve computational efficiency from both the modeling and control standpoints. Throughout this presentation, we will illustrate the effectiveness of these advanced ML modeling and control methods using examples from nonlinear chemical processes.
References:
[1] E. Terzi, F. Bonassi, M. Farina, and R. Scattolini, "Learning model predictive control with long short-term memory networks," International Journal of Robust and Nonlinear Control, vol. 31, no. 18, pp. 8877-8896, 2021.
[2] Z. Wu, A. Tran, D. Rincon, and P. D. Christofides, "Machine-learning-based predictive control of nonlinear processes. Part II: Computational implementation," AIChE Journal, vol. 65, no. 11, p. e16734, 2019.
[3] W. C. Wong, E. Chee, J. Li, and X. Wang, "Recurrent Neural Network-Based Model Predictive Control for Continuous Pharmaceutical Manufacturing," Mathematics, vol. 6, no. 11, p. 242, 2018.
[4] J. Luo, B. Çıtmacı, J. B. Jang, F. Abdullah, C. G. Morales-Guio, and P. D. Christofides, "Machine learning-based predictive control using on-line model linearization: Application to an experimental electrochemical reactor," Chemical Engineering Research and Design, vol. 197, pp. 721-737, 2023.
[5] J. Berberich and F. Allgöwer, "An Overview of Systems-Theoretic Guarantees in Data-Driven Model Predictive Control," Annual Review of Control, Robotics, and Autonomous Systems, early access, doi: 10.1146/annurev-control-030323-024328.
[6] Y. M. Ren et al., "A tutorial review of neural network modeling approaches for model predictive control," Computers & Chemical Engineering, vol. 165, p. 107956, 2022.
[7] F. Bonassi, M. Farina, J. Xie, and R. Scattolini, "On Recurrent Neural Networks for learning-based control: Recent results and ideas for future developments," Journal of Process Control, vol. 114, pp. 92-104, 2022.
[8] L. Brunke et al., "Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning," Annual Review of Control, Robotics, and Autonomous Systems, vol. 5, pp. 411-444, 2022.
[9] C. Gonzalez, H. Asadi, L. Kooijman, and C. P. Lim, "Neural networks for fast optimisation in model predictive control: A review," arXiv preprint arXiv:2309.02668, 2023.
[10] R. Nian, J. Liu, and B. Huang, "A review On reinforcement learning: Introduction and applications in industrial process control," Computers & Chemical Engineering, vol. 139, p. 106886, 2020.
[11] Z. Wu et al., "A tutorial review of machine learning-based model predictive control methods," Reviews in Chemical Engineering, early access, doi: 10.1515/revce-2024-0055.
[12] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, "Physics-informed machine learning," Nature Reviews Physics, vol. 3, no. 6, pp. 422-440, 2021.
[13] S. J. Pan and Q. Yang, "A Survey on Transfer Learning," IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp. 1345-1359, 2010.
[14] A. Thebelt, J. Wiebe, J. Kronqvist, C. Tsay, and R. Misener, "Maximizing information from chemical engineering data sets: Applications to machine learning," Chemical Engineering Science, vol. 252, p. 117469, 2022.