Machine Learning and Model Predictive Control
Organizers and Speakers:
Ali Mesbah (Department of Chemical and Biomolecular Engineering, University of California-Berkeley, USA)
Joel Paulson (Department of Chemical and Biomolecular Engineering, The Ohio State University, USA)
Model predictive control (MPC) is an optimization-based control strategy that has become hugely popular due to its ability to handle systems with multivariable dynamics, nonlinearities, and constraints. Furthermore, due to its versatility in terms of its ability to provide robustness guarantees (by design) as well as consideration of economically oriented control objectives, MPC is increasingly being applied in emerging chemical process systems, energy systems, and biomedical systems, to name a few. Yet, many MPC applications face important challenges related to the difficulty of modeling complex systems and the need for MPC strategies to provide provably safe and robust performance with low online computational and memory requirements.
The last years have witnessed an enormous interest in the use of machine learning techniques in different fields, including control systems, which is partly driven by the demonstrated success of machine learning methods in the field of computer science, but also by the increasing availability of data as well as new computation, sensing and communication capabilities. This pre-conference workshop will focus on how recent advances in machine learning can be leveraged to develop and implement improved MPC schemes. The workshop will focus on the following topics:
Approximation of complex MPC laws using machine learning
Machine learning for MPC with adaptive uncertainty models
Bayesian optimization for MPC tuning and model learning
This half-day workshop will consist of two primary parts.
The first part will provide an overview of how machine learning approaches can be used for learning-based MPC of processes with significant plant-model mismatch, approximating MPC problems towards fast and embedded MPC applications, and automated tuning of MPC problems via constrained Bayesian optimization.
The second part of the workshop will entail demonstrating the application of several methods for learning- based MPC, approximate MPC, and automated tuning of MPC methods using various benchmark chemical engineering problems. The workshop will conclude by discussing the open research problems in this area.
All workshop notes and slides will be provided, and open-source codes for selected examples and applications will be made available on the workshop’s website.
The workshop is designed for researchers with a basic knowledge of model predictive control and machine learning who intend to gain an overview of how machine learning methods can be leveraged for learning-based and approximate predictive control of uncertain and nonlinear systems.
Tentative Schedule (as of February 2021):
Part 1: General Introduction and Concepts (125 min)
1.1) Opening remarks and introduction to machine learning for MPC (20 min)
1.2) Learning-based MPC (35 min)
The fundamental paradigm of learning-based MPC for dealing with problems with significant plant-model mismatch (e.g., due to unmolded and/or time-varying process dynamics) will be introduced. We will then discuss various learning-based MPC methods that use Gaussian process regression and Bayesian neural networks for online model adaptation.
1.3) Approximate MPC (35 min)
The notion of approximating (robust/stochastic) MPC laws with cheap-to-evaluate surrogates will be introduced. In particular, we will focus on deep learning approaches for approximating MPC laws, and discuss the theoretical and practical considerations for the application of approximate MPC strategies for safety-critical systems.
1.4) Automatic tuning of MPC controllers under uncertainty (35 min)
Recent developments in automated controller tuning will be discussed. We will then discuss Bayesian optimization methods for systematic and efficient tuning of MPC controllers under system constraints and uncertainties. We will specifically focus on performance-oriented model learning from data.
Part 2: Applications and hands-on problems (90 min)
We will demonstrate the concepts and methods introduced in Part 1 for predictive control of various uncertain and nonlinear processes. The application examples range from biomanufacturing and materials processing applications to integrated design and control of energy systems. The open-source codes for the hands-on problems will be distributed.
Part 3: Wrap-up of the workshop and discussion of open problems (15 min)
Ali Mesbah is Associate Professor of Chemical and Biomolecular Engineering at the University of California at Berkeley. Before joining UC Berkeley, Dr. Mesbah was a senior postdoctoral associate at MIT. He holds a Ph.D. degree in Systems and Control from Delft University of Technology. Dr. Mesbah is a senior member of the IEEE and AIChE. He serves on the IEEE Control Systems Society Conference Editorial Board and IEEE Control Systems Society Technology Conference Editorial Board, and is a subject editor of Optimal Control Applications and Methods and IEEE Transactions on Radiation and Plasma Medical Sciences. Dr. Mesbah is recipient of the Best Application Paper Award of the IFAC World Congress in 2020, the AIChE's 35 Under 35 Award in 2017, the IEEE Control Systems Outstanding Paper Award in 2017, and the AIChE CAST W. David Smith, Jr. Publication Award in 2015. His research interests lie at the intersection of optimal control, machine learning, and applied mathematics, with applications to learning-based analysis, diagnosis, and predictive control of manufacturing and biomedical systems.
Joel Paulson is an Assistant Professor of Chemical and Biomolecular Engineering at The Ohio State University. He joined OSU in 2019 after completing his Ph.D. and M.S. degrees in Chemical Engineering at the Massachusetts Institute of Technology and a subsequent postdoctoral appointment at the University of California, Berkeley. Dr. Paulson is a recipient of the Best Application Paper Award at the 2020 IFAC World Congress and is an active member of both IEEE and AIChE. His research group develops optimization, machine learning, and multi-scale simulation methods to improve the quality, efficiency, and sustainability of engineered products and processes. The developed strategies have been applied to a broad range of systems including continuous pharmaceutical manufacturing, colloidal self-assembly, and non-equilibrium plasma jets.