| Time: | May 16, 2023 |
|---|---|
| Download as iCal: |
|
Prof. Sebastien Gros
Department of Engineering Cybernetics
Norwegian University of Science and Technology
Trondheim, Norway
Tuesday 2023-05-16 4:00 p.m.
IST Seminar Room 2.255 - Pfaffenwaldring 9 - Campus Stuttgart-Vaihingen
Abstract
Machine Learning and AI are quickly permeating all aspects of our society, and all technical applications. Control and optimization do not escape that trend. In “general” optimal control, the most broadly visible AI tools are based on Reinforcement Learning (RL), with some very public successes such as AlphaZero and ChatGPT today. Despite these successes, pure AI-based tools face certain difficulties when approaching demanding applications where, e.g., safety or FEAT (fairness, explainability, accountability, transparency) are necessary. In that context, the more classic tools central in control communities are highly relevant. In particular, MPC is a central tool for safety and FEAT.
In this talk, we will introduce the necessary concepts to establish a synergy between MPC and RL. In particular, we will discuss how RL tools allow us to view MPC as "model of optimality” rather than a model-based policy. That view makes it possible to generate learning-based optimal policies using MPC, even if the predictions (or model) underlying the MPC scheme are inaccurate. Conversely, this view allows to introduce structure in RL. We will then discuss the role of the MPC model in that learning context. Relating MPC to RL requires establishing strong connections between MPC and Markov Decision Processes (MDPs). This talk will also briefly discuss some novel results deriving from that connections.
Biographical Information
Sebastien Gros has received his PhD degree from EPFL in 2008, focusing on data-driven optimization and control. After a bike trip from Switzerland to the Everest base camp, and a one year experience in the wind industry, he has joined KU Leuven for a postdoc in numerical optimization and real-time MPC. He became assistant Prof. in 2013 at Chalmers University of Tech., Sweden, and associate Prof. in 2016. In 2019 he joined the Dept. Of Cybernetic, NTNU, Norway as a full Prof. He has become head of that Dept. in 2022. His research interests include MPC, Optimal Control, Markov Decision Processes, Stochastic Optimal Control and Reinforcement Learning. His applications focus on energy-related problems.