Event-triggered Learning
Topic: Identification | All
Date : 22/06/2023 De 16h00 A 17h00
Lien / Link : https://minesparis-psl-eu.zoom.us/j/91924174451?pwd=ZW1sSmk0aTdkakZIRmwyN253cENLQT09
ID de la réunion / Meeting ID : 91924174451
Mot de passe / Password : 867812
Speaker : Friedrich Solowjow, Intelligent Control Systems group, Max Planck Institute for Intelligent Systems, Stuttgart, Germany
Abstract: Learning-control systems are conceptually different from classical machine learning tasks and there are several key challenges.
Excitation and exploration play a major role, which leads to non-iid settings. On the one hand, we usually aim for controllers that stabilize a system with the goal of avoiding deviations from a setpoint or reference. However, we also need informative data for learning, which is often not the case when controllers work well. Therefore, there is a problem due to the opposing objectives of many control theoretical tasks and the requirements for successful learning outcomes. Additionally, change in dynamics or other conditions is often encountered for control systems. For example, new tasks, changing load conditions, or different external effects have a substantial influence on the underlying distributions. Learning can provide the flexibility to adapt to these events. Since learning has to be applied with sufficient excitation there are many practical situations that hinge on the following problem:
When to trigger learning updates in learning-control systems?
This is the core question of my research. As a potential answer, we propose and develop a new paradigm for principled decision-making on when to learn, which we call event-triggered learning (ETL). At the heart of ETL, we design dedicated statistical tests that leverage kernel-based techniques to compare dynamical systems.
Bio : Friedrich Solowjow received B. Sc. degrees in Mathematics and Economics from the University of Bonn in 2014 and 2015, respectively, and a M. Sc. degree in Mathematics also from the University of Bonn in 2017. He is currently a Ph. D. student in the Intelligent Control Systems Group at the Max Planck Institute for Intelligent Systems, Stuttgart, Germany and a member of the International Max Planck Research School for Intelligent Systems. His main research interests are in systems and control theory and machine learning.
Lien / Link : https://minesparis-psl-eu.zoom.us/j/91924174451?pwd=ZW1sSmk0aTdkakZIRmwyN253cENLQT09
ID de la réunion / Meeting ID : 91924174451
Mot de passe / Password : 867812
Speaker : Friedrich Solowjow, Intelligent Control Systems group, Max Planck Institute for Intelligent Systems, Stuttgart, Germany
Abstract: Learning-control systems are conceptually different from classical machine learning tasks and there are several key challenges.
Excitation and exploration play a major role, which leads to non-iid settings. On the one hand, we usually aim for controllers that stabilize a system with the goal of avoiding deviations from a setpoint or reference. However, we also need informative data for learning, which is often not the case when controllers work well. Therefore, there is a problem due to the opposing objectives of many control theoretical tasks and the requirements for successful learning outcomes. Additionally, change in dynamics or other conditions is often encountered for control systems. For example, new tasks, changing load conditions, or different external effects have a substantial influence on the underlying distributions. Learning can provide the flexibility to adapt to these events. Since learning has to be applied with sufficient excitation there are many practical situations that hinge on the following problem:
When to trigger learning updates in learning-control systems?
This is the core question of my research. As a potential answer, we propose and develop a new paradigm for principled decision-making on when to learn, which we call event-triggered learning (ETL). At the heart of ETL, we design dedicated statistical tests that leverage kernel-based techniques to compare dynamical systems.
Bio : Friedrich Solowjow received B. Sc. degrees in Mathematics and Economics from the University of Bonn in 2014 and 2015, respectively, and a M. Sc. degree in Mathematics also from the University of Bonn in 2017. He is currently a Ph. D. student in the Intelligent Control Systems Group at the Max Planck Institute for Intelligent Systems, Stuttgart, Germany and a member of the International Max Planck Research School for Intelligent Systems. His main research interests are in systems and control theory and machine learning.