Emmanuel Rachelson is professor of machine learning and optimization at ISAE-SUPAERO. Prior to joining ISAE-SUPAERO (2011), he held a researcher position at EDF Research and Development (2009, 2011) and postdoctoral positions at the University of Liège (2010) and the Technical University of Crete (2009). He earned a PhD in artificial intelligence (2009) and the Habilitation degree (2020) from the University of Toulouse.
As a teacher, he has been involved in promoting artificial intelligence education. He has been responsible for the Intelligent Decision Systems minor track (MS, 2012) and has founded the Data and Decision Sciences major track (MS, 2015) in the ISAE-SUPAERO curriculum. He also co-founded the Artificial Intelligence and Business Transformation executive master program (2021) and co-organized the international Reinforcement Learning Virtual School (2021). He teaches machine learning and optimization to master and PhD students, and in continuing education programs.
His research is in the field of reinforcement learning and related topics. He created the ISAE-SUPAERO Reinforcement Learning Initiative (SuReLI, 2016) which fosters interaction between PhD students, postdocs and permanents researchers on reinforcement learning topics and their interplay with other disciplines. He investigates the reliability of reinforcement learning methods from different points of view such as statistical generalization, robustness to uncertainty, transfer, simulation to reality, etc. He is also interested in the practical applications of reinforcement learning such as fluid flow control, parameter control in optimization problems, unmanned vehicles, air traffic management, software testing, or therapeutic planning.
He is a member of the ACM and the AFIA, is an ANITI member, has published papers and is a reviewer in the main machine learning and artificial intelligence conferences and journals.
A Few Perspectives on Reliability in Reinforcement Learning
Human (dis)trust in artificial intelligence has multiple causes and can be linked to various subjective factors. Although objectively quantifying these within a single criterion does not seem appropriate, one can try exploring what makes good reliability arguments when learning control strategies for dynamical systems. In this talk, I will try to cover different notions of reliability in the output of reinforcement learning algorithms. Should we trust an agent because it finds good strategies on average (what happens when it does not)? This top-performing AI plays this video game really well, but can I trust it to play new levels? Through recent work on transfer between learning tasks, mitigation of observational overfitting, and robustness to a span of environments, I will explore some of the formal criteria and properties that might lead to better reliability when learning control strategies for dynamical systems.