On May 7th 2025, Paul-Antoine le Tolguenec successfully defended his PhD, entitled “Exploration Methods for Reinforcement Learning Applied to Critical System Testing”. This industrial (CIFRE) thesis with Airbus was directed by Emmanuel Rachelson with my co-supervision. Paul-Antoine studied exploration methods, both in evolutionary algorithms and reinforcement learning, to test critical systems. The application of this thesis was a specific aircraft system, and the collaboration with Airbus enabled a detailed study on a real test case of their aircraft. The thesis abstract is below and the manuscript will be public shortly. I’m looking forward to seeing what Paul-Antoine does in the future, after having demonstrated a keen understanding of both theoretical subjects and their real applications.
Exploration Methods for Reinforcement Learning Applied to Critical System Testing
Critical systems in fields such as aviation, healthcare, and industrial automation require rigorous validation to ensure their reliability. However, in many cases, the spaces representing all possible configurations of these systems are extremely vast, complex, and high-dimensional, which limits the effectiveness of traditional software testing approaches. This thesis explores the use of reinforcement learning to automate testing, focusing on exploration methods that make it possible to identify vulnerabilities in these systems more effectively.
This work proposes several methodologies aimed at improving the exploratory capabilities of deep reinforcement learning algorithms. The application of these approaches to the testing of a critical avionic system demonstrates the potential of such techniques to detect rare failure scenarios. The knowledge gained paves the way for a new range of optimization problems, which could enable a rational automation of system testing through reinforcement learning.