Public thesis defense Adrien Banse - ICTEAM
sst |
Data-driven memory-dependent abstractions of dynamical systems
Monday June 29th, 2026 - 5pm - Auditorium BARB93 - Place Sainte-Barbe 1, 1348 Louvain-la-Neuve
With the increasing ubiquity of safety-critical autonomous systems operating in uncertain environments, there is a need for methods that provide behavioural guarantees in the presence of epistemic uncertainty, i.e. lack of knowledge about the system, and potentially aleatoric uncertainty, i.e. intrinsic uncertainty of the system. Toward formally verifying properties of complex systems, methods based on system approximations - abstractions - have surged in recent years.
This thesis studies memory-dependent abstraction-based methods for both deterministic and stochastic systems (modelling aleatoric uncertainty) where only partial observations are available (modelling epistemic uncertainty). Instead of representing the system through memoryless discrete states, the proposed approach builds abstractions whose states encode observation histories. For deterministic systems, two frameworks are developed: uniform memory, where abstractions encode fixed-length output sequences, and adaptive memory, a more flexible setting in which memory is allocated locally through a refinement procedure. The thesis then extends memory-dependent abstractions to stochastic systems, a technically challenging extension that necessitates different theoretical tools.
Additionally, we introduce the Cantor-Kantorovich distance, a novel metric between labelled Markov chains. This distance arises from the natural topology over infinite symbolic traces, and is useful for our memory refinement procedure. Beyond its direct role in this thesis, the Cantor-Kantorovich distance exhibits strong theoretical properties and holds potential for broader applications, making it a transversal yet central contribution of this thesis.
Jury members
Prof. Raphaël Jungers (UCLouvain), Supervisor
Prof. Alessandro Abate (Oxford University, UK), Supervisor
Prof. Philippe Lefèvre (UCLouvain), Chairperson
Prof. Gianluca Bianchin (UCLouvain), Secretary
Prof. Sadegh Soudjani (University of Birmingham, (UK) & Max Planck Institute, Germany)
Prof. Manuel Mazo Jr. (TU Delft, The Netherlands)
Prof. Duarte Antunes (TU Eindhoven, The Netherlands)