Location
MC 5417
Candidate
Hannah Berin-Costain | Applied Mathematics, University of Waterloo
Title
Verifiable State-Estimation Error Bounds for Neural Network Based Observers
Abstract
In this seminar, I will discuss the use of neural networks for nonlinear observer design and their robustness guarantees, with a focus on physics-informed neural networks (PINNs) for Kazantzis-Kravaris/Luenberger (KKL) observer synthesis. A challenge in the implementation of KKL observers is constructing an injective map that satisfies the KKL partial differential equation. Recent work has addressed this problem by using PINNs to approximate this map. While theoretical and probabilistic state-estimation error bounds have been established, these results depend on non-computable quantities and therefore cannot serve as robustness certificates. We have derived a computable error bound that depends only on quantities that can be certified over a prescribed region using neural network verification. I will discuss our approach and propose directions for future research.