Room 206 (2nd floor, badged access)
3 October 2025 - 10h00
Formal Analysis of Optimization-based Controllers Approximated by ReLU Neural Nets
Eduardo Camponogara from Federal University of Santa Catarina
invited by Thao DANG
3 October 2025 - 10h00
Formal Analysis of Optimization-based Controllers Approximated by ReLU Neural Nets
Eduardo Camponogara from Federal University of Santa Catarina
invited by Thao DANG
Abstract: Abstract: This talk presents a novel reachability analysis method for closed-loop systems in which an optimization-based controller is approximated by a ReLU neural network. The algorithm performs exact reachability analysis for discrete-time LTI systems controlled by ReLU neural networks, as well as for systems both described and controlled by such networks. The proposed approach reformulates ReLU neural network controllers into a Piecewise Affine (PWA) representation, which enables the formal verification of positive invariance and contractivity in controlled systems. Using this PWA reformulation, we can rigorously assess whether a system is safe, positively invariant, and stable. Unlike approximate methods in the literature, our approach achieves exact verification, even for bifurcating systems. Finally, we demonstrate the method through numerical simulations of a collision avoidance scenario with an optimization-based controller.
