OVERT: Verification of nonlinear dynamical systems with neural network controllers via overapproximation
Author
Abstract
Neural network verification is an important tool for reasoning about the correct behavior of safety-critical deep learning-based systems. We contribute a methodology for verification of closed-loop systems with neural network controllers. Our method uses a sound overapproximation constructed with bounded dynamics models. It may be implemented with any neural network verification tool, which allows our method to trade off the benefits and drawbacks of different tools depending on the specific problem at hand.
Year of Publication
2019
Conference Name
Workshop on Safe Machine Learning, International Conference on Learning Representations
URL
https://drive.google.com/uc?export=download\&id=17dj7HXXzjGymoh1VUm4sfb7ynQ_R0Xem