Narrowing the Gap: Towards Analyzable and Realistic Simulators for Safety Analysis of Neural Network Control Systems

Vivian Lin, James Weimer, Insup Lee

Abstract: Although neural network control systems have been used in many applications, they are still challenged by the unreliability of neural networks in the face of unseen data. Thus, it is important to formally verify and assure the safety properties of neural network control systems. A major challenge of this task is a pervading gap between realistic simulators and analyzable simulators of system plants. To narrow this gap, we propose a method to closely approximate a realistic simulator with a less complex simulator. Our approach relies on generative adversarial networks to learn the error statistics between an existing analyzable simulator and an existing realistic simulator. We additionally present a general probabilistic guarantee on the input-output error of an approximation obtained by any method, including our own. We apply our GAN-based approximation technique in a case study of glucose control for type 1 diabetics and evaluate the resulting approximator against data simulated from real patient parameters. We observe that our approach more closely approximates a realistic insulin-glucose simulator than a baseline method.