Generating probabilistic safety guarantees for neural network controllers

Neural networks serve as effective controllers in a variety of complex settings due to their ability represent expressive policies. The nature neural networks, however, makes output difficult verify and predict, which limits use safety-critical applications. While simulations provide insight into the performance network controllers, they are not enough guarantee that controller will perform safely all scenarios. To address this problem, recent work has focused on formal methods properties outputs. For we can dynamics model determine must hold for operate safely. In work, develop method results from verification tools probabilistic safety guarantees controller. We an adaptive approach efficiently generate overapproximation policy. Next, modify traditional formulation Markov decision process (MDP) checking overapproximated policy given stochastic model. Finally, incorporate techniques state abstraction reduce error during process. show our is able meaningful aircraft collision avoidance loosely inspired by Airborne Collision Avoidance System X (ACAS X), family systems formulates problem partially observable (POMDP).