NeuralStoc: A Tool for Neural Stochastic Control and Verification

We present NeuralStoc, a tool for neural controller synthesis and verification in discrete-time stochastic dynamical systems. The tool implements and builds upon the first learner-verifier framework for neural stochastic control with certificates, by jointly learning and/or formally verifying a neural controller together with a neural supermartingale certificate of its correctness. NeuralStoc provides a unified interface for analyses with respect to reachability, safety, reach-avoidance, and stability specifications. We also propose a number of optimizations, which lead to significant improvements in practical performance and scalability of the framework. Notably, our tool is the first to be able to solve neural stochastic control and verification tasks in 4-dimensional environments (4D state space + 2D control input space).