Efficient Verification of Control Systems with Neural Network Controllers

Abstract

Recently, many state-of-art machine learning methods have been applied to Autonomous cyber-physical systems (CPS) which need high safe insurance. This paper develops an effective way to approximate the reachable set of a closed-loop discrete linear dynamic system with a Neural network(NN) controller, whose activation function is Rectified Linear Unit(ReLU). In our method, we choose SHERLOCK, a valid NN verification tool, to estimate the output set of NN and adopt initial state set partitioning to improve the total performance. The approach is evaluated on numerical examples and shows evident superiority to the method before refined.