Reachability Analysis of Neural Network Control Systems With Tunable Accuracy and Efficiency

Abstract

The surging popularity of neural networks in controlled systems underscores the imperative for formal verification to ensure the reliability and safety of such systems. Existing set propagation-based approaches for reachability analysis in neural network control systems encounter challenges in scalability and flexibility. This letter introduces a novel tunable hybrid zonotope-based method for computing both forward and backward reachable sets of neural network control systems. The proposed method incorporates an optimization-based network reduction technique and an activation pattern-based hybrid zonotope propagation approach for ReLU-activated feedforward neural networks. Furthermore, it enables two tunable parameters to balance computational complexity and approximation accuracy. A numerical example is provided to illustrate the performance and tunability of the proposed approach.