Reachability Set Analysis of Closed-Loop Nonlinear Systems with Neural Network Controllers
Résumé
A forward reachability analysis method for the safety verification of nonlinear systems controlled by neural networks is presented. The proposed method relies on abstracting the activation functions in the neural networks (NN) by quadratic constraints (QCs) resorting to local sector bounds. To tackle the system nonlinearity, the nonlinear model is embedded into a linear parameter varying (LPV) representation. An outer-approximation of the forward reachable set of the closed-loop system is obtained using semidefinite programming. A numerical example clearly demonstrates the applicability of the proposed method. Comparison with some available methods reveals that the provided approach may potentially lead to less conservative results.