Falsification of cyber-physical systems through multi-fidelity stochastic optimization

Abstract

This talk will review some recent results on requirements driven search-based testing for Cyber-Physical Systems (CPS). Namely, we will present an extension to a Bayesian optimizer that makes use of models of different complexity (and accuracy) that allow obtaining several approximations of the guiding cost function. The goal of the method is to enable falsification of computationally expensive simulation models through the use of simpler but much faster simulation models.