Towards Data-driven Verification of Unknown Dynamical Systems

John Jackson, Luca Laurenti, Eric Frew, Morteza Lahijanian

Abstract: The verification of a complex autonomous system is made more challenging when the system dynamics are partially unknown. This paper summarizes an abstraction-based formal verification framework that uses Gaussian process regression to learn unmodelled dynamics from a given data set of noisy system measurements. We summarize our approach and present preliminary results that show great potential for this framework.

Keywords: formal verification, safe learning, machine learning, stochastic process, formal methods, gaussian process

TL;DR: In this paper, we present a framework for data-driven verification of systems using Gaussian process regression and abstraction using uncertain Markov decision processes.