 |
Center for Nonlinear Studies |
We discuss solutions to the problem of deriving safety guarantees for verification problems where accurate models are either not available or not practical for reasoning about safety. In addition to discussing the general theory of deriving data-driven probabilistic guarantees, we will examine two applications of data-driven verification: estimating the forward reachable sets of a controlled dynamical system, and verifying the range of outputs obtained by a neural network.
The solutions we discuss are applicable to any model from which we can obtain a finite (but potentially large) number of point evaluations or simulations, and that possesses the property of measurability in the sense of measure theory-- these are absolute minimal conditions under which probabilities can still be defined. Under these limited circumstances it turns out to be possible to construct safety guarantees using a data-driven approach. These guarantees are probabilistic, reflecting the reality that limited information constrains the possible accuracy of our verification and the degree to which we can place confidence in them.
Dr. Rosalyn A. Devonport was born in Phoenix, AZ, USA, in 1993. She received her B.S. degree in electrical engineering from Arizona State University, Tempe, AZ, in 2016 and her Ph.D. degree in electrical engineering & computer sciences at the University of California, Berkeley.From 2014 to 2016 she worked as a research assistant in the Goldwater Engineering center at Arizona State university, Tempe, AZ, USA, and from 2016 to 2017 as a research staff member. In 2019 she was a research intern at the Technical University of Munich, Munich, Germany. She is currently a postdoctoral scholar at the University of New Mexico with Prof. Meeko Oishi.Her research interests include data-driven methods in control theory, particularly for reachability analysis and robust stochastic control, and applications thereof to aerospace systems.
Host: Juston Moore (XCP-AI4ND)