Date & Time: October 29th, 14:00-15:00
Location: Seminar Room M 3170-73 in the M-building, LTH
Speaker:  Geir E. Dullerud, University of Minnesota
Title: Learning for Safety and Control Design in Dynamical Systems

Abstract:
AI-based methods have tremendous potential for impacting the performance of 
autonomous aerospace and robotic systems. Such systems include drones, 
ground- and water-based vehicles, and limbed robots for instance. A barrier to 
the wide deployment of AI-powered methods in such applications is the risk or 
unpredictability of algorithm performance. In this presentation we consider the 
development of safe machine learning (ML) methods for control that provide 
guarantees about their convergence and performance. Specifically, the
presentation will focus on two distinct topics involving the application of learning 
techniques to analysis of dynamical systems. First: we present an algorithm and 
a tool for statistical model checking (SMC) of continuous state space Markov 
chains initialized to a prescribed set of states. We observe that it can be 
formulated as an X-armed bandit problem, and therefore, can be solved using 
hierarchical optimistic optimization. Our experiments, using our tool HooVer, 
suggest that the approach scales to realistic-sized problems and is often more 
sample-efficient compared to other existing tools. Second: we present recent 
results on the global convergence of policy gradient methods for quadratic 
optimal control of discrete-time Markovian jump linear systems (MJLS); switching 
is a common feature in systems that are comprised of interacting software and 
physical processes, and MJLS are models in which discrete states evolve 
according to a finite Markov chain and continuous states evolve according to 
linear dynamics specified by these discrete states. We study the optimization 
landscape of direct policy optimization for MJLS. Numerical examples are 
presented to illustrate the application of this theory. This work brings new insights 
for understanding the performance of policy gradient methods on the Markovian 
jump linear quadratic control problem. Third: We will introduce a new class of 
quadratic constraints that are satisfied by the graph of the Repeated ReLU, an 
activation function commonly used in neural networks. We show that this class is 
the largest that the Repeated ReLU satisfies, and further that only the Repeated 
ReLU and one other function satisfy the class. Hardware: presented will be the 
HoTDeC multi- vehicle testbed, which consists of indoor airborne and groundbased vehicles. 

Bio:
Geir E. Dullerud is a Professor in Electrical and Computer Engineering at the 
University of Minnesota, where he is Head of the Department and the Centennial 
Chair in Electrical Engineering. Prior to this he was at the University of Illinois at 
Urbana-Champaign where as a member of the Coordinated Science Laboratory
he served as the Director of the Decision and Control Laboratory, and 
subsequently the Director of the Illinois Center for Autonomy. He has held visiting 
positions in Electrical Engineering KTH, Stockholm (2013), and Aeronautics and 
Astronautics, Stanford University (2005-2006). Earlier he was on faculty in 
Applied Mathematics at the University of Waterloo (1996-1998), after being a 
Research Fellow at the California Institute of Technology (1994-1995), in the 
Electrical Engineering Department. He holds a PhD in Engineering from 
Cambridge University. He has published two books: "A Course in Robust Control 
Theory", Texts in Applied Mathematics, Springer, and "Control of Uncertain 
Sampled-data Systems", Birkhauser. His areas of current research interest 
include autonomy and cooperative robotics, convex optimization in control, 
cyber-physical system security, stochastic simulation, and hybrid dynamical 
systems. In 1999 he received the CAREER Award from the National Science 
Foundation, and in 2005 the Xerox Faculty Research Award at UIUC. In 2018 he 
was awarded the UIUC Engineering Council Award for Excellence in Advising. 
He is a Fellow of both IEEE (2008) and ASME (2011). He was the General Chair 
of the IFAC workshop Distributed Estimation and Control in Networked Systems 
(NECSYS2019). He is currently a Senior Editor for the IEEE Transactions on 
Automatic Control.