Dear all,
As previously announced at the information coffee, Olle's dissertation
on October 11 will be preceded by a seminar series in the afternoon of
October 10. See below.
Welcome!
/Anders
----------------------------------------------------------------------
October 10 at 13:15-14:00 in the Seminar Room M:3170-73 in the
M-building, LTH.
Jack Umenberger, Oxford University
Title:
No such thing as a model-free lunch? Model-free search and reliable
decision making.
Abstract: Inspired by breakthrough results in the processing of complex
data, machine learning is being increasingly applied to problems in
decision-making and control. However, to be suitable for deployment in
applications, we require learning-based algorithms that come with
guarantees of reliability, robustness, and safety.
This talk will focus on the reliability of model-free policy search,
addressing the question: when do such methods find optimal solutions,
and when do they get trapped in poor local minima? Existing work has
considered static policies; however, for dynamic policies that remember
past observations - necessary for optimal decision making in many
applications - these questions have hitherto remained unanswered.
Focusing on the classic control-theoretic problem of output estimation,
I will present the first model-free policy search algorithm for dynamic
policies guaranteed to converge to the optimal solution.
Along the way, I’ll also describe my path toward working on this
problem, highlighting some of my contributions to model-based approaches
for safe and reliable control, including data-driven robust control,
system identification, and trajectory optimization. I will offer my
perspective on the strengths and weaknesses of model-free and
model-based methods, as well as the ways in which they complement each
other.
----------------------------------------------------------------------
October 10 at 14:05-14:50
Håkan Hjalmarsson, KTH
Title:
Data-driven control: A statistical decision theory perspective
There are many different methods for deriving a controller from a given
data.set but little guidance is provided in the literature in terms of
which methods are preferrable from a performance point of view. In this
talk we focus on the information aspects of the problem. We use
existing results from decision theory to establish necessary and
sufficient conditions for a data-driven controller that guarantees that
there is no other data-driven controller that uniformly (regardless of
the underlying system) can outperform it in terms of average control
cost. We discuss issues like direct vs indirect data-driven control from
this perspective.
----------------------------------------------------------------------
October 10 at 15:30-16:15
Valentina Breschi, TU Eindhoven
Title:
Philosophies, promises and challenges of direct data-driven control
Abstract:
Direct data-driven control (DDC) has become a central focus in control
research, advocating the direct use of data to achieve desired control
objectives and, eventually, satisfy design constraints. By exploring a
selected set of DDC techniques, this talk provides a (personal) take on
the philosophy behind data-driven control, highlighting some of the
promises that DDC approaches have already fulfilled and spotlighting
some of the challenges yet to be faced.
Bio: Valentina Breschi received her B.Sc. in Electronic and
Telecommunication Engineering and her M.Sc. in Electrical and Automation
Engineering from the University of Florence (Italy) in 2011 and 2014,
respectively. She received her Ph.D. in Control Systems from IMT School
for Advanced Studies Lucca (Italy) in 2018. From January to July 2017,
she was a visiting scholar at the Department of Aerospace Engineering,
University of Michigan (United States). From 2018 to 2020, she was a
post-doctoral researcher at Politecnico di Milano (Italy), where she
then held a fixed-term position as a junior Assistant Professor from
2020 to 2023. In 2023, she joined the Department of Electrical
Engineering at Eindhoven University of Technology (The Netherlands) as
an Assistant Professor. Her main research interests include data-driven
control, system identification, with a focus on piecewise affine and
switching systems, collaborative learning, and human-centered policy
design, focusing on mobility systems.
----------------------------------------------------------------------
October 10 at 16:20-17:05
Babak Hassibi, Caltech
Title:
Distributionally Robust Control
Abstract:
Traditional methods in control either assume that the disturbances are
random processes with known statistics (stochastic or H_2 control) or
that they are adversarial (robust or H_infinity control). Recently,
there has been growing interest in a framework, called distributional
robustness, that can interpolate between these two extremes. In
distributionally robust control one assumes that the disturbances have a
distribution that lies in a given uncertainty set and the goal is to
design a controller that minimizes the worst case expected control cost
over all distributions in this set. A not-very-well-known fact is that,
if the uncertainty set is a Kullback-Liebler ball around a nominal
Gaussian distribution, then distributionally-optimal controllers are
given by central H_infinity controllers. These days Wasserstein-2 balls
are mostly considered for distributional robustness because, unlike KL
divergence, they do not require the support of all the distributions
within the ball to be identical, and because of their connections to
optimal transport theory. In this talk we consider the problem of
designing distributionally robust optimal (DRO) controllers in the
Wasserstien-2 metric. We characterize the optimal control policy and
show that, although it lacks a finite-order state-space realization
(i.e., it is non-rational), it can be characterized by a
finite-dimensional parameter. Leveraging this, we develop an efficient
frequency-domain algorithm to compute the optimal control policy and
present a convex optimization method to construct a rational state-space
controller that best approximates the optimal non-rational controller in
the H_infinity norm. This approach leads to efficiently computable
control strategies and we demonstrate its advantages over conventional
H_2 and H_infinity control through several examples.
