(Ali Jadbabaie, MIT speaking) Good morning everyone my name is Ali
Jadbabaie. I'm a faculty member here at MIT, and one of the co-organizers of of this
conference. I'm delighted to welcome you all to MIT for our inaugural workshop
on Learning for Dynamics and Control. I'm just gonna go quickly over the schedule.
So first, Ben, my co-organizer here, is gonna say a few remarks for a few
minutes. Then we'll have our first session. We'll have a coffee break at
10:30. There's no food or drinks allowed in the
room so you can have the coffee outside. We're gonna have poster sessions at noon both days and lunch is served during the poster session. The posters will be
outside. Bathrooms are just when you walk out to your right. There are other
bathrooms in the Koch Institute just across the street. And if you have
questions please use the mics on both sides of the room. All talks are recorded.
All posters will be, in addition to talks, will be posted on on our website. I'm
delighted that you're all here and now Ben, please get us started.
Thank you. (Ben Recht, UCBerkeley speaking) All right. Welcome to the first meeting of the conference on Learning
for Dynamics and Control. Would you say we're still not quite sure what the
acronym is for, but we're working, we're working it together, maybe what that will
be something that we'll discuss later? But the acronym will stay L4DC. What
those letters mean – up for debate. So anyway, thank you all for coming. I mean
this is amazing turnout for 8:45 (am) on a cloudy day in Boston. Thank you everybody for being here. This meeting grew out of a fantastically successful workshop at
last year's CDC, again the DNC, who knows what those things mean? That was a great –
great workshop on the intersections of learning and control. It was organized by
George Pappas, Pramod Khargonekar, Manfred Morari, Konstantinos Gatsis. And that workshop, we already kind of found that they had to
shut the doors and not let people in. Which was crazy, such an overwhelming
support there! So the organizers here kind of – we were just inspired by the
success of that workshop and we decided to try to do something a little bit
bigger, and to create kind of a new home, to be fully shared by all these
different communities. Whether they be machine learning, control, decision theory.
The goal was to bring together these researchers because they all kind of
converged on a very set of similar research problems and research
challenges. So we thought it was a bit ambitious to try to organize this on
such short notice. I think we did, we had, about a three
month lead time. And I know that, I think we figured it's late May
everyone's gonna be busy, traveling in Europe, doing all these things. But to all
of our surprise the response and the enthusiasm were just overwhelming. And
you could see it here today. It's just, it's just astounding. Almost everyone we
asked to speak said yes and before we knew it we had 400 people registered. Including a lot of great young researchers, graduate students, and others.
It's just really very inspiring. So over the next two days what we're gonna hope
to learn is what exactly is everybody excited about. I think we'd argue, as the
organizers would argue, that the excitement stems from the fact that kind
of everything in machine learning, which of course has had its a tremendous
amount of excitement over the past few years, but everything in machine learning
now needs to consider feedback. Whether this, you know, autonomous cars obviously are control systems. But any contemporary machine learning system that you connect
in a complex feedback interaction with with people is also a complex dynamical
feedback system. Similarly, every complex control system that starts to push
against the limits of what we know from classical control needs to consider data
in some capacity. And so there are these new and exciting tasks and autonomy that
require managing very unstructured data streams, high dimensional data streams,
and understanding how to do that in the appropriate way
is an exciting challenge. So there's been a bit of a singularity, if you will,
that's brought control and machine learning back together. And we're all
here passionate to work at this intersection, and for whatever it's worth,
I personally can't remember being so excited about something happening in
research. Machine learning and control there's a natural compliment between the
two of them, right? Machine learning uses data to learn about how to act,
predict, decide, things about the world. This data-driven approach allows an
engineer to, kind of, obviate a lot of complexity that comes when you're in
different environments, and different sensing modalities, and in very complex
dynamical models. Controls on the other hand uses feedback to mitigate on the
effects of dynamic uncertainty. And that uncertainty might arise in the
environment, in your sensing, or in your model. And so together there's got to be
this kinda merged perspective that can help us build new, safe, and agile
autonomous systems. Now recently, there has, of course, been a bit of an intellectual split between machine learning and control. and I'm only gonna argue that this has been a
recent phenomenon over the last, maybe, twenty years. And now a
contemporary machine learning researcher can download a PYtorch, import a few
packages, and then immediately be solving complex motor control benchmarks. And
they might never learn why negative feedback around an integrator is a good
idea. On the other hand, a control engineer
might understand all sorts of intricate complexities of loop shaping but they
might never appreciate how you can use, like, the laws of large numbers to enable
sharp quantification of uncertainty in their models or sensors. But I think if
we think back historically control and machine learning have had very strong
historical connections. Much of the progress on adaptive control, on model
predictive control, and on reinforcement learning has been made by researchers
who has historically straddled the boundaries between control,
and between machine learning. And in fact, some of the earliest breakthroughs in
machine learning such as margin theory came from people who have originally
identified themselves as control theorists. Our program over the next few
days features 19 stellar speakers whose researcher, research, is trying to rebuild
those bridges between these communities, between machine learning and control. We also have fabulous posters that were submitted by many researchers here
and we hope that you will have plenty of time to see all of those. And this
afternoon, I think this is very important, we're gonna try to have a discussion
about the future plans for this community. You know, we hope to make this
conference a regular occurrence, and we invite everybody who's here to help us
shape what the future of this nascence community would look like. Okay, so before
we proceed very importantly we'd like to thank all the excellent support staff
at MIT who has made this possible. This is Doro Unger-Lee, Flavia Cardarelli,
Laura Dorson, Brian Jones, Amy Kaczur, Jeremy Rossen, Erin Schenck and Kim
Strampel. Now all of them committed tireless work, and they all were
instrumental in making this a reality. The logistics here would've been
impossible without their dedication. We'd also like to thank the ARO, the AFOSR
the NSF, and the ONR who all supported this event and are helping to create a
new community. In particular I would like to thank Kishan Baheti, Fariba Fahroo, Sam Stanton, Behzad Kamgar-Parsi, Brian Sadler, and Fred Lev for their
advocacy and support. All right. So thank you all for coming. And here's to a
fantastic two days of exploring the interface between learning and control.
So with that I'd like to hand the podium over to Melanie Zellinger who will
introduce the first session, thank you. (Applause, Melanie approaches.)
