News!
Thank you!
.
We would like to thank all our fantastic speakers for their enormous efforts and inspiring talks! We hope you enjoyed it and learned a lot, we did for sure! Please stay tuned for a variety of videos appearing on our YouTube channel. Nobody knows what the future will bring but lets hope our next meeting will be in person!
The Summer School is Fast Approaching
.
We are really looking forward to welcoming all of you at the Summer School!
The event is fast approaching and we have received many registrations from all over the globe. We are really looking forward to welcoming all of you and our great speakers!
Make sure you have received our email with the technical information such as zoom link and password, YouTube links for the lives etc. by Friday the 4th of June. In case you haven't, please first check your Spam folder and then contact us at {fmg.data.driven.control}@gmail.com. Last but not least, in order to fully enjoy the event, make also sure that you have a stable Internet connection.
YouTube Channel
.
Check out here our new YouTube Channel!
We are going to live stream and then upload all the lectures and talks of the Summer School on YouTube to make the material available to everyone.
We already uploaded Prof. Marco Campi his pre-recorded lectures on Data-Driven Decision Making and the Scenario Approach, enjoy! On Tuesday, June 15 at 10:40 we host a Q&A session, with Prof. Campi, on these lectures.
On the Future of the Field
.
The summer school is organized by students of the groups of Prof. John Lygeros (IfA-ETH), Prof. Daniel Kuhn (RAO-EPFL) and Prof. Giancarlo Ferrari Trecate (DECODE-EPFL). So, to set the stage, we asked them to share their views on the future of data-driven control.
Which results in the field of data-driven control excite you at the moment and what do you hope to see in the future?
John: “I am intrigued by recent work on the connections between System Level Synthesis and data driven control. System Level Synthesis provides a nice framework for addressing, for example, robustness in the data driven setting, with interesting links to other methods, such as disturbance feedback in robust or stochastic MPC.
Going forward, a coherent framework for regret-optimal control would be a promising direction to explore. Results are emerging, but the picture is not nearly as clear as for on-line optimisation and learning.”
In your opinion, what are the most promising directions to bridge the gap between theory and practice in data-driven control?
Giancarlo: “When we imagine systems driven in real time by data streams, critical questions immediately pop into our heads. ‘Will they be dangerous?’ ‘How can one be sure they won't be hacked or compromised by viruses?’ There's a very strong need to build trust in data-driven control systems. That's why I think that research about security, reliability, and resilience to cyberattacks and failures is very important in this field. It is an indispensable step in bridging the gap between theory and practice and for enabling the massive deployment of control systems merging models and data.”
John: “Technology transfer, founding start-ups to exploit research results, or identifying ‘customers’ to licence these to. Getting there would require intermediate steps. One is the development of efficient computational tools to allow the deployment of novel and existing methods to practical systems. Then, there are many benchmarks to test methods in simulation, but for data-driven control this is just part of the story. As important are physical test-beds were methods can be validated in a convincing setting; for example the ‘living labs’ that are becoming available for energy and mobility applications.
These issues are more a concern for academic research, researchers in the tech companies have a head start in this respect!”
In the recent years we have witnessed the raise of deep learning in control: what is your opinion on that and can you comment on the hype vs reality dilemma?
Giancarlo: “In computer science, deep learning has achieved extraordinary success in many complex problems such as face recognition and content generation. It's only natural that enthusiasm for these methods is going through the roof! However, research in deep learning is also highlighting the limitations of these algorithms and applications where more traditional learning approaches work better. I believe that we will see the same trajectory in controls and that research will allow us to draw an increasingly clearer boundary between applications for which deep learning will be fundamental and applications for which efficiency and reliability of classical control methods will be hard to beat. This level of maturity and awareness will clarify once for all what today is hype and what is reality!”
Presentations and Discussions of Some Relevant Papers in the Field
