I attended the launch event for MIT’s new Institute for Data, Systems, and Society (IDSS). IDSS is a new entity at MIT with the goal to bring together researchers in data science, statistics, systems, economics, and social science to address big societal challenges. Examples include the modernization of transportation networks, energy generation/distribution, healthcare, and financial markets. For each of these areas, rapid growth is outpacing our ability to control, operate, and manage. The big question is: how can the judicious and principled use of data help us ensure efficiency, robustness, safety, security, and privacy moving forward?

The two-day launch event featured talks from leading economists, social scientists, medical doctors, statisticians, engineers, and other researchers, building the case for why IDSS is not only important in general, but necessary right now. Nate Silver even dropped in to talk about the future of elections! Overall a wonderful event and I look forward to seeing the research products of IDSS in the years to come! Attached is the view of the Charles River from the sixth floor of the MIT Media Lab.

I attended the 2016 International Conference on Continuous Optimization (ICCOPT) in Tokyo, Japan. This was my first optimization conference, and also my first time visiting Tokyo. Both were impressive! I attended many high-quality talks from top names in optimization, met many new people, and learned a great deal. I gave a talk about my recent work on using robust control to analyze and design iterative optimization algorithms. This talk was part of a larger invited session titled “Notions of robustness and dynamics in convex optimization”, organized by Ben Recht and Pablo Parrilo. If you’re interested, my slides are available here.

I also spent a couple days exploring Tokyo after the conference. Admittedly most of that time was spent seeking out delicious food, but I also did a bit of sightseeing. The photo on the right is the view from the Skytree Tower, the tallest tower in Tokyo. Its observation deck is at the same height as that of the CN Tower in Toronto, even though Skytree is much taller. Fun fact: the Greater Tokyo region has a population slightly larger than that of Canada!

I attended the 2016 American Control Conference in Boston. A wonderful city, beautiful weather, and as usual, a well organized event. The banquet at the New England Aquarium was a particularly nice touch!

On the research side of things, in addition to the many great talks I attended, I particularly enjoyed two semi-plenary lectures on robotics. The first, by Neville Hogan (MIT), was about the challenges of building and controlling therapeutic robots, e.g. for helping a stroke patient re-learn how to walk. It turns out the robots are only part of the equation — the real mystery is that we still don’t really understand the human brain’s role in walking. This makes re-learning even more difficult! The second talk, by Aaron Ames (Georgia Tech) was about having robots do the walking themselves. This is also a very challenging problem and it’s encouraging to see the tremendous progress that has been made even over the past decade. Aaron’s talk featured many videos of walking robots from his lab and even a live demonstration!

I recently attended the IMA Workshop on Optimization and Parsimonious Modeling at the University of Minnesota. This was my first opportunity to attend an optimization workshop and interact with a new community of researchers. Although the controls and optimization communities do intersect, there were many new faces here. Thanks to the great venue and moderate size of the event, I had the opportunity to meet and have great discussions with many new researchers outside my usual field. I particularly enjoyed the workshop format; a single-track of talks with frequent breaks for discussion. Kudos to the organizers for a job well done!

I also had the opportunity to present a poster about my work on analyzing optimization algorithms using integral quadratic constraints. The poster can be found here.

I attended the 2015 Conference on Decision and Control in Osaka, Japan. This was a special trip for me as it was my first conference attending as an assistant professor. I attended many excellent talks and had several productive discussions. It was also my first time ever visiting Japan! I stayed a couple extra days after the conference to do some sightseeing around Osaka and Kyoto. The photo on the right was taken at Kiyomizu temple in Kyoto. The weather was very pleasant, the food was amazing, and I can’t wait for my next visit to Japan!

At the conference, my co-author Ross Boczar (from Berkeley) presented our paper (also co-authored by Ben Recht) “Exponential convergence bounds using integral quadratic constraints”. This work shows how the classical theory of Integral Quadratic Constraints can be generalized to provide certificates for exponential convergence rather than just input-output stability. This tool was applied with great success in a recent paper by me, Ben Recht, and Andy Packard to prove tight convergence bounds on the convergence of many different iterative algorithms commonly used in large-scale optimization.

I recently attended the IMA Workshop on Distributed Control and Decision Making over Networks at the University of Minnesota. The workshop was very well organized; I feel that the perfect balance was struck between seminars, guided group discussions, and free blocks of time well suited for one-on-one discussions. Overall a very positive and productive experience, and hopefully I will be attending more IMA workshops in the future!

I am excited to announce that I have accepted a faculty position at the University of Wisconsin–Madison! Starting in a couple weeks, I will join the Optimization Group at the Wisconsin Institute for Discovery. My tenure home department will be Electrical and Computer Engineering. If you’re curious about my research and how I’ll fit in at Wisconsin, the WID published this profile about me.

The WID is a very collaborative place, so my hope is that the new position will be an opportunity for me to learn a great deal more about optimization, machine learning, and new engineering applications. At the same time, I will contribute my own expertise in control theory to help broaden the group’s reach and impact. Of course, I am also looking forward to teaching classes, mentoring students, and helping shape future research and teaching at Wisconsin–Madison. I feel very lucky and privileged to have this opportunity, and I could not have made it this far had it not been for the support of my friends, family, colleagues and mentors. Thank you all, and if you’re ever in the Madison area, drop me a line!

The paper “Optimal Control of Two-Player Systems with Output Feedback” by me and Sanjay Lall will appear in the August 2015 issue of the IEEE Transactions on Automatic Control. The paper derives an analytical expression for the optimal LQG controller in perhaps the simplest decentralized architecture: two players with unidirectional information sharing between them. The centralized (single-player) version of this problem leads to a classical result in control theory: the separation principle. Although this principle does not hold for decentralized problems, we nevertheless recover a very elegant structure for the optimal controller. The controller’s state dimension is twice that of the plant, since two different state estimates must be maintained. Surprisingly, the computational effort required to find the optimal decentralized controller is comparable to the effort required for the centralized case. I am currently working on extensions of this result to architectures involving multiple players. A more spaced-out and legible version of the paper is also available on arXiv, and the final published version of the paper is available here.

I have created a code section on this website, and it is now live! The idea is to make code accessible that allows other researchers to reproduce my published work and build upon it. For this reason, the code I post is written with an emphasis on accessibility and ease of use; I made no effort to optimize for speed or numerical performance. At the moment, I only have one project in the code section: some recent work on IQC analysis techniques applied to algorithm analysis and interconnected systems. I hope to add more in the future!

O’Reilly Media’s chief data scientist Ben Lorica (@bigdata) posted an article on the popular Radar Blog covering some of our research on analyzing/tuning optimization algorithms! As Ben points out, machine learning projects in practice consist of pipelines — complex sequences of algorithms that are tailored to the particular task at hand. As the pipelines become larger and more complex, it becomes increasingly difficult to maintain them and to predict how changes in one component will affect other components downstream. Our work (see arxiv paper here) takes important steps towards the analysis and design of such pipelines by bringing to bear tools from robust control theory.