I attended the 2019 American Control Conference in Philadelphia, Pennsylvania. The conference was held in downtown Philadelphia, which is filled with beautiful historical buildings, such as the Union League building, pictured on the right.
This was a conference of firsts. Together with my former postdoc Bin Hu, who is now faculty at UIUC, I co-organized my first ACC workshop! Here is a link to the workshop website. It was a full-day workshop on the interplay between optimization, control, and machine learning. Our morning session focused on how tools from controls, differential equations, and dynamical systems could be used to better analyze and design first-order optimization algorithms, which are a key tool in machine learning. Our afternoon session focused on how ideas from machine learning, reinforcement learning, and complexity theory could be used to address large-scale and computationally challenging control problems. We were lucky to have many great speakers, and our workshop was the most popular one at ACC this year — check out the full room on the right! If you’re interested in these topics, the workshop website linked contains info for all the talks that took place (authors, titles, and slides!).
In addition to the workshop, my student Akhil Sundararajan and postdoc Bryan Van Scoy joined me at ACC and Akhil presented a paper by the three of us titled: “A canonical form for first-order distributed optimization algorithms”. In distributed optimization, the goal is for a number of agents (computer servers, robots, mobile phones, etc.) to jointly solve an optimization problem. The problem is global, but the data collection and processing happens locally. The challenge is to design an algorithm that mediates how each agent processes its local data and what is communicated with the other agents so that each agent eventually learns the solution to the global problem. There have been many such algorithms proposed in recent years, but little insight on what makes one algorithm better than another, or how to go about actually designing such an algorithm. To make matters more confusing, there are many equivalent sets of equations that can be used to describe the same algorithm. So it might happen (and it has!) that two researchers invent the same algorithm without realizing it. Our paper makes an effort to classify distributed algorithms, by providing a canonical form that uses as few parameters as possible yet parameterizes a broad class of algorithms (including many existing published algorithms). Our unifying parameterization is a first step toward a more systematic and principled analysis and design of distributed optimization algorithms.
My student Akhil, who is also first author on the paper, presented our work. As a side note, this was Akhil’s first conference presentation, and he did a fantastic job! (see photo on the right). If you’re interested in seeing Akhil’s slides, you can download them here.