In the background of all of our electronic machines and devices are systems that sense and automatically compensate for noise, disturbances, and other unwanted behavior. These mitigations are the basis of feedback control systems, the principal focus of Laurent Lessard, who joined the Department of Electrical and Computer Engineering as an assistant professor in fall 2015.
Lessard, who grew up in Toronto, received his master’s and PhD in aeronautics and astronautics (in 2005 and 2011, respectively) from Stanford University before traveling to Lund University in Sweden to work as a postdoctoral researcher in its Department of Automatic Control.
There, Lessard was able to work with a group of about 60 researchers who all specialized in controls, an area that spans many engineering disciplines, but doesn’t have a basis in any specific field. He says he enjoys the opportunities his field presents.
“I like not being tied down to one particular problem at a time,” he says. “I like to think about what these various disciplines have in common—and more recently, where I think they’re going to go in the future.”
This is certainly evident in Lessard’s background. He has worked in fields that include adaptive optics at Stanford, mechanical engineering at the University of California, Berkeley, and now electrical and computer engineering at UW-Madison. At the university, Lessard also is a controls theoretician within the Wisconsin Institute for Discovery (WID) optimization group.
He is excited by the opportunity to work at UW-Madison, and with colleagues at WID, because the environment is highly collaborative and will allow him to pursue a variety of projects that interest him. The WID optimization group also maintains close ties with industry, ensuring that his research continues to have relevant applications.
Lessard already has made advances in decentralized control theory, an area that essentially has been on hiatus for the past 40 years—in part because the problems are especially difficult. In decentralized systems, there may be multiple controllers and feedback loops. The controllers may be spatially separated and measure different information about the system. Nevertheless, the goal is to find strategies that allow the different controllers to effectively coordinate their efforts.
In Lessard’s experience, however, not all of these problems are hard. In fact, what makes a decentralized control problem difficult has less to do with how many controllers there are and more to do with how the controllers share information with one another. The network architecture makes all the difference.
“Let’s say there are two mechanical arms on a robot, each with a separate controller, and the goal is to have the arms work together to carry an object,” Lessard says. “If the controllers don’t share any information at all, that’s not a tractable architecture. That’s very difficult to solve. But if the controllers are able to share sensor measurements with one another in a particular way, the problem can be solved very efficiently. The key is to discover which types of information-sharing lead to tractable problems.”
Although Lessard’s work focuses on theory, it provides a platform for further research into decentralized control, and allows engineers to build large-scale systems more efficiently.
While he will continue research in decentralized control at UW-Madison, he currently is focusing on problems at the intersection of control theory and optimization.
Modern machines are changing dramatically: The self-driving car is a perfect example of how controls are evolving as these changes take place. Although cars already have control systems, a self-driving car requires powerful computers that can perform image recognition on-the-fly and learn a vehicle’s surroundings as they change. Engineers from several disciplines are working on improving the self-driving car, and each type of engineer approaches the problem differently. While computer scientists consider it to be an optimization challenge, for example, control theorists view it as a feedback problem. Lessard sees himself somewhere in the middle of the two.
In many ways, a self-driving car represents the sudden union of optimization and control. In traditional large-scale optimization problems such as personalized advertising and movie recommendation engines, large-scale algorithms can afford to be aggressive because the cost of making errors is low. With self-driving cars, errors are unacceptable, and engineers are forced to be risk-averse. That’s where controls come in.
“In controls, performance is a luxury, and robustness is key,” Lessard says. “Optimization is the other way around: Performance is key, and robustness has been a luxury. How do you design for these systems of the future that are going to involve not only a lot of feedback loops, but also a lot of very intense computation? You can think of a self-driving car as an example of wanting both high performance and luxury.”
Lessard won’t be working on the next self-driving car—but the example demonstrates his far-ranging goals: To build a strong connection between optimization and control experts, therefore broadening the possibilities and capabilities of future systems.
“I think you’re going to have a lot more complicated systems being built,” he says. “Will you be able to control them the way you would like, and will they do what you think they are going to do? That’s the question.”