Lisa Li wins IEEE TCNS Best Paper Award for work on guarantees for model predictive control

Lisa Li, assistant professor of Electrical and Computer Engineering (ECE), received the 2024 IEEE Transactions on Control of Network Systems (TCNS) Best Paper Award. The award recognizes one paper, from those published in the journal within the last two years, based on “originality, potential impact on the foundations of network systems, importance and practical significance in applications, and clarity.” The selected paper, “Distributed and Localized Model Predictive Control—Part II: Theoretical Guarantees,” showed that stability and safety guarantees can be made on distributed model predictive control systems.

Model predictive control involves using sensor data to determine what action a system should take to keep a variable of interest within certain bounds. A simple example of this is the cruise control on a car—the throttle is adjusted based on speed sensor data in real time to ensure that the car’s speed stays constant.

“What we did was take this type of algorithm and scale it up to really large networks with incomplete information access,” explained Li.

The example that Li uses in her work is a power network, consisting of many nodes that supply power to houses on a shared electricity grid. Each node tries to supply a certain amount of power while remaining stable; if a tree falls on a powerline in the next neighborhood over, the system shouldn’t propagate that instability through the whole network as a grid failure. 

In practice, the nodes in existing power networks are controlled by a central communication mechanism. Li models an alternative type of system, where the nodes are distributed and communicate with each other, rather than a central hub.

“We apply the algorithm at each node in the graph, so now we’re saying, instead of an centralized entity that communicates to every single house, each house is just going to communicate with its neighbors,” she said. “And that’s actually going to be enough for the system to stay safe and stable.”

Li’s paper shows that the same guarantees can be made on both centralized and distributed networks. Compared to centralized networks, distributed networks can decrease the number of variables under computation at each node, allowing them to be more efficient, protected against system-wide failure, and flexible to structural changes. A related paper that she published in IEEE Open Journal of Control Systems found that these benefits don’t always have to come at the expense of network performance.

“There are certain ways you can take advantage of existing network structure that actually lets you have scalability, efficiency, and guarantees—all essentially for free,” she said.

These findings were surprising even to Li, who expected to find trade-offs between distributed communication as a constraint on information access, and performance.

Li did this work as part of her PhD work at the California Institute of Technology, before joining U-M ECE as an assistant professor in 2023. As the PI of her current research group, she plans to investigate how she can apply distributed control theory to neuroscience.

“Distributed and Localized Model Predictive Control—Part II: Theoretical Guarantees” was co-authored with Carmen Amo Alonso, Nikolai Matni, and James Anderson.