Linear dynamical systems research by Yanxi Chen and H. Vincent Poor wins outstanding paper at machine learning conference

July 27, 2022

Research on linear dynamical systems by Yanxi Chen, a doctoral student in Princeton University’s Department of Electrical and Computer Engineering, and H. Vincent Poor, the Michael Henry Strater University Professor, won the outstanding paper award at this year’s International Conference on Machine Learning, which was held in Baltimore, Maryland from July 17 to 23.

The paper, “Learning Mixtures of Linear Dynamical Systems,” uses a machine learning approach to understand and learn at once a mixture of multiple linear dynamical systems. These systems can encompass a wide range of complex phenomena such as a heterogeneous landscape that a robot must navigate, viral disease outbreaks, tracking multiple patients who have discrete symptoms and require different treatment plans, weather patterns, and other complex, dynamic systems that change over time and have latent variables.

Diagram from Yanxi Chen and H. Vincent paper on linear dynamical systems

A figure from the paper, “Learning Mixtures of Linear Dynamical Systems,” by Yanxi Chen and H. Vincent Poor.

Chen said they developed an algorithm to learn these dynamical systems and validated their theoretical studies with numerical experiments, thus showing that the algorithm is successful. They outlined how the algorithm works in the paper and showed how it is computationally and statistically efficient.

“It breaks new ground in being able to learn a mixture of dynamical systems,” said Poor. “It looks at the situation where you have a mixture of multiple dynamical processes going at the same time, with data coming from some or all of them, but you don’t know which are at work at a given time.”

Chen added, “You might have some measurements across a certain period of time, and the signal or data might be driven by a few different underlying physical laws or underlying models. And we don't know which model is driving which part of the signal. Our algorithm is able to understand and learn the different models in these complex systems.”

Applications for the algorithm could perhaps be deployed in robotics, driverless cars, healthcare settings and weather forecasting.

Poor noted that the conference had about 1,000 research papers presented and only 10, including their paper, won the award.

The paper can be accessed here.

The website for the International Conference on Machine Learning can be found at this link.