Machine Learning For Modeling Real-World Dynamical Systems

Real-world sensor data in robotics and related domains is sequential, high-dimensional, noisy, and collected in a raw and unstructured form. In order to interpret, track, predict, or plan with such data, we often assume that it is generated by some underlying dynamical system model. Although we can sometimes use extensive domain knowledge to write down a dynamical system, specifying a model by hand can be a time-consuming process. This motivates an alternative approach: learning the model directly from sensor data. This is a formidable problem. Revealing the dynamical system that governs a complex time series is often not just difficult, but provably intractable. Popular maximum likelihood strategies for learning dynamical system models are slow in practice and often get stuck at poor local optima, problems that greatly limit the utility of these techniques when learning from real-world data. Although these drawbacks were long thought to be unavoidable, recent work has shown that progress can be made by shifting the focus of learning to realistic instances that rule out the intractable cases.

In this talk, I will present an overview of my work on modeling a range of robotics problems as dynamical systems. I will then focus on several related computational approaches for learning dynamical system models directly from high-dimensional sensor data. The key insight is that low-order moments of observed data often possess structure that can be revealed by powerful spectral decomposition methods, and, from this structure, model parameters can be recovered. Based on this insight, we design highly effective algorithms for learning parametric models like Kalman Filters and Hidden Markov Models, as well as nonparametric models via reproducing kernels, and new models based on predictive state representations. Unlike maximum likelihood-based approaches, these new learning algorithms are statistically consistent, computationally efficient, and easy to implement using established linear-algebra techniques. The result is a set of tools for learning dynamical system models with state-of-the-art performance on video, robotics, and biological modeling problems.
Byron Boots is an Assistant Professor in the School of Interactive Computing and the College of Computing at Georgia Tech. He directs the Georgia Tech Robot Learning Lab, which is affiliated with the Center for Machine Learning, the Institute for Data Engineering and Science, and the Institute for Robotics and Intelligent Machines. His research focuses on developing theory and systems that integrate perception, learning, and decision making. His work on learning models of dynamical systems received the Best Paper award at the International Conference for Machine Learning (ICML) in 2010. His research is supported by a NSF CISE Research Initiation Initiative award (CRII), a NSF National Robotics Initiative award (NRI), and BMW Manufacturing. Prior to joining Georgia Tech, Dr. Boots was a postdoctoral researcher working with Dieter Fox in the Robotics and State Estimation Lab at the University of Washington. He received his Ph.D. in Machine Learning from Carnegie Mellon University advised by Geoff Gordon.