Learning to control the linear quadratic regulator
Abstract:
Classical control theory and machine learning have similar goals: acquire data about the environment, perform a prediction, and use that prediction to positively impact the world. However, the approaches they use are frequently at odds. Controls is the theory of designing complex actions from well-specified models, while machine learning makes intricate, model-free predictions from data alone. For contemporary autonomous systems, some sort of hybrid is essential in order to fuse and process the vast amounts of sensor data recorded into timely, agile, and safe decisions.
In this talk, I will propose a framework that operates at a midpoint between the precise-physical-models of classical robust control and the model-free-but-performance-uncertain approaches taken by recent successes in reinforcement learning. Using the standard linear quadratic regulator as a case study, I will explore how learning and control can be combined, and discuss issues of safety, constraint satisfaction, and exploration. I will highlight how our techniques relate to existing methods in both the machine learning and control communities and how questions of interest to both communities can be explored through different aspects of the canonical LQR problem.