Simultaneous Inverse Optimal Control and System Inference

Many systems operate under control processes that guide their dynamics. In neuroscience, for example, one may record brain regions subject to control from other brain regions, or observe an animal's behavior performing optimally under some control objective. In most of these data-driven settings, neither the system itself nor the control objective is known. Can we infer both these components simultaneously from partial observations of the dynamics? Historically, these questions have been tackled separately: work in system identification deals with inferring the system, while inverse optimal control focuses on uncovering the underlying objective. However, the combined problem—addressing both at once—has received much less attention. In this talk, I will describe some of our recently published work on approaching this problem in a standard control framework, linear-quadratic Gaussian (LQG) control.