Machine Learning, Probabilistic Inference, System Identification and Control

Abstract

In this talk, I will explore the use of recent techniques from probabilistic non-parametric machine learning to system identification and control. In this framework, the simultaneous tasks of inferring dynamics and designing a controller is thought of as a statistical inference problem. Conventionally, stochastic models (such as Gaussians) have been used ubiquitously to characterise noise and disturbances; here I will show how the Gaussian (slightly extended to a Gaussian process), can also naturally be used to model the (partially unknown) system dynamics.Thus, the inference problem is solved with Gaussian processes, using no task specific prior knowledge. Traditional wisdom may suggest that learning without strong prior information will be impractically slow, but I show that provided that uncertainties in the learned dynamics are carefully characterised, learning can be extremely rapid, and result in highly practical tools. The methods are demonstrated on a variety of control problems.