RKHS Methods for Control and Time Series Analysis

29^th November 2012, 16:00 Ashton Lecture Theatre
Dr. Steffen Grünewälder
Computational Statistics and Machine Learning
University College London

Abstract

Reproducing Kernel Hilbert Spaces (RKHSs) are one of the key tools in machine learning with which to perform inference. I will show why they are so useful and will give an overview of some of my recent work in which RKHS techniques are used to derive agent learning algorithms and to process time series data. The application of RKHS methods to agent learning problems has been something of a niche topic, but I will argue that we are close to a major breakthrough. Two of the major challenges in agent learning are: the processing of real world sensor data, like camera images, into low dimensional representations; and learning hierarchical splits of control problems into simpler sub-problems. I will explain why RKHS methods appear so promising in tackling both problems. Finally, I will discuss an application of RKHS methods to the processing of real world time series data. In my case, this is data from wild animals in Botswana. This is an area of ecology in which new, affordable hardware has been deployed and is generating very considerable amounts of data. The analytical methods in use at present are not suited to the data volume and there is a significant opportunity to develop new analytical techniques that can change the nature of the science that can be done.