Learning and Using Multimodal Stochastic Models : A Unified Approach

Abstract

This dissertation presents a principled approach to representing and using instance-based knowledge. Perceptions and actions are probabilistically modelled in a unified structure which allows for simultaneous perception modelling and reasoning about desired actions. In particular, a new method for online instance-based learning of such models is presented and analyzed. This method, called Dynamic Gaussian Mixture Estimation (DGME), adapts a model's complexity to the process being modelled. The models produced by DGME are evaluated on several classification, prediction, and control applications, and its characteristics are compared with other state-of-the-art methods. In the context of control applications, an additional novel method, Gaussian Mixture Control (GMC), is introduced for precisely controlling systems that exhibit multimodality.