Learning and Using Multimodal Stochastic Models: A Unified Approach

This talk introduces a principled approach to representing and using instance-based knowledge called Dynamic Gaussian Mixture Estimation (DGME).  With this method, continuous streams of perceptions and actions can be probabilistically modelled in a unified structure, which allows for simultaneous perception modelling and reasoning about desired actions.  Examples will be presented in which models produced by DGME are applied to classification, prediction, and control problems. In the context of control, a novel method is introduced for precisely controlling systems that exhibit multimodality.