Learning Models of World Dynamics Using Bayesian Networks
Edward Tolson · 2002
Dynamic Bayesian networks (DBN's) provide a framework for learning in the context of stochastic time-series. DBN's are factored and allow learning of hidden state; in fact both hidden Markov models and Kalman filters can be thought of as subclasses of DBN's. In this thesis, I explore the use of DBN's to learn models of world dynamics. DBN's are learned both under assumption of the Markov property, and in the pursuit of learning non-Markovian dynamics, and the resulting models are evaluated.