|
Abstract:
|
A discrete-time Markov process can be compactly modeled as a dynamic Bayesian network (DBN)--a graphical model with nodes representing random variables and directed arcs indicating causality between variables. Each node has a probability distribution, conditional on the variables represented by the parent nodes. A DBN's graphical structure encodes fixed conditional dependencies between variables. But in real-world systems, conditional dependencies between variables may vary over time. Model errors can result if the DBN fails to capture all possible interactions between variables. Thus we propose adaptive DBNs, whose structure and parameters can change: a distribution's parameters and its set of conditional variables are dynamic. This work builds on nonparametric Bayesian modeling and machine learning methods, such as structural EM. We show its advantages in a complex plant modeling task.
|
