Dynamic Bayesian networks provide compact representations of
probabilistic dynamic systems.
An essential problem that arises is how to monitor the state of a
probabilistic dynamic system represented by a dynamic Bayesian
network.
This is a difficult problem.
Although there are a few structures for which this can be done exactly
in an efficient manner (see the paper on sufficiency), in
general exact inference is intractable, so approximate inference
algorithms must be used.
Two general families of approximate inference methods have been
developed.
The first uses sampling.
The second factors the state distribution into components.
The paper on factored particles shows how to combine the good
properties of the two algorithms.
One unresolved problem is how to factor the state distribution.
Currently this must be done by hand.
We are currently investigating methods for performing the factoring
automatically.

<u>
<li>
<a href = "../Papers/fpsm.ps">
Factored Particles for Scalable Monitoring,</a>
B. Ng and L. Peshkin and A. Pfeffer,
<i>Uncertainty in Artificial Intelligence</i>, 
2002.
</li>

<u>
<li>
<a href = "../Papers/separability.ps">
Sufficiency, Separability and Temporal Probabilistic Models,</a>
<i>Uncertainty in Artificial Intelligence</i>, 
2001.
</li>