Department of Electrical Engineering and Computer Science
University of Wisconsin Milwaukee.
|
- BNG - Bayesian Network Generator BNG is a tools for constructing structurally minimal Bayesian Networks from universally quantified probability logic statements with optional context contrainsts. |
| Investigators |
|
|
| Current Status |
|
BNG is a system for knowledge-based construction of Bayesian
networks. A class of Bayesian networks is specified with a
knowledge base of rules. Rules may contain universally
quantified temporal and non-temporal variables, as well as
context constraints. Context constraints are deterministic
information used to index probabilistic relations. Given a
knowledge base, a set of context information, some evidence,
and a query, BNG constructs the structurally minimal Bayesian
network to compute the posterior probability of the query given
the evidence within the given context. This is done by first
constructing the network, as indexed by the context and then
performing fast d-separation based pruning. The pruning is made
efficient by incorporating it in the network construction
process.
BNG provides both representational and computational advantages over the use of traditional Bayesian networks. The incorporation of quantified variables in the knowledge base allows the representation of information not expressible in traditional Bayesian networks. Because only a small portion of a large probabilistic model may be relevant to the computation of a particular posterior probability, use of BNG can result in significant computational savings. These savings can be particularly large with temporal Bayesian networks. The knowledge base representation language has a declarative semantics and the network construction algorithm has been proven sound and complete. BNG is written in CommonLisp and interfaces to IDEAL. A user manual and example files are provided. |
| Software |
|
| Publications |
|