| Summary: | In recent years, researchers in the A l domain have used Bayesian belief networks to build
models of expert opinion. Though computationally expensive deterministic algorithms have been devised,
it has been shown that exact probabilistic inference in belief networks, especially multiply connected
ones, is intractable. In view of this, various approximation methods based on stochastic simulation
appeared in an attempt to perform efficient approximate inference in large and richly interconnected
models. However, due to convergence problems, approximation in dynamic probabilistic networks
has seemed unpromising. Reversing arcs into evidence nodes can improve convergence performance
in simulation, but the resulting exponential increase in network complexity and, in particular,
the size of the conditional probability tables (CPTs) can often render this evidence reversal
method computationally inefficient.
In this thesis, we describe a structured simulation algorithm that uses the evidence reversal
technique based on a tree-structured representation for CPTs. Most real systems exhibit a large
amount of local structure. The tree can reduce network complexity by exploiting this structure to keep
CPTs in a compact way even after arcs have been reversed. The tree also has a major impact on improving
computational efficiency by capturing context-specific independence during simulation. Experimental
results show that in general our structured evidence reversal algorithm improves convergence
performance significantly while being both spatially and computationally much more efficient
than its unstructured counterpart.
|
|---|
