A second way to compile a Bayesian network, built on variable elimination and algebraic decision diagrams, that exploits local structure far better than earlier elimination methods and can answer many queries at once.
Compiling a network once had proven an effective route to inference that uses both its global and its local structure. This paper defines a new way to compile, based on variable elimination and algebraic decision diagrams. It exploits local structure much more effectively than earlier elimination methods, lets any variant of variable elimination answer multiple queries at the same time, and makes a large body of work on structured factor representations useful in many more settings.
In experiments, elimination exploited local structure as effectively as the best conditioning-based algorithms on the networks tested, and sometimes compiled much faster.
One of the routes to a compiled model. The point that matters for everything on this page is the same: compile once, then answer every later query fast and exactly.
A methods paper. No spacecraft or product claims attach to it.