A third way to parallelizing the exact computation of probability in
Bayesian networks (besides table split up and local conditioning)
consists in exploiting the fact that, at each moment of the
propagation, there are usually several messages that can be computed
independently at different processors.
Further details can be found in
http://www.dia.uned.es/~fjdiez/papers/distrib1.html
The method is similar to Pearl's proposal for the polytree, but
differs mainly in three points:
--it can be applied to networks with loops (either via
clustering or via local conditioning);
--it does not need a processor for every node;
--the messages corresponding to a certain node may be computed by
different processors.
Javier Diez