Results
Weather Network
Junction tree proves to be the fastest inference technique. Pearl's
algorithms takes approximately five times as long, and variable
elimination takes dramatically
longer. The fact that many
iterations after the first 5 to 10 take effectively zero time
is a result of the fact that some
inputs do not add any new
information to the network, in
which case they can be discarded
without performing any significant
work. Variable elimination
actually performs faster than Pearl
when simply adding data to the
network, but when attempting to
extract marginal data it is
dramatically slower.
Intrusion Detection Network
The intrusion detection network is significantly larger than the
weather network, and the effect on the variable elimination method is
remarkable. Variable elimination is now much faster than Pearl in the
inference step, but in the marginal step it is massively slower than
the other two techniques. I would speculate that the variable
elimination is degenerating into exponential behavior while the other
methods remain more efficient. Junction tree remains the fastest of
the methods, beating Pearl's algorithm 4.96 seconds to 5.68 seconds.
Insurance Network
The insurance network is similar to the intrusion detection network in
size, and it gives similar results. Junction tree is nearly twice as
fast as Pearl, but both are approximately 100 times faster than
variable elimination.
Conclusions
I have omitted any discussion of accuracy from the results above
because my tests failed to discover any significant difference in
predictive accuracy between these methods. Junction tree was clearly
the fastest of the techniques that were tested. However, because it is
the default inference technique for this toolkit, it has been optimized
further than the other techniques. Therefore, it is impossible to
conclude that junction tree is superior to Pearl's algorithm outside
the context of this toolkit. However, it is quite clear that both of
these techniques have an advantage over variable elimination,
especially when extraction of marginal data is important.
