Information gain calculation oddities

[lobstericus]

Dear SMILE Team,

I have a network where 'A ' and 'C' are 'target's and
'B' and 'D' are observations

A-> B

C-> {D,B}


If I use the 'test view' from the diagnosis menu I see that both
'D' and 'B' are informative.

If I break this into two networks, I should in theory still be able to
still compute entropy gain as 'B' tells me about 'A' and 'D' tells me
about 'C'.

A->B

C->D

However, testview (and direct calls to a smile engine running this net)
claim that observing 'B' now provides 0 bits information. Why does
the engine arbitrarily decide 'B' now provides no information?

I attached a sample network for the problem.

Does this make sense? Am I missing something?

Regards,

Bob.

Version 2.0.3470.0 7/2/2009

[shooltz[BayesFusion]]

The information gain is calculated for the pursued fault(s); these are displayed in red. You can change the pursued fault set by selecting them and invoking the context menu. GeNIe/SMILE can pursue more than one fault (single pursued fault is used by default when you open the diagnostic window).

When you change the pursued fault from A to C, you'll also see that the node with nonzero information gain changes from B to D. If you pursue A and C, both B and D will have nonzero gain.

BTW, I also think you should only keep the FAIL states as diagnostic targets in A and C (there should be only one checkbox checked in the 'Target' column of the grid displayed in the node properties/general page).

[lobstericus]

Thanks for your reply. I didn't realize that you could select multiple targets in the diagnosis view!