Impossible outcomes, and how to deal with them

[s.arikan]

Greetings,
I'm using Genie & Smile for my PhD, and I've come across an interesting situation which must be a common case.
I learn the parameters of my network using some training data, and then I test the performance of the network by predicting the outcomes using test data.
Some of the observations in the test data are rejected by Genie (and I'm sure Smile will do the same) with error stateX is of nodeY impossible.
Obviously, this is due to training data assigning zero probability to something that exists in the test data.

1) Is there a recommended way of dealing with this situation in Genie & Smile?
2) I am very interested in references to papers/books about this particular problem, in the context of Bayesian Networks. (I've got some pointers, but I'd really appreciate some good starting points)

Regards
Seref

[mark]

Can you tell me how you invoked EM?

[s.arikan]

Hi,
In GENIE I've learned the structure of the network from data using PC. Then in Genie, I've selected learn parameters from the network menu, arranged the order of some data items so that they match states in nodes, and clicked next. That is it.

[mark]

It is possible that EM converges to 0 probabilities if the priors counts are set to zero. This is controlled by the confidence parameter when you invoke EM. Did you set this to a number larger than 0?

[s.arikan]

Thanks a lot! I'll check this out, and share the results here.