Hidden node states and learning from data

[mh]

Hi,

I am trying to learn a BN from data which does not contain any information on one network variable. When I learn the model e.g. with PC+EM, this hidden variable ends up having a single state (I am using Hugin v7.01). I would like it to have several states. Is it possible to somehow define the states of a hidden variable for the learning, if I know them? Or should I just learn the structure first, then manually add more states to the hidden variable, and then use EM separately for learning the probability distributions?

-mh

[Anders L Madsen]

Hi

Is it possible to somehow define the states of a hidden variable for the learning, if I know them?

You will have to do structure learning and manually add the states of a hidden node before running EM. It is not possible to add states to a node in the structure learning wizard.

Or should I just learn the structure first, then manually add more states to the hidden variable, and then use EM separately for learning the probability distributions?

Yes.