Query on EM learning in HUGIN

[Sid Ghosh]

Hi,

I have a few queries on using Hugin which I wish to address.

1). I am trying to apply EM learning to a dataset and I keep getting this exception

"The state range of the node is insufficient for the chosen standard distribution"

Could you clarify why might I be getting this error?

2). Secondly, I am trying to model a problem where I have 3 discrete variables A, B and C. Assume each has three levels and the DAG is

A->B and B->C

Assume we know the CPTs of p(A) and p(C|B). Let B be a hidden variable. Firstly, how can I specify that B is hidden.

Secondly, if I have observations of the states of A and C, is it possible to learn the CPT for p(B|A) from data?

I would appreciate any help.

Kind regards,
Sid

[Martin]

"The state range of the node is insufficient for the chosen standard distribution"

This usually means that some values in your data cannot be mapped to node states.
You must correct your data before you can continue.
Look for red cells in the EM leraning data window. A red cell means that the value cannot be mapped to a state.

[Frank Jensen]

2). Secondly, I am trying to model a problem where I have 3 discrete variables A, B and C. Assume each has three levels and the DAG is

A->B and B->C

Assume we know the CPTs of p(A) and p(C|B). Let B be a hidden variable. Firstly, how can I specify that B is hidden.

Secondly, if I have observations of the states of A and C, is it possible to learn the CPT for p(B|A) from data?

There is no need to specify explicitly that a variable is hidden.

A hidden variable can be learned, but the EM algorithm works best if you start with a non-uniform initial distribution for p(B|A).

The EM algorithm uses the presence and absence of experience tables to control which conditional distributions will be learned.  In this case, A and C should not have experience tables, but B should.