Template subnets? Class nodes?

[a.s]

I hope my question isn't too stupid (I am rather new to Bayesian network analysis).

I am trying to model a set of students doing peer evaluations.
I am modelling each student as a subnet, with the same CPTs used in the nodes representing their behaviours/decisions.
In some sense, I am defining a subnet as if it were a general Student class, with the CPTs defined at class level, to be instantiated in several copies to be connected depending on their actual interactions.
Then I would like to learn the "class CPTs" from actual data.

Is there some "templating" mechanism in Genie that I could use to simulate this?
(perhaps I could use a temporal plate to emulate the class, provided that the interactions among instances were topologically orderable, but my cases contains normally cycles)

Else, at a lower level:
How could I constrain the CPTs of some nodes not to change during learning? (I am using some nodes for well-known decisions)
How could I constrain the CPTs of a set of nodes to be the same across all their copies? (to simulate "class CPTs")

Thanks

[shooltz[BayesFusion]]

Is there some "templating" mechanism in Genie that I could use to simulate this?

No, there's no such feature other than copy/paste (obviously this only works one way and does not keep track of the source nodes).

How could I constrain the CPTs of some nodes not to change during learning? (I am using some nodes for well-known decisions)

Use the 'Fixed nodes' feature in EM. In GeNIe you can access it in the 'Match Network and Data' dialog.

How could I constrain the CPTs of a set of nodes to be the same across all their copies? (to simulate "class CPTs")

You can't.

[a.s]

Thanks for your fast answer!

Do you think I could use some kind of trick to emulate several instance nodes, copies of a single one containing the "true" CPT ?

The only trick coming to my mind now is using temporal plates, mapping each one of the student's subnet to a different temporal moment...
Yet this would constrain my network only to regular interconnection patterns, without loops ...

This is starting to work ... even if it constrain me to use a regular network.
Now the network contains only one 3 CPT instead than 30 ... not bad.

... let see ... but how to use Genie to learn the CPTs? (the dialog is all different from the normal one)
... and how to select fixed nodes?

[shooltz[BayesFusion]]

... let see ... but how to use Genie to learn the CPTs? (the dialog is all different from the normal one)

The first dialog displayed for parameter learning on a DBN is used to assign data columns to their corresponding time series (nodes on the temporal plate). This is done automatically if column names follow the pattern "var", "var_1", "var_2", ... (note that 0th slice has no numeric suffix).

... and how to select fixed nodes?

The 'Fixed nodes' button is in the 'Match nodes and data' dialog.

[a.s]

It works a charm, the temporal plate emulates a class almost exactly as I wanted!
Thanks!