DBN question

[JohnSandiford]

Hi,
DBN support is a great addition to the software. I do however have the following question/issue.

I have a network, part of which contains the nodes:

A (Initial/Anchor)
B (Temporal)

with the following links:
A->B
B->B (order 2)

When I come to specify the network parameters for node B, I only get distributions at t=0 and t=2. This is strange because at t=1 node B has no parents, and so is different to t=0 and t=2. If I unroll the network the distribution at t=1 is a truncated version of the distribution at t=0.

Is this by design, or is it a bug?

Regards

[marek [BayesFusion]]

DBN support is a great addition to the software. I do however have the following question/issue.

I have a network, part of which contains the nodes:

A (Initial/Anchor)
B (Temporal)

with the following links:
A->B
B->B (order 2)

When I come to specify the network parameters for node B, I only get distributions at t=0 and t=2. This is strange because at t=1 node B has no parents, and so is different to t=0 and t=2. If I unroll the network the distribution at t=1 is a truncated version of the distribution at t=0.

Is this by design, or is it a bug?

John,

It is by design, although I agree that the behavior is funny. My inclination is to blame it on the funny model -- it does not make much sense to have an order 2 link without order 1 link. You might want to redefine the time ticks and have just order 1 link. How would you imagine an ideal behavior of the DBN feature for this model?
Cheers,

Marek

[JohnSandiford]

Thanks for the reply Marek.

I was really thinking of the example given as part of a larger network, for example we could add:

C (Temporal)

and the additional links:

C->B
C->C (order 1)

(see attached file)

The example is purely hypothetical, and it may be that in practice one would most likely include an additional link of order 1 to B, however I was surprised to see the truncated distribution.

I would like to reiterate that I am delighted to see DBN support in the library and I am a big fan of your library.

[marek [BayesFusion]]

Thanks for the reply Marek.

I was really thinking of the example given as part of a larger network, for example we could add:

C (Temporal)

and the additional links:

C->B
C->C (order 1)

(see attached file)

The example is purely hypothetical, and it may be that in practice one would most likely include an additional link of order 1 to B, however I was surprised to see the truncated distribution.

I would like to reiterate that I am delighted to see DBN support in the library and I am a big fan of your library.

Thanks for the example. You have got a point -- the enrolled network is funny in that it has two independent strains of C. The question is whether the network makes sense or not and, if ot does, what would be a theoretically and practically sound way of enrolling it. If you have any ideas here, please do let us know -- we will give it a serious thought.

And thanks for your kind words -- we are trying hard .

Marek