DBN, inference algorithms and unrolling

[Vin]

Hello,

I would like to know if the inference algorithms are applied on unrolled dynamic bayesian network (so that they are effectively applied on a static bayesian network), or if they use original dynamic bayesian network. I ask because when changing time step count I can see a comment that "Enter the number of time steps to which dynamic network will be unrolled during inference and after invoking the "Unroll" command.".

Regards,

Vin

[shooltz[BayesFusion]]

DBN inference runs on the unrolled network (the unrolling is performed automatically).

[saifjamal]

Hi all

I am using GeNIe 2.0, I have performed the inference in first order DBN using Polytree. Is that the same Forward backward in the HHM?

[Martijn]

Hi,

Polytree is the algorithm proposed by Pearl:
Here's a reference:

Judea Pearl. (1986). Fusion, Propagation, and Structuring in Belief Networks. Artificial Intelligence, 29(3), 241-288
Pdf: http://www.ece.tamu.edu/~bjyoon/ecen689 ... l_1986.pdf

[saifjamal]

Hi

Thank you for your reply, i mean if i perform the inference using this algorithm will it work and the results will be ok

[saifjamal]

i have another question, if i perform inference in 3 time slices will get different results if i perform it in 6 timeslices. in both cases i set the same evidences. i have noticted that the system is updating the belief in all time slices which means that the polytree ( message passing pearl) is working in a DBN exactly as forward backward propagation. am i correct?

i want a reply please

[shooltz[BayesFusion]]

if i perform inference in 3 time slices will get different results if i perform it in 6 timeslices. in both cases i set the same evidences. i have noticted that the system is updating the belief in all time slices which means that the polytree ( message passing pearl) is working in a DBN exactly as forward backward propagation.

SMILE unrolls the DBN then applies the Bayesian inference algorithm to the unrolled network. The choice of algorithm (Pearl vs clustering etc) should not have any impact on the beliefs, assuming there's enough memory to complete the inference call. Is there any specific reason you've chosen Pearl instread of the default (clustering) ?

Since unrolling to 3 slices creates different BN than unrolling to 6 slices, it's possible to obtain different results.

i want a reply please

Such comments are unnecessary and, in fact, reduce the attention the forum question gets from us.

[saifjamal]

Hi
Thank you for your reply and sorry for the comments i don't mean anything by that comments.
i am new in using genie so I have understand that polytree algorithm is the same as clustering instead the clustering work for multiply connected and my network is singly connected. will i get different results if use clustering?

I have designed my network and i have one hypothesis node with 2 parents and 5 children the only dynamic node have temporal arc order 1 is the hypothesis node. i have understand that the polytree perform the diagnostic and predictive inference by considering the node at the previous time slice as a parent node in time t. am i right?

[shooltz[BayesFusion]]

i am new in using genie so I have understand that polytree algorithm is the same as clustering instead the clustering work for multiply connected and my network is singly connected. will i get different results if use clustering?

You should get identical results. If the beliefs you see are different when you change from polytree to clustering, post the network here so we can investigate the issue.

i have understand that the polytree perform the diagnostic and predictive inference by considering the node at the previous time slice as a parent node in time t. am i right?

No, the polytree doesn't have any knowledge of the time slices. All inference algorithms in GeNIe work on non-dynamic network, which is created in the background based on DBN and slice count.

[saifjamal]

Thanks for reply
Sorry for asking too much but i just want to understand exactly so i can justify my work.
I am getting the same results for both algorithms.
Can i know how the network takes into account the hypothesis node at the previous time slice while computing the posterior probability of the hypothesis node at time t.

[saifjamal]

I have read that the inference in DBN can be done by either considering the unrolled network as exactly as static network and we can perform polytree or converting the network to a HMM and perform forward backward. In this case i have to say that we consider the dynamic network as unrolled static network and we perform poly tree. is that right?

[shooltz[BayesFusion]]

In this case i have to say that we consider the dynamic network as unrolled static network and we perform poly tree. is that right?

Yes, that's correct.

[saifjamal]

thank you very much for your help
i really like the software and your support

[saifjamal]

just want to understand how the network will take into account the hypothesis node at the previous time slice during infering the hypothesis node at current time slice

[saifjamal]

is the first order temporal arc means that the relationship between two time slices modeled by first order hidden markov model. thats what made me confused