How to learn DBN parameters with incomplete data

JohnYu · Post by **JohnYu** » Thu Mar 27, 2008 10:11 am

Hi,
I created a dynamic bayesian network and want to learn the parameters from data(time series). My network contains a hidden node(discrete state), so it's not observable. How can I do this in GeNIe ?
Thanks.

Best Regards,
JohnYu

mark · Post by **mark** » Thu Mar 27, 2008 7:38 pm

Parameter learning is only possible for static networks at the moment. We are planning to extend it to dynamic networks as well, but I don't know when that will be completed.

JohnYu · Post by **JohnYu** » Mon Mar 31, 2008 4:21 am

Thanks for your reply.

I don't know how DBN parameters learning works.
May I expand the dynamic model for required time slice and preprocess the data for the format and then use GeNIe's parameter learning?
For example, if I expand the model to 3 time slice, and the raw data is
" 3 4 2 1 5 6 7 2 4 5 3", then the data need to be transformed as
"3 4 2", "2 1 5", "5 6 7", " 7 2 4", "4 5 3", totally five records.
If the model have markov property, the the model will be expand to two time slice. Is that right?
Thanks.

mark · Post by **mark** » Mon Mar 31, 2008 3:05 pm

I will actually implement parameter learning for DBNs somewhere in the next two weeks. Do you have time to wait for that?

JohnYu · Post by **JohnYu** » Mon Apr 07, 2008 7:02 am

Thank you.
I hope it won't spare you too much time.
I'm not in a hurry.