Asked by:
Can Infer.net support Dynamic Bayes Network and continuous hidden variable in Dynamic Bayes Network ?
Question

Hello . I'm new to Infer.net .
I am wondering if Infer.net support Dynamic Bayes Network or continuous hidden variable/ layers in Dynamic Bayes Network ?
Best
Wednesday, May 30, 2012 3:17 PM
All replies

Hi Heng,
The answer to your question is yes, and it can be found in the FAQ  Does Infer.NET support Hidden Markov Models (HMMs)? You might want to look at the following examples  click through model and influence model, or search the forum for others.
Yordan
 Proposed as answer by Yordan ZaykovMicrosoft employee Friday, June 1, 2012 12:46 PM
Friday, June 1, 2012 12:45 PM 
Hi Yordan :
Thanks for the reply . But what I mentioned "Dynamic Bayes Network with continuous hidden variable/ layers" is more like HMM that to be unroll ?
I mean the "with random variable arrays for the variables along the HMM" , which is not supported in current version ?
Best
Heng Lu
Tuesday, September 4, 2012 9:53 AM 
Heng
Can you be more explicit about your model so that we can focus our advice? In theory you can build pretty much any flavour of HMM via the modeling API  you can have continuous or discrete, array or scalar, and as many layers as you want.
John
Monday, September 10, 2012 2:55 PMOwner 
Hi John :
Yes, my question is like this : Can we build a Dynamic byesian network like this : We have continuous munerical observations , and the observation itself is dependent on continuous hidden variables (more than one ).
For example, we have acoustic features of the speech database as observation , and we assume this continuous acoustic features is dependent on the trajectory of the phone (based ) which is a hidden continuous variable , and on the trajectory of the context information which is also a hidden continuous variable. What we know is the class of the phone and the context as discrete values as the label of the database. And by the way , the observed acoustic features obey the Gaussian distribution , can be modeled as GMM / HMM .
Simply , can we model both phone and context info using some distribution , and acoustic feature using GMM /HMM , and the GMM /HMM is dependent on the trajectory of the phone and context distribution ?
Best
Heng Lu
Heng Lu
Monday, September 10, 2012 4:15 PM 
Hi Heng
You can certainly define such a model via the Infer.NET API  but it is likely that the inference would not be accurate. The reason is this  if you want the conditional distributions representing the transitions to be Gaussian Mixture Models, you would really need to use VMP rather than EP, but as VMP does not propagate uncertainty correctly, this is very problematic in long chains. So I cannot recommend doing this. Is there a paper that you are looking at that you are trying to implement?
John
Tuesday, September 11, 2012 1:56 PMOwner 
Hi John :
Thank you . I see what you mean. I am not trying to implement some paper but to try some ideas of my own .
I have tryed some other toolkit , but most of them do not support continuous hidden variables .
thank you for the suggestion , I will think about it and try some experiments .
Best
Heng Lu
Heng Lu
Wednesday, September 12, 2012 6:08 AM 
The link for click through model is not working :(Thursday, November 1, 2018 10:17 PM