I am trying to build a DBN using SAMIAM . It uses sample to train by EM . But since in reality the value of unobserved/latent variable is not known , I used N/A but it gives just the prior already specified for latent.
I was wondering where I go wrong?
EM is specially designed for the case of unobserved variables. Their value is never explicitly computed, but the probability distribution over their domain is estimated (that's the E step). Given this estimation over latent variables, the most probable model is then computed (that's the M step, a maximization over parameters). That's the general idea.
Concerning the practical implementation in SAMIAM however, since I do not know the software, I cannot help you more specifically. But my guess is that you should not try to set values for unobserved variables, since that's the EM's job. Hope this helps!
Thank you Julien Diard, I could perform it using the Murphy bayes net toolbox for Matlab.Your comment was really helpful and gave an overall idea of what is going on.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series