BayesFusion Support Forum

There could be several reasons, but no data is the most likely one.

You can check the GeNIe documentation (http://genie.sis.pitt.edu/wiki/Main_Page) but I'm afraid it's lacking a bit at the moment.

Do you have any experience with Dynamic Bayesian Networks? It sounds to me they do exactly what you want.

I don't understand what you're trying to do but have you thought about using a Dynamic Bayesian Network to handle time?

You are right and we will make the change. Thanks for the feedback.

EM algorithm seems can solve problem about Missing Value .. if there is no Missing Value in my data file? [Run EM Algorithm]step will calculate the probability of nodes? EM estimates the parameters of all the conditional distributions (CPTs) in the network regardless if any of the values are missing.

1. It looks at the relative improvement of the log likelihood. When it's below a certain threshold, learning is stopped. This criteria cannot be changed at the moment.

2. It is the data log likelihood of the learned parameters.

GeNIe uses the basic EM, no special alterations.

I haven't studied the paper in-depth, but it seems to me this is some sort of modified EM algorithm which would explain the differences. The 'plain' EM algorithm has no learning rate parameter.

What do you mean by learning rate in the context of EM? Also, the initial parameters influence the final parameter values as EM may get stuck in a local maximum (at least in the case of incomplete data).

It is possible that EM converges to 0 probabilities if the priors counts are set to zero. This is controlled by the confidence parameter when you invoke EM. Did you set this to a number larger than 0?

Can you tell me how you invoked EM?

Please refer to Heckerman's "A Tutorial on Learning With Bayesian Networks" for an explanation of GreedyThickThinning.

It look ok, although many entries in the CPTs do not seem to be updated because of a lack of data. There is a big difference between learning with and without unrolling. If you unroll, the CPTs at each time step are learned separately. However, these CPTs are usually assumed to be identical and then...

The right way is to perform learning on the original network and never on the unrolled network. This way your predictions should be much more accurate. Can you give this a try?