About Parameter Learning
About Parameter Learning
When I learn the parameter of a constructed Bayesian network with a complete learning dataset, the default learning algorithm is still EM algorithm with uniform distribution. How does the EM algorithm deal with the complete dataset? I found the result is slightly different from the one estimated by Maximum Likelihood Estimation, which I originally expect the results will be the same.

 Site Admin
 Posts: 282
 Joined: Tue Dec 11, 2007 4:24 pm
Re: About Parameter Learning
The EM algorithm replaces all missing values by the most likely values given the values of all the other variables and the current model and then learns the network parameters using MLE. It repeats this process until the change between the values of the parameters between two subsequent runs is smaller than a predefined threshold. When the data file contains no missing values, the algorithm deteriorates to MLE, so for a data set with no missing values, EM should result in learning similar parameters.
I hope this helps,
Marek
I hope this helps,
Marek