Hello,
I learned the parameters of my Bayes-net using the EM-algorithm. I have no a priori knowledge about the parameters. Is it better to start with a uniform distribution or to runthe algorithm n-times using the randomize-initial-parameters-option and choose the one with the highest log(p)-value?
If I use the randomize-initial-parameters-option: is it also better to start with the uniform distribution since for some combinations exist no training data.
For combinations where no training data exist the EM-algorithm even with randomizing of initial-parameters seems nothing to change. So it is also with using randomizing important for the result to start with plausible start parameters. Is starting with a uniform distribution here the best?
By the way im surprised that even with randomizing parameters the same result is achieved for the combinations where training data exists. I used around 2000 data records and exactly the same parameters (until to last digit after the dot) were the result.
samoth
initial parameters
Re: initial parameters
To answer your questions, can you tell me how much of the data is missing and are there variables that have no data at all?
Re: initial parameters
Hallo,
my training data sets are all complete.
samoht
my training data sets are all complete.
samoht
Re: initial parameters
Then these settings don't really matter. If there is no data available for a certain parameters in a CPT, then these won't be updated and will keep their original uniform or randomized values.