What is the learning rate that is used by default in GeNIe.

[nisha269]

Hi,
I am a newbie to Genie and have been learning the tutorials given. I am trying to learn parameters for the alarm-burglary network (attached) with 2 hypothetically created dataset (attached). When I Learn parameters, Genie says it is successfully learnt and gives me a resultant network with CPT tables. But when I calculate manually for the 2 datasets using the EM Algorithm with a learning rate 0.5, the CPT tables donot seem to match with the ones that come from Genie.
In Genie I also checked the randomize initial parameters when I did learn parameters.

My question is what is the learning rate that Genie uses to compute the CPTs for a given data and how can I change this rate?

Hope someone can help me asap.

Nisha

[mark]

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).

[nisha269]

By learning rate, I mean the eta value used by the algorithm to calculate the CPT table parameters. From the attached paper, I got the formula (Eqn 6) which I am using to calculate manually the CPT table values for 2 records data-set. But when I try learning in Genie the answers do not match up.

Is there anything wrong with my understanding?
Please let me know.

Nisha.

[mark]

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.

[nisha269]

Thanks for letting me know.

Is my understanding that GeNIe uses an EM algorithm to compute CPTs correct?? If yes, can you let me know as to which EM algorithm does it exactly use, I mean there are other modified Em algorithms like Voting EM etc as well.

Also how many records do we need to have the CPT converge to their real probability as close as possible.

Nisha

[mark]

GeNIe uses the basic EM, no special alterations.

[nisha269]

Thanks. That helps.