node is not consistent

The front end.
Post Reply
Yan
Posts: 22
Joined: Fri Mar 04, 2022 5:26 am

node is not consistent

Post by Yan »

Hello,

I met some problems about the inconsistent node. Please see the attached example. After I used EM algorithm to learn the parameter, it said sum of probabilities should be 1 (please see fig.1). However, after I changed the value of 2.2185907e-25 to 0.0000000001, this problem was solved (see fig.2). I would like to know why this happens. And 2.2185907e-25 is much smaller than 0.0000000001, why after I changed the value, this problem can be solved?

Many thanks.

Kind regards,
Yan
Attachments
fig.1
fig.1
Picture1.png (10.24 KiB) Viewed 2321 times
fig.2
fig.2
Picture2.png (9.99 KiB) Viewed 2321 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: node is not consistent

Post by shooltz[BayesFusion] »

I can't reproduce the problem. I have typed the numbers in your screenshots and got no warning icon with the 2e-25 value.
Can you send me your xdsl file (use a forum private message if model contains confidential information).

BTW: you can hover the mouse cursor over the warning icon sign and the program will show a tooltip with the additional information about the problem.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: node is not consistent

Post by shooltz[BayesFusion] »

We tracked down this problem to GeNIe's algorithm for epsilon selection. If a column in the CPT has a very small non-zero probability (like 2e-25 in your example), it will be used for comparison of the total probabilities in the column with 1.0. The default epsilon is on the order of 1e-6.

This algorithm was implemented before GeNIe supported parameter learning. We'll be switching to constant 1e-6 epsilon for all consistency checks in GeNIe.
Yan
Posts: 22
Joined: Fri Mar 04, 2022 5:26 am

Re: node is not consistent

Post by Yan »

shooltz[BayesFusion] wrote: Wed Apr 06, 2022 12:57 pm We tracked down this problem to GeNIe's algorithm for epsilon selection. If a column in the CPT has a very small non-zero probability (like 2e-25 in your example), it will be used for comparison of the total probabilities in the column with 1.0. The default epsilon is on the order of 1e-6.

This algorithm was implemented before GeNIe supported parameter learning. We'll be switching to constant 1e-6 epsilon for all consistency checks in GeNIe.
Thanks Schooltz. Can I understand it as "2e-25" is much smaller than the default smallest value in the algorithm, it will not affect the result of the BN model too much, so we can ignore this issue?
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: node is not consistent

Post by shooltz[BayesFusion] »

You can ignore the issue. Your model is closer to the perfect sum=1.0 distribution that our default epsilon.

Before we release new version, you can consider using the 'Complement' command in the columns with warnings to get perfect 1.0 total.
Post Reply