Discretization problem

The front end.
Post Reply
charlie
Posts: 65
Joined: Wed Aug 09, 2017 10:55 pm

Discretization problem

Post by charlie » Thu Feb 21, 2019 6:17 am

Dear administrator

I have a network with equation nodes and got the following message (excerpted) in the Output window:
Discretization problem in node NumLights: Overflow samples: 156, max=124.02 hiBound=100 Total valid samples: 9844 of 10000
Discretization problem in node Best_ESC_price: Underflow samples: 250149, min=5.00009 loBound=15 Total valid samples: 2749851 of 3000000 CPT configs with no valid samples: 20 of 300

What does it mean by "underflow" and "overflow"?

Thanks

Charles

shooltz[BayesFusion]
Site Admin
Posts: 1202
Joined: Mon Nov 26, 2007 5:51 pm

Re: Discretization problem

Post by shooltz[BayesFusion] » Fri Feb 22, 2019 11:37 am

During the discretization process the CPTs for the equation nodes are derived by sampling. Some of these samples may fall outside of the bounds specified for the node. If the sample value is less than a lower node bound, it's rejected and a count of underflow samples is increased. If the sample value is greater than an upper bound, it's also rejectd and a count of overflow samples is increased. After discretization is complete and under/overflow counts are nonzero, GeNIe displays the appropriate message in its output pane to notify an user about possible problem with the variable domain specified by the bounds.

In your case, NumLights variable has an upper bound of 100, but there were 156 samples with values larger than 100, the highest being 124.02. For Best_ESC_price there were 250149 samples with values less than 15.

Generally, this kind of message is a warning, not an error.

charlie
Posts: 65
Joined: Wed Aug 09, 2017 10:55 pm

Re: Discretization problem

Post by charlie » Fri Feb 22, 2019 8:00 pm

Thank you for the explanation.
Charlie

Post Reply