Diagnosis

The front end.
Post Reply
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Diagnosis

Post by samoht »

Hello,

I use the diagnosis tool for my Bayes-Net
When I try to change the Entropy/Cost-Ratio neither the ranking of the tests nor the absolute diagnostic values are changing.Does it mean that there is something wrong with my bayes-net and where can be the error? I defined costs for my tests greater zero. If I execute a test the diagnostic values are changing.

Second question: my diagnostic values are absolutely never higher than 0.25 often only between 0.1 and 0.2. So if i have three tests one with diagnostic value 0.14 one with 0.17 and one with 0.15. Does it mean that none of the tests is worth executing? When can I say that a diagnostic value is high enough to execute the test? Does an absolut diagnostic value of 0.14 mean that the probablility of the target state of the target node is expected to rise by 0.14 if the test will be executed? Is the cost of the test also expresssed in the diagnostic value and how? Or is only the relative value important and it gives me only a ranking?

regards
samoht
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Diagnosis

Post by shooltz[BayesFusion] »

When I try to change the Entropy/Cost-Ratio neither the ranking of the tests nor the absolute diagnostic values are changing.Does it mean that there is something wrong with my bayes-net and where can be the error? I defined costs for my tests greater zero. If I execute a test the diagnostic values are changing.
Do you have the observation costs defined in the node properties for ranked observation nodes? The default cost is zero and does not change the diagnostic value regardless of the Entropy/Cost ratio setting.

Second question: my diagnostic values are absolutely never higher than 0.25 often only between 0.1 and 0.2. So if i have three tests one with diagnostic value 0.14 one with 0.17 and one with 0.15. Does it mean that none of the tests is worth executing?
The diagnostic value for given observation is entropy-based (and obviously the costs will affect it if defined) and measures the change on the posteriors of fault states in pursued fault node(s) if observation is instantiated. There's no absolute meaning for this value, but it can be used to order observations so you can choose the most 'valuable'/least 'costly' one and instantiate it first.
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Re: Diagnosis

Post by samoht »

hi,

thanks for the reply.
But if there is no absolute meaning what is the difference between:
case 1:
Test1: 0.15
Test2: 0.10
and case 2:
Test1: 0.85
Test2: 0.80 ?
In both cases Test1 would be preferred and the difference to Test2 is in both cases 0.05.

second question: when is a test rated as not beneficial?
In two cases:
1. if the costs are higher than the value of information (depending of hte entropy/cost ratio)
2. if the test cannot contribute to improve the probability of the target hypothesis.
In all other cases there should be a gain to execute the test even if the diagnostic vaue is very low.
Do you have the observation costs defined in the node properties for ranked observation nodes? The default cost is zero and does not change the diagnostic value regardless of the Entropy/Cost ratio setting.
yes i have defined costs greater 0 in the observation tab for the ranked observation node. Even for a simple example with two test nodes and one target node i see no effect (no change of the diagnostic values) when i change the ratio with the slider.

Samoht
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Diagnosis

Post by shooltz[BayesFusion] »

samoht wrote:But if there is no absolute meaning what is the difference between:
case 1:Test1: 0.15Test2: 0.10
and case 2:Test1: 0.85Test2: 0.80 ?
There's no difference from decision maker's POV. The first test to perform is Test1.
second question: when is a test rated as not beneficial?
GeNIe marks tests as not beneficial when their diagnostic values is negative. This can only happen when costs are defined, so your answer #1 is the right one.
yes i have defined costs greater 0 in the observation tab for the ranked observation node. Even for a simple example with two test nodes and one target node i see no effect (no change of the diagnostic values) when i change the ratio with the slider.
Can you post your network here?
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Re: Diagnosis

Post by samoht »

Hello,

thanks for reply.
here just the example in the genie help. If I change the slider there is only change from not beneficially to 0 but no change in the diagnostic value or ranking.


samoht
Attachments
example.xdsl
(3.56 KiB) Downloaded 324 times
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Diagnosis

Post by shooltz[BayesFusion] »

Your model has uniform distributions in all CPT columns of observation nodes. Instantiating the observations does not change the fault posteriors, so the calculated entropy is zero. Using observation costs drives the diagnostic value even lower - you'll need a strictly positive value for the test to become beneficial.
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Re: Diagnosis

Post by samoht »

Hi,

thanks for reply. I understood and it works now for example.xdsl. The diagnostic values are changing if i use different probabilities for this example.

But in example2.xdsl the diagnostic values dont change if i move the slider. Why?

Another question: I cannot use explicit utility nodes in the diagnosis like in influence diagramms. I use conditional costs with cost arcs for my tests. That means the costs of my test depends of the state of other nodes. Where can I see what is the current cost of my test after updating (similar as I can see the costs i an utility node in the value tab) ? Or is it only possible to get this test cost with ObservationInfo.getCost() via smile-api? Or is it generally not possible to use cost arcs to define the costs of a test in the diagnosis tool?

regards
samoht
Attachments
example2.xdsl
(32.26 KiB) Downloaded 275 times
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Re: Diagnosis

Post by samoht »

I debugged my program and I found out that ObservationInfo.cost is always 0 , that explains that the slider move doent change anthing. But as you can see in my example2.xls I defined costs != 0. so why ObservationInfo.cost = 0?
samoht
Posts: 25
Joined: Wed Apr 28, 2010 1:41 pm

Re: Diagnosis

Post by samoht »

Hi ,

urgent question: If I use conditional costs for a test in the diagnosis tool and move the slider: first the diagnostic value doesn't change. Only after I set evidence for all nodes which influence the cost of the test node moving the slide has an effect.

But why only after setting evidence???
Also without evidence there should be a change because the estimated cost depend on the probability and can be calculated also without evidence.

best regards
samoth
Post Reply