Negative Value of Information

[xanax]

Hello everybody,

I have a question regarding the calculation that Genie does to obtain the VOI of the nodes of an Influence Diagram. I know that the VOI of a certain variable is calculated as the difference between the expected utility of the best decision (that is the one that maximizes the utility) observing that variable and without observing it. Logically, the VOI should be always positive or zero in the worst case, but a negative value does not make any sense since you cannot worse off by observing new evidence.

My problem is that in the network I have designed I am getting negatives values for the VOI of most of the nodes! I have tried to calculate the VOI by hand with smaller examples and it's fine, but this one is bigger and it's too cumbersome to do it manually.

I have observed that by changing the probability distributions of some nodes (making them more accurate) the VOI turns positive in some cases, but I do not understand the relation.

Does anyone know what could be the reason for the negative values? I really don't understand it because I though that was not possible...

Thank you in advance for any comments...

[shooltz[BayesFusion]]

Can you post your model so we can investigate?

[xanax]

Hallo!

Here is the example network.

Thanks!

[shooltz[BayesFusion]]

On my machine VOI is calculated as -2.84e-014. This value is very close to zero and is probably a result of precision loss in VOI algorithm.

[xanax]

Hi!

Thank you very much for your prompt answer. You were right and it seems that is just and accuracy problem.

Anyway, it looks strange to me that the VOI of all the variables of the network is zero...Could it be related to the probability distributions (maybe they are not accurate enough and the action of consulting them is not profitable...) or may it be because of the utility function I have defined...?

Thanks again!