Difficulty with interpretation of a network SOS

The front end.
Post Reply
Melenia
Posts: 4
Joined: Tue Jul 04, 2017 1:38 pm

Difficulty with interpretation of a network SOS

Post by Melenia »

Ηallo everyone!

I am facing an issue concerning a network I created. Attached you may find it.

I just wanted to ask if I could predict a situation starting from the end, from the last node.
For example: Giving the node <<success>> the value 100 and predict what the quantity of granulometry, porous, quality etc would be.
Does it mean that the possibilities for granulometry of getting low is 36%, of getting medium is 40% and high 24% ?
All the above possibilities happen simultaneously? How can this network be read?

Hope my question to be understood.
Thank you in advance.
Attachments
teliko.xdsl
(5.7 KiB) Downloaded 443 times
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Difficulty with interpretation of a network SOS

Post by marek [BayesFusion] »

Hi Melenia,

I'm afraid it is not possible to do what you want, i.e., enter evidence in a utility node and reason backwards from it to see what settings decision nodes should have. I can see in your model that you have tried/intended to have the four nodes on the left hand side (granulometry, porous, hydraulics and geopurification) to be decision nodes but noticed that this will not work and made them chance nodes. To obtain what you want, I advise you to turn the utility node (success) into a chance node, discretize it into as many intervals as you feel is needed, and construct its CPT. Then you can observe the "success" and reason about the parents/decisions so that you can optimize their values. Please note that there may be many combinations of the decision variables that will produce any given value of the node success.

Given what you are working on, I advise you to look at the hybrid capacity of GeNIe, which allows you to make some nodes continuous. If you know the physics of what you are modeling and can specify interactions by means of equations, this will save you a lot o feffort. If you discretize the continuous nodes (a special tab in the node properties; please also look at the manual), you will effectively produce discrete distributions that you can interpret as histograms the underlying continuous distributions. Try it and you will love it!
I hope this helps,

Marek
Post Reply