Influence diagrams and Diagnosis

[samoht]

Hello,

why its not allowed to use nodes from influence-diagrams like utility/decision nodes in the diagnosis-tool?
Is it just not yet implemented to use this nodes in diagnosis now or is there a special conceptional or mathematical reason not to allow these node types?

I want to use diagnosis tool for suggestion if it is resonable collecting more information or not and after enough information for decison making is collected I want to use decision node to take action. But it looks taht I have to use two BN one for diagnosis (without decision nodes) and one for decision making. Is do all in one network and without having to do without the diagnosis tool?

A problem for me is also I want to define the costs in the tab "observation cost" in the diagnosis tool but I had to calulate this cost from the output from other nodes. In Influence diagrams i have utility and MAU/ALU node types which allow to define a formula to calculate two the total cost from two or more cost faktors. What can I use in diagnosis tool to do this task?

regards

[marek [BayesFusion]]

why its not allowed to use nodes from influence-diagrams like utility/decision nodes in the diagnosis-tool?
Is it just not yet implemented to use this nodes in diagnosis now or is there a special conceptional or mathematical reason not to allow these node types?

Influence diagrams do not need tools like this -- they have decision options and utilities. In practice, however, you will find that creating correct IDs will be very laborious. BNs for diagnosis is just a very handy and useful simplification of the problem.

I want to use diagnosis tool for suggestion if it is resonable collecting more information or not and after enough information for decison making is collected I want to use decision node to take action. But it looks taht I have to use two BN one for diagnosis (without decision nodes) and one for decision making. Is do all in one network and without having to do without the diagnosis tool?

You need an influence diagram for this problem. IDs have built in VOI computation. If information is free, it is always useful. If it comes at a price, you collect information as long as it gives you VOI>0. I suggest a good decision analysis textbook for this -- you can get only so far by intuition.

A problem for me is also I want to define the costs in the tab "observation cost" in the diagnosis tool but I had to calulate this cost from the output from other nodes. In Influence diagrams i have utility and MAU/ALU node types which allow to define a formula to calculate two the total cost from two or more cost faktors. What can I use in diagnosis tool to do this task?

Try conditional costs in BNs.
I hope this helps.
Cheers,

Marek

[samoht]

Hi Marek,

thank you for your reply.

If I understood it right the diagnosis tool is just a special influence diagram which simplifies some tasks. You write that utilities and decision options are already included in diagnosis BN, so the cost tab in the test variable replaces the utility nodes. And the observation variables have the function of the decision nodes. E.g. instead of a test node for a measurement I could use a chance node for the measurement which is influenced by a decision node (for the decision to take the measurement).
In my example I have different sensor types (=tests) and have to decide if I want to wait for the arrival of a special sensor or if i have already enough information. Waiting for the sensor costs time which has to be weighted against the gain of information. So far all is realizable with the diagnosis tool. After having collected all information which has a VoI>0
I have to take a decision to take an action (e.g alarm or not alarm).
I cannot use a decision node in the diagnosis tool, so I eigther have to abandon the use of this tool and model all as ID (what should be possible since VoI can also be calculated in ID, just without ranking display of the tests) or I have to model this alarm-decision as an additional test so I can model all in the diagnosis tool.

Try conditional costs in BNs.
I hope this helps.

I know about conditional costs, but the utility nodes support to use arithmetice to give in a formula for the costs and the conditional costs in the diagnosis tool are no replacement for it. So if i have a mathematical formula for test costs how can i model it?

Best Regards Thomas

why its not allowed to use nodes from influence-diagrams like utility/decision nodes in the diagnosis-tool?
Is it just not yet implemented to use this nodes in diagnosis now or is there a special conceptional or mathematical reason not to allow these node types?

Influence diagrams do not need tools like this -- they have decision options and utilities. In practice, however, you will find that creating correct IDs will be very laborious. BNs for diagnosis is just a very handy and useful simplification of the problem.

I want to use diagnosis tool for suggestion if it is resonable collecting more information or not and after enough information for decison making is collected I want to use decision node to take action. But it looks taht I have to use two BN one for diagnosis (without decision nodes) and one for decision making. Is do all in one network and without having to do without the diagnosis tool?

You need an influence diagram for this problem. IDs have built in VOI computation. If information is free, it is always useful. If it comes at a price, you collect information as long as it gives you VOI>0. I suggest a good decision analysis textbook for this -- you can get only so far by intuition.

A problem for me is also I want to define the costs in the tab "observation cost" in the diagnosis tool but I had to calulate this cost from the output from other nodes. In Influence diagrams i have utility and MAU/ALU node types which allow to define a formula to calculate two the total cost from two or more cost faktors. What can I use in diagnosis tool to do this task?

Try conditional costs in BNs.
I hope this helps.
Cheers,

Marek