Hello,

are you going to develop the network from the data or manually?

If the former is the case, dataset could be enough to give you the values for your probabilities (given there is enough combination of various cases in the data).

If the latter is true, then you need to feed the network manually, getting the probabilities from the sources that you have, literature, experiments, experts, etc.

Regards

Jakub

Thanks for your valuable contribution,

If we assume the structure of the network and the relationship between the variables have been defined. My method of collecting data bases on expert judgments through questionnaire to define the probabilities of the root nodes (they were asked to define the probability of occurrence based on five point likert scale). The method used in the work above is, the root nodes have been grouped into many groups (intermediate nodes). For a specific group (G1), lets say we have three parent nodes (N1,N2,N3) and the average values of probability obtained through questionnaire for each node are (3.9, 2.7, 2.3). Now we normalize and calculate the contribution of each root node in the intermediate node (0.44, 0.3, 0.26). In the study above, only these values were used to calculate the CPTs and so on for other nodes. And this is all for the network.

My first question is, there is no learning step, no learning from cases or data or file. just the average values were used then after normalization the values were used to define CPTs. Is this applicable? to learn the network from the average values? its like learning the network from on case but with average.

The second question I provide an example to be more clearer

the picture in the file represents the CPTs for Design Risk node (intermediate node)

we already know

-P(Design changes= T)= 0.277

-P(Design errors =T)= 0.313

-P(insufficient detailing= T)= 0.410

is it as easy as appears for defining CPTs?

for example: P(Design Risk = T/T,T,T)= sum all relevant probabilities when T= 1

p(Design Risk= T/ T,T,F)= P(Design changes= T)+P(Design errors =T)=0.277+0.313=0.59

Best regards

BTW you can combine the expert judgements and real database in to BBN, or firstly build the structure manually and then updating the structure by using evidences which are real databas. Hope it help.

Qing