Sensitivity analysis in GeNIe v 4

[nzsk]

In earlier versions of GeNIe, it was possible to discretise a hybrid network and then perform sensitivity analysis. This functionality does not appear to be there in the latest version of GeNIe. Is it still possible to turn a hybrid network into a discrete network or is this no longer possible? If not, what would you recommend as an alternative way or performing sensitivity analysis on a hybrid network?

[shooltz[BayesFusion]]

In GeNIe 4.0 we have modified the inference algorithm for the hybrid models. Instead of a complete discretized copy (which you can get in GeNIe 3.x), we use 'virtualized' CPTs for equation nodes. Additionally, we always take into account all possible reductions in the CPTs, such as the presence of the evidence in the parent. This makes the explicit discretization option difficult, because the output depends on the current evidence.

As a simple workaround, you can change all equation nodes to CPTs in a copy of your model. This will create the discrete model, which can be used with sensitivity.

[nzsk]

Thank you, I have done that and it works. Just another question - is it possible to call the hybrid network in R using rSMILE and then connect it to a global sensitivity analysis package in R - thus performing sensitivity analysis on the hybrid network itself?

[shooltz[BayesFusion]]

is it possible to call the hybrid network in R using rSMILE and then connect it to a global sensitivity analysis package in R

We don't have any experience with R's sensitivity analysis packages. Can you post a link to the package you're referring to?

[nzsk]

I'm considering either using the SAFE toolbox https://www.safetoolbox.info/ for R or the R Sensobol package https://cran.r-project.org/web/packages ... index.html

Many thanks

[shooltz[BayesFusion]]

This is an interesting idea and we have made a note of it but it is unlikely we will provide a connection between SMILE and SAFE toolbox/R Sensobol very soon. Given that you have SMILE and the other packages, you will have to prepare the input for the external packages using R. Sorry for not being immediately accommodating - we'll be happy to answer any SMILE-specific questions, of course.

[nzsk]

Thank you, I'll see what I can do :-)