Hybrid BN

The engine.
Post Reply
nour
Posts: 3
Joined: Wed Mar 23, 2022 8:35 pm

Hybrid BN

Post by nour »

I would like to know if genie smile treats hybrid bayesian networks or not. I started to create my first bayesian network (hybrid), I noticed a problem with the arcs, some of them were in blue and others in grey, I've noticed that there is something strange and when I tried to apply the em algorithm it showed me an error that says that it is not possible to mix continuous and discrete nodes even though I did the descritization, I'm not sure if there is another algorithm that can be applied instead of the em algorithm or if genie smile doesn't deal with hybrid networks at all (although I've found a lot of examples of hybrid bayesian networks in genie smile) so i think it is my bad i didn't know how to use it. Can you please help me by explaining me how to create correctly an hybrid bayesian network(without the arcs problem ) and how to do a parameter learning and if there is another tool or solution that replaces the use of the em algorithm please it's very important for me, thank you in advance.
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Hybrid BN

Post by marek [BayesFusion] »

Hi Nour,

SMILE has everything that GeNIe has, so it certainly implements hybrid networks. Grey color denotes arcs that are not necessary in the sense of having no impact on the child variable. Whenever you create a new arc, it will have grey color because the CPT in the node contains (identical) uniform distributions and, hence, the parent does not impact the child.

EM will work if the nodes in your network are discrete and their states correspond to the values in the data file. If this does not work, please post a simple subset of your model for which EM does not work -- I will be happy to look at it.

EM will not support a mixture of discrete and continuous variables, so practically you cannot use in hybrid networks (yet -- we will look at it at some point). You will have to use it outside of GeNIe/SMILE and then copy the learned distributions to your model in GeNIe or SMILE. I assume you know what you want, so it will not be a theoretical problem to do the learning?
I hope this helps,

Marek
nour
Posts: 3
Joined: Wed Mar 23, 2022 8:35 pm

Re: Hybrid BN

Post by nour »

thank you very much Marek, your answer really helped me, I guess the problem of grey arcs is due to the equations written at the nodes, I will check them. I have another question, can you suggest me another tool in which I can apply the em algorithm on my HBN, or can you suggest me another algorithm of jsmile which can do a parameter learning instead of EM
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Hybrid BN

Post by marek [BayesFusion] »

Hi Nour,

I'm afraid I am not aware of any tool that can do that. Parameter (and equation?) learning in a hybrid system is a complex problem and it depends very much on the form of the equation, distributions, etc. No general solution. When we move into that direction, which we definitely will, we will at best solve a special case of this problem.
I hope this helps,

Marek
nour
Posts: 3
Joined: Wed Mar 23, 2022 8:35 pm

Re: Hybrid BN

Post by nour »

Thank you very much Marek for your help and I apologize for asking a lot of questions because it's the first time I work with BN and the first time I use genie smile I'm going to do a dataset cleaning to have complete data to avoid the use of the em algorithm which doesn't work with HBN and I have heard of an algorithm called statistical estimation (maximum likelihood) is it available on genie smile if not what can I use as an algorithm for learning parameters in the case of complete data in genie smile ? thank u in advince.
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: Hybrid BN

Post by marek [BayesFusion] »

You are most welcome! SMILE always uses EM, which more or less (I'm skipping the fine details of the implementation) deteriorates to simple counting when the data set contains all values.
Cheers,

Marek
Post Reply