Set node state as evidence

The engine.
Post Reply
rmorales
Posts: 14
Joined: Mon Oct 29, 2018 11:21 pm

Set node state as evidence

Post by rmorales » Wed Oct 31, 2018 12:16 am

Given the fact that the method to set the evidence in a node only allows to provide a given value, not a probability distribution, I wonder if there is a way to set a node state as evidence; that is to say, so that updating the beliefs do not modifies it but uses it as input to the updating process process.

marek [BayesFusion]
Site Admin
Posts: 306
Joined: Tue Dec 11, 2007 4:24 pm

Re: Set node state as evidence

Post by marek [BayesFusion] » Wed Oct 31, 2018 12:02 pm

I'm not sure I fully understand your question. GeNIe and SMILE allow for virtual evidence, which is specifying evidence as a probability distribution. It should be quite intuitive at the GUI but please search the manual for "virtual evidence" in case you cannot figure it out. Is this what you want?

Also, entering evidence in GeNIe but also in any Bayesian network package that I know is precisely what you are describing -- indicating a state of a variable that has been observed and then propagating this value in the updating algorithm without modifying the network.
Does this help?

Marek

rmorales
Posts: 14
Joined: Mon Oct 29, 2018 11:21 pm

Re: Set node state as evidence

Post by rmorales » Wed Oct 31, 2018 4:52 pm

Thank you for responding. What I am trying to do is to implement a dynamic bayesian network in the following way:
  • I create the original network and update beliefs.
  • I create the 'past nodes', link them, define them with the values got in the original network (previously saved), and update the definitions of the nodes in the original network, so they depend on the 'past nodes'. At this stage, I cannot set virtual evidence in the 'past nodes'.
  • I update the beliefs, but then my 'past nodes' get modified (and they should not, as far as I know, as the represent the evidence from the past).
  • Ignoring this situation, in the next step I set the 'past nodes' as virtual evidence using the values of the original network (updated).
  • I update the beliefs, and my virtual evidence gets modified too, but they should not,should they?
Regards,
Rafael

marek [BayesFusion]
Site Admin
Posts: 306
Joined: Tue Dec 11, 2007 4:24 pm

Re: Set node state as evidence

Post by marek [BayesFusion] » Tue Nov 20, 2018 4:51 pm

I'm not sure if I fully understand what you are doing but it seems that you are trying to replicate the formalism of Dynamic Bayesian Networks :-). I understand that you are using only first order influences. Please note that with GeNIe/SMILE you can have any time-order influences. Coming back to your question, perhaps to understand how virtual evidence works, you should implement it yourself within GeNIe. It amounts essentially to having a child node that is being observed and has parameters that set the virtual evidence in the child. One thought that I had in the context of your attempts is that perhaps you may want to use "Control Value" rather than "Observe Evidence" but I am not 100% sure if this is going to achieve what you want.
I hope this helps,

Marek

Post Reply