## Virtual evidence

The front end.
charlie
Posts: 65
Joined: Wed Aug 09, 2017 10:55 pm

### Virtual evidence

Why setting virtual evidence for a node that has child nodes has no effect? If the node has no child node everything is all right.

Thanks

Charlie

marek [BayesFusion]
Posts: 282
Joined: Tue Dec 11, 2007 4:24 pm

### Re: Virtual evidence

Why setting virtual evidence for a node that has child nodes has no effect? If the node has no child node everything is all right.
I think it absolutely does have an effect in both cases. Would you perhaps attach an example where you think it should work but does not?
Cheers,

Marek

charlie
Posts: 65
Joined: Wed Aug 09, 2017 10:55 pm

### Re: Virtual evidence

Please see attached a simple net with three nodes. If I set virtual evidence to the middle node say state0=0.5, state1=0.5, state2=0, and state3=0, the actual states shown are different: state0=0.7, state1=0.3, state2=0, and state3=0 after execution.
Thanks
Charlie
Attachments
VirtualEvidence.xdsl

marek [BayesFusion]
Posts: 282
Joined: Tue Dec 11, 2007 4:24 pm

### Re: Virtual evidence

I have examined the model. Everything is working fine, as expected. Virtual evidence does not set the node in question to a specified posterior probability distribution but rather offers uncertain evidence (specified by the virtual evidence distribution) for its states. The marginals are a function of the definition, evidence coming from parents (in this case Node3), and the virtual evidence. When the node in question has no parents, virtual evidence pretty much sets the posterior. What you may intuitively in mind is "soft evidence" (not implemented in GeNIe) rather than "virtual evidence."
I hope this helps,

Marek

charlie
Posts: 65
Joined: Wed Aug 09, 2017 10:55 pm

### Re: Virtual evidence

Thanks Marek for your patience. Surely you're right I misunderstood it as "soft evidence" in your word and the marginals should be the virtual evidence. Is there any more detailed introduction of virtual evidence that could help me gain a more thorough understanding?

I was quite bewildered always why it's called virtual evidence or "soft evidence". Could it be real and solid when the evidence is based on proportions of a population falling into different states? Say if a node represents the occurrence of a sickness and eventually a research fund actually 15% of the target population have the sickness and the rest are healthy, can I set a virtual evidence of 15% and 85%? Are they not real?

It seems I incorrectly assumed once an evidence, real or virtual, is set to a node, the node becomes independent of its parent nodes, which allows us to study the influence propagation beyond this point, ignoring all the parents.