Genie Sampling and fixed nodes

The front end.
techmec
Posts: 21
Joined: Sun Feb 05, 2012 2:53 pm

Re: Genie Sampling and fixed nodes

Post by techmec »

but why are hard evidences not sampled? (in GENIE, not in SMILE there I'm in the beginning, you know :) )
Martijn
Posts: 76
Joined: Sun May 29, 2011 12:23 am

Re: Genie Sampling and fixed nodes

Post by Martijn »

Well, hard evidence means basically your are absolute sure.
Let's say in a weather model you observe it's raining. There's no discusion and setting this evidence 'raining' in the node translates into giving it a probability distribution of 1.0 for 'raining' and 0 for any other option here

In Genie you click the state it should be in the set evidence list (right mouse button click on node)

Do you see what I mean?
techmec
Posts: 21
Joined: Sun Feb 05, 2012 2:53 pm

Re: Genie Sampling and fixed nodes

Post by techmec »

Yes, that means that as long as I'm not sure about my evidence, the sampling helps me to verify.

So, in my model e.g. I have the prior of which colors are possible and how often. This is exact knowledge.
The node for that has as much as states as possible colors.
That means, I cannot set a single color to hard evidence, because the other colors also have a value of their occurences.

But if I remodel the BN in that way, that I get single nodes for the specific colors, I can just set the colors to true or false, but I cannot model the (co)relation of the colors.

Thats it. Exactly.
Martijn
Posts: 76
Joined: Sun May 29, 2011 12:23 am

Re: Genie Sampling and fixed nodes

Post by Martijn »

Yes, that means that as long as I'm not sure about my evidence, the sampling helps me to verify.
I don't quite understand what you mean by this.
Why and how would sampling help you verify the evidence?

Maybe you somehow relate the sampling to the evidence you provide, I don't think that's right.

Maybe it helps if you see the inference algorithm you use as a black box that get the network and the set evidence as input and outputs posterior distributions for the nodes of the network.
You shouldn't worry about how the calculations are done, just that both (exact/sampling) are providing you with the posterior conditional distributions.
techmec
Posts: 21
Joined: Sun Feb 05, 2012 2:53 pm

Re: Genie Sampling and fixed nodes

Post by techmec »

Why and how would sampling help you verify the evidence?
In that way that it varies the evidence too. Not important I think I understand you.
Maybe you somehow relate the sampling to the evidence you provide, I don't think that's right.

Maybe it helps if you see the inference algorithm you use as a black box that get the network and the set evidence as input and outputs posterior distributions for the nodes of the network.
You shouldn't worry about how the calculations are done, just that both (exact/sampling) are providing you with the posterior conditional distributions.
Okay I try that
I have no other solution yet, so I trust :)
Thank you so much
Post Reply