How is soft evidence implemented in SMILE?

The engine.
Post Reply
ronnie
Posts: 7
Joined: Wed Jan 20, 2010 10:39 am
Location: Skövde, Sweden
Contact:

How is soft evidence implemented in SMILE?

Post by ronnie »

In a previous post I mentioned that I have implemented a kind of "uncertain evidence" in Java on top of jsmile. One member of the forum notified me that SMILE provides a soft evidence function and asked whether they are equivalent. I tried to compare my implementation (which is based on the ideas of Jeffrey and Pearl and which adds an additional node to a network) with the one of SMILE and they turned out to yield different results.

In my implementation (consistent with Jeffrey's idea I believe), soft evidence is a probability function which is "inscribed" into a node. If no other observations (hard or soft) are made, the node in question will get the exact probability values as specified by the soft evidence. The soft evidence function in SMILE seems to work differently though, as the soft evidence will be affected by the current probability values of the node in question (the soft evidence of SMILE seems to indicate for which node outcomes the probabilities should increase and decrease rather than their desired values).

As I haven't found any documentation about the implementation or usage of soft evidence for SMILE, can you please let me know how it is supposed to work (and possibly where I can find the implemented algorithm)?
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: How is soft evidence implemented in SMILE?

Post by shooltz[BayesFusion] »

Check this post for more info on SMILE's soft evidence implementation:

http://genie.sis.pitt.edu/forum/viewtop ... 1516#p1516
marek [BayesFusion]
Site Admin
Posts: 430
Joined: Tue Dec 11, 2007 4:24 pm

Re: How is soft evidence implemented in SMILE?

Post by marek [BayesFusion] »

Hi Ronnie,

First of all, I woul dlike to appologize -- SMILE and GeNIe implement "virtual" and not "soft" evidence. I have understood the difference after we have added this functionality SMILE. We are actually planning to add "soft evidence". Our implementation of virtual evidence is by adding a child node to the node in which you observe virual evidence. That new node's CPT makes sure that the evidence received by the parent is precisely what is specified by the user. Does this explain the difference between SMILE's and your implementation? If not, let's talk more about it, perhaps off-line? My Email address is marek@sis.pitt.edu.
Cheers,

Marek
Post Reply