Error -5 running EPIS

The engine.
Post Reply
opcode
Posts: 1
Joined: Fri May 09, 2008 10:10 am

Error -5 running EPIS

Post by opcode »

I have been trying out the EPIS algorithm in Bayesian networks generated from relational templates, and while it has worked quite well on smaller examples, it is frequently giving me the following error in my actual application:

SMILE error -5 in function UpdateBeliefs. Logged information: No useful samples have been generated for some of nodes.

(I am using the jSMILE wrapper.)

Can you perhaps elaborate on the circumstances under which this error is likely to occur?

Best regards,
opcode
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Error -5 running EPIS

Post by shooltz[BayesFusion] »

opcode wrote:Can you perhaps elaborate on the circumstances under which this error is likely to occur?
(disclaimer: not a sampling expert here)

Most likely reason is that probability of the evidence you've set is very small. Did you increase the number of samples above the default value with Network.setSampleCount?

EPIS can be fine tuned with additional parameters. There's no jSMILE API for this feature, but you can do that in GeNIe - see network properties dialog. EPIS parameters are stored in .xdsl file, so your jSMILE app will run the inference with EPIS params specified in GeNIe.

You can also try to call Network.setTarget on nodes for which you need the beliefs. This may (but doesn't have to) reduce the subnets created by relevance layer and passed down to inference algorithm. This trick should work for clustering as well.
Post Reply