Error while dealing with Gaussians
-
Donnerstag, 9. Juni 2011 05:13
I have written an inference model which involves Gaussian random variables. The model compiles correctly, but while inferring, I get the following run-time exception:
The distribution is improper (Gaussian(m/v=-1.159, 1/v=-0.1572)). Cannot compute expectations.
How do I understand this exception? What can be the possible causes of this and what should be the approach to debug it?
Please note that I am making profuse use of ProductOfGaussians operator which is low on quality band. I am displayed a warning regarding this while the model compiles. I have not been able to find a workaround for this warning because I cannot avoid the product of Gaussian variables in the model.
Alle Antworten
-
Donnerstag, 9. Juni 2011 09:45Besitzer
For general comments on this, see http://social.microsoft.com/Forums/en-US/infer.net/thread/9393550c-9342-4849-8318-46596ad792fc and also the FAQ.
When you are working with products of random variables, try to use Variational Message Passing which will not cause improper messages; this is a difficult factor for Expectation Propagation. If other parts of your model require factors which don't support VMP, either try to reformulate those, or you may end up having to do a hybrid model.
John
