Sign in
Microsoft.com
 
Microsoft
 
 
 

locked
Is it feasible to create custom InferenceEngines? RRS feed

 > 

  • Question

  • Question
    Sign in to vote
    0
    Sign in to vote
    I'm very interested in using Infer.NET. I am particularly interested in using MCMC techniques, but I didn't see anything discussing creation of custom InferenceEngines, is that possible? If it is, is it possible to get log posterior and log posterior gradient evaluations through the framework?
    Tuesday, June 21, 2011 6:07 PM

All replies

  • Question
    Sign in to vote
    1
    Sign in to vote
    Infer.NET does not provide an API for defining arbitrary new inference algorithms.  At most you can change the message updates used for particular factors, as described in the user guide.  Do you have a model in mind where the built-in algorithms will not work?
    Wednesday, June 22, 2011 11:16 AM
  • Question
    Sign in to vote
    1
    Sign in to vote

    No, I don't, as I am not very familiar with the algorithms Infer.NET implements. I will have to do some reading.

    I am particularly interested in implementing Hamiltonian (or Hybrid) Monte Carlo (HMC) and related algorithms. Unlike other MCMC techniques, which scale proportional to ndimensions**2, these algorithms scale proportional to ndimensions**1.3 or ndimensions**1 which makes them much more feasible for large problems than other MCMC algorithms while remaining very general. HMC is not very complex and I've coded it before, which is why I was interested. I am sad to hear it is not possible to define a new algorithm. Thanks for the information.

    Wednesday, June 22, 2011 3:20 PM