search

kaskr / adcomp

AD computation with Template Model Builder (TMB)
Other
175 stars 80 forks source link

Anyone working on AD variational Bayes? #178

Open James-Thorson opened 8 years ago

James-Thorson commented 8 years ago

Hi all,

Given Cole Monnahan's success replicating Hamiltonian MCMC into TMB (based on the NUTS algorithm from STAN)), I was wondering if anyone was interested in continuing to explore STAN features within TMB. In particular, variational Bayes is discussed as an important tool for Gaussian process models for large data sets:

http://arxiv.org/abs/1603.00788

and it appears like it could be done in TMB. Is anyone working on this already? Any interest in discussing?

cheers, jim

skaug commented 8 years ago

When you consider how to improve the Laplace approximation you are naturally led in the direction of variational methods. There is however a "difficult expectation" buried in the formulas. Maybe the integration technique in Thorson & Kristensen (2016, Fisheries Research) can be applied.

Anyway, TMB should be a good platform for implementing variational methods. It would be nice to have a few examples. Maybe this can be done during the June workshop.