Todo / Ideas

[kadeng]

Finalize improved Metropolis Hastings Sampler

• Make it possible to calculate the logp of just a subset of the variables
• Extract variable dependency graph from theano graph
• Create a networkx graph from that, with the option to convert it to a dotx graphic
• Calculate accept / reject probability over the markov blanket of a variable subset
• Implement all changes from pull request critique

  Finalize Wishart & Inverse Wishart Distributions

• Unit test their logp against scipy implementation
• Provide "random" method for Wishart & Inverse Wishart
• Provide conjugate_update method for both
• Finalize the huang_wang Prior implementation
• Sample precision and covariance matrices from them and see if they behave as expected

  Look into implementing SMC and RJMCMC for PyMC3

• See http://www.kev-smith.com/tutorial/rjmcmc.php for RJMCMC Tutorial
• See http://www.kev-smith.com/papers/smith-thesis.pdf for RJMCMC Details
• See Gelman, Bayesian Data Analysis for a short intro to RJMCMC
• See http://ebilionis.github.io/pysmc/tutorial.html#what-is-sequential-monte-carlo for SMC

[kadeng]

See also the paper at https://www.seas.harvard.edu/courses/cs281/papers/hastie-green-2011.pdf about Model selection using RJMCMC

[kadeng]

Some wishart sampler implementations I could find

https://code.google.com/p/haines/source/browse/gcp/wishart.py http://metaoptimize.com/qa/questions/14628/sampling-from-a-normal-inverse-wishart-in-python https://github.com/stan-dev/stan/blob/develop/src/stan/prob/distributions/multivariate/continuous/wishart.hpp https://code.google.com/p/haines/source/browse/gcp/wishart.py

[kadeng]

emcee integration

http://twiecki.github.io/blog/2013/09/23/emcee-pymc/ http://dan.iel.fm/emcee/current/ http://msp.org/camcos/2010/5-1/p04.xhtml