Supporting multichain/population-samplers

[michaelosthege]

Currently, multiple chains run sequentially (sampling.py:434) but multichain-samplers such as DE-MCMC require information from all chains at every iteration.

I am currently trying to analyze what changes would have to be made and how hard it would be:

Requirements & Progress Tracker

• maintain support for multiple samplers
• modify the sampling procedure to iterate chains in parallel
• pass chain points to the astep method of the samplers
• (optional) track chain potentials (this is currently not known to the sampling iterators)

Right now I am trying to set up some sort of benchmark scenario with the banana distribution.

I have created a benchmark example that compares the effective sample size of different samplers on a correlated bivariate normal distribution.

[michaelosthege]

I am making some progress on my branch: https://github.com/michaelosthege/pymc3

A side note: The njobs/chains arguments of sample() are somewhat unintuitive. Usually I would assume that

• nchains run on one process
• each of njobs runs nchains
• total number of chains would then be njobs * nchains

I'm bringing this up because with population samplers, one would expect all nchains on one process to interact and the njobs as completely independent.