search

dflemin3 / approxposterior

A Python package for approximate Bayesian inference and optimization using Gaussian processes
https://dflemin3.github.io/approxposterior/
MIT License
41 stars 9 forks source link

refactor burn-in estimation #16

Closed dflemin3 closed 5 years ago

dflemin3 commented 6 years ago

Since the user has access to the entire emcee sampler object, and hence the full MCMC chains, the user could (and probably should) run their own burn-in/convergence diagnostics. However, it would be useful to give the user some more in-house burn-in estimation procedures, e.g. the Gelman-Rubin statistic, to help their analysis. I should also write a test for these.

dflemin3 commented 6 years ago

It would be a good idea to use emcee's new fancy autocorrelation techniques, like tau = sampler.get_autocorr_time()

dflemin3 commented 5 years ago

Now using emcee v3's integrating autocorrelation time method as the default, but in general, user should post-process and examine their chains to confirm convergence.

dflemin3 commented 5 years ago

Added the ability for the user to use pretty good estimates for burn in times, or they could not do that and post-process chains. Changes are on the dev branch and will make it to master soon.