Skip to content

DEV: update custom dynesty proposals

Colm Talbot requested to merge dynesty-differential into master

This is my latest attempt to improve the sampling used within our wrapper of dynesty. I've wound up having three prescriptions for the MCMC length that have different benefits/disadvantages. I also added differential evolution proposals that seem to perform much better than the ellipsoidal proposals, especially in multimodal spaces.

  • the existing default behaviour is minimally changed. Long term, I would like to remove this, but I think it is best to leave it for a while.
  • a more stable tracking of the average number of accepted steps. This is an incremental improvement to the existing method for a few reasons, (1) it sets the MCMC length once for each parallel process, whereas the current one has independent MCMC lengths for each process, (2) it is more honest that the quantity being tracked in the expected number of accepted steps rather than the autocorrelation length
  • track the autocorrelation time using emcee by running the MCMC for a long time and building a cache.

Merge request reports