search

joshspeagle / dynesty

Dynamic Nested Sampling package for computing Bayesian posteriors and evidences
https://dynesty.readthedocs.io/
MIT License
347 stars 76 forks source link

Clarification on Parallelisation of batches #223

Closed philastrophist closed 3 years ago

philastrophist commented 3 years ago

Hi,

I'm trying to use dynesty to map a highly mulimodal, computationally-intensive posterior (a gaussian mixture) and I'd like to know how to parallelise. I understand that we can use a multiprocess/MPI pool and I do it that way atm, but you also mention that we can add batches of runs together post-hoc. I can't find any references which talk about what effect that has on the posterior/evidence. Does it just increase resolution or do separate runs extend exploration (this seems doubtful).

Basically, are run-batch-combinations equivalent to just running longer?

I'm just trying to work out if it would be beneficial for me to break up into many batches.

joshspeagle commented 3 years ago

It essentially just increases resolution. Combining N runs with K live points is equivalent to running a single run with NK live points.

philastrophist commented 3 years ago

Right, and I'm guessing there is no reason to expect a combination to map modes that an extended run did not find?

joshspeagle commented 3 years ago

Not as much, especially since you're more likely to miss modes with runs that use fewer live points initially.

philastrophist commented 3 years ago

ok, that's what I thought, thanks!