Add a mean-log-likelihood method to improve the ACT estimation
Suggestion from @colm.talbot. I've then added a number of other tweaks in addition.
Merge request reports
Activity
added 2 commits
added 20 commits
-
deb39eac...96cedb44 - 18 commits from branch
master - d95b627a - Merge branch 'master' into add-mean-log-like-to-ptemcee
- bee9274b - Fix tests
-
deb39eac...96cedb44 - 18 commits from branch
added 9 commits
-
bee9274b...5428c2ee - 8 commits from branch
master - 4a3693c8 - Merge branch 'master' into add-mean-log-like-to-ptemcee
-
bee9274b...5428c2ee - 8 commits from branch
mentioned in merge request !869 (closed)
added 2 commits
added 47 commits
-
28d6caad...47dc99b3 - 46 commits from branch
master - a7d6e4e1 - Merge branch 'master' into add-mean-log-like-to-ptemcee
-
28d6caad...47dc99b3 - 46 commits from branch
added 1 commit
- 92ddfe9a - Update method: add use config of convergence testing
added 1 commit
- 1cfb250f - Improve plotting and better explain the steps
changed milestone to %Future
added Sampling label
added 27 commits
-
4abe2f28...3cd2bff8 - 26 commits from branch
master - dc09f82a - Merge branch 'master' into add-mean-log-like-to-ptemcee
-
4abe2f28...3cd2bff8 - 26 commits from branch
-
Okay, this current implementation is "close" and I think it is worth merging this now. I believe we can get things working by merging this then figuring out the right settings.
- Here is a comparison with dynesty for the 4s TDP-marginalized run
- Here the is the config
-
sampler-kwargs:
sampler-kwargs={nsamples=5000, ntemps=20, walkers=100, burn_in_fixed_discard=500, mean_logl_frac=0.1, log10beta_min=-2, safety=5}Timing (with 16 cpus)
- dynesty: 10 hrs 45 mins
- ptemcee: 7hrs 11 mins
Edited by Gregory Ashton 
Thinking about this more, I think that perhaps the quantity we should be looking at is the posterior probability, rather than the likelihood, as non-trivial priors could change the ensemble to change the mean likelihood by a small amount during sampling. Although perhaps that would be a sign of not using a large enough ensemble?
added 2 commits
added 2 commits
-
@colm.talbot I've just updated to use the log posterior rather than log likelihood. Rerunning tests to make sure I didn't break anything.
Also, for people's reference, these are the trace plots: the grey points are the samples discarded because of instability (e.g. the mean log-likelihood hasn't stabilized)
Another useful plot that is created is the mean log-likelihood (soon to be mean log post). This shows where it achieves stability mixing between modes.
- Resolved by Moritz Huebner
I've run with no marginalisation. It's definitely not a good match, but a start.
- Last reply by Gregory Ashton
mentioned in commit 655699ed