Skip to content

DRAFT: Rationalize the ACT running average computation in the custom dynesty sampler.

Gregory Ashton requested to merge will-farr/bilby-nmcmc into master

COPY OF !953 (closed) UPDATED TO MASTER

  • No longer issue a warning when there are repeated likelihood values in the sampler output.

  • Add adapt_tscale: the running average ACT estimate will average over this number of livepoint replacements using an exponentially-decaying weight. Default is 100.

  • Get aggressive with nact = 1 by default now that the ACT computation is stabilized. Each MCMC live point replacement will run for one estimated ACT. This means, on average, two accepted MCMC steps, so a ~10% (= exp(-2)) chance of zero acceptances and a repeated live point.
    This does not bias the sampler (see http://arxiv.org/abs/1805.03924).

  • The ACT estimate is an exponentially-weighted running mean of 2/P_accept - 1 over each live point replacement; 2/P_accept - 1 is the ACT of an ideal rejection sampler with acceptance rate P_accept, so is a lower-bound for the ACT of the MCMC.

Merge request reports