Changes the default number of walks to 30x ndim (previously 10x ndims).
The default settings are insufficient for the pp-tests of the high-spin 128s basis to pass (#388 (closed)). This changes fixes that. The downside of this is it is setting out default settings by CBC inference. So people doing "easier" problems will need to tone down the settings.
Merge request reports
Activity
changed milestone to %0.5.4
added Sampling label
Something I hadn't appreciated in #388 (closed) is that the default settings are in fact not a fixed number of walks, but a multiplicative factor of the number of dimensions. The current default is 10xndim. Here I show the PP test for the 20x and 30x the number of dimensions. The sampling time is correspondingly increased, but it improves the robustness of the result.
PP plot Meta data 1000nlive 20xndim 1000nlive 30xndim Edited by Gregory AshtonOkay, so running with
dlogz=0.5
we havePP plot Meta data 1000nlive 30xndim dlogZ=0.5 This passes the test, but still looks slightly ratty. However, this doesn't combine multiple runs so in practise I expect things to be cleaned up quite substantially on real runs.
Edited by Gregory Ashtonmentioned in commit e3319cd3
mentioned in issue #388 (closed)