Open davidcerny opened 2 months ago
is this fixed by PR #407 ?
No, it isn't.
If I remember correctly, the problem is that somehow changing the parameters does not propagate all the way down the likelihood. I was trying to follow this in the debugger and tracked the propagation part way down but didn't quite find the point where the propagation quits early.
I think the PR helped, but was not quite sufficient.
Describe the bug Because of the "stochaministic" nature of
dnUPP()
(the uniform partitioning distribution), changes in parameter values are not correctly propagated through the model DAG. This results in analyses that either run seemingly fine but produce invalid results, or crash when RevBayes is compiled with theDEBUG_MCMC
flag. With a sufficiently complex model, the distribution can cause MCMC to fail for seemingly unrelated reasons even when this flag is not enabled; see #293 for an example.To reproduce Run this script. (One absolute path has to be changed; the
bears_cytb.nex
file is the one from our test suite and dating tutorials). This is a heavily stripped down version of the script from #293. When I run it after enablingDEBUG_MCMC
, I get:Expected behavior Reversible-jump MCMC analyses should be able to employ
dnUPP
without crashing.Computer info macOS 14.5 (Sonoma), ARM64 architecture, RevBayes 1.2.5-preview, latest version of the
development
branch (c46173f).Additional context This bug may have affected papers published by @wf8 and @dpaluh:
Corresponding Rev scripts: