Open alashworth opened 5 years ago
Comment by sakrejda
Monday Apr 03, 2017 at 21:05 GMT
If we allow users to set the threshold reporting when it falls below the threshold at each iteration seems like the best. Almost an aside: could we detect and distinguish the two sources of low step size? I think the answer is yes and it would be useful to report but I doubt we could do it reliably.
Comment by wds15
Wednesday Apr 05, 2017 at 14:02 GMT
Making it configurable as suggested is certainly nice. Among the three options you present I would opt for the second one; reporting when it transitions across the threshold.
Description:
Currently adaption makes no attempt to warn the user when the step size falls to extremely small values. Technically such small values are okay and indicate an intermediate state where the sampler is trying to adapt to a posterior with a poorly-posed covariance. Because those models are not ideal, however, it may be useful to warn the user. Additionally, inaccurate gradients also force the adaptation to extremely small step sizes.
A somewhat-informative warning can be displayed to users, but when should it be displayed?
Current Version:
v2.14.0