Open florianhartig opened 7 years ago
This seems to be more complicated, added a warning in https://github.com/florianhartig/BayesianTools/commit/1bb636af652129ce7276c2fb244b059fdddf7abf
I checked the Rtwalk package and the same problem occurs there, maybe it would be better/faster to rewrite the whole sampler
is our code comparable in structure to this package? Because when I remember correctly, @stefan-paul had implemented it differently than the code from the publication ...
No, it's similar implemented...
A colleague told me that he had the feeling the https://cran.r-project.org/web/packages/Rtwalk/index.html package works - could you try out if the sampler in the T-walk package works where ours fails?
Sorry, I just saw that Max had already tried this - but maybe try if this is still so. Also, we should maybe tell the guys that they have a problem in their package!
Did anyone check the sampler in the original paper? If I remember correctly, there was code included with the paper!
@JohannesOberpriller - when I looked at the code, I was relatively sure the problem is in the Gfun (if I remember correctly, or one of the helper functions) - the weird jumps occur when one of these functions creates very high values. Didn't look like an overflow though, seemed the values were realistic!
There is a numeric problem in the T-walk. For the simple test densities this did not occur, but for the VSEM, if I run the following example, the algorithm jumps sometimes into areas of the posterior space that are practically impossible in terms of the log posterior differences.
I suspect that there is somewhere some under/overflow issue in the jump probabilities, maybe in the g function, but I didn't see an obvious problem so far.