On Sun, 24 Mar 2013, Maik Stiebler wrote:
Hi folks,
I tried to do some analysis using the Jacobs-Trice 100 MET and got a
surprising result.
Upon investigation, I found that in non-variance-reduced, non-truncated
rollouts, the rollout match equity given by gnubg did not match my hand
calculation based on the cube statistics info. I suspect that on inversion of
the match equity table, the gammon values don't get inversed. This seems to
influence not only rollouts, but also evals. I will refrain from an awkward
description of my debugging process until you tell me this is not a known bug
and just give you a symptomatic example (sorry for the German, I hope it's
easy enough to follow):
I don't think this is a known bug. I reproduced your case with 0-ply
evaluations (where looking at what happens step by step should be easier)
and the results from the two positions differ as well.
I took a quick look at the code and didn't find anything obvious, except
that the problem is unlikely to be in the inverting but rather in the
handling of asymetric METs in general.
You don't write it explicitly but does the first result, with the table
not inverted, match your hand calculation or do they both seem wrong ? I
would expect the latter.
_______________________________________________
Bug-gnubg mailing list
Bug-gnubg@gnu.org
https://lists.gnu.org/mailman/listinfo/bug-gnubg