Hi folks,
weird things are going on. I made a few tests with a certain position
regarding the MT bug (standard error no/double).
I played a moneygame (attached, it's not that bad ;-) ) and made the
following rollouts at the redouble/take position:
Current snapshot, all rollouts 2-ply, cube 2ply, 108 games, untruncated,
dice Mersenne Twister, seed 1, I restarted gnubg after each rollout!
1) Twice at the take with one processor
2) Twice at the take with two processors
3) Twice at the double with one processor
4) Twice at the double with two processors
Here are the numbers:
nd std.err take std.error
1a +0.569 0.013 +0.654 0.033
1b exactly the same results (what I expected)
2a +0.569 0.046 +0.654 0.033
2b +0.569 0.045 +0.654 0.033
3a exactly the same as 1a and 1b
3b exactly the same as 1a and 1b
4a +0.569 0.048 +0.654 0.033
4b +0.569 0.046 +0.654 0.033
Question 1: The std.error is not shown as big as it was when I first
noticed it, but it's still a little bit higher than with using only one
processor. Why?
Question 2: It differs! Why? I expected the same numbers because of the
same seed.
Assumption 1: The rollout results themselves seem to be reliable (well,
somehow), but I'll crosscheck this with other positions also.
Assumption 2: Richard Stallman did issue the blessing on gnubg after we
shifted to GPLv3 ;-).
If someone wants to debug this, I have saved all rollouts and results.
Ciao
Achim
PS. The behaviour, that gnubg with 2 threads somehow freezes (without
error message at the command line) when stopping a rollout in between
still remains.
_______________________________________________
Bug-gnubg mailing list
[email protected]
http://lists.gnu.org/mailman/listinfo/bug-gnubg
