Hi All,
I'm of the opinion that given the stability of the current builds on
Unix and Windows platforms, and what appears as a measurable increase in
strength and better performance on 3ply (Depreli studies suggest it),
that we should consider ourselves at 1.0 . If we aren't at 1.0 yet, then
we probably will never have a reason to be.
We have a stable product, yes it has some known bugs, but in general I
think we can finally pull it out of testing after more than a decade.
Anyone have any objections?
If there aren't any sizable objections, it may require some
documentation tweaking (to up issue from 0.91) to 1.0, and possibly move
the version of the weights file to 1.0 as well.
Questions, comments etc, please feel free to make them known.
Preliminary (not complete) Depreli results are attached using XG2
rollouts as of 20130427. A regression (using v0.91) still has to be done
against all decisions in the original match files to see if more
positions need rollouts over what is in the current Depreli list. This
would likely increase the totals, but I don't believe it will be
substantial.
--
Michael Petch
GNU Backgammon Developer
OpenPGP FingerPrint=D81C 6A0D 987E 7DA5 3219 6715 466A 2ACE 5CAE 3304
Preliminary
Depreli Results for GNUBG
Using Depreli
positions from 04-27-2013
(5427
Decisions)
Bot Description Chequer Error Missed Double Wrong
Double Wrong Take Wrong Pass All Cube Errors All Errors
--------------------------------------------------------------------------------------------------------------------------------------------------------------
GNUBG v0.91 4-ply 9.1071(1138) 1.1315( 47)
0.5550( 30) 1.4345( 36) 0.3077( 3) 3.4287( 116)
12.5358(1254)
GNUBG v0.91 3-ply Grandmaster 12.8651(1391) 0.7775( 27)
1.8143( 76) 0.8975( 25) 0.4208( 6) 3.9101( 134)
16.7752(1525)
GNUBG v0.91 2-ply CH(LF)/3-ply CU 15.8649(1545) 0.7775( 27)
1.8143( 76) 0.8975( 25) 0.4208( 6) 3.9101( 134)
19.7750(1679)
GNUBG v0.91 Supremo 15.8649(1545) 1.0627( 39)
1.2697( 58) 1.5792( 35) 0.4201( 3) 4.3317( 135)
20.1966(1680) Incomplete/Close
GNUBG v0.91 2-ply WorldClass 16.6188(1559) 1.0627( 39)
1.2697( 58) 1.5792( 35) 0.4201( 3) 4.3317( 135)
20.9505(1694)
GNUBG v0.90 4-ply 16.0467(1414) 3.2647( 108)
0.8644( 27) 4.5741( 58) 0.0969( 2) 8.8001( 195)
24.8468(1609)
GNUBG v0.90 Supremo 25.5800(1805) 3.4119( 97)
1.8000( 53) 5.0106( 59) 0.1567( 3) 10.3792( 212)
35.9592(2017) Incomplete/Close
GNUBG v0.90 2-ply WorldClass 25.9465(1814) 3.4119( 97)
1.8000( 53) 5.0106( 59) 0.1567( 3) 10.3792( 212)
36.3257(2026)
GNUBG v0.90 3-ply Grandmaster 22.0227(1794) 0.8041( 19)
10.0117( 187) 1.1159( 19) 5.9247( 44) 17.8564( 269)
39.8791(2063)
GNUBG v0.90 2-ply CH(LF)/3-ply CU 25.5800(1805) 0.8041( 19)
10.0117( 187) 1.1159( 19) 5.9247( 44) 17.8564( 269)
43.4364(2074)
--------------------------------------------------------------------------------------------------------------------------------------------------------------
Notes:
1. Incomplete: There are 7 moves that need to be rolled out for supremo, but
the final result won't change much.
2. Positions were also analysed with GNUBG v0.14.3. They were almost identical
to v0.90. This is to be expected given that both
versions are based on the same neural net weights file (v0.15). Differences
between the two are changes in the
crashed/contact/race net processing and other small changes over the years
to the neural net code.
3. 2-ply CH(LF)/3-ply CU is an experiment. 2-ply checker with large filter,
3-ply cube.
4. All plies have pruning enabled for chequer and cube decision.
_______________________________________________
Bug-gnubg mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/bug-gnubg
