Huh and hah!
That was the first time i was digging that deep to discover a gap which possibly does not exist at all. Hopefully i was able to keep myself from pushing forward that insane experiment toward infinity. There's no gold in this pit man!
Given that i'm not wired enough to compete with Stockfish testing network, 2448 games per engine should be more than enough for me and my modest tablet fleet. All 306 openings of TCEC-7 were played with both sides between all participating engines.
The chart shown below summarizes the result of 6120 games played by 3 Stockfish 7 builds, Stockfish 6 and Komodo 9.3.
There is no clear reason to replace the build compiled by Peter Österlund, released with Droidfish 1.60 & 1.61 with any of the builds which came from Jim Ablett. They have proven their quality by playing exactly at the same level but they are not stronger at 300+1 time control at least, on a RK3188 processor running at 4 cores x 1 GHz.
There is no clear reason to replace the build compiled by Peter Österlund, released with Droidfish 1.60 & 1.61 with any of the builds which came from Jim Ablett. They have proven their quality by playing exactly at the same level but they are not stronger at 300+1 time control at least, on a RK3188 processor running at 4 cores x 1 GHz.
Program Elo + - Gams Score Oppo Draws
1 Stockfish 7 DF160 : 3368 8 8 2448 56.9 % 3320 66.8 %
2 Stockfish 7 Beta2 JA : 3364 8 8 2448 56.2 % 3321 66.6 %
3 Stockfish 7 JA : 3364 8 8 2448 56.2 % 3321 66.7 %
4 Stockfish 6 : 3290 9 9 2448 43.1 % 3339 61.2 %
5 Komodo 9.3 32-bit : 3259 10 10 2448 37.6 % 3347 48.6 %
Once again, Komodo 9.3 suffered badly and took the last place, even behind Stockfish 6. One should remember that even Komodo 9.2 used to perform ~40 ELO ahead of SF6. What an unexpected regression...
In short, i will keep Stockfish7.DF160.arm7 in Rapidroid and won't care about the lack of an official build of Stockfish 7. The latter will never be released as confirmed in written by Daylen Lang of Stockfish helpdesk. Thus, the Droidfish build is the closest to deserve "official" label since the main app is mentioned in the official site.
Screenshots of 5 stages of the huge round robin based on lot of 50 openings each (56 for the last one to reach 306 total openings):
TCEC-7: Openings 1 to 50 of 306
TCEC-7: Openings 51 to 100 of 306
TCEC-7: Openings 101 to 150 of 306
TCEC-7: Openings 151 to 200 of 306
TCEC-7: Openings 201 to 250 of 306
TCEC-7: Openings 251 to 306 of 306
In case you wanna check how low was the quality of the games compared to Rapidroid, the games in PGN compressed with 7z (7.8 MB) can be downloaded: HERE
P.S.: The latest development version i've posted yesterday was not available during above tourney. Although i'm not sure i can bear another deep experiment to find out how stronger is the develeopment verison, i may come back to the same subject later on because i'm rather interested to see how Komodo 9.3 with contempt 0 will deal with Stockfish 7 and both questions can be answered with a mixed double gauntlet extension.
No comments:
Post a Comment