A lot of science fiction on abrok.eu/stockfish

!!!



???

R_Silvian wrote:

!!!

???

If you read the results in the right way, all fits very well:

Fishtest plays Selfplay-Tests, only. And the last progression-tests shows indeed a nice progress, compared to Stockfish 16:
https://github.com/official-stockfish/Stockfish/wiki/Regression-Tests

The progress is +19 Elo here.


Now, we look in the single head-to-head results of Stockfish 231231 in my testrun versus Stockfish 16:
https://www.sp-cc.de/files/programs.dat

Because this is a (small) Selfplay-Test, too:
1000 (+285,=489,-226), 53.0 %
means: Stockfish 231231 scored 53.0% in its 1000 games head-to-head versus Stockfish 16. And 53.0% score means: +21 Elo


So, the Selfplay-Elo results fit very well... And the complete ratinglist-result of Stockfish 231231 in my UHO-Top15 Ratinglist is not selfplay-Elo.
So, what we can learn here, is, selfplay-Elo is not the same as Ratinglist-Elo, even though, both use UHO-openings.
And, we can learn, reading results carefully and in the right way, can make some fiction very real.

pohl4711 wrote:

So, the Selfplay-Elo results fit very well... And the complete ratinglist-result of Stockfish 231231 in my UHO-Top15 Ratinglist is not selfplay-Elo.
So, what we can learn here, is, selfplay-Elo is not the same as Ratinglist-Elo, even though, both use UHO-openings.
And, we can learn, reading results carefully and in the right way, can make some fiction very real.

Of course, selfplay-Elo is not the same as Ratinglist-Elo. It is a well-known fact. Almost all the authors present the gain of the new versions in selfplay Elo......and in the rating lists (games between different engines - Stefan Pohl RL, CCRL, CEGT & so on) the progress shown is about 50% of that presented by the authors. The science fiction part appears when a progress in selfplay (actually a whole chain of temporal progress) is not found as even a minimal progress in the actual game (between engines).