Fat Fritz 2

[MikeB]

Any reputable rating organization should view this as an acknowledge derivative of Stockfish by the authors and the rating should be part of Stockfish ratings. End of story.

[gaard]

Legal concerns aside for a moment, I do think it's a bit dishonest they compare F.F to SF 12 rather than using the clearly stronger SF Dev.

When you write an article claiming to be the strongest chess entity, the closest rivals need to be included and SF 12 isn't it anymore.

So what was the result of this larger neural network with deeper chess knowledge? In a large test at 1 minute plus 1 second, Fat Fritz 2 scored +42 Elo over Stockfish 12. Click to download the PGN games.

Code: Select all

Score of Fat Fritz 2 vs Stockfish 12:

    286 wins / 99 losses / 1167 draws

    Elo difference: 42.1 +/- 8.5, LOS: 100.0 %, DrawRatio: 75.2 %
    1552 of 1552 games finished.

Code: Select all

Score of Stockfish 200902 vs stockfish 201212: 48 - 164 - 812 [0.443]
...      Stockfish 200902 playing White: 8 - 140 - 364  [0.371] 512
...      Stockfish 200902 playing Black: 40 - 24 - 448  [0.516] 512
...      White vs Black: 32 - 180 - 812  [0.428] 1024
Elo difference: -39.5 +/- 9.5, LOS: 0.0 %, DrawRatio: 79.3 %
1024 of 1024 games finished.

It gets worse. It seems that they've gained 2.6 Elo over SF-dev from December 12, 2020, perhaps less given the error margins. It's likely already well behind SF-dev of today.

[MikeB]

Any reputable rating organization should view this as an acknowledge derivative of Stockfish by the authors and the rating should be part of Stockfish ratings. End of story.

In fact , with respect to Stockfish, is is more clone like ( more carbon copy) than any other Stockfish clone I am aware of. Basically , the only difference is the net. They really should be just selling the net and offering the modified version of Stockfish for free. If someone wants the Fritz bells and whistles, they can pay the $99, and perhaps others would pay $49 for the net only . But right now, I'm not paying a dime.

[Guenther]

Legal concerns aside for a moment, I do think it's a bit dishonest they compare F.F to SF 12 rather than using the clearly stronger SF Dev.

When you write an article claiming to be the strongest chess entity, the closest rivals need to be included and SF 12 isn't it anymore.

So what was the result of this larger neural network with deeper chess knowledge? In a large test at 1 minute plus 1 second, Fat Fritz 2 scored +42 Elo over Stockfish 12. Click to download the PGN games.

Code: Select all

Score of Fat Fritz 2 vs Stockfish 12:

    286 wins / 99 losses / 1167 draws

    Elo difference: 42.1 +/- 8.5, LOS: 100.0 %, DrawRatio: 75.2 %
    1552 of 1552 games finished.

Code: Select all

Score of Stockfish 200902 vs stockfish 201212: 48 - 164 - 812 [0.443]
...      Stockfish 200902 playing White: 8 - 140 - 364  [0.371] 512
...      Stockfish 200902 playing Black: 40 - 24 - 448  [0.516] 512
...      White vs Black: 32 - 180 - 812  [0.428] 1024
Elo difference: -39.5 +/- 9.5, LOS: 0.0 %, DrawRatio: 79.3 %
1024 of 1024 games finished.

It gets worse. It seems that they've gained 2.6 Elo over SF-dev from December 12, 2020, perhaps less given the error margins. It's likely already well behind SF-dev of today.

This can also be seen at current CCRl testing, where it appears as an engine of the 'FatFritz family', which can now be referred as the Addams family of computerchess ;-)

Here it is currently mere 3 rating points above old SF12.

This is from the complete Blitz list (formerly 40/4 now 40/2) after 493 games
http://ccrl.chessdom.com/ccrl/404/rating_list_all.html

Code: Select all

1       Stockfish 12 64-bit 8CPU        3693    +14     -13     72.9%   -155.2  52.0%   1806
2       Dragon by Komodo 64-bit 8C      3671    +15     -15     65.8%   -98.8   60.8%   1313
3       Lc0 0.26.3 t40-1541 RTX2080     3648    +20     -19     63.5%   -84.1   59.4%   763
        Stockfish 2019-10-09 64-bit 8C  3646    +21     -20     79.9%   -204.4  38.6%   900
        Lc0 0.25.1 t40-1541 RTX2080     3643    +15     -15     64.4%   -94.2   59.4%   1389
4       Fat Fritz 2 64-bit              3634    +26     -25     71.9%   -146.1  53.3%   493
5       Stoofvlees II a16 RTX2080       3632    +18     -18     55.5%   -34.3   65.2%   868
        Fat Fritz w471 RTX2080          3631    +11     -11     60.0%   -59.4   68.7%   2490
        Stockfish 12 64-bit             3631    +12     -12     78.5%   -216.8  37.9%   2699

[Ozymandias]

3 Elo points above SF12? Much ado about nothing, I guess.

[Alayan]

It's shameful that Fat Fritz (Leela variant and Stockfish variant) is marked as an independent engine family on CCRL.

[Graham Banks]

It's shameful that Fat Fritz (Leela variant and Stockfish variant) is marked as an independent engine family on CCRL.

It is marketed with Fritz in the name.

The SF connection can be mentioned in the engine notes.

[twobeer]

More honest would be be to name all SF deriveraties Stockfish(name) for example Stockfish(FF2 network) in List..

[twobeer]

A question back to Dkappe -- Is it a violation of (whomever; Albert?) trained the Network to reverse engineer the binary and extract the weights into a typical Stockfish form? To distribute that?

I would suggest anyone extracting or downloading then change some weights by a single RL run and then release it for free, . call it "Slim Shady Fritz", and then let see CB lawyers try to "prove" these new weights net infringe their "copyright" -IPs of the net...

Similar thinking here as the excellent work with the Rocket (houndini) engine.. "Slim Fritoz 2"

[dkappe]

A question back to Dkappe -- Is it a violation of (whomever; Albert?) trained the Network to reverse engineer the binary and extract the weights into a typical Stockfish form? To distribute that?

I would suggest anyone extracting or downloading then change some weights by a single RL run and then release it for free, . call it "Slim Shady Fritz", and then let see CB lawyers try to "prove" these new weights net infringe their "copyright" -IPs of the net...

Similar thinking here as the excellent work with the Rocket (houndini) engine.. "Slim Fritoz 2"

Well, inciting others to break the law doesn’t strike me as very clever.

I don’t know that FF2 behaves the same way, but it’s very easy to detect that Night Nurse is trained from mcts data rather than ab data. Also, RL training a mcts trained net will make it weaker in a hurry.

So good luck with that...from prison.