Fat Titz 2 released!

[Sopel]

I'm proud to present Fat Titz 2, which uses an improvement network, trained with help of Farseer, and showing an overall improvement of 60 elo over Fat Titz 1*! Additionally more patches have been ported from Stockfish, though some were omitted.

Code: Select all

   # PLAYER        :  RATING  ERROR  POINTS  PLAYED   (%)
   1 fat_titz_2    :  2331.5    2.3  5888.0   10000    59
   2 fat_titz      :  2268.5    2.3  4112.0   10000    41

You can learn about unique features of Fat Titz here, here, and here

You can find windows binaries (not PGO, you may still benefit from compiling yourself) and source on the github's releases page. The network file can be found here or in the github releases page, and must be placed in the same directory as the FT2 executable.

Bench: 5591016

For those curious, no NEON support yet, it might or might not happen in the future.

* 10s+0.1s TC, UHO_XXL_+0.90_+1.19.epd, 2xIntel(R) Xeon(R) Gold 6230R CPU @ 2.10GHz

[chrisw]

I'm proud to present Fat Titz 2, which uses an improvement network, trained with help of Farseer, and showing an overall improvement of 60 elo over Fat Titz 1*! Additionally more patches have been ported from Stockfish, though some were omitted.

Code: Select all

   # PLAYER        :  RATING  ERROR  POINTS  PLAYED   (%)
   1 fat_titz_2    :  2331.5    2.3  5888.0   10000    59
   2 fat_titz      :  2268.5    2.3  4112.0   10000    41

You can learn about unique features of Fat Titz here, here, and here

You can find windows binaries (not PGO, you may still benefit from compiling yourself) and source on the github's releases page. The network file can be found here or in the github releases page, and must be placed in the same directory as the FT2 executable.

Bench: 5591016

For those curious, no NEON support yet, it might or might not happen in the future.

* 10s+0.1s TC, UHO_XXL_+0.90_+1.19.epd, 2xIntel(R) Xeon(R) Gold 6230R CPU @ 2.10GHz

Congratulations on the release. Also on the size of the accumulator!

[dkappe]

Fat Titz 2 by Stockfish. Hooray!!

[Eduard]

It seems the engine is very strong.

[mehmet123]

+63 elo is a very important difference. The only problem is that this test was done at 10 sec 0.1 sec time control with an unbalanced book. What is important is how Fat Titz 2 will perform in rapid and longer time tests.


Fat Titz vs Stockfish Dev:

10 sec + 0.2 TC: - 44 elo
1 min + 0.5 TC: -20 elo
10 min + 0 TC: +1 elo

forum3/viewtopic.php?f=6&t=74518&start=240

[mbabigian]

If you want to test FatTitz against Stockfish, you cannot use hyper-bullet speed. The network is much larger and slower. FatTitz will blunder tactically at those speeds.

Elo compression begins when the draw rate rises. So increasing the TC time until the draw rate climbs is how I determine a fair TC on my hardware. In my tests, the draw rate didn't go up until I went from 120+1.2 to 240+2.4. So the minimum fair TC to test FatTitz on my hardware is greater than or equal to 120+1.2, but less than 240+2.4. (1 Thread)

Here is my 240+2.4 test against SFdev

Code: Select all

   # PLAYER      :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W     D    L  D(%)
   1 SFdev       :     0.0   ----  1635.5    3206    51      99  843  1585  778    49
   2 FatTitz2    :    -9.5    8.3  1570.5    3206    49     ---  778  1585  843    49

My two cents,
Mike (Farseer)

[ChickenLogic]

+63 elo is a very important difference. The only problem is that this test was done at 10 sec 0.1 sec time control with an unbalanced book. What is important is how Fat Titz 2 will perform in rapid and longer time tests.


Fat Titz vs Stockfish Dev:

10 sec + 0.2 TC: - 44 elo
1 min + 0.5 TC: -20 elo
10 min + 0 TC: +1 elo

forum3/viewtopic.php?f=6&t=74518&start=240

Of course 10 seconds + 0.1 with an unbalanced book is a valid test. And that's regardless of NN size or whatever you care about. In the end, Chess can be solved with enough time. That means every algorithm naturally comes closer and closer to the solution with more time so the draw rate increases and the elo gap shrinks. Given that we don't have an infinite amount of time the faster algorithm is better. You wouldn't argue that a pure brute force algorithm without TT is the same strength as SF just because they in theory converge to the same solution (ignoring GHI). So stop arguing the same for a weaker net that needs substantially more time to find solutions of SF dev's caliber.

Constantly shifting the goal post "but with this extremely long time control this engine is actually better" and relying on the fact that nobody can feasibly test with such long time controls is just poor. Both Leela and Stockfish managed to prove elo gain at these low time controls after introducing a substantially larger net compared to their previous one so I really don't get your argument. Either a net is better or it isn't.

I challenge you to run a SPRT test via cutechess-cli at 10mins + 0 TC of SF dev vs FT2 (the time control you used to 'prove' that "FT > SF dev") and give us the result and PGN. Since Fishtest is running two tests with 95% confidence each I'll be extremely generous and ask for a single test with 96% confidence (alpha and beta of 0.04).
If you can't provide such a test maybe stop using that TC to 'test'. I'd recommend calling it 'tournament' as that makes it clear that the result isn't meant to be taken as a significant one and it indicates that you mostly do this for the fun of it (which I get). It would be a 100 times more honest.

[mbabigian]

I think you over estimate my interest in your nonsense. If you don't like my test indicating FT is weaker. Run all the tests you want.

[ChickenLogic]

I think you over estimate my interest in your nonsense. If you don't like my test indicating FT is weaker. Run all the tests you want.

Wasn't replying to you. Your test is fine as it is.

[Eduard]

Fette Brezel.bin (ZIP):
My new polyglot book for fat games on the servers.

Book Depth = 100 Halfmoves
Best Move = OFF

https://filehorst.de/d/enuxGzbu

First test with Fat Titz 2 on PlayChess: It's competitive, unbeaten after 6 hours.

Fette = Fat

Brezel =