Fat Fritz defeats Stockfish in 100-game matches

by ChessBase
11/6/2019 – Our recent article on Fat Fritz elicited some criticism and predictions. In one comment a reader concluded "Stockfish is STILL the champion." While we look forward to testing that proposition in formal competition soon, we have naturally "run the experiment" so to speak, already: We ran two engine vs engine matches, a pre-release build of Fat Fritz against Stockfish 8 and then against Stockfish 10. The result? Fat Fritz won both! Here are some of the games along with grandmaster analysis from GM ELSHAN MORADIABADI.

ChessBase 15 - Mega package ChessBase 15 - Mega package

Find the right combination! ChessBase 15 program + new Mega Database 2020 with 8 million games and more than 80,000 master analyses. Plus ChessBase Magazine (DVD + magazine) and CB Premium membership for 1 year!

More...

Advantage Fat Fritz

Our recent article "Fat Fritz seemed to be from an entirely different plane of existence!" elicited some criticism and predictions. Here's one of the comments we received:

Karbuncle 11/4/2019 12:56: FYI Leela was defeated by Stockfish in the TCEC Cup knockout tournament. Stockfish also won the TCEC Season 16 Super Final. Leela played so poorly against weaker engines that it didn't even qualify for the Super Final. So while GMs and other may be jumping on the NN bandwagon, Stockfish is STILL the champion.

Well, we wanted to see for ourselves, and staged two matches between Fat Fritz and Stockfish. The computer we used for both Fat Fritz and Stockfish was a Ryzen 7 3700X with an NVIDIA RTX 2080, 32 GB RAM. Ponder was switched off, the Silver Openings Suite was used to ensure maximum opening diversity, and the time control was 20m+10s.

Stockfish ran on 16 threads at 16 million nodes per second, while the pre-release version of Fat Fritz ran at a little over 11 thousand NPS (for the Leela specialists this was done with the slower CUDNN backend). The speed conditions are actually a good deal worse than AlphaZero, which had around a 900 to 1 NPS speed difference, while here Stockfish was running at 1450 times faster in NPS, so if anything, it favored Stockfish. 

We conducted matches against Stockfish 8 and Stockfish 10. Here are the results:

So Fat Fritz defeated Stockfish 8 with 62.5%, and Stockfish 10 by 52.5%. Readers will wonder why Stockfish 8 at all? The reason was to be able to compare with AlphaZero to get a rough baseline for comparison. We only know how strong AlphaZero was by virtue of its results against Stockfish 8, which it beat by a bit over 30 Elo on average in long games. In roughly similar conditions, the pre-release build of Fat Fritz scored nearly 90 Elo over it, using a varied set of balanced openings even in speed conditions that were a bit worse. While hardly proof, it does strongly suggest Fat Fritz is stronger than AlphaZero in equal conditions.

Still, not content to just post a number of results, we sent the games to a strong grandmaster to look at them. There have been numerous posts in feedback on how moves shown were all typical fare for the open-source project Leela Chess Zero. GM Elshan Moradiabadi is a player who has analyzed extensively with it, so it was interesting to see whether he agreed with this assessment. He did not. 

In fact, after just one hour with the 200 games we gave him, he messaged back, "I have 100 things to do, but this is so much fun, I hate you!". When was the last time you ever heard a GM say he was loving looking at a game between engines?

Ok, but what did he think now that he was able to see how it played?

"Unlike Leela, Fat Fritz has a sweet tooth for complexity, and while Leela deals with positional sacrifices as they happen (as a matter of necessity or not), Fat Fritz looks for those complex lines, making it a great companion for those who take a liking to the games of Kasparov and Tal!"

Remember, this was not his opinion based on a couple of carefully chosen games given to him. This was his assessment after reviewing 200 games. We'll be releasing all the games and GM Elshan Moradiabadi's comments on moments he highlighted.

GM Elshan Moradiabadi lives in Texas and studied in NODET Schools (National Organization for Development of Exceptional Talents) from 1996 to 2003. NODET is an organization that recruits students for middle and high schools through and is aimed to provide a unique educational environment for the exceptionally talented students.

Elshan is a currently a member of the USCF and has played for the Texas Tech University, winning the Pan-American Team Championship for the first time in 2015. Originally he is from Iran, where he made an international mark by winning the 2001 Iranian Chess Championship with a score of 10/11, ahead of Ehsan Ghaem Maghami. He was just 16 years old at the time.

Elshan is very well-versed in computer chess and has been working intensely with chess engines for many years now. After receiving the games we had played with Fat Fritz against Stockfish he wrote us his impressions.


Where human intuition trains AI

By GM Elshan Moradiabadi

I think it is better to start the article with the most distinctive characteristic of FF, something that separates it from Leela, showing why these two phenomenal engines are in fact two very distinct entities. Take the following game in a classical King's Indian Defense where Fat Fritz is Black.

[Event "Rapid match 20m+10s"] [Site ""] [Date "2019.10.28"] [Round "100"] [White "Stockfish 10 64 POPCNT"] [Black "Fat Fritz"] [Result "1/2-1/2"] [ECO "E99"] [Annotator "Elshan"] [PlyCount "165"] [EventDate "2019.??.??"] [TimeControl "1200+10"] 1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5 7. O-O Nc6 8. d5 Ne7 9. Ne1 Nd7 10. Be3 f5 11. f3 f4 12. Bf2 g5 {This is my favorite game of this match. You might be surprised why, as not only does FF fail to win the game but it is nothing more than a repetition of a well-tested line whose soundness has been proven over the board and in correspondence chess. However, we need to remember that the openings suite ends here, yet it follows the best line, simply finding it during this rapid game.} 13. Nd3 {0.93/32 97} Nf6 { 0.03/14 20} 14. c5 {0.98/31 62 (a4)} Ng6 {0.00/16 26} 15. Rc1 {1.04/30 51 (c6)} Rf7 {-0.02/21 36} 16. Kh1 {0.78/30 30 (Sb5)} (16. a4 {is another well-probed path.}) 16... h5 {0.19/20 47 Again, the most principled. Bf8 is also a sound option.} (16... Bf8 17. Rg1 Rg7 18. b4 h5 19. Qd2 g4 20. g3 fxg3 21. Rxg3 a5 22. a3 axb4 23. axb4 c6 24. dxc6 bxc6 25. cxd6 Qxd6 26. Bc5 Qd8 27. Bxf8 Qxf8 28. fxg4 Bxg4 29. h3 Bxe2 30. Qxe2 Raa7 31. Rf1 Nf4 32. Rxg7+ Rxg7 33. Nxf4 exf4 34. Qc4+ Qf7 35. Qxf7+ Kxf7 36. Rxf4 Rg3 37. Kh2 Rxc3 38. e5 Ke6 39. exf6 Kf7 40. h4 c5 41. bxc5 {1/2-1/2 (41) Lenderman,A (2599)-Nakamura,H (2787) Saint Louis 2018}) 17. Nb5 {0.98/28 17} g4 {0.20/25 34} 18. cxd6 {0.88/33 133} cxd6 {0.15/29 17} 19. Nxa7 {0.35/34 140} (19. Qc2 Ne8 20. Nxa7 Bd7 21. Qb3 g3 22. Bg1 {is a line you need to test with FF, once you obtain it!}) 19... Bd7 { 0.09/29 24} 20. Qb3 {0.36/30 13 [#]} Nxe4 $3 {0.04/29 22 The only move to keep Black in the game. Correspondence games have shown that this move holds the ground for Black, yet Fat Fritz found this on its own after just 22 seconds. Astonishing!} 21. fxe4 {0.15/34 60} f3 {0.04/26 0} 22. Bb6 {0.00/35 57} (22. gxf3 gxf3 23. Bd1 Bh3 24. Rg1 Bg2+ 25. Rxg2 fxg2+ 26. Kg1 Qg5 $4 {Doctor Nunn is definitely among the greatest chess minds of the 20th century, but here he chose a wrong path, after} (26... Bh6 27. Rc3 (27. Rc2 Rxa7 28. Bxa7 $19 Rf1+) 27... Qf6 28. Qb6 Nh4 29. Rc8+ Rxc8 30. Nxc8 Nf3+ 31. Kxg2 Qg5+ 32. Bg3 h4 33. Bxf3 hxg3 34. hxg3 Qd2+ 35. Nf2 Be3 36. Qd8+ Kg7 37. Kh3 Qxf2 38. Bg2 {Black obtains a better endgame after Qf6. I am sure FF would have come up with an even better line!}) 27. Rc8+ Rxc8 28. Nxc8 Bf8 29. Qc2 Nh4 30. Nb6 Nf3+ 31. Bxf3 Rxf3 32. Nc4 h4 33. Qe2 Rf7 34. h3 Be7 35. Ne3 Kh8 36. Nxg2 Rg7 37. Nde1 Qg6 38. Qf3 Bg5 39. Qf5 Qxf5 40. exf5 Rc7 41. Nxh4 Kg7 42. Nhf3 Kf6 43. Bb6 Rc8 44. Kf2 Kxf5 45. Nxg5 Kxg5 46. Ke3 Kf5 47. h4 e4 48. a4 Rg8 49. Nc2 Ke5 50. Bd4+ Kxd5 51. b3 Rg3+ 52. Kf4 Rf3+ 53. Kg4 Rxb3 54. h5 Rb1 55. Bf2 Rh1 56. Ne3+ Ke5 57. Bg3+ Ke6 58. Bh4 Rg1+ 59. Kf4 d5 60. h6 Rg8 61. Nc2 Kf7 62. Ne3 Ke6 63. Bg5 d4 64. Nc2 d3 65. Nd4+ Kd5 66. Nf5 d2 67. Ne3+ Kd4 68. Bf6+ Kd3 69. h7 Rf8 70. h8=Q Rxh8 71. Bxh8 b6 72. Be5 d1=Q 73. Nxd1 Kc2 74. Nc3 {1-0 (74) Merry,A (2429)-Nunn,J (2572) London 2018}) 22... Qh4 {-0.01/30 71} 23. Bxf3 {0.00/35 11 } gxf3 {-0.01/40 11} 24. Rxf3 {0.00/37 15} Rxf3 {0.00/32 0} (24... Qxe4 25. Rxf7 Kxf7 26. Nf2 Qe2 27. Qf3+ Qxf3 28. gxf3 Nf4 29. Rc7 Ke8 30. Ne4 Nxd5 31. Nxd6+ Ke7 32. Nf5+ Kf6 33. Rxd7 Nxb6 34. Rxb7 Kxf5 35. Rxg7 Kf4 {1/2-1/2 (35) Brugger,A (2523)-Zugrav,W (2559) ICCF email 2016}) 25. gxf3 {0.00/38 24} Rf8 { -0.04/31 65} 26. Qd1 {0.00/40 20 The merits of the move Qg5 have been proven in the past, here FF suggests a new equally good route.} Bh6 {-0.04/17 0} 27. Rc7 {-0.10/41 94} Bh3 {-0.04/16 0} 28. Ne1 {0.00/36 14 (Tc3)} Kh8 {-0.38/21 144 (Sf4)} 29. Rc2 {0.00/39 26} Bd7 {-0.38/19 0} 30. Rf2 {-0.09/37 22} Nf4 { -0.39/22 54 (Dh3)} 31. Rf1 {0.00/41 27 (Sg2)} Qh3 {-0.31/23 248} 32. Rg1 { 0.00/40 25} h4 {-0.32/25 0} 33. Qc2 {0.00/42 27 (a4)} Nh5 {-0.36/19 83 (Le8)} 34. Rg2 {0.00/39 17} Nf4 {-0.34/20 32 (Lf4)} 35. Rg1 {0.00/41 35} Nh5 {-0.34/ 21 14 (Kh7)} 36. Rg2 {0.00/43 35} Bf4 {-0.29/20 66 (Kh7)} 37. Kg1 {0.00/41 17} Kh7 {-0.29/24 0 (Tf7)} 38. a4 {0.00/40 40} Bh6 {-0.29/21 0 (Lg3)} 39. Kh1 { 0.00/37 17} Nf4 {-0.30/21 68 (Ta8)} 40. Rg1 {0.00/41 31} Be8 {-0.29/23 38 (Sh5) } 41. Qc7+ {0.00/38 24 (Dc4)} Bf7 {-0.13/30 54} 42. Rg4 {0.00/41 18} Ne2 { -0.09/34 76} 43. Rg2 {0.00/43 43} Nd4 {-0.10/34 0} 44. Qe7 {0.00/42 33} Nxf3 { -0.08/33 8} 45. Rf2 {0.00/43 16} Ng5 {-0.07/30 21} 46. Nb5 {0.00/43 16} Qb3 { -0.08/15 0} 47. Rf5 {0.00/44 27} Qd1 {-0.07/20 0} 48. Bf2 {0.00/43 27 (La5)} Qe2 {-0.20/29 91} 49. Ng2 {0.00/41 13} Rg8 {-0.20/30 0} 50. Rxf7+ {0.00/41 16 (Sc3)} Nxf7 {-0.35/27 76} 51. Qxf7+ {0.00/42 23} Rg7 {-0.30/27 72} 52. Qf5+ { 0.00/44 19} Kh8 {-0.28/26 28 (Tg6)} 53. Qf8+ {0.00/45 27 (Lg3)} Kh7 {-0.26/21 32} 54. Qf5+ {0.00/44 35} Rg6 {-0.24/21 17} 55. Bg3 {0.00/45 13} hxg3 {-0.23/ 21 25} 56. hxg3 {0.00/46 18} Be3 {-0.22/20 22 (Dd1+)} 57. Nxe3 {0.00/43 25} Qxe3 {-0.21/19 6} 58. Qd7+ {0.00/49 12} Kh8 {-0.19/17 65 (Kh6)} 59. Qd8+ { 0.00/49 29 (Dc8+)} Kg7 {-0.17/15 22 (Kh7)} 60. Qd7+ {0.00/52 22} Kg8 {-0.14/14 18 (Kh8)} 61. Qe8+ {0.00/52 15 (Dc8+)} Kh7 {-0.13/13 12 (Kg7)} 62. Qd7+ { 0.00/51 10} Rg7 {-0.11/12 12 (Kh6)} 63. Qf5+ {0.00/50 20} Kg8 {-0.07/12 22 (Kh6)} 64. Qc8+ {0.00/53 19 (De6+)} Kh7 {-0.06/10 5} 65. Qh3+ {0.00/54 99 (Df5+)} Kg6 {-0.10/13 17} 66. Qg4+ {0.00/56 18} Kh6 {-0.08/13 10} 67. Qh3+ { 0.00/53 30 (Dh4+)} Kg5 {-0.08/11 18 (Kg6)} 68. Qh4+ {0.00/51 5 (Df5+)} Kg6 { -0.09/12 0} 69. Qg4+ {0.00/59 32} Kf6 {-0.06/13 27 (Kh6)} 70. Qf5+ {0.00/46 12} Ke7 {-0.04/11 0} 71. Qe6+ {0.00/46 6} Kf8 {-0.03/11 4} 72. Qc8+ {0.00/49 11} Ke7 {-0.02/8 7} 73. Qc7+ {0.00/49 17 (De6+)} Kf6 {-0.04/17 19 (Kf8)} 74. Qxd6+ {0.00/45 4} Kf7 {-0.04/14 3} 75. Qd7+ {0.00/48 10 (De6+)} Kg8 {-0.06/15 27 (Kf8)} 76. Qd8+ {0.00/46 30 (De6+)} Kh7 {-0.10/16 9 (Kf7)} 77. Qh4+ {0.00/52 4} Kg6 {-0.08/13 14 (Kg8)} 78. Qg4+ {0.00/50 8} Kh6 {-0.06/11 11 (Kh7)} 79. Qh3+ { 0.00/51 14 (Dh4+)} Kg5 {-0.07/13 12 (Kg6)} 80. Qf5+ {0.00/52 20 (Dh4+)} Kh6 { -0.11/11 0} 81. Qh3+ {0.00/54 11} Kg6 {-0.03/6 21} 82. Qf5+ {0.00/52 5} Kh6 { -0.01/3 0} 83. Qh3+ {0.00/57 15 If it were a battle among two top-GMs , you could say it was a moral victory for Black, as White struggled his way (although never worse) to maintain equality. In my opinion, this is the most distinctive game where we can see why Leela and FF- although in essence representative of deep positional ideas while banking on domination and closed pawn structures- are essentially very different.} 1/2-1/2

FF seems to be more keen toward tactical lines, making it somewhat faster at calculation compared to Leela (I cannot comment on accuracy), while its choice for most principled candidate moves, which may be just a short-term observation, makes FF more vulnerable than Leela which instead prefers positional principles before initiative. In my humble opinion, this is the most visible difference between the engines, yet this alone makes the two engines absolutely separate entities despite their common roots.

Now, one may ask what are those strategic cues which differentiate neural networks from brute force engines. The following game, is a great example where both Leela (I checked Leela’s moves on the cloud) and FF show deep positional understanding and outplay SF10 in a closed position.

Game 15: The Stone Wall game where FF demonstrates a fantastic idea after trading the dark square bishops.

 

This game in fact inspired me a lot, as I had a similar game back in 2006, winning with the same maneuver.

[Event "Rapid match 20m+10s"] [Site ""] [Date "2019.10.24"] [Round "15"] [White "Fat Fritz"] [Black "Stockfish 10 64 POPCNT"] [Result "1-0"] [ECO "A90"] [Annotator "Elshan"] [PlyCount "227"] [EventDate "2019.??.??"] [TimeControl "1200+10"] 1. d4 f5 2. g3 Nf6 3. Bg2 e6 4. c4 d5 5. Nf3 c6 6. O-O Bd6 7. b3 Qe7 {Last book move.} 8. Ne5 {0.47/12 21} O-O {0.00/29 44 (Sbd7)} 9. Bf4 {0.59/13 25 (a4) } b6 {0.09/31 31} 10. Nd2 {0.60/15 26 (Sc3)} Bb7 {0.00/30 12} 11. Nd3 {0.63/15 22 (h4)} Bxf4 {0.01/31 25 (La3)} 12. Nxf4 {0.67/16 36} g5 {0.00/32 15} 13. Nd3 {0.66/17 22} Nbd7 {0.25/34 41} 14. cxd5 {0.58/18 107} exd5 {0.36/34 29} 15. e3 {0.59/19 32 (Sf3)} a5 {0.34/33 133 (Se4)} 16. Nf3 {1.01/14 27 (Tc1)} h6 { 0.55/31 27 (Se4)} 17. Rc1 {1.04/12 59} Rac8 {0.40/32 13} 18. Nde5 {1.02/13 28 (Te1)} Qe6 {0.31/28 18} 19. Nxd7 {1.01/12 17 (Dc2)} Nxd7 {0.39/29 42} 20. Qd3 { 1.01/11 24 (De2)} Rc7 {0.49/32 49 (Tce8)} 21. Rc2 {1.11/12 53 (Tfd1)} Ra8 { 0.44/33 33 (c5)} 22. Rfc1 {1.08/12 34} Ba6 {0.60/37 38 (Sf6)} 23. Qd1 {1.19/13 36} Rac8 {0.60/40 19} 24. h4 {1.19/13 27} g4 {0.60/41 17 (Df6)} 25. Ne1 { 1.28/13 41} Nf6 {0.60/42 38 [#] This is where Leela and FF are very similar: Fantastic positional maneuvers.} 26. Bh1 $3 {1.29/13 28 You can start with Rc3. } Nh5 {1.03/39 204 (Kg7)} 27. Rc3 {1.28/11 88 (Sg2)} Kg7 {1.03/36 24} 28. Ng2 $1 {1.30/12 27} Bb7 {1.03/36 21 (Dd6)} 29. Nf4 {1.69/17 30 (Dc2)} Nxf4 { 0.90/33 14} 30. exf4 {1.67/18 15} Re7 {0.90/36 27} 31. Re3 {1.67/19 20} Qd6 { 0.85/38 55 (Df6)} 32. Re5 {1.73/17 49 A positional masterpiece. This actually reminded me of a game of mine, I played long-time ago. So, Fat Fritz can give you nostalgia too!} Rf8 {1.05/37 64} 33. Bg2 {1.85/16 50 (Kh2)} Rxe5 {1.22/33 50 (h5)} 34. fxe5 {1.66/14 41} Qe7 {1.29/37 30 (Dg6)} 35. Qe2 {1.96/13 40} Qe6 {1.29/38 27 (Df7)} 36. Qd2 {1.94/18 38 (f4)} Rf7 {1.29/40 22 (f4)} 37. Kh2 { 2.18/12 52} Qe7 {1.30/38 15} 38. Qf4 {2.21/13 21 (f4)} Qe8 {1.31/37 35 (De6)} 39. Bf1 {2.40/12 40 (Dd2)} Bc8 {1.31/38 11} 40. a3 {2.35/13 69} Ra7 {1.40/41 29 (Ld7)} 41. Bd3 {2.46/10 37 (Tc2)} Qe6 {1.50/38 30} 42. Rc2 {2.30/10 41} Ra8 {1.53/41 50 (Ld7)} 43. Kg2 {2.20/11 34 (f3)} Bd7 {1.31/37 10} 44. Rc1 {2.14/12 16 (De3)} Ra7 {1.31/35 13 (a4)} 45. Kg1 {2.18/10 51 (Tc2)} Rc7 {1.31/35 24 (Ta8)} 46. Qd2 {2.27/11 23} Ra7 {1.49/38 65 (c5)} 47. Qe3 {2.19/9 36 (b4)} Rb7 {1.69/40 39 (Ta8)} 48. Kh2 {2.37/9 30 (Dd2)} Rc7 {1.69/37 18 (Ta7)} 49. Rc2 { 2.29/10 25 (Kg1)} Qe7 {1.69/37 7} 50. Qf4 {2.21/12 10} Qe6 {1.69/40 10} 51. Qe3 {2.19/12 25 (Dc1)} Ra7 {1.69/42 10 (De7)} 52. Qf4 {2.14/10 36 (Kg1)} Bc8 { 1.69/42 15 (Ta8)} 53. Kg2 {2.32/9 27 (Kg1)} a4 {1.69/43 34 (Ld7)} 54. b4 { 2.17/10 12} b5 {1.69/46 9} 55. f3 {2.18/10 19 (Tc5)} gxf3+ {1.79/41 17 (Tf7)} 56. Qxf3 {2.28/10 23} Rb7 {2.12/42 29 (Tf7)} 57. Kh2 {2.38/9 24 (Df4)} Rf7 { 2.46/40 27} 58. Rf2 {2.30/9 16} Bd7 {2.56/40 15} 59. Qh5 {2.29/10 10} Rf8 { 2.56/40 22} 60. Rf4 {2.24/11 9} Qf7 {2.56/43 10} 61. Qd1 {2.20/11 12 (De2)} h5 {2.56/43 10} 62. Be2 {2.18/10 10 (Tf3)} Be8 {2.56/45 13 (Th8)} 63. Rf1 { 2.16/10 31 (Ld3)} Qg6 {2.56/41 9} 64. Kg2 {2.12/10 16 (Ld3)} Qh6 {2.56/42 19} 65. Qc1 {2.13/11 4 (Tf3)} Qxc1 {2.13/43 14 (Dg6)} 66. Rxc1 {2.00/12 10} Kg6 { 2.13/47 7 (Tg8)} 67. Kf2 {2.03/10 19 (Tc3)} Bd7 {2.13/45 10 (Th8)} 68. Ke3 { 2.09/10 13 (Tf1)} Rh8 {1.92/43 6} 69. Kf4 {1.96/10 14 (Tc5)} Kf7 {1.92/50 8} 70. Rc3 {1.83/10 14} Ke7 {1.92/48 7 (Ke6)} 71. Re3 {1.79/10 11 (Ld1)} Be6 { 1.92/50 11 (Ke6)} 72. Bd1 {1.74/11 11 (Tc3)} Bd7 {1.92/50 11} 73. Bc2 {1.69/12 11 (Te1)} Rf8 {1.92/51 9 (Le6)} 74. Bd3 {1.73/12 9 (Ld1)} Be6 {1.92/55 8 (Ke6)} 75. Be2 {1.63/13 10} Rh8 {1.92/54 10} 76. Rc3 {1.62/14 12} Bd7 {1.92/56 14} 77. Bd1 {1.61/13 6 (Tc5)} Ke6 {1.92/52 8} 78. Bc2 {1.60/15 5 (Tc5)} Ke7 {1.92/53 12 (Tf8)} 79. e6 {2.03/19 10 (Tc5)} Kxe6 {1.67/35 6} 80. Bxf5+ {1.88/24 5} Kd6 {2.75/40 29} 81. Bxd7 {1.83/21 15} Kxd7 {3.03/42 54} 82. Kg5 {1.83/20 0} Kc7 { 3.68/41 30 (Kd6)} 83. g4 {2.24/16 14 (Te3)} hxg4 {1.85/35 3} 84. h5 {2.24/20 2} Rg8+ {3.63/36 17} 85. Kh4 {2.12/25 10 (Kf4)} Kd7 {2.39/38 6} 86. Re3 {2.05/26 9 (Tg3)} c5 {1.90/35 4} 87. dxc5 {2.01/25 7} d4 {1.91/40 6} 88. Rg3 {2.00/24 5} Ke6 {1.91/42 7} 89. Rxg4 {1.98/23 6} Rd8 {1.91/44 6} 90. c6 {1.98/22 7 (Tg6+)} d3 {1.35/36 5} 91. c7 {1.97/20 7} Rc8 {1.90/43 19} 92. Rd4 {1.95/19 6} Rxc7 { 1.91/46 18} 93. Rxd3 {1.94/17 5} Kf6 {1.91/39 4} 94. Rf3+ {1.92/16 8} Kg7 { 1.91/46 6} 95. Kg5 {1.91/14 12 (Tg3+)} Rc2 {3.52/37 30} 96. Re3 {1.88/14 36 (Kf4)} Rf2 {1.41/42 5 (Th2)} 97. Rd3 {1.88/11 41 (Tg3)} Rf1 {1.41/45 7 (Kh7)} 98. Rg3 {1.88/10 10 (Td7+)} Rf2 {3.07/43 22} 99. Re3 {1.83/10 10 (Th3)} Rf1 { 3.69/38 12 (Kh7)} 100. Rc3 {1.81/10 10 (Te7+)} Kh7 {3.69/40 6} 101. Rg3 { 1.77/10 9 (Td3)} Rf2 {3.88/40 14} 102. Rh3 {1.76/10 10 (Tc3)} Kg7 {3.98/45 10} 103. Rc3 {1.74/10 9 (Te3)} Re2 {3.98/40 3 (Kh7)} 104. Rc7+ {1.72/11 9 (Tg3)} Kg8 {3.98/42 7} 105. Rc3 {1.70/12 4} Kf7 {4.08/41 7 (Kh7)} 106. Rh3 {1.97/13 11 (Tf3+)} Rf2 {4.61/36 22 (Kg7)} 107. h6 {2.17/14 8} Kg8 {5.15/35 12} 108. Kg6 {2.15/14 6} Rf1 {5.49/34 10 (Kh8)} 109. Re3 {2.18/14 13 (Th5)} Rg1+ {5.77/31 8} 110. Kf5 {2.21/13 8 (Kf6)} Rh1 {4.78/35 11 (Kh7)} 111. Ke6 {2.39/13 9 (Kg6)} Rxh6+ {6.67/28 11} 112. Kd7 {2.32/13 7} Rh5 {7.50/29 10 (Kg7)} 113. Rf3 { 2.23/11 13 (Te8+)} Kg7 {7.59/33 10} 114. Kc7 {2.13/11 22 (Kc6)} 1-0

The game is a positional gem, with a noteworthy point to mention: FF demonstrates very different technical skills in converting a winning position than Leela. I find it quite amusing, despite the fact that at some point there were more than two paths to victory. FF’s choices make it a perfect training partner for those (surprisingly someone like me) who struggle with the accurate conversion of dynamic decisive advantages.


Check back tomorrow for more commentary from GM Moradiabadi, plus all match games!

We'll also be happy to answer questions and feedback you may have on this new beast.


Links




Reports about chess: tournaments, championships, portraits, interviews, World Championships, product launches and more.
Discussion and Feedback Join the public discussion or submit your feedback to the editors


niteguide27 niteguide27 11/9/2019 10:58
What a misleading article!
caliche2016 caliche2016 11/9/2019 08:18
What is the point of comparing Fat Fritz with Stockfish if anybody with an nVidia video card has access to Leela Chess Zero for FREE? If FF could produce the same high quality analysis without using an expensive video card, things would be different, but as they are, there is absolutely no reason to spend a single dollar on this engine.

Unless you can prove that FF can also beat LCZero in the analysis of difficult positions, all this FF vs Stockfish fuss is simply absurd.
SunriseK SunriseK 11/9/2019 01:36
Oh, and in my previous wishlist I forgot to mention the Larsen Opening, which is also missing from the suite.
I believe it should stay instead, at least with one peculiar variation.
sovaz1997 sovaz1997 11/8/2019 03:25
@langit7 Fa fritz based on Leela
langit7 langit7 11/8/2019 10:21
Leela ia just 2 years old girls, don't put to hard on her
CAIMANETERO CAIMANETERO 11/8/2019 08:46
my question is ...if i get this fat fritz can i use it online ...i mean...can i use it in playchess engines room???
SunriseK SunriseK 11/8/2019 12:52
@Albert Silver:
"The basic suite is comprised of 50 positions, which are designed to cover the basic range of openings, and opening structures, in proportions roughly comparable to their popularity in Mega"

Yes, I've read also your previous article 'Test your engine - the Silver Suite'; so I already knoq this point.
And of course (IMHO) your suite is much better than Nunn's one.
Still I feel unsatisfied: there are a lot of very interesting opening's variations that are missing! I believe a 100 initial position set (i.e. double size of your suite, thus getting a 200 games match between the engines, which should not be a problem at all) would be much more interesting and accurate.

"I have larger versions of the suite that are not published, also handmade, which do have the Budapest."

Happy to know that. And I would be inclined to believe that a handmade set would be even better than "blindly" select the openings according with the frequency they are played in Mega Database 2010.
For example, even if they are not played so often, I would like to see also variations from open games like:
- some Vienna Game variations! (not only the Vienna gambit that is already in your SilverGambitSuite, but for example the Fianchetto g3 variation and the 2 main variations)
- 2 or 3 Two Knights variations!
- at least 1 Bishop's Opening variation
- 1 Centre Game (not the Danish Gambit already present in your SilverGambitSuite; for example 3. Qxd4 or 3. Nf3)
- a couple of Four Knights variations!
- 1 more Petrov variation (for example the one with 3. d4)
One Berlin Wall (I know is boring, but it's among the most played variations and anyway engines can't get bored) :-)
About the semi - open games, Nimzowitsch Defence is missing.
About the closed games, London System and Trompowsky are missing.
Then, with 100 starting positions available, you would have around still 35 of them to add interesting variations to the openings already in.
RaoulBertorello RaoulBertorello 11/7/2019 08:57
If you are so sure about Fat Fritz superiority against Stockfish, why don’t you, Fat Fritz makers, take part into next TCEC competition ? That would be the best way to clear the field from any dispute about which engine is stronger. TCEC plays about thrice a year, and I think it’s very easy for professionals like you to join it: just think what amateurs like Stockfish keepers have been able to achieve so far. Any reason put forward to justify Fat Fritz absence from that contest would be taken as an admission of obvious inferiority.
Rusty Razor Rusty Razor 11/7/2019 07:47
A low blow & cheapskate marketing ploy for something 99% don't need, nor ever will. Like trying to sell snow to an Eskimo. Try harder please!
siddorm siddorm 11/7/2019 03:40
This is in my view the most misleading article I have ever seen on Chessbase. Shameful :(
ThunderWinkle ThunderWinkle 11/7/2019 02:28
xD RTX 2080. Y tho. Who are you people. I hope you then went on to run Battlefield V at 4k
sovaz1997 sovaz1997 11/7/2019 01:41
ChessBase again deceives people and uses the most dirty methods of making money.
Fat Fritz is Lc0 LOL...
physica physica 11/7/2019 12:10
Whoa!

To put this shortly: a ghetto version of AlphaZero-SF match. But a nice promo for FF.

SF getting only 16Mn/s on Ryzen 7 3700X (8 cores/16 threads) while various benchmarks almost double that! Seriously? And 32 GB RAM with SF is vastly overstated (impairs performance) with this hardware as shown i.e. at TCEC (circa 3 times faster vs. 16 Mn/s!?) and as basic knowledge of current status of alpha-beta engines.

P.S. Congrats for finding yet another setup to make NNs great again!
NYTed NYTed 11/7/2019 10:49
Is there a link to all 200 games that I am missing? The article does state, "Here are all the games and GM Elshan Moradiabadi's comments on moments he highlighted."
brabo_hf brabo_hf 11/7/2019 08:41
It is a pure coincidence but a month ago I got exactly the same score on my own new portable with Leela against Stockfish 10. I reported about it on https://www.chesspub.com/cgi-bin/chess/YaBB.pl?num=1561210232/68. I think not having to use any internetconnection for running an engine is an advantage as I always consider it annoying to set up a hotspot via the mobile when no wifi is available.
Scorpion29 Scorpion29 11/7/2019 07:47
Bullkitten if you can't accept the fact that given similar conditions Fat Fritz is already clearly stronger than fat Fritz then God Bless you! I can safely say that in your conditions as well Fat Fritz will not lose as badly as you claim. I believe this test was more fair than the Alpha Zero-Stockfish match. So there is no substance in your 'garbage' allegation.
Bullkitten Bullkitten 11/7/2019 07:23
These tests are very easy to manipulate. One engine uses CPU, the other GPU. I can give Fat Fritz a weak GPU and Stockfish a very powerful CPU, and Stockfish will crush Fat Fritz. Or vice versa. Stop comparing apples and oranges. Unless you use top of the line GPU vs top CPU, this is gargage.
Marselos Marselos 11/7/2019 06:59
I follow Chessbase.
I follow Chessbase from time of Fritz 1 on dos.
I don't make a substantial difference between Fritz or Stockfish.
Today the computer:
- perfectly analyzes the tactics,for tournaments games.
-It is ,practically, unbeatable.
-we are full of cheater...
Open Questions :
why do you want a better engine?
why create thousands and thousand of engine ?
Do you want to solve chess?
Why not look for the stupid "artificial intelligence" elsewhere ..
Albert Silver Albert Silver 11/7/2019 04:04
@Keshava - Stockfish 10 was released on Nov 30, 2018, so less than one year ago. While I have no doubt the development versions are stronger, they are also changing on an almost daily basis. When Stockfish 11 is released, it will be duly tested of course. Even Fat Fritz is still in development, and within weeks after its release you can expect a newer and stronger version to be available via a software update. Stockfish is not the only engine that improves... :-)
Albert Silver Albert Silver 11/7/2019 03:59
@sunrisek - The basic suite is comprised of 50 positions, which are designed to cover the basic range of openings, and opening structures, in proportions roughly comparable to their popularity in Mega. I have larger versions of the suite that are not published, also handmade, which do have the Budapest.
@Besler - The match uses 50 positions, all of which are played twice with colors reversed the second time. In other words, for every position Stockfish had white in one and black in the other, and same goes for Fat Fritz.
Keshava Keshava 11/7/2019 03:56
Stockfish 10 was released in 2018; it's almost two years old!
Stockfish has had a lot of updates since then.
That is why the Stockfish team don't compete with SF 10.
For strong rapid play please retest against this version:
231019 64 BMI2
That engine has been chewing up all other engines at the
c----.com/cccc
Only then you will know how strong Fat Fritz is
(assuming comparable hardware and optimum parameters for Stockfish).
besler besler 11/7/2019 02:53
Very interesting!
I feel like the following experiment may also help illuminate the fundamental differences in the performance of the two engines:
1. Take an instructive game won by FF.
2. Swap the color assignments of the engines (e.g. if FF was white, FF is now black).
3. Start from the final move of the game and see if SF plays the same move as FF (given a reasonable choice of dwell time).
4. Repeat step 3) while decrementing the move number until you find the move at which SF prefers a different move than FF.
5. Allow SF to finish the game against FF from that point onwards.
6. Merge both games (the original and “branched” version) into one pgn file and compare them.
7. Perform the same experiment for a game won by SF.
This could help pinpoint the exact moment at which a positional concept or tactical idea was found by one engine and not the other!
SunriseK SunriseK 11/7/2019 02:06
Interesting article, but... no Budapest Gambit in the suite? Not even in gambit suite??? :-o
And just 8 (if I'm not wrong) open games in the silver suite?
Rambus Rambus 11/7/2019 01:20
The sole purpose of this article seems to be to boost $ASLE$
Albert Silver Albert Silver 11/6/2019 11:55
@Classique - It doesn't matter if it is 16 million or 80. What matters is that the relative speed of the opponent is proportionate. In this case Fat Fritz is running much slower compared to TCEC, meaning that the TCEC speed is actually more advantageous to the neural network than this test. The hardware described is what Stockfish ran on: a Ryzen 7 3700X. I see now the text did not make this clear and will fix it.
Classique Classique 11/6/2019 11:54
Sixteen million nodes per second is very slow. Stockfish 10 can be seen, as I speak, on the TCEC website at over five times that speed.

We are told about the hardware Fat Fritz was run on, but not about the hardware Stockfish was run on. Fat Fritz may be stronger than Stockfish 10, but this redacted report doesn't demonstrate it.
Tom Box Tom Box 11/6/2019 10:29
In my opinion all the recent articles on Fat Frtiz seem to be motivated by commercial considerations and don't appear to be objective. The people at Stockfish should be be involved to guarantee a fair contest.
Albert Silver Albert Silver 11/6/2019 09:52
@OlivierEvan Ce n'est pas necessaire. Non seulement Windows sait administrer ce genre de chose, mais en plus, à l'installation de l'engine tu as la possibilité de specifier "Below normal priority" de manière à ne jamais laisser la machine congelé par manque de ressources.
arjo1982 arjo1982 11/6/2019 09:39
"We only know how strong AlphaZero was by virtue of its results against Stockfish 8, which it beat by a bit over 30 Elo on average in long games. In roughly similar conditions, the pre-release build of Fat Fritz scored nearly 90 Elo over it, using a varied set of balanced openings even in speed conditions that were a bit worse. While hardly proof, it does strongly suggest Fat Fritz is stronger than AlphaZero in equal conditions."

This doesnt proof or suggest anything... Even less if you use an openingbook
OlivierEvan OlivierEvan 11/6/2019 09:36
Bonjour CB !

Vous faites une terrible erreur avant que le match commence....Faire fonctionner SF avec :
??...Stockfish ran on 16 threads ??.

Vous devez laisser 1 thread pour la gestion du système d'exploitation, video, Ram....
Réessayez avec cette configuration et vous verrez la différence.

Olivier.
Albert Silver Albert Silver 11/6/2019 09:28
@Metaphysicianx - The proportion of speed is a factor, so if a faster computer were given to Stockfish, then a faster machine would also need to be provided to Fat Fritz. The PGNs of both matches will be provided in the second part tomorrow.
@Conwycastle - Yes, it was run on the same computer, though Fat Fritz only used two threads, not sixteen.
ConwyCastle ConwyCastle 11/6/2019 09:13
Just to be clear - was Stockfish using the same hardware?
Metaphysicianx Metaphysicianx 11/6/2019 09:12
I wonder how Stockfish 10 (or even better, a current development version) would have done on more than just 16 threads.
And may we have a pgn file of all the games please?
Thank you.
1