Opponents for handicap match dragon-humans

[lkaufman]

I am not sure if the humans are going to do better in the second round.

Part of them play 2 games so they may be tired from the first game.
One of them is going to play active chess in an earlier tournament at the same day so I doubt he is going to have a lot of time to learn from the first game.

I see that only 3 of the 9 players were born before 2002, so we're mostly playing teenagers, who are likely to be under-rated. I hadn't taken that into consideration, probably the average actual strength of the field will be close to 2200 FIDE. Especially since during the Pandemic, many young players played almost exclusively online, getting stronger while their FIDE ratings remained stagnant. It would be interesting to look up the chess.com (and/or LiChess) Rapid ratings of the players, if they have any. Chess.com Rapid ratings run close to FIDE standard level on average, LiChess Rapid ratings are more than a class inflated relative to FIDE.

[lkaufman]

Unfortunately the strongest player IM Ariel Erenberg will not be able to participate so we will have only knight handicap games and not knight for pawn f7 handicap games.

I expect other games with full knight odds to be played 14:00 pm and 17:30 pm israel time.
Total number of games I expect is 13 games.

FM Ori Kochavi fide rating 2365(first game)
https://ratings.fide.com/profile/2816873

FM guy Levin fide rating 2354(only second game)
https://ratings.fide.com/profile/2824213

Nadav Zemach fide rating 2238(2 games)
https://ratings.fide.com/profile/2820927

WFM Michelle Katkov fide rating 2208(only first game)
https://ratings.fide.com/profile/2816393

FM Pavel Chernomordik fide rating fide rating 2178(2 games)
https://ratings.fide.com/profile/2804530

Guy Gov fide rating 2149(2 games)
https://ratings.fide.com/profile/2820200

Itay Sitbon(2 games) fide rating 2062
https://ratings.fide.com/profile/2820706

Alon Cohen(second game) fide rating 2016
https://ratings.fide.com/profile/2801809

Oded Ross(second game does not have fide rating except blitz but I believe at the level of fide rating 2000 at long time control)
https://ratings.fide.com/profile/2811189

Average FIDE rating (incl. est 2000 for last player) 2169. This is about 250 elo below the performance at knight odds in Rapid for Dragon; here the humans will be playing classical, which may raise their level by nearly a class, and with the simul handicap thrown in it should really be a tossup. The humans might do better in the second round, since although they will individually switch knights between games, they may learn something from the games of the other players at the same knight handicap.

In the first session Dragon won two games, drew one, and lost three. Interesting was the fact that all of its wins and draw were with g1 odds, while it lost all three with b1 odds. It sure is starting to look like if given the choice, White should remove g1 knight.

[Uri Blass]

link for the games
https://www.chess.com/member/playkomodo

Humans won 6 games lost 3 games and drew 4 games.

KomTest1 is guy gov
komTest2 is guy Levin.

players above 2300 fide rating played one game and won their games

komtest2 (Guy Levin) and oriking(Ori Kochavi)

[lkaufman]

Unfortunately the strongest player IM Ariel Erenberg will not be able to participate so we will have only knight handicap games and not knight for pawn f7 handicap games.

I expect other games with full knight odds to be played 14:00 pm and 17:30 pm israel time.
Total number of games I expect is 13 games.

FM Ori Kochavi fide rating 2365(first game)
https://ratings.fide.com/profile/2816873

FM guy Levin fide rating 2354(only second game)
https://ratings.fide.com/profile/2824213

Nadav Zemach fide rating 2238(2 games)
https://ratings.fide.com/profile/2820927

WFM Michelle Katkov fide rating 2208(only first game)
https://ratings.fide.com/profile/2816393

FM Pavel Chernomordik fide rating fide rating 2178(2 games)
https://ratings.fide.com/profile/2804530

Guy Gov fide rating 2149(2 games)
https://ratings.fide.com/profile/2820200

Itay Sitbon(2 games) fide rating 2062
https://ratings.fide.com/profile/2820706

Alon Cohen(second game) fide rating 2016
https://ratings.fide.com/profile/2801809

Oded Ross(second game does not have fide rating except blitz but I believe at the level of fide rating 2000 at long time control)
https://ratings.fide.com/profile/2811189

So the final result was a comfortable win for the human side, six human wins, three dragon wins, four draws, so 8 to 5. The three FMs provided the margin of victory by scoring 3.5 out of 4 games. But the results were really remarkable in that they were drastically different depending on which knight was given! With the b1 knight removed, White scored merely a single draw out of six games against a 2164 field for a pathetic 1747 performance rating. But giving the g1 knight, White scored a remarkable 3 wins, 1 loss, and 3 draws against a 2172 field for an excellent 2274 performance! Of course, samples of 6 to 7 games are very small, but surely a difference in performance of 527 elo cannot be written off as sample error. With the b1 knight missing, White never got any compensation for the knight out of the opening, whereas with g1 off White usually managed to get at least a preferable position if you don't count the pieces! The difference is that with b1 gone, White shouldn't open 1.e4 due to 1...d5! 2.exd5 Qxd5 when 3.Nc3 is illegal. So he has to settle for a quiet move like 1.Nf3. But with g1 gone there is no drawback to 1.e4, which at least gives White some hope for a fight. In past matches with IMs and GMs it hasn't been so clear, although I've generally preferred the positions White gets with g1 missing. But after this event, there can be no doubt; White should always choose to remove the g1 knight given the choice, which per tradition is White's prerogative. So knight odds remains viable for future events, but it seems it needs to be the g1 knight removed.

[Uri Blass]

I think dragon did a mistake by allowing black to trade queens in the only game it lost with knight g1 handicap
18.Bc3 was probably better.

Maybe Guy Levin could win in every case but I see no reason to allow the opponent to trade queens so easily
Stockfish suggests 18.Bc3

[pgn][Event "Live Chess - Odds Chess"]
[Site "Chess.com"]
[Date "2022.05.07"]
[Round "?"]
[White "PlayKomodo"]
[Black "KomTest2"]
[Result "0-1"]
[BlackElo "300"]
[WhiteElo "3400"]
[TimeControl "5400+30"]
[SetUp "1"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[Termination "normal"]
[PlyCount "168"]
[WhiteType "human"]
[BlackType "human"]

1. e4 c5 2. Nc3 Nc6 3. Bb5 a6 4. Bxc6 dxc6 5. Qh5 Qd6 6. b3 Nf6 7. Qe2 e5
8. O-O Be7 9. a4 a5 10. d3 O-O 11. Bd2 b6 12. Nd1 Bg4 13. f3 Be6 14. Ne3
Nh5 15. g3 Bg5 16. Rf2 Bxe3 17. Qxe3 Nf6 18. Rff1 Qd4 19. h4 Nd7 20. g4 f6
21. Qf2 Qxf2+ 22. Kxf2 Nb8 23. Ke3 Na6 24. Rab1 Rfd8 25. Rb2 Nb4 26. Rfb1
Rd7 27. Ke2 Kf7 28. h5 g6 29. hxg6+ hxg6 30. Rh1 Kg8 31. g5 fxg5 32. Bxg5
Rf8 33. Be3 Rh7 34. Rxh7 Kxh7 35. Bg1 g5 36. c3 Na6 37. Ke3 Kg6 38. Bh2 Kf6
39. Ke2 Rh8 40. Kf1 Rh7 41. Bg1 Rh3 42. Kg2 g4 43. fxg4 Rxd3 44. Rf2+ Kg5
45. Bh2 Rxc3 46. Bxe5 Rxb3 47. Rf6 Bxg4 48. Rf8 Be6 49. Re8 Bf7 50. Rf8 Kg6
51. Rd8 c4 52. Bf4 Nc5 53. Rd6+ Kh7 54. Rh6+ Kg7 55. Rxc6 Nxe4 56. Be5+ Kh7
57. Rc8 c3 58. Kf3 Nc5 59. Ke3 c2+ 60. Kd2 c1=Q+ 61. Kxc1 Nd3+ 62. Kd2 Nxe5
63. Rb8 Ra3 64. Rb7 Kg7 65. Kc2 Nc4 66. Rc7 Kf6 67. Rc8 Bg6+ 68. Kd1 Bd3
69. Ke1 Ra2 70. Rd8 Ne5 71. Kd1 Kf5 72. Rxd3 Ra1+ 73. Ke2 Nxd3 74. Kxd3
Rxa4 75. Ke3 Rc4 76. Kd3 b5 77. Kd2 a4 78. Kd1 a3 79. Kd2 a2 80. Ke2 a1=Q
81. Kf2 Qb2+ 82. Ke3 Qb3+ 83. Kf2 Rc2+ 84. Ke1 Qb1# 0-1
[/pgn]

[lkaufman]

I think dragon did a mistake by allowing black to trade queens in the only game it lost with knight g1 handicap
18.Bc3 was probably better.

Maybe Guy Levin could win in every case but I see no reason to allow the opponent to trade queens so easily
Stockfish suggests 18.Bc3

[pgn][Event "Live Chess - Odds Chess"]
[Site "Chess.com"]
[Date "2022.05.07"]
[Round "?"]
[White "PlayKomodo"]
[Black "KomTest2"]
[Result "0-1"]
[BlackElo "300"]
[WhiteElo "3400"]
[TimeControl "5400+30"]
[SetUp "1"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[Termination "normal"]
[PlyCount "168"]
[WhiteType "human"]
[BlackType "human"]

1. e4 c5 2. Nc3 Nc6 3. Bb5 a6 4. Bxc6 dxc6 5. Qh5 Qd6 6. b3 Nf6 7. Qe2 e5
8. O-O Be7 9. a4 a5 10. d3 O-O 11. Bd2 b6 12. Nd1 Bg4 13. f3 Be6 14. Ne3
Nh5 15. g3 Bg5 16. Rf2 Bxe3 17. Qxe3 Nf6 18. Rff1 Qd4 19. h4 Nd7 20. g4 f6
21. Qf2 Qxf2+ 22. Kxf2 Nb8 23. Ke3 Na6 24. Rab1 Rfd8 25. Rb2 Nb4 26. Rfb1
Rd7 27. Ke2 Kf7 28. h5 g6 29. hxg6+ hxg6 30. Rh1 Kg8 31. g5 fxg5 32. Bxg5
Rf8 33. Be3 Rh7 34. Rxh7 Kxh7 35. Bg1 g5 36. c3 Na6 37. Ke3 Kg6 38. Bh2 Kf6
39. Ke2 Rh8 40. Kf1 Rh7 41. Bg1 Rh3 42. Kg2 g4 43. fxg4 Rxd3 44. Rf2+ Kg5
45. Bh2 Rxc3 46. Bxe5 Rxb3 47. Rf6 Bxg4 48. Rf8 Be6 49. Re8 Bf7 50. Rf8 Kg6
51. Rd8 c4 52. Bf4 Nc5 53. Rd6+ Kh7 54. Rh6+ Kg7 55. Rxc6 Nxe4 56. Be5+ Kh7
57. Rc8 c3 58. Kf3 Nc5 59. Ke3 c2+ 60. Kd2 c1=Q+ 61. Kxc1 Nd3+ 62. Kd2 Nxe5
63. Rb8 Ra3 64. Rb7 Kg7 65. Kc2 Nc4 66. Rc7 Kf6 67. Rc8 Bg6+ 68. Kd1 Bd3
69. Ke1 Ra2 70. Rd8 Ne5 71. Kd1 Kf5 72. Rxd3 Ra1+ 73. Ke2 Nxd3 74. Kxd3
Rxa4 75. Ke3 Rc4 76. Kd3 b5 77. Kd2 a4 78. Kd1 a3 79. Kd2 a2 80. Ke2 a1=Q
81. Kf2 Qb2+ 82. Ke3 Qb3+ 83. Kf2 Rc2+ 84. Ke1 Qb1# 0-1
[/pgn]

Well, I agree with you, but Stockfish 15 doesn't choose to trade the queens by 18...Qd4 as Black after the dubious 18.Rff1? so it apparently is choosing Bc3 for some other reason. It seems that these NN engines don't evaluate piece-down positions very well, since they are all losing anyway, so they don't fully appreciate the principle of "don't trade pieces when down in material". Contempt helps to solve this, but apparently not enough. Something we can work on, I've noticed other examples myself.

[lkaufman]

Unfortunately the strongest player IM Ariel Erenberg will not be able to participate so we will have only knight handicap games and not knight for pawn f7 handicap games.

I expect other games with full knight odds to be played 14:00 pm and 17:30 pm israel time.
Total number of games I expect is 13 games.

FM Ori Kochavi fide rating 2365(first game)
https://ratings.fide.com/profile/2816873

FM guy Levin fide rating 2354(only second game)
https://ratings.fide.com/profile/2824213

Nadav Zemach fide rating 2238(2 games)
https://ratings.fide.com/profile/2820927

WFM Michelle Katkov fide rating 2208(only first game)
https://ratings.fide.com/profile/2816393

FM Pavel Chernomordik fide rating fide rating 2178(2 games)
https://ratings.fide.com/profile/2804530

Guy Gov fide rating 2149(2 games)
https://ratings.fide.com/profile/2820200

Itay Sitbon(2 games) fide rating 2062
https://ratings.fide.com/profile/2820706

Alon Cohen(second game) fide rating 2016
https://ratings.fide.com/profile/2801809

Oded Ross(second game does not have fide rating except blitz but I believe at the level of fide rating 2000 at long time control)
https://ratings.fide.com/profile/2811189

So the final result was a comfortable win for the human side, six human wins, three dragon wins, four draws, so 8 to 5. The three FMs provided the margin of victory by scoring 3.5 out of 4 games. But the results were really remarkable in that they were drastically different depending on which knight was given! With the b1 knight removed, White scored merely a single draw out of six games against a 2164 field for a pathetic 1747 performance rating. But giving the g1 knight, White scored a remarkable 3 wins, 1 loss, and 3 draws against a 2172 field for an excellent 2274 performance! Of course, samples of 6 to 7 games are very small, but surely a difference in performance of 527 elo cannot be written off as sample error. With the b1 knight missing, White never got any compensation for the knight out of the opening, whereas with g1 off White usually managed to get at least a preferable position if you don't count the pieces! The difference is that with b1 gone, White shouldn't open 1.e4 due to 1...d5! 2.exd5 Qxd5 when 3.Nc3 is illegal. So he has to settle for a quiet move like 1.Nf3. But with g1 gone there is no drawback to 1.e4, which at least gives White some hope for a fight. In past matches with IMs and GMs it hasn't been so clear, although I've generally preferred the positions White gets with g1 missing. But after this event, there can be no doubt; White should always choose to remove the g1 knight given the choice, which per tradition is White's prerogative. So knight odds remains viable for future events, but it seems it needs to be the g1 knight removed.

Actually, reviewing the 21 Rapid games played by Dragon giving knight odds to GMs, in those games Dragon performed 160 elo better giving g1 than giving b1 knight, so combined with this event we have 34 games with a performance advantage for g1 odds of 300 elo! That's really shocking. I'm sure that part of the 300 elo is sample error, but it looks clear that "knight odds" should mean "g1 odds". Although the majority of the historical games at knight odds were with b1 off, I suspect that this was either because most players didn't know about 1e4 d5 (at b1 odds) or because there was some agreement (explicit or implied) that 1.e4 should be met by 1...e5. After Morphy's playing career, g1 odds became more common.

[Uri Blass]

I think dragon did a mistake by allowing black to trade queens in the only game it lost with knight g1 handicap
18.Bc3 was probably better.

Maybe Guy Levin could win in every case but I see no reason to allow the opponent to trade queens so easily
Stockfish suggests 18.Bc3

[pgn][Event "Live Chess - Odds Chess"]
[Site "Chess.com"]
[Date "2022.05.07"]
[Round "?"]
[White "PlayKomodo"]
[Black "KomTest2"]
[Result "0-1"]
[BlackElo "300"]
[WhiteElo "3400"]
[TimeControl "5400+30"]
[SetUp "1"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[Termination "normal"]
[PlyCount "168"]
[WhiteType "human"]
[BlackType "human"]

1. e4 c5 2. Nc3 Nc6 3. Bb5 a6 4. Bxc6 dxc6 5. Qh5 Qd6 6. b3 Nf6 7. Qe2 e5
8. O-O Be7 9. a4 a5 10. d3 O-O 11. Bd2 b6 12. Nd1 Bg4 13. f3 Be6 14. Ne3
Nh5 15. g3 Bg5 16. Rf2 Bxe3 17. Qxe3 Nf6 18. Rff1 Qd4 19. h4 Nd7 20. g4 f6
21. Qf2 Qxf2+ 22. Kxf2 Nb8 23. Ke3 Na6 24. Rab1 Rfd8 25. Rb2 Nb4 26. Rfb1
Rd7 27. Ke2 Kf7 28. h5 g6 29. hxg6+ hxg6 30. Rh1 Kg8 31. g5 fxg5 32. Bxg5
Rf8 33. Be3 Rh7 34. Rxh7 Kxh7 35. Bg1 g5 36. c3 Na6 37. Ke3 Kg6 38. Bh2 Kf6
39. Ke2 Rh8 40. Kf1 Rh7 41. Bg1 Rh3 42. Kg2 g4 43. fxg4 Rxd3 44. Rf2+ Kg5
45. Bh2 Rxc3 46. Bxe5 Rxb3 47. Rf6 Bxg4 48. Rf8 Be6 49. Re8 Bf7 50. Rf8 Kg6
51. Rd8 c4 52. Bf4 Nc5 53. Rd6+ Kh7 54. Rh6+ Kg7 55. Rxc6 Nxe4 56. Be5+ Kh7
57. Rc8 c3 58. Kf3 Nc5 59. Ke3 c2+ 60. Kd2 c1=Q+ 61. Kxc1 Nd3+ 62. Kd2 Nxe5
63. Rb8 Ra3 64. Rb7 Kg7 65. Kc2 Nc4 66. Rc7 Kf6 67. Rc8 Bg6+ 68. Kd1 Bd3
69. Ke1 Ra2 70. Rd8 Ne5 71. Kd1 Kf5 72. Rxd3 Ra1+ 73. Ke2 Nxd3 74. Kxd3
Rxa4 75. Ke3 Rc4 76. Kd3 b5 77. Kd2 a4 78. Kd1 a3 79. Kd2 a2 80. Ke2 a1=Q
81. Kf2 Qb2+ 82. Ke3 Qb3+ 83. Kf2 Rc2+ 84. Ke1 Qb1# 0-1
[/pgn]

Well, I agree with you, but Stockfish 15 doesn't choose to trade the queens by 18...Qd4 as Black after the dubious 18.Rff1? so it apparently is choosing Bc3 for some other reason. It seems that these NN engines don't evaluate piece-down positions very well, since they are all losing anyway, so they don't fully appreciate the principle of "don't trade pieces when down in material". Contempt helps to solve this, but apparently not enough. Something we can work on, I've noticed other examples myself.

If the NN evaluation is not good in some winning positions then it is better not to use NN in these positions.

In other words you can use NN as long as the NN evaluation is close enough to draw but when it is not close enough to draw to use a different evaluation.

[Fritz 0]

I'm impressed how much difference a time control makes. As I remember, GM Ben Finegold (and you are not!) struggled at 15+10. Yesterday much lower rated players scored equally well, or maybe even better, at 90+30.

[lkaufman]

I think dragon did a mistake by allowing black to trade queens in the only game it lost with knight g1 handicap
18.Bc3 was probably better.

Maybe Guy Levin could win in every case but I see no reason to allow the opponent to trade queens so easily
Stockfish suggests 18.Bc3

[pgn][Event "Live Chess - Odds Chess"]
[Site "Chess.com"]
[Date "2022.05.07"]
[Round "?"]
[White "PlayKomodo"]
[Black "KomTest2"]
[Result "0-1"]
[BlackElo "300"]
[WhiteElo "3400"]
[TimeControl "5400+30"]
[SetUp "1"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKB1R w KQkq - 0 1"]
[Termination "normal"]
[PlyCount "168"]
[WhiteType "human"]
[BlackType "human"]

1. e4 c5 2. Nc3 Nc6 3. Bb5 a6 4. Bxc6 dxc6 5. Qh5 Qd6 6. b3 Nf6 7. Qe2 e5
8. O-O Be7 9. a4 a5 10. d3 O-O 11. Bd2 b6 12. Nd1 Bg4 13. f3 Be6 14. Ne3
Nh5 15. g3 Bg5 16. Rf2 Bxe3 17. Qxe3 Nf6 18. Rff1 Qd4 19. h4 Nd7 20. g4 f6
21. Qf2 Qxf2+ 22. Kxf2 Nb8 23. Ke3 Na6 24. Rab1 Rfd8 25. Rb2 Nb4 26. Rfb1
Rd7 27. Ke2 Kf7 28. h5 g6 29. hxg6+ hxg6 30. Rh1 Kg8 31. g5 fxg5 32. Bxg5
Rf8 33. Be3 Rh7 34. Rxh7 Kxh7 35. Bg1 g5 36. c3 Na6 37. Ke3 Kg6 38. Bh2 Kf6
39. Ke2 Rh8 40. Kf1 Rh7 41. Bg1 Rh3 42. Kg2 g4 43. fxg4 Rxd3 44. Rf2+ Kg5
45. Bh2 Rxc3 46. Bxe5 Rxb3 47. Rf6 Bxg4 48. Rf8 Be6 49. Re8 Bf7 50. Rf8 Kg6
51. Rd8 c4 52. Bf4 Nc5 53. Rd6+ Kh7 54. Rh6+ Kg7 55. Rxc6 Nxe4 56. Be5+ Kh7
57. Rc8 c3 58. Kf3 Nc5 59. Ke3 c2+ 60. Kd2 c1=Q+ 61. Kxc1 Nd3+ 62. Kd2 Nxe5
63. Rb8 Ra3 64. Rb7 Kg7 65. Kc2 Nc4 66. Rc7 Kf6 67. Rc8 Bg6+ 68. Kd1 Bd3
69. Ke1 Ra2 70. Rd8 Ne5 71. Kd1 Kf5 72. Rxd3 Ra1+ 73. Ke2 Nxd3 74. Kxd3
Rxa4 75. Ke3 Rc4 76. Kd3 b5 77. Kd2 a4 78. Kd1 a3 79. Kd2 a2 80. Ke2 a1=Q
81. Kf2 Qb2+ 82. Ke3 Qb3+ 83. Kf2 Rc2+ 84. Ke1 Qb1# 0-1
[/pgn]

Well, I agree with you, but Stockfish 15 doesn't choose to trade the queens by 18...Qd4 as Black after the dubious 18.Rff1? so it apparently is choosing Bc3 for some other reason. It seems that these NN engines don't evaluate piece-down positions very well, since they are all losing anyway, so they don't fully appreciate the principle of "don't trade pieces when down in material". Contempt helps to solve this, but apparently not enough. Something we can work on, I've noticed other examples myself.

If the NN evaluation is not good in some winning positions then it is better not to use NN in these positions.

In other words you can use NN as long as the NN evaluation is close enough to draw but when it is not close enough to draw to use a different evaluation.

I think that NN eval is still better in general than heuristic eval even when a piece down, but it doesn't seem to understand the principle about simplifying when ahead. Dragon has performed much better than Komodo (pre-Dragon, pre-NNUE) in giving knight odds, so simply turning off NNUE when a piece down would almost surely make it play worse on average.