You want to see why human can NOT compete vs engines in Bullet Games?

[Chessqueen]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

[lkaufman]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

Obviously no human has a chance against a top engine in bullet chess without a handicap, that's been clear for many years. But it's not obvious to me which side to bet on if Nakamura played Komodo Dragon at 1' + 1" bullet with knight odds. Maybe a study of the centipawn errors in games he played at that level would be a guide as to whether they add up to the roughly 400 or so needed to lose a knight odds game, but then errors might be smaller at knight odds since he can avoid unclear lines.

[mwyoung]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

Obviously no human has a chance against a top engine in bullet chess without a handicap, that's been clear for many years. But it's not obvious to me which side to bet on if Nakamura played Komodo Dragon at 1' + 1" bullet with knight odds. Maybe a study of the centipawn errors in games he played at that level would be a guide as to whether they add up to the roughly 400 or so needed to lose a knight odds game, but then errors might be smaller at knight odds since he can avoid unclear lines.

I ran the Nakamura win in the latest Chess.com Speed Championship 2020. A few days ago. This will give you are start. Not all games are at 1+1. This is the results for the whole match.

Results.

Code: Select all

Hikaru:   56/98/65/41/36/13/76/44/56/2/16/58/32/20/40/36/1/56/24/12/47/39/31/28/64/39/12/13/10/9/4  => Average=0.37
LyonBeast:   55/100/73/51/60/15/22/56/65/3/15/24/56/44/22/42/3/21/76/47/9/42/38/33/8/107/48/33/10/75/67  => Average=0.42

Games Link.

[lkaufman]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

Obviously no human has a chance against a top engine in bullet chess without a handicap, that's been clear for many years. But it's not obvious to me which side to bet on if Nakamura played Komodo Dragon at 1' + 1" bullet with knight odds. Maybe a study of the centipawn errors in games he played at that level would be a guide as to whether they add up to the roughly 400 or so needed to lose a knight odds game, but then errors might be smaller at knight odds since he can avoid unclear lines.

I ran the Nakamura win in the latest Chess.com Speed Championship 2020. A few days ago. This will give you are start. Not all games are at 1+1. This is the results for the whole match.

Results.

Code: Select all

Hikaru:   56/98/65/41/36/13/76/44/56/2/16/58/32/20/40/36/1/56/24/12/47/39/31/28/64/39/12/13/10/9/4  => Average=0.37
LyonBeast:   55/100/73/51/60/15/22/56/65/3/15/24/56/44/22/42/3/21/76/47/9/42/38/33/8/107/48/33/10/75/67  => Average=0.42

Games Link.

Can you explain exactly what the string of numbers mean? Apparently each number is a measure of error in centipawns for each game, but they don't seem to be average (unsigned) errors, because some of the numbers are 1,2,3,3, and 4, and it doesn't seem likely that a human player would play an entire blitz or bullet game averaging just 1 to 4 centipawn errors. Maybe they are only counted as errors if they differ from the supposedly correct score by more than X? If they are centipawn errors, if a mate is missed, the error would be huge, so perhaps they are capped at some limit? Or maybe they are median errors, not average? Any details would be helpful.

[mwyoung]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

Obviously no human has a chance against a top engine in bullet chess without a handicap, that's been clear for many years. But it's not obvious to me which side to bet on if Nakamura played Komodo Dragon at 1' + 1" bullet with knight odds. Maybe a study of the centipawn errors in games he played at that level would be a guide as to whether they add up to the roughly 400 or so needed to lose a knight odds game, but then errors might be smaller at knight odds since he can avoid unclear lines.

I ran the Nakamura win in the latest Chess.com Speed Championship 2020. A few days ago. This will give you are start. Not all games are at 1+1. This is the results for the whole match.

Results.

Code: Select all

Hikaru:   56/98/65/41/36/13/76/44/56/2/16/58/32/20/40/36/1/56/24/12/47/39/31/28/64/39/12/13/10/9/4  => Average=0.37
LyonBeast:   55/100/73/51/60/15/22/56/65/3/15/24/56/44/22/42/3/21/76/47/9/42/38/33/8/107/48/33/10/75/67  => Average=0.42

Games Link.

Can you explain exactly what the string of numbers mean? Apparently each number is a measure of error in centipawns for each game, but they don't seem to be average (unsigned) errors, because some of the numbers are 1,2,3,3, and 4, and it doesn't seem likely that a human player would play an entire blitz or bullet game averaging just 1 to 4 centipawn errors. Maybe they are only counted as errors if they differ from the supposedly correct score by more than X? If they are centipawn errors, if a mate is missed, the error would be huge, so perhaps they are capped at some limit? Or maybe they are median errors, not average? Any details would be helpful.

Look at the games. The will give the results of each move, and the weighted value. Yes Naka did have some very low values. When using Stockfish in the analysis. But also some very high values. And a .37 overall score for the match at blitz and bullet time controls. And that is what you would expect at this time control over many games.

I like using Dragon much better when using centipawn analysis. I will run this match again using Dragon, and this will show us the difference.

Results for Long time controls. 60 to 30 for club players. And 30 to 15 for Strong GM's at long time controls. Not blitz and bullet.

60 to 30 is looking to be the Range for fast time controls. And worse for bullet time controls.

But the expected values can be discovered for these time controls by running many players and matches.

[Chessqueen]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

Obviously no human has a chance against a top engine in bullet chess without a handicap, that's been clear for many years. But it's not obvious to me which side to bet on if Nakamura played Komodo Dragon at 1' + 1" bullet with knight odds. Maybe a study of the centipawn errors in games he played at that level would be a guide as to whether they add up to the roughly 400 or so needed to lose a knight odds game, but then errors might be smaller at knight odds since he can avoid unclear lines.

I ran the Nakamura win in the latest Chess.com Speed Championship 2020. A few days ago. This will give you are start. Not all games are at 1+1. This is the results for the whole match.

Results.

Code: Select all

Hikaru:   56/98/65/41/36/13/76/44/56/2/16/58/32/20/40/36/1/56/24/12/47/39/31/28/64/39/12/13/10/9/4  => Average=0.37
LyonBeast:   55/100/73/51/60/15/22/56/65/3/15/24/56/44/22/42/3/21/76/47/9/42/38/33/8/107/48/33/10/75/67  => Average=0.42

Games Link.

Can you explain exactly what the string of numbers mean? Apparently each number is a measure of error in centipawns for each game, but they don't seem to be average (unsigned) errors, because some of the numbers are 1,2,3,3, and 4, and it doesn't seem likely that a human player would play an entire blitz or bullet game averaging just 1 to 4 centipawn errors. Maybe they are only counted as errors if they differ from the supposedly correct score by more than X? If they are centipawn errors, if a mate is missed, the error would be huge, so perhaps they are capped at some limit? Or maybe they are median errors, not average? Any details would be helpful.

Look at the games. The will give the results of each move, and the weighted value. Yes Naka did have some very low values. When using Stockfish in the analysis. But also some very high values. And a .37 overall score for the match at blitz and bullet time controls. And that is what you would expect at this time control over many games.

I like using Dragon much better when using centipawn analysis. I will run this match again using Dragon, and this will show us the difference.

Results for Long time controls. 60 to 30 for club players. And 30 to 15 for Strong GM's at long time controls. Not blitz and bullet.

60 to 30 is looking to be the Range for fast time controls. And worse for bullet time controls.

But the expected values can be discovered for these time controls by running many players and matches.

I noticed that as the Blitz games expand in time control beyond 5 Minutes Blitz MVL gets better results than Nakamura, therefore, the centipawns score should benefit MVL as the results from the speed championship indicate, but what does it really means and if you were to compare them to giving Knight Odds to Dragon in Bullet games, of course Carlsen and Nakamura would be the only one to be competitive or to have a chance, but I still favor Dragon unless Naka is given 1' + 3" or even 1' +5"

[mwyoung]

For those that want Andrews Tang vs Komodo match in Bullet, Komodo will simply destroy any human in Bullet chess, take a look at these game Nakamura vs Maxime Vachier-Lagrave and see if you can detect the errors all the blunders check it with centipawns score if you want https://en.chessbase.com/post/nakamura- ... ampionship

Andrews Tang versus the Next World Champion in Bullet

Obviously no human has a chance against a top engine in bullet chess without a handicap, that's been clear for many years. But it's not obvious to me which side to bet on if Nakamura played Komodo Dragon at 1' + 1" bullet with knight odds. Maybe a study of the centipawn errors in games he played at that level would be a guide as to whether they add up to the roughly 400 or so needed to lose a knight odds game, but then errors might be smaller at knight odds since he can avoid unclear lines.

I ran the Nakamura win in the latest Chess.com Speed Championship 2020. A few days ago. This will give you are start. Not all games are at 1+1. This is the results for the whole match.

Results.

Code: Select all

Hikaru:   56/98/65/41/36/13/76/44/56/2/16/58/32/20/40/36/1/56/24/12/47/39/31/28/64/39/12/13/10/9/4  => Average=0.37
LyonBeast:   55/100/73/51/60/15/22/56/65/3/15/24/56/44/22/42/3/21/76/47/9/42/38/33/8/107/48/33/10/75/67  => Average=0.42

Games Link.

Same games with Dragon as the analysis engine.

Step analysis

Code: Select all

Hikaru:   45  => Average=0.45
LyonBeast:   49  => Average=0.49
LyonBeast:   49/4  => Average=0.27
Hikaru:   45/4  => Average=0.25
Hikaru:   45/4/14  => Average=0.21
LyonBeast:   49/4/13  => Average=0.22
LyonBeast:   49/4/13/29  => Average=0.24
Hikaru:   45/4/14/60  => Average=0.30
Hikaru:   74/45/4/14/60  => Average=0.39
LyonBeast:   25/49/4/13/29  => Average=0.24
LyonBeast:   25/49/4/13/29/35  => Average=0.25
Hikaru:   74/45/4/14/60/34  => Average=0.38
Hikaru:   74/45/4/14/26/60/34  => Average=0.37
LyonBeast:   25/49/4/13/37/29/35  => Average=0.27
LyonBeast:   25/49/4/13/37/29/56/35  => Average=0.30
Hikaru:   74/45/4/14/26/60/21/34  => Average=0.35
Hikaru:   74/45/4/14/26/60/21/34/17  => Average=0.33
LyonBeast:   25/49/4/13/37/29/56/35/35  => Average=0.31
LyonBeast:   49/25/49/4/13/37/29/56/35/35  => Average=0.33
Hikaru:   45/74/45/4/14/26/60/21/34/17  => Average=0.35
Hikaru:   45/17/74/45/4/14/26/60/21/34/17  => Average=0.33
LyonBeast:   49/15/25/49/4/13/37/29/56/35/35  => Average=0.31
LyonBeast:   49/15/25/49/4/13/37/32/29/56/35/35  => Average=0.31
Hikaru:   45/17/74/45/4/14/26/34/60/21/34/17  => Average=0.33
Hikaru:   45/17/74/45/4/14/26/34/60/21/34/17/4  => Average=0.32
LyonBeast:   49/15/25/49/4/13/37/32/29/56/35/35/54  => Average=0.32
LyonBeast:   49/15/25/49/4/13/37/28/32/29/56/35/35/54  => Average=0.32
Hikaru:   45/17/74/45/4/14/26/38/34/60/21/34/17/4  => Average=0.32
Hikaru:   45/17/74/45/4/14/26/38/34/60/21/37/34/17/4  => Average=0.33
LyonBeast:   49/15/25/49/4/13/37/28/32/29/56/9/35/35/54  => Average=0.30
LyonBeast:   49/15/25/49/4/13/37/28/32/4/29/56/9/35/35/54  => Average=0.29
Hikaru:   45/17/74/45/4/14/26/38/34/4/60/21/37/34/17/4  => Average=0.31
Hikaru:   45/17/74/45/4/14/26/38/34/4/60/21/37/34/40/17/4  => Average=0.31
LyonBeast:   49/15/25/49/4/13/37/28/32/4/29/56/9/35/52/35/54  => Average=0.30
LyonBeast:   49/15/25/49/4/13/37/49/28/32/4/29/56/9/35/52/35/54  => Average=0.31
Hikaru:   45/17/74/45/4/14/26/13/38/34/4/60/21/37/34/40/17/4  => Average=0.30
Hikaru:   45/17/74/45/4/14/50/26/13/38/34/4/60/21/37/34/40/17/4  => Average=0.31
LyonBeast:   49/15/25/49/4/13/22/37/49/28/32/4/29/56/9/35/52/35/54  => Average=0.30
LyonBeast:   49/15/25/49/4/13/22/37/49/28/32/4/29/56/9/35/52/35/14/54  => Average=0.30
Hikaru:   45/17/74/45/4/14/50/26/13/38/34/4/60/21/37/34/40/17/9/4  => Average=0.30
Hikaru:   45/17/74/45/4/14/50/26/13/38/34/4/60/21/37/34/40/57/17/9/4  => Average=0.31
LyonBeast:   49/15/25/49/4/13/22/37/49/28/32/4/29/56/9/35/52/9/35/14/54  => Average=0.29
LyonBeast:   49/92/15/25/49/4/13/22/37/49/28/32/4/29/56/9/35/52/9/35/14/54  => Average=0.33
Hikaru:   45/89/17/74/45/4/14/50/26/13/38/34/4/60/21/37/34/40/57/17/9/4  => Average=0.35
Hikaru:   45/89/17/74/45/61/4/14/50/26/13/38/34/4/60/21/37/34/40/57/17/9/4  => Average=0.36
LyonBeast:   49/92/15/25/49/71/4/13/22/37/49/28/32/4/29/56/9/35/52/9/35/14/54  => Average=0.34
LyonBeast:   49/92/15/25/49/71/4/13/22/37/49/28/32/4/29/56/9/35/52/9/81/35/14/54  => Average=0.36
Hikaru:   45/89/17/74/45/61/4/14/50/26/13/38/34/4/60/21/37/34/40/57/32/17/9/4  => Average=0.36
Hikaru:   45/89/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/40/57/32/17/9/4  => Average=0.35
LyonBeast:   49/92/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/52/9/81/35/14/54  => Average=0.36
LyonBeast:   49/92/60/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/52/9/81/35/14/54  => Average=0.37
Hikaru:   45/89/41/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/40/57/32/17/9/4  => Average=0.35
Hikaru:   45/89/41/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/40/57/32/11/17/9/4  => Average=0.34
LyonBeast:   49/92/60/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/52/9/81/40/35/14/54  => Average=0.37
LyonBeast:   49/92/60/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/52/9/81/40/35/14/66/54  => Average=0.38
Hikaru:   45/89/41/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/40/57/32/11/17/9/7/4  => Average=0.34
Hikaru:   45/89/41/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/33/40/57/32/11/17/9/7/4  => Average=0.34
LyonBeast:   49/92/60/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/38/52/9/81/40/35/14/66/54  => Average=0.38
LyonBeast:   49/92/77/60/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/38/52/9/81/40/35/14/66/54  => Average=0.39
Hikaru:   45/89/65/41/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/33/40/57/32/11/17/9/7/4  => Average=0.35
Hikaru:   45/89/65/41/50/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/33/40/57/32/11/17/9/7/4  => Average=0.36
LyonBeast:   49/92/77/60/45/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/38/52/9/81/40/35/14/66/54  => Average=0.40

Done

Code: Select all

Hikaru:   45/89/65/41/50/17/74/45/61/4/14/50/26/13/38/34/4/60/21/9/37/34/33/40/57/32/11/17/9/7/4  => Average=0.36
LyonBeast:   49/92/77/60/45/15/25/49/71/4/13/22/37/49/28/32/4/29/56/33/9/35/38/52/9/81/40/35/14/66/54  => Average=0.40

Link to games with Dragon analysis.

[mwyoung]

I checked games of both Fischer and Carlsen at or near their peak. Playing both classical and blitz time controls.

Carlsen blitz results.

Code: Select all

Carlsen Magnus:   57/14/28/25/29/57/28/24/6/65/40/79/10/28/16/21/30/15/16/0  => Average=0.31

Fischer blitz results.

Code: Select all

Fischer Robert James:   86/101/31/32/44/37/23/24/33/79/48/31/38/50/67/30/17/76/19/11  => Average=0.46

Carlsen classic results.

Code: Select all

Carlsen Magnus:   3/15/64/8/66/8/3/54/22/13/22/12/11/2/7/2/10/5/12/4  => Average=0.18

Fischer classic results.

Code: Select all

31/29/14/29/86/9/10/18/12/5/16/9/12/15/12/20/15/12/19/3  => Average=0.20

Games
https://drive.google.com/drive/folders/ ... sp=sharing

[lkaufman]

I checked games of both Fischer and Carlsen at or near their peak. Playing both classical and blitz time controls.

Carlsen blitz results.

Code: Select all

Carlsen Magnus:   57/14/28/25/29/57/28/24/6/65/40/79/10/28/16/21/30/15/16/0  => Average=0.31

Fischer blitz results.

Code: Select all

Fischer Robert James:   86/101/31/32/44/37/23/24/33/79/48/31/38/50/67/30/17/76/19/11  => Average=0.46

Carlsen classic results.

Code: Select all

Carlsen Magnus:   3/15/64/8/66/8/3/54/22/13/22/12/11/2/7/2/10/5/12/4  => Average=0.18

Fischer classic results.

Code: Select all

31/29/14/29/86/9/10/18/12/5/16/9/12/15/12/20/15/12/19/3  => Average=0.20

Games
https://drive.google.com/drive/folders/ ... sp=sharing

One point to consider when looking at these average error numbers: Some portion of the average error is due to error by the analysis engine. I don't know the depth or time limit used for the analysis, but I imagine that it is a very short amount of time, so the engine error is not trivial. This doesn't much matter if we are just comparing one player to another, but to interpret the actual number, it does. For example, Fischer's average classic error is 0.20. On average, if the engine says 0.00 and Fischer's move is -0.20, then the truth is probably something like -.05, in between the two but much closer to the engine. So I think 0.05 is a good estimate of the amount to deduct from the averages to estimate the average "true" error. If so then the average "true" error in classic is 0.13 to 0.15 for Carlsen and Fischer, and 0.26 to 0.41 in blitz for same players, somewhat over double on average, which seems quite reasonable. Presumably Rapid (15' + 10") errors would be about halfway between these numbers. Based solely on this math even Carlsen would have no chance vs. top engine in blitz at knight odds and little chance even in Rapid (the errors would add up to way more than 400 cp for forty moves), but that's clearly wrong. Carlsen would just play safely to minimize the chance of big errors. Anyway it does appear that the errors are smaller with the top players now than fifty years ago, and that results and ratings do correlate with the errors reported. I suppose that if we knew what rating player averaged 0.31 error at classic time limit, we would be able to estimate the classic rating level of Carlsen's blitz play (for example).

[mwyoung]

I checked games of both Fischer and Carlsen at or near their peak. Playing both classical and blitz time controls.

Carlsen blitz results.

Code: Select all

Carlsen Magnus:   57/14/28/25/29/57/28/24/6/65/40/79/10/28/16/21/30/15/16/0  => Average=0.31

Fischer blitz results.

Code: Select all

Fischer Robert James:   86/101/31/32/44/37/23/24/33/79/48/31/38/50/67/30/17/76/19/11  => Average=0.46

Carlsen classic results.

Code: Select all

Carlsen Magnus:   3/15/64/8/66/8/3/54/22/13/22/12/11/2/7/2/10/5/12/4  => Average=0.18

Fischer classic results.

Code: Select all

31/29/14/29/86/9/10/18/12/5/16/9/12/15/12/20/15/12/19/3  => Average=0.20

Games
https://drive.google.com/drive/folders/ ... sp=sharing

One point to consider when looking at these average error numbers: Some portion of the average error is due to error by the analysis engine. I don't know the depth or time limit used for the analysis, but I imagine that it is a very short amount of time, so the engine error is not trivial. This doesn't much matter if we are just comparing one player to another, but to interpret the actual number, it does. For example, Fischer's average classic error is 0.20. On average, if the engine says 0.00 and Fischer's move is -0.20, then the truth is probably something like -.05, in between the two but much closer to the engine. So I think 0.05 is a good estimate of the amount to deduct from the averages to estimate the average "true" error. If so then the average "true" error in classic is 0.13 to 0.15 for Carlsen and Fischer, and 0.26 to 0.41 in blitz for same players, somewhat over double on average, which seems quite reasonable. Presumably Rapid (15' + 10") errors would be about halfway between these numbers. Based solely on this math even Carlsen would have no chance vs. top engine in blitz at knight odds and little chance even in Rapid (the errors would add up to way more than 400 cp for forty moves), but that's clearly wrong. Carlsen would just play safely to minimize the chance of big errors. Anyway it does appear that the errors are smaller with the top players now than fifty years ago, and that results and ratings do correlate with the errors reported. I suppose that if we knew what rating player averaged 0.31 error at classic time limit, we would be able to estimate the classic rating level of Carlsen's blitz play (for example).

I am using 4s a move, and this is longer then recommended for a fast system.

Finding the Classic Elo of .31 should not too be hard. As I can filter games by Fide Elo. I will start with 2500 to 2525. Low end GM, and adjust from there.