GM Awonder Liang vs Leela Knight Odds Rapid match

[lkaufman]

Why are odds matches played with Leela? Are Komodo or Stockfish not suitable?

Expanding a little bit on Uri's answer Leelaknightodds is currently using a specialized network tuned on knight odds games which has shown improvements at odds over the previous use of wdl contempt (itself significantly better than stockfish or komodo at knight odds). This is better at odds because it leads to Leela playing moves which maximize expected score against a human strength opponent rather than moves best against engine level opponents. Uri has provided an example position showing this, you can also look at some of its games on lichess to get a feel for its playing style and see how well it is doing in practice. One final note Leela currently has to limit search in order to maintain this style of play.

Komodo Dragon was the first engine to train on knight odds, but it was only a tiny percentage of the training mixed with the normal training, so not nearly as effective as what Leela is now doing. It was sufficient to reach a 33% score in 15'10" games against "Threshold" GMs (rated near the 2500 threshold), which at the time seemed amazing. When we stopped including odds positions in training, results in odds play dropped off (in testing vs engines) so we had to use the older nets for odds matches. But a net fully trained to play the odds positions is much better of course. Also Komodo used a mixture of MCTS and alpha-beta, while Leela doesn't use alpha-beta at all, which is probably the right decision for odds play.

[cbash]

Komodo Dragon was the first engine to train on knight odds, but it was only a tiny percentage of the training mixed with the normal training, so not nearly as effective as what Leela is now doing. It was sufficient to reach a 33% score in 15'10" games against "Threshold" GMs (rated near the 2500 threshold), which at the time seemed amazing. When we stopped including odds positions in training, results in odds play dropped off (in testing vs engines) so we had to use the older nets for odds matches. But a net fully trained to play the odds positions is much better of course. Also Komodo used a mixture of MCTS and alpha-beta, while Leela doesn't use alpha-beta at all, which is probably the right decision for odds play.

Wow, that's really interesting. Thanks for history of engine odds play.

[lkaufman]

I just discovered that a couple days ago, the LeelaQueenOdds bot played 13 bullet games (1'0") against Wei Yi, the number 1 Chinese GM and FIDE World number 5 Rapid player! I had seen the games but didn't recognize the username LoveVae1999, which just said Chinese GM rated 3130 Lichess bullet. But I noticed today that he uses the same username on chess.com, where he shows his name (I do the same). The 1999 is his birth year. So no doubt that it was him. The strongest player ever to play any of the odds bots on LiChess without hiding his identity. The result was 9 wins for Leela (playing White with no queen), 2 wins for Wei Yi, and 2 draws. Of course bullet chess is nothing like Rapid chess, and queen odds is way more than double knight odds, so it's not too relevant for our upcoming match with Liang, but nice to see such a great player willing to play the bot on record.

[lkaufman]

Match set to start in half an hour. Watch Matthew Sadler's Silicon Road youtube channel for coverage, also perhaps "Chess-Network".

[lkaufman]

So the final result was six wins for Leela, seven draws, and one win for Awonder. However that's a bit misleading; Awonder won by 2.5 to 1.5 playing Rapid (1.5 to 0.5 at 10'5" and two draws at 8'3"), but then switched to blitz (3'2") and didn't win any games (4 draws, six losses). So we still can't say that we won a "Rapid" match against a top-fifty GM. The games were quite exciting and GM Matthew Sadler did a great job of commenting them live on his Silicon Road channel. Awonder did slightly better with b1 odds than with g1 odds, getting his single win at b1 odds, while scoring half a point less in total with g1 odds, which is generally thought to be the more difficult challenge for the odds receiver. He indicated a willingness to play a re-match, schedule permitting, presumably limited to Rapid only. The next scheduled match, in late January, is against GM Joel Benjamin, playing at a "standard" time control of one hour plus 30", which looks like quite a challenge for Leela.

[Father]

So the final result was six wins for Leela, seven draws, and one win for Awonder. However that's a bit misleading; Awonder won by 2.5 to 1.5 playing Rapid (1.5 to 0.5 at 10'5" and two draws at 8'3"), but then switched to blitz (3'2") and didn't win any games (4 draws, six losses). So we still can't say that we won a "Rapid" match against a top-fifty GM. The games were quite exciting and GM Matthew Sadler did a great job of commenting them live on his Silicon Road channel. Awonder did slightly better with b1 odds than with g1 odds, getting his single win at b1 odds, while scoring half a point less in total with g1 odds, which is generally thought to be the more difficult challenge for the odds receiver. He indicated a willingness to play a re-match, schedule permitting, presumably limited to Rapid only. The next scheduled match, in late January, is against GM Joel Benjamin, playing at a "standard" time control of one hour plus 30", which looks like quite a challenge for Leela.

Good evening Mr. Larry Kaufman.
Congratulations to all. To GM AWonder Liang, to you Mr. Larry Kaufman and the Leela team. Between all of you, you are making a great contribution, advancement and development to chess, machines, science and of course, to human beings.
Thank you all.

[lkaufman]

So the final result was six wins for Leela, seven draws, and one win for Awonder. However that's a bit misleading; Awonder won by 2.5 to 1.5 playing Rapid (1.5 to 0.5 at 10'5" and two draws at 8'3"), but then switched to blitz (3'2") and didn't win any games (4 draws, six losses). So we still can't say that we won a "Rapid" match against a top-fifty GM. The games were quite exciting and GM Matthew Sadler did a great job of commenting them live on his Silicon Road channel. Awonder did slightly better with b1 odds than with g1 odds, getting his single win at b1 odds, while scoring half a point less in total with g1 odds, which is generally thought to be the more difficult challenge for the odds receiver. He indicated a willingness to play a re-match, schedule permitting, presumably limited to Rapid only. The next scheduled match, in late January, is against GM Joel Benjamin, playing at a "standard" time control of one hour plus 30", which looks like quite a challenge for Leela.

Good evening Mr. Larry Kaufman.
Congratulations to all. To GM AWonder Liang, to you Mr. Larry Kaufman and the Leela team. Between all of you, you are making a great contribution, advancement and development to chess, machines, science and of course, to human beings.
Thank you all.

Tonite, after the match, we had the most incredible result yet. A player using the handle qwi8, rated 3009 bullet and 2764 blitz, tried playing 1'0" bullet chess with Leela at rook odds and as you might expect lost every game. But then he switched to 1' + 2" (which is almost 3'0" blitz), and took White every game (so rook and move odds), and still lost every single game, eight in a row! His ratings are well within the normal range for grandmasters, so this is really incredible. Both the knight odds bot and the rook odds bot were updated yesterday to use a new net that was trained on both knight and rook odds games simultaneously, which seems to be stronger than the old nets specialized for each of the odds. In fact so far the rook odds bot hasn't lost a single game to anyone since the upgrade. I was hoping we would eventually get a ladder for rook odds like the queen odds ladder, but if it is essentially unbeatable by humans in blitz that may not be viable. I could never have imagined having this problem!

[Father]

So the final result was six wins for Leela, seven draws, and one win for Awonder. However that's a bit misleading; Awonder won by 2.5 to 1.5 playing Rapid (1.5 to 0.5 at 10'5" and two draws at 8'3"), but then switched to blitz (3'2") and didn't win any games (4 draws, six losses). So we still can't say that we won a "Rapid" match against a top-fifty GM. The games were quite exciting and GM Matthew Sadler did a great job of commenting them live on his Silicon Road channel. Awonder did slightly better with b1 odds than with g1 odds, getting his single win at b1 odds, while scoring half a point less in total with g1 odds, which is generally thought to be the more difficult challenge for the odds receiver. He indicated a willingness to play a re-match, schedule permitting, presumably limited to Rapid only. The next scheduled match, in late January, is against GM Joel Benjamin, playing at a "standard" time control of one hour plus 30", which looks like quite a challenge for Leela.

Good evening Mr. Larry Kaufman.
Congratulations to all. To GM AWonder Liang, to you Mr. Larry Kaufman and the Leela team. Between all of you, you are making a great contribution, advancement and development to chess, machines, science and of course, to human beings.
Thank you all.

Tonite, after the match, we had the most incredible result yet. A player using the handle qwi8, rated 3009 bullet and 2764 blitz, tried playing 1'0" bullet chess with Leela at rook odds and as you might expect lost every game. But then he switched to 1' + 2" (which is almost 3'0" blitz), and took White every game (so rook and move odds), and still lost every single game, eight in a row! His ratings are well within the normal range for grandmasters, so this is really incredible. Both the knight odds bot and the rook odds bot were updated yesterday to use a new net that was trained on both knight and rook odds games simultaneously, which seems to be stronger than the old nets specialized for each of the odds. In fact so far the rook odds bot hasn't lost a single game to anyone since the upgrade. I was hoping we would eventually get a ladder for rook odds like the queen odds ladder, but if it is essentially unbeatable by humans in blitz that may not be viable. I could never have imagined having this problem!

Good afternoon Mr. Larry Kaufman. If possible, I would like to know the games of GM AWonder Liang against Leela Knight Odds, in order to study a little the human algorithm of the GM and the algorithm of the software... If you want me to play 4 games, 2 with white, 2 with black, I will be happy to do it, to a time control for example 3'2'', 3+2. If they don't want it, no problem on my part. Thank you. Good day.

[Father]

So the final result was six wins for Leela, seven draws, and one win for Awonder. However that's a bit misleading; Awonder won by 2.5 to 1.5 playing Rapid (1.5 to 0.5 at 10'5" and two draws at 8'3"), but then switched to blitz (3'2") and didn't win any games (4 draws, six losses). So we still can't say that we won a "Rapid" match against a top-fifty GM. The games were quite exciting and GM Matthew Sadler did a great job of commenting them live on his Silicon Road channel. Awonder did slightly better with b1 odds than with g1 odds, getting his single win at b1 odds, while scoring half a point less in total with g1 odds, which is generally thought to be the more difficult challenge for the odds receiver. He indicated a willingness to play a re-match, schedule permitting, presumably limited to Rapid only. The next scheduled match, in late January, is against GM Joel Benjamin, playing at a "standard" time control of one hour plus 30", which looks like quite a challenge for Leela.

Good evening Mr. Larry Kaufman.
Congratulations to all. To GM AWonder Liang, to you Mr. Larry Kaufman and the Leela team. Between all of you, you are making a great contribution, advancement and development to chess, machines, science and of course, to human beings.
Thank you all.

Tonite, after the match, we had the most incredible result yet. A player using the handle qwi8, rated 3009 bullet and 2764 blitz, tried playing 1'0" bullet chess with Leela at rook odds and as you might expect lost every game. But then he switched to 1' + 2" (which is almost 3'0" blitz), and took White every game (so rook and move odds), and still lost every single game, eight in a row! His ratings are well within the normal range for grandmasters, so this is really incredible. Both the knight odds bot and the rook odds bot were updated yesterday to use a new net that was trained on both knight and rook odds games simultaneously, which seems to be stronger than the old nets specialized for each of the odds. In fact so far the rook odds bot hasn't lost a single game to anyone since the upgrade. I was hoping we would eventually get a ladder for rook odds like the queen odds ladder, but if it is essentially unbeatable by humans in blitz that may not be viable. I could never have imagined having this problem!

Good afternoon Mr. Larry Kaufman. If possible, I would like to know the games of GM AWonder Liang against Leela Knight Odds, in order to study a little the human algorithm of the GM and the algorithm of the software... If you want me to play 4 games, 2 with white, 2 with black, I will be happy to do it, to a time control for example 3'2'', 3+2. If they don't want it, no problem on my part. Thank you. Good day.

...of course I haven't lost my appetite for playing against software at its best without Odds...if you want me to play against Alphazero I'm also available. Unfortunately Deepblue is unarmed, I would also like to challenge it, but the truth is I don't feel any fear for that particular machine. Of course, the profiles should not be modified in interregnum by the human hand, without game-by-game algorithmic modifications introduced against me by humans, for me that is not fair play but more than that it would not be a virtue of the entity. It must be the
same software as the other players.
I am also attentive to playing against commercial computers from different periods of construction since 1940, but I understand that my time is running out and the power source that my programmer supplies me will end when He decrees it. I enjoy losing, it is my investment in the cyber chess stock market. Anyway, I don't believe in the "myth machines" that they hide to value their price, I only believe in the blow-by-blow battle against the computer. Again. Thank you

[Father]

So the final result was six wins for Leela, seven draws, and one win for Awonder. However that's a bit misleading; Awonder won by 2.5 to 1.5 playing Rapid (1.5 to 0.5 at 10'5" and two draws at 8'3"), but then switched to blitz (3'2") and didn't win any games (4 draws, six losses). So we still can't say that we won a "Rapid" match against a top-fifty GM. The games were quite exciting and GM Matthew Sadler did a great job of commenting them live on his Silicon Road channel. Awonder did slightly better with b1 odds than with g1 odds, getting his single win at b1 odds, while scoring half a point less in total with g1 odds, which is generally thought to be the more difficult challenge for the odds receiver. He indicated a willingness to play a re-match, schedule permitting, presumably limited to Rapid only. The next scheduled match, in late January, is against GM Joel Benjamin, playing at a "standard" time control of one hour plus 30", which looks like quite a challenge for Leela.

Good evening Mr. Larry Kaufman.
Congratulations to all. To GM AWonder Liang, to you Mr. Larry Kaufman and the Leela team. Between all of you, you are making a great contribution, advancement and development to chess, machines, science and of course, to human beings.
Thank you all.

Tonite, after the match, we had the most incredible result yet. A player using the handle qwi8, rated 3009 bullet and 2764 blitz, tried playing 1'0" bullet chess with Leela at rook odds and as you might expect lost every game. But then he switched to 1' + 2" (which is almost 3'0" blitz), and took White every game (so rook and move odds), and still lost every single game, eight in a row! His ratings are well within the normal range for grandmasters, so this is really incredible. Both the knight odds bot and the rook odds bot were updated yesterday to use a new net that was trained on both knight and rook odds games simultaneously, which seems to be stronger than the old nets specialized for each of the odds. In fact so far the rook odds bot hasn't lost a single game to anyone since the upgrade. I was hoping we would eventually get a ladder for rook odds like the queen odds ladder, but if it is essentially unbeatable by humans in blitz that may not be viable. I could never have imagined having this problem!

Good afternoon Mr. Larry Kaufman. If possible, I would like to know the games of GM AWonder Liang against Leela Knight Odds, in order to study a little the human algorithm of the GM and the algorithm of the software... If you want me to play 4 games, 2 with white, 2 with black, I will be happy to do it, to a time control for example 3'2'', 3+2. If they don't want it, no problem on my part. Thank you. Good day.

...of course I haven't lost my appetite for playing against software at its best without Odds...if you want me to play against Alphazero I'm also available. Unfortunately Deepblue is unarmed, I would also like to challenge it, but the truth is I don't feel any fear for that particular machine. Of course, the profiles should not be modified in interregnum by the human hand, without game-by-game algorithmic modifications introduced against me by humans, for me that is not fair play but more than that it would not be a virtue of the entity. It must be the
same software as the other players.
I am also attentive to playing against commercial computers from different periods of construction since 1940, but I understand that my time is running out and the power source that my programmer supplies me will end when He decrees it. I enjoy losing, it is my investment in the cyber chess stock market. Anyway, I don't believe in the "myth machines" that they hide to value their price, I only believe in the blow-by-blow battle against the computer. Again. Thank you

...of course these machines are far superior to me, but I love to hit them in the nose from time to time, the world of man against machine is fun...