Navara vs Leela Knight Odds Match

[lkaufman]

A knight odds match has been announced between GM David Navara (FIDE Rapid rating 2701, world number 24 Rapid) and LeelaKnightOdds on LiChess this Saturday. 1400 UTC is 10 a.m. in eastern US, 3 pm CET. Alternating between b1 and g1 knight odds, initial time control 5'3"; after every pair of games time control may be raised or lowered depending on the results to make for a competitive match. David was over 2750 in both Classical and Rapid (and over 2800 FIDE blitz) within the past decade, with a peak classical ranking as world number 13. He is the strongest player ever to take on a computer at knight odds in a public match. Note that the three hour duration is only an estimate; we don't know what the actual time limits will be for the last eight games so it could be anywhere from two to four hours roughly depending on results.

Based on results vs. other strong players, mostly in blitz or bullet, I would not expect David to be able to reduce the time limit, but who knows? Rather than predicting the score, the question is to predict what will be the final time limit for the last two games, given that it will probably be increased if he fails to win either game of a pair? 2+1, 3+2, 5+3, 10+2, 10+5, 15+5, 15+10? Can an engine really give knight odds successfully to a top-25 Rapid GM in a serious Rapid match, or is that possible only in blitz? We shall see.

[CornfedForever]

Two observations:
1. Wave enough $$ under a GM's nose and they will do anything with 32 wooden pieces and board (take that as far as "Freestyle" chess...)
2. When you take away a piece, you alter a bit of what a GM relies on - pattern recognition and such....so one might make the argument that you are offering him a bit less than 'knight odds'.

[lkaufman]

Two observations:
1. Wave enough $$ under a GM's nose and they will do anything with 32 wooden pieces and board (take that as far as "Freestyle" chess...)
2. When you take away a piece, you alter a bit of what a GM relies on - pattern recognition and such....so one might make the argument that you are offering him a bit less than 'knight odds'.

Regarding point 1, Navara declined an option to play a two day 20 game match for twice the money, so clearly he isn't primarily doing this for money, he just relishes the challenge and is curious as to the outcome.

Regarding point 2, it should be noted that Navara is perhaps the most experienced chess960 grandmaster and a very strong proponent of it, so to the extent that knight odds is not quite standard chess he is perhaps the ideal opponent for the engine. Also Rapid seems to be his best time control. Also I don't think that he will go on "tilt" after a bad start as Nakamura did in his two pawns odds match with Komodo Dragon; he is not under any illusion that this will be an easy match win for him.

[Chessqueen]

A knight odds match has been announced between GM David Navara (FIDE Rapid rating 2701, world number 24 Rapid) and LeelaKnightOdds on LiChess this Saturday. 1400 UTC is 10 a.m. in eastern US, 3 pm CET. Alternating between b1 and g1 knight odds, initial time control 5'3"; after every pair of games time control may be raised or lowered depending on the results to make for a competitive match. David was over 2750 in both Classical and Rapid (and over 2800 FIDE blitz) within the past decade, with a peak classical ranking as world number 13. He is the strongest player ever to take on a computer at knight odds in a public match. Note that the three hour duration is only an estimate; we don't know what the actual time limits will be for the last eight games so it could be anywhere from two to four hours roughly depending on results.

Based on results vs. other strong players, mostly in blitz or bullet, I would not expect David to be able to reduce the time limit, but who knows? Rather than predicting the score, the question is to predict what will be the final time limit for the last two games, given that it will probably be increased if he fails to win either game of a pair? 2+1, 3+2, 5+3, 10+2, 10+5, 15+5, 15+10? Can an engine really give knight odds successfully to a top-25 Rapid GM in a serious Rapid match, or is that possible only in blitz? We shall see.

I Believe that 5+3 is the most that Leela would be able to give, after that Leela will NOT be able to win a single game

[lkaufman]

A knight odds match has been announced between GM David Navara (FIDE Rapid rating 2701, world number 24 Rapid) and LeelaKnightOdds on LiChess this Saturday. 1400 UTC is 10 a.m. in eastern US, 3 pm CET. Alternating between b1 and g1 knight odds, initial time control 5'3"; after every pair of games time control may be raised or lowered depending on the results to make for a competitive match. David was over 2750 in both Classical and Rapid (and over 2800 FIDE blitz) within the past decade, with a peak classical ranking as world number 13. He is the strongest player ever to take on a computer at knight odds in a public match. Note that the three hour duration is only an estimate; we don't know what the actual time limits will be for the last eight games so it could be anywhere from two to four hours roughly depending on results.

Based on results vs. other strong players, mostly in blitz or bullet, I would not expect David to be able to reduce the time limit, but who knows? Rather than predicting the score, the question is to predict what will be the final time limit for the last two games, given that it will probably be increased if he fails to win either game of a pair? 2+1, 3+2, 5+3, 10+2, 10+5, 15+5, 15+10? Can an engine really give knight odds successfully to a top-25 Rapid GM in a serious Rapid match, or is that possible only in blitz? We shall see.

I Believe that 5+3 is the most that Leela would be able to give, after that Leela will NOT be able to win a single game

You could be right, but Navara has struggled a bit in a few practice games at that level. My guess is that 10+2 will be the balanced time control for knight odds. So far no GM has scored better than around 20% in blitz, but those games are mostly 3'2" or faster and mostly with GMs with FIDE ratings starting with 25.

[AndrewGrant]

Do we know if Leela team will be offering up a net specially trained for Knight odds? Or if this will simply be using their mainline series of nets? I have to imagine that with some work for odds, it should be easy to gain a little edge. At least this holds true for NNUE engines. Training Leela is a much more difficult ordeal.

[Uri Blass]

Do we know if Leela team will be offering up a net specially trained for Knight odds? Or if this will simply be using their mainline series of nets? I have to imagine that with some work for odds, it should be easy to gain a little edge. At least this holds true for NNUE engines. Training Leela is a much more difficult ordeal.

The problem is how to test if some change is an advantage with knight odds.
It is easy to test against engines but it is not clear that being better against engines with knight odds is equivalent to being better against humans with knight odds.

[Uri Blass]

A knight odds match has been announced between GM David Navara (FIDE Rapid rating 2701, world number 24 Rapid) and LeelaKnightOdds on LiChess this Saturday. 1400 UTC is 10 a.m. in eastern US, 3 pm CET. Alternating between b1 and g1 knight odds, initial time control 5'3"; after every pair of games time control may be raised or lowered depending on the results to make for a competitive match. David was over 2750 in both Classical and Rapid (and over 2800 FIDE blitz) within the past decade, with a peak classical ranking as world number 13. He is the strongest player ever to take on a computer at knight odds in a public match. Note that the three hour duration is only an estimate; we don't know what the actual time limits will be for the last eight games so it could be anywhere from two to four hours roughly depending on results.

Based on results vs. other strong players, mostly in blitz or bullet, I would not expect David to be able to reduce the time limit, but who knows? Rather than predicting the score, the question is to predict what will be the final time limit for the last two games, given that it will probably be increased if he fails to win either game of a pair? 2+1, 3+2, 5+3, 10+2, 10+5, 15+5, 15+10? Can an engine really give knight odds successfully to a top-25 Rapid GM in a serious Rapid match, or is that possible only in blitz? We shall see.

2+1, 3+2, 5+3, 10+2, 10+5, 15+5, 15+10?


I think it is more logical to increase both the base and the increment in all cases so more logical may be something like

2+1, 3+2, 5+3, 8+4, 10+5, 13+7, 15+10

[lkaufman]

Do we know if Leela team will be offering up a net specially trained for Knight odds? Or if this will simply be using their mainline series of nets? I have to imagine that with some work for odds, it should be easy to gain a little edge. At least this holds true for NNUE engines. Training Leela is a much more difficult ordeal.

Leela will be using their standard BT4 Net, not trained for knight odds. They use smaller nets for higher handicaps. However, what is special is that they are using WDL Contempt which seems to be highly effective for odds play. I imagine this would also work in NNUE engines if they were trained to give win, loss, and draw statistics. The point is that the odds giver tries hard to avoid drawish positions, since they presumably have a much smaller probable error per move. I suppose Leela could be even stronger at odds play if it also trained a net specially for this, given that Komodo Dragon showed this, but it is pretty amazing as is. Tonite it played 11 games at 5' + 3" against two players rated near 2700 and 2800 in blitz on LiChess, and scored 9.5 to 1.5, performing over 3000. With David rated about 2900 blitz on LiChess, he looks to be an underdog at that time control, might be favored in the Rapid games.

[lkaufman]

Do we know if Leela team will be offering up a net specially trained for Knight odds? Or if this will simply be using their mainline series of nets? I have to imagine that with some work for odds, it should be easy to gain a little edge. At least this holds true for NNUE engines. Training Leela is a much more difficult ordeal.

The problem is how to test if some change is an advantage with knight odds.
It is easy to test against engines but it is not clear that being better against engines with knight odds is equivalent to being better against humans with knight odds.

It's true that you can't assume that a rating for odds play based on engine play will hold against humans, but in general I've found that if you make a change that helps vs. engines it will normally also help vs humans, just not by the same number of elo points.