LeelaKnightOdds

[lkaufman]

There is now a bot on LiChess called "LeelaKnightOdds", which is Lc0 running on an RTX 4090 always giving knight odds (b1 or g1, your choice), Rapid,Blitz, or Bullet, as long as there is at least one second increment. Parameters have been gradually tuned so it is getting better. Results are pretty impressive. Especially today it has scored 10 wins and one draw in slow blitz (5' + 2 or 3 sec inc) against players over 2350 (including me), with a performance rating of nearly 3000, roughly the top of the scale (I think only Magnus Carlsen, Artemiev, and one other player are over 3000). I recommend any strong players reading this to try it, it's remarkable. Is it possible that an engine can now give knight odds to any human (except perhaps Magnus and Hikaru) in "slow blitz"? Hard to believe.

[OneTrickPony]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

[lkaufman]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

Yes, at the time limit you chose, 10' + 5", a strong GM might be favored, but not in blitz, I think. The opening seems to be critical; Leela is playing somewhat randomly for the first few moves in order to provide sufficient variety, but this hurts its results when it chooses bad lines. Probably the opening randomness is set too high. With best or near-best openings, it is not easy to beat at knight odds.

[Chessqueen]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

Yes, at the time limit you chose, 10' + 5", a strong GM might be favored, but not in blitz, I think. The opening seems to be critical; Leela is playing somewhat randomly for the first few moves in order to provide sufficient variety, but this hurts its results when it chooses bad lines. Probably the opening randomness is set too high. With best or near-best openings, it is not easy to beat at knight odds.

If GM Nakamura can draw versus Torch playing even at 5 minutes game, he will destroy LeelaKnight at Knight Odds in Blitz https://www.youtube.com/watch?v=0NNn7943_uw

[lkaufman]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

Yes, at the time limit you chose, 10' + 5", a strong GM might be favored, but not in blitz, I think. The opening seems to be critical; Leela is playing somewhat randomly for the first few moves in order to provide sufficient variety, but this hurts its results when it chooses bad lines. Probably the opening randomness is set too high. With best or near-best openings, it is not easy to beat at knight odds.

If GM Nakamura can draw versus Torch playing even at 5 minutes game, he will destroy LeelaKnight at Knight Odds in Blitz https://www.youtube.com/watch?v=0NNn7943_uw

I checked into that, it seems he was playing against Torch running in the browser (so way below true strength, though still far above 3000), with moves relayed, but apparently without knowing move history. He lost the first game as Black, then in the second game as White he lost a rook for just a pawn but achieved a very blocked up position where he could hope for a draw. Without knowing move history, the engine just allowed a repetition. Whether it would have won knowing move history is unclear, but even if Hikaru had obtained a draw in this way, he was just doing the same thing "Father" (or "Grandfather") has done here many times, getting a draw by fooling the engine into blocking up the game too much. It doesn't tell us anything about who might win a knight odds match, where playing for a draw won't be helpful. Having said that, I would bet on Hikaru at this time limit (which was 5' + 2", not 5 min) at knight odds vs the Leela bot, but not on anyone below 2700 FIDE (regular or blitz). A three move opening book was added to the bot tonight to provide variety while avoiding bad moves; this should make it noticeably stronger. Hopefully we'll see some more strong players try it soon.

[Uri Blass]

There is now a bot on LiChess called "LeelaKnightOdds", which is Lc0 running on an RTX 4090 always giving knight odds (b1 or g1, your choice), Rapid,Blitz, or Bullet, as long as there is at least one second increment. Parameters have been gradually tuned so it is getting better. Results are pretty impressive. Especially today it has scored 10 wins and one draw in slow blitz (5' + 2 or 3 sec inc) against players over 2350 (including me), with a performance rating of nearly 3000, roughly the top of the scale (I think only Magnus Carlsen, Artemiev, and one other player are over 3000). I recommend any strong players reading this to try it, it's remarkable. Is it possible that an engine can now give knight odds to any human (except perhaps Magnus and Hikaru) in "slow blitz"? Hard to believe.

Did you test KomodoDragon in these conditions?
What is the performance of KomodoDragon against humans in these conditions for comparison?

[lkaufman]

There is now a bot on LiChess called "LeelaKnightOdds", which is Lc0 running on an RTX 4090 always giving knight odds (b1 or g1, your choice), Rapid,Blitz, or Bullet, as long as there is at least one second increment. Parameters have been gradually tuned so it is getting better. Results are pretty impressive. Especially today it has scored 10 wins and one draw in slow blitz (5' + 2 or 3 sec inc) against players over 2350 (including me), with a performance rating of nearly 3000, roughly the top of the scale (I think only Magnus Carlsen, Artemiev, and one other player are over 3000). I recommend any strong players reading this to try it, it's remarkable. Is it possible that an engine can now give knight odds to any human (except perhaps Magnus and Hikaru) in "slow blitz"? Hard to believe.

Did you test KomodoDragon in these conditions?
What is the performance of KomodoDragon against humans in these conditions for comparison?

Best comparison is that GM Alex Lenderman played a lot of knight odds games with Komodo Dragon at various fast time limits. I think that the break-even time limit was close to 5' + 2". I would guess that he would be something like 2700 to 2800 Lichess blitz, I don't know if he actually plays there. We'll need more data against GM level players for the Leela bot to tell if it can actually perform above Lenderman's level at such time controls. My feeling is yes, it can, now that it has an opening book too (as did Dragon). We'll see.

[lkaufman]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

Today, playing with an opening book, it won 3 out of 3 slow Rapid games (15' + 10") with a player rated 2545 Blitz (no meaningful rapid rating), plus one more win at 5'2. Total score today 15 out of 15. Someone needs to get at least a draw from it to be able to calculate a performance rating, otherwise it is infinite! All A/B engines play "correct" chess at knight odds, trying to pick up a pawn here or there and hoping to hold a draw (winning if opponent blunders), but this Lc0 (using an original form of Contempt) just plays for mate, quite effectively it seems.

[Uri Blass]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

Today, playing with an opening book, it won 3 out of 3 slow Rapid games (15' + 10") with a player rated 2545 Blitz (no meaningful rapid rating), plus one more win at 5'2. Total score today 15 out of 15. Someone needs to get at least a draw from it to be able to calculate a performance rating, otherwise it is infinite! All A/B engines play "correct" chess at knight odds, trying to pick up a pawn here or there and hoping to hold a draw (winning if opponent blunders), but this Lc0 (using an original form of Contempt) just plays for mate, quite effectively it seems.

I wonder if this strategy also works better than "correct" chess against relatively weak chess engines.

[lkaufman]

It's awesome! I've lost 3 first games to it and won 4th one. I think it would take a bit of practice to beat it consistently. I don't think an active strong GM would encounter problems though.

Today, playing with an opening book, it won 3 out of 3 slow Rapid games (15' + 10") with a player rated 2545 Blitz (no meaningful rapid rating), plus one more win at 5'2. Total score today 15 out of 15. Someone needs to get at least a draw from it to be able to calculate a performance rating, otherwise it is infinite! All A/B engines play "correct" chess at knight odds, trying to pick up a pawn here or there and hoping to hold a draw (winning if opponent blunders), but this Lc0 (using an original form of Contempt) just plays for mate, quite effectively it seems.

I wonder if this strategy also works better than "correct" chess against relatively weak chess engines.

I think that the answer is yes, but I'll run more tests tonite to confirm. For this purpose, I use the elo levels of Komodo Dragon 3.3 but with NNUE turned OFF. I discovered that, although elo 2700 with and without NNUE are about the same strength in normal chess, the no NNUE version is MUCH stronger when receiving knight odds. Probably that's because the net isn't really much better than no net in abnormal positions, but the NNUE setting does a much shorter search than the no NNUE setting to be equivalent in normal chess. Anyway, bottom line is that the currently used Lc0 version and settings comes out slightly ahead of Komodo 2700 elo no NNUE at knight odds (using chrisW opening book for variety) at 4'2"; I don't think any A/B engine can do this but I'll check tonite. This supposedly means that LeelaKnightOdds should be even or plus with a FIDE 2700 at 10 + 5" if our estimates are correct, but of course this assumes that the human and engine are equally good at playing knight odds as well. I don't quite believe this, considering that such great players as Gukesh, Radjabov, Aronian, and Mamedyarov are all near 2720 now, but it's not impossible.