Can Komodo give knight odds in standard chess to a master?

[duncan]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

did you ever say something like komodo knight handicap in self play is 800-1000 elo advantage. cannot remember time limit.

[lkaufman]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

did you ever say something like komodo knight handicap in self play is 800-1000 elo advantage. cannot remember time limit.

I think knight odds was more like 1300 elo or so in self-play, but it's very hard to measure, because with scores of more than 99% the elo gap varies greatly depending on whether you assume normal or logistic distribution, and also because results depend heavily on things like Contempt settings. As I recall Kai initially advocated Normal, then switched to advocating logistic, which makes for a much higher elo spread for say a 99% score. It also depends greatly on the level of the odds receiver. If he is strong enough (whatever that means), then the elo value of a knight may be infinite, since even perfect play (or "perfectly tricky" play) won't beat him. Another way to estimate it is that we have reasonably good data that the f7 handicap is worth close to 500 elo in both self play and komodo vs GM play at our standard time control. Knight odds is roughly 2.5 times as big a handicap, so that would give 1250. But in amateur human play it is a much smaller handicap. I used to give knight odds successfully to players rated up to about 1800 USCF (1700 FIDE) in games with rapid time controls.

[Lyudmil Tsvetkov]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

Since Komodo is facing a human (75 yrs old at that) you'd have to tune Komodo to take more calculated risks in such Knight-odds games. One thing that makes humans uncomfortable is double-edged, complicated, chaotic positions where it feels like walking a minefield, despite being up a piece.
This is probably easier said than done, since being down a piece effectively gets in the way of creating those counter chances it needs.

Nonetheless, since you're starting from a lost position, with "nothing to lose", taking increased risks seems like the only approach that has a realistic chance of succeeding. After all, Komodo was able to overcome such a deficit in chess960/FRC, if memory serves correctly.

I suggest Gilden plays at least 2700.

Then I must be close to 2900 myself, based on our head to head results in tournaments this century. So Carlsen should be seeking a match with me

might very well be, Larry, might very well be...

if you ask me, the primary reason for this sweep is, of course, the unplayable nature of the odds.

no matter how K and SF would evaluate it, N odds is at least 2.5 times easier than 2Ps odds.

simple material calculation might bring it at just 1.5 times or so, 300cps vs 200cps, but, actually, the distinction is way way bigger.

so, almost everyone should be able to win with N odds, but winning 2Ps odds is incredibly harder.

would be very interesting to see GM Gilden playing Komodo at the same 2Ps odds you used to play against the 2500 GMs last year.

this could confrim one of 2 hypotheses:

- Gilden is indeed champion
- N odds is exponentially easier than 2Ps odds

btw., not to Larry, but anyone able to post here replayable games?(I guess many more people will follow that way)

[lkaufman]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

Since Komodo is facing a human (75 yrs old at that) you'd have to tune Komodo to take more calculated risks in such Knight-odds games. One thing that makes humans uncomfortable is double-edged, complicated, chaotic positions where it feels like walking a minefield, despite being up a piece.
This is probably easier said than done, since being down a piece effectively gets in the way of creating those counter chances it needs.

Nonetheless, since you're starting from a lost position, with "nothing to lose", taking increased risks seems like the only approach that has a realistic chance of succeeding. After all, Komodo was able to overcome such a deficit in chess960/FRC, if memory serves correctly.

I suggest Gilden plays at least 2700.

Then I must be close to 2900 myself, based on our head to head results in tournaments this century. So Carlsen should be seeking a match with me

might very well be, Larry, might very well be...

if you ask me, the primary reason for this sweep is, of course, the unplayable nature of the odds.

no matter how K and SF would evaluate it, N odds is at least 2.5 times easier than 2Ps odds.

simple material calculation might bring it at just 1.5 times or so, 300cps vs 200cps, but, actually, the distinction is way way bigger.

so, almost everyone should be able to win with N odds, but winning 2Ps odds is incredibly harder.

would be very interesting to see GM Gilden playing Komodo at the same 2Ps odds you used to play against the 2500 GMs last year.

this could confrim one of 2 hypotheses:

- Gilden is indeed champion
- N odds is exponentially easier than 2Ps odds

btw., not to Larry, but anyone able to post here replayable games?(I guess many more people will follow that way)

Gilden did play one game at the largest possible two pawn handicap (f7 + g7) a year or so ago, but he lost. He did play much better this time. I think you are right about knight odds being very much larger than two pawns, much more so than the evals suggest. But I don't know how to modify the evals to reflect this without losing elo points. Another puzzle.

[Lyudmil Tsvetkov]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

Since Komodo is facing a human (75 yrs old at that) you'd have to tune Komodo to take more calculated risks in such Knight-odds games. One thing that makes humans uncomfortable is double-edged, complicated, chaotic positions where it feels like walking a minefield, despite being up a piece.
This is probably easier said than done, since being down a piece effectively gets in the way of creating those counter chances it needs.

Nonetheless, since you're starting from a lost position, with "nothing to lose", taking increased risks seems like the only approach that has a realistic chance of succeeding. After all, Komodo was able to overcome such a deficit in chess960/FRC, if memory serves correctly.

I suggest Gilden plays at least 2700.

Then I must be close to 2900 myself, based on our head to head results in tournaments this century. So Carlsen should be seeking a match with me

might very well be, Larry, might very well be...

if you ask me, the primary reason for this sweep is, of course, the unplayable nature of the odds.

no matter how K and SF would evaluate it, N odds is at least 2.5 times easier than 2Ps odds.

simple material calculation might bring it at just 1.5 times or so, 300cps vs 200cps, but, actually, the distinction is way way bigger.

so, almost everyone should be able to win with N odds, but winning 2Ps odds is incredibly harder.

would be very interesting to see GM Gilden playing Komodo at the same 2Ps odds you used to play against the 2500 GMs last year.

this could confrim one of 2 hypotheses:

- Gilden is indeed champion
- N odds is exponentially easier than 2Ps odds

btw., not to Larry, but anyone able to post here replayable games?(I guess many more people will follow that way)

Gilden did play one game at the largest possible two pawn handicap (f7 + g7) a year or so ago, but he lost. He did play much better this time. I think you are right about knight odds being very much larger than two pawns, much more so than the evals suggest. But I don't know how to modify the evals to reflect this without losing elo points. Another puzzle.

well, if he played a previous match/game against Komodo, then certainly he has prepared very well.

it is possible he has been training every day vs Komodo for the past year or so.

[lkaufman]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

Since Komodo is facing a human (75 yrs old at that) you'd have to tune Komodo to take more calculated risks in such Knight-odds games. One thing that makes humans uncomfortable is double-edged, complicated, chaotic positions where it feels like walking a minefield, despite being up a piece.
This is probably easier said than done, since being down a piece effectively gets in the way of creating those counter chances it needs.

Nonetheless, since you're starting from a lost position, with "nothing to lose", taking increased risks seems like the only approach that has a realistic chance of succeeding. After all, Komodo was able to overcome such a deficit in chess960/FRC, if memory serves correctly.

I suggest Gilden plays at least 2700.

Then I must be close to 2900 myself, based on our head to head results in tournaments this century. So Carlsen should be seeking a match with me

might very well be, Larry, might very well be...

if you ask me, the primary reason for this sweep is, of course, the unplayable nature of the odds.

no matter how K and SF would evaluate it, N odds is at least 2.5 times easier than 2Ps odds.

simple material calculation might bring it at just 1.5 times or so, 300cps vs 200cps, but, actually, the distinction is way way bigger.

so, almost everyone should be able to win with N odds, but winning 2Ps odds is incredibly harder.

would be very interesting to see GM Gilden playing Komodo at the same 2Ps odds you used to play against the 2500 GMs last year.

this could confrim one of 2 hypotheses:

- Gilden is indeed champion
- N odds is exponentially easier than 2Ps odds

btw., not to Larry, but anyone able to post here replayable games?(I guess many more people will follow that way)

Gilden did play one game at the largest possible two pawn handicap (f7 + g7) a year or so ago, but he lost. He did play much better this time. I think you are right about knight odds being very much larger than two pawns, much more so than the evals suggest. But I don't know how to modify the evals to reflect this without losing elo points. Another puzzle.

well, if he played a previous match/game against Komodo, then certainly he has prepared very well.

it is possible he has been training every day vs Komodo for the past year or so.

Since he doesn't even own a computer or a cellphone, I don't think there is much chance of that!

[Laskos]

I recall Kai initially advocated Normal, then switched to advocating logistic, which makes for a much higher elo spread for say a 99% score.

I had older indications that Gaussian fits better, but newer, more comprehensive results from one year ago posted in CCC on wide range of ELO span and several engines gave this plot:



which showed logistic as more consistent with high statistical significance.

[Lyudmil Tsvetkov]

In view of the match loss, I ran a test to see if Komodo can give knight odds to a decent engine. I chose Fritz 11 running on 1 core, with Komodo able to use all 4 cores on my laptop. Time limit was 15 min plus 5 seconds, one third of the human match level but still much slower than blitz. Contempt set at 100 (I used 150 vs. Gilden). Alternating knight handicap.

Fritz 11 on 1 pcu is rated 2853 on the CCRL 40/40 list, which actually seems a bit low since it is several versions improved from the Fritz that defeated then-World Champion Kramnik in a match. I would bet on Fritz 11 1 cpu in a match with Magnus Carlsen, though it might be close.

Much to my surprise, Komodo (same version that played Gilden) won by 13 to 7 (12 wins, 6 losses, 2 draws). So in rapid chess, Komodo can decisively defeat a Magnus Carlsen level engine giving it knight odds, but lost all three games to a nearly 75 year old player with a FIDE rating of 2114. Pretty amazing difference. Perhaps Fritz 11 doesn't know to simplify when ahead, I don't know about that, but this surely isn't the whole answer. In most of the winning games Komodo whipped up a strong attack on the king, something that was never even a dream against Gilden.

Must be something to learn from this.

Since Komodo is facing a human (75 yrs old at that) you'd have to tune Komodo to take more calculated risks in such Knight-odds games. One thing that makes humans uncomfortable is double-edged, complicated, chaotic positions where it feels like walking a minefield, despite being up a piece.
This is probably easier said than done, since being down a piece effectively gets in the way of creating those counter chances it needs.

Nonetheless, since you're starting from a lost position, with "nothing to lose", taking increased risks seems like the only approach that has a realistic chance of succeeding. After all, Komodo was able to overcome such a deficit in chess960/FRC, if memory serves correctly.

I suggest Gilden plays at least 2700.

Then I must be close to 2900 myself, based on our head to head results in tournaments this century. So Carlsen should be seeking a match with me

might very well be, Larry, might very well be...

if you ask me, the primary reason for this sweep is, of course, the unplayable nature of the odds.

no matter how K and SF would evaluate it, N odds is at least 2.5 times easier than 2Ps odds.

simple material calculation might bring it at just 1.5 times or so, 300cps vs 200cps, but, actually, the distinction is way way bigger.

so, almost everyone should be able to win with N odds, but winning 2Ps odds is incredibly harder.

would be very interesting to see GM Gilden playing Komodo at the same 2Ps odds you used to play against the 2500 GMs last year.

this could confrim one of 2 hypotheses:

- Gilden is indeed champion
- N odds is exponentially easier than 2Ps odds

btw., not to Larry, but anyone able to post here replayable games?(I guess many more people will follow that way)

Gilden did play one game at the largest possible two pawn handicap (f7 + g7) a year or so ago, but he lost. He did play much better this time. I think you are right about knight odds being very much larger than two pawns, much more so than the evals suggest. But I don't know how to modify the evals to reflect this without losing elo points. Another puzzle.

well, if he played a previous match/game against Komodo, then certainly he has prepared very well.

it is possible he has been training every day vs Komodo for the past year or so.

Since he doesn't even own a computer or a cellphone, I don't think there is much chance of that!

he is using a tablet.

[jd1]

Just a colloquial observation, but I am not that surprised by this result. I am 2100 FIDE, and (in my experience) can beat any engine with knight odds fairly consistently. It's just relatively straightforward to play simple openings and trade down.

Two pawns I lose nearly every single time, regardless of who moves first, which pawns, etc. I suspect three pawns, or exchange plus pawn(s) would be the same. I also suspect that it would be much harder to consistently convert a knight handicap against a super GM for example.

My theory is that:
1. The path to win from a whole knight up is far simpler to find, certainly for humans, than other close or "equivalent" (in terms of cp) handicaps.
2. Top engines don't realise that they need to complicate the position against a human player with an extra knight. A grandmaster does. I don't think this knowledge is particularly useful in computer vs computer games (although surely complicating when you are losing is a good idea in theory, but AFAIK no one has made it work in terms of elo), but it would likely help in this sort of handicap match.

[Nordlandia]

The general consensus is that most skilled players beat Komodo with knight odds. An attempt to balance it is to disable castling rights. Even then human can castle artificially. So i don't see the reason for not trying this.

Komodo evaluates knight odds without castling rights about -2.50.