Next Komodo vs GM match - 2 pawns

[Ozymandias]

I happen to think that human-computer in correspondence chess can be stronger than the same computer alone if the human is competent.

Me too, I was talking about old school corr players, who left the field because they don't like to have to use the computer so much to stay competitive. I don't think there's many of those left, and I'm sure no one is successful against standalone engines.

[Laskos]

I don't think GMs would be offended by this idea, it's no more embarassing than playing with material odds. Rather, the problem is that I don't think GMs will have any confidence that takebacks will help much. I don't think your simulation has much to do with the Komodo vs GM situation. Usually GMs make small mistakes, not outright blunders. Sometimes they are apparent after the computer responds, sometimes not.
I suppose any takeback handicap would stipulate that a player may retract his move any time until he makes another move, and if he does so he does not get back any time but Komodo should get back the time of the last move (this may be impractical though). It should probably also stipulate that you cannot take back the same move number more than once. I don't know how to simulate all this, but I think that your numbers are way too optimistic for the human side. In my opinion, even with unlimited takebacks (subject to my above rules) an ordinary GM would have little chance to win a match vs. Komodo. Maybe a top ten player might have a fair chance; even then he might need some other edge like White pieces or short opening book. I'm willing to try the idea if some GM has enough confidence to play knowing he won't make much money if he loses all the games.

You are of the opinion that GMs constantly make small mistakes? I would change just a bit your rules by the stipulation that a player may take back several times (subject to total) the same move, but this is a detail. If you are not sure that a good GM can sometimes beat Komodo taking back all the moves once, you can organize a match, favoring in your view Komodo.

I don't say that the GM won't win any games that way, just not a match. Taking back the same move more than once might be too embarassing or seem too unreasonable, and obviously if we allow unlimited number of total takebacks then this wouldn't work. I'm willing to try your idea if I find a GM who is confident enough to try. Maybe I'll ask some GM opinions. So far we only have one (mine). I think this idea would have to be played from my home on a real chessboard, not just on chess.com as some have been.

It is indeed hard to model human mistakes and you surely must know better. I tried right now to see the value of depth=2 mistake and depth=5 mistake at 2700 ELO level, where random move mistake is 127 ELO points. The depth=2 mistake is 42 ELO points, so if a typical GM mistake is modeled well, then the 2700 GM would need 3*5=15 takebacks. The depth=5 mistake is about 20 ELO points at 2700 level, and a 2700 GM would need 30 takebacks of this sort of mistakes. Lower level players would indeed need pretty much unlimited takebacks. But I guess you are right, and this sort of modeling is naive.

[Lyudmil Tsvetkov]

Excellent initiative, but something seems to be lacking.
All 4 games were, to say the least, plain dull.
In sharp contrast to, say, almost all Kasparov-machine encounters.
I really do not know who Erenburg is. To make things interesting, find someone with a strong chess philosophy, for example Bareev, Morozevich, or why not Finegold? If not anything else, there will at least be fun on the board.

[mehmet karaman]

Back to those odds already? Too bad.

I'm guessing you would have preferred another match like the last one but with larger time odds? I didn't like this so much because making Komodo play much faster than 3' + 1" seems pointless, since hardly anyone uses it with just one or two seconds thinking time I imagine. Or did you have another idea in mind?

I have to say that watching Komodo play real chess vs a human GM in the last match was totally fascinating. If you want to level the playing field more you probably need someone a bit stronger than GM Erenburg, rather than giving more time to the human or further slowing down Komodo.

Material handicaps on the other hand, present little spectator interest as far as I'm concerned. Once you've seen one, it feels like you've seen them all.
With an equal starting position, you're giving Komodo an opportunity to shine at real chess, which never gets old. More people can probably relate to that, I would imagine.

CL

I agree you.

We want to see real chess vs GM (2600+, 2700+ elo) with same time odds (1-4 core cpu, 8-12 move opening book and more contempt)

[Laskos]

Back to those odds already? Too bad.

I'm guessing you would have preferred another match like the last one but with larger time odds? I didn't like this so much because making Komodo play much faster than 3' + 1" seems pointless, since hardly anyone uses it with just one or two seconds thinking time I imagine. Or did you have another idea in mind?

I have to say that watching Komodo play real chess vs a human GM in the last match was totally fascinating. If you want to level the playing field more you probably need someone a bit stronger than GM Erenburg, rather than giving more time to the human or further slowing down Komodo.

Material handicaps on the other hand, present little spectator interest as far as I'm concerned. Once you've seen one, it feels like you've seen them all.
With an equal starting position, you're giving Komodo an opportunity to shine at real chess, which never gets old. More people can probably relate to that, I would imagine.

CL

I agree you.

We want to see real chess vs GM (2600+, 2700+ elo) with same time odds (1-4 core cpu, 8-12 move opening book and more contempt)

Yes, most of the matches will end 4-0.

[mehmet karaman]

Back to those odds already? Too bad.

I'm guessing you would have preferred another match like the last one but with larger time odds? I didn't like this so much because making Komodo play much faster than 3' + 1" seems pointless, since hardly anyone uses it with just one or two seconds thinking time I imagine. Or did you have another idea in mind?

I have to say that watching Komodo play real chess vs a human GM in the last match was totally fascinating. If you want to level the playing field more you probably need someone a bit stronger than GM Erenburg, rather than giving more time to the human or further slowing down Komodo.

Material handicaps on the other hand, present little spectator interest as far as I'm concerned. Once you've seen one, it feels like you've seen them all.
With an equal starting position, you're giving Komodo an opportunity to shine at real chess, which never gets old. More people can probably relate to that, I would imagine.

CL

I agree you.

We want to see real chess vs GM (2600+, 2700+ elo) with same time odds (1-4 core cpu, 8-12 move opening book and more contempt)

Yes, most of the matches will end 4-0.

You are right. Maybe a match vs. a 2600+ GM with same conditions at Erenburg match except time control (90'+30'' vs 10''+0.01" ) is more exciting

[lkaufman]

Back to those odds already? Too bad.

I'm guessing you would have preferred another match like the last one but with larger time odds? I didn't like this so much because making Komodo play much faster than 3' + 1" seems pointless, since hardly anyone uses it with just one or two seconds thinking time I imagine. Or did you have another idea in mind?

I have to say that watching Komodo play real chess vs a human GM in the last match was totally fascinating. If you want to level the playing field more you probably need someone a bit stronger than GM Erenburg, rather than giving more time to the human or further slowing down Komodo.

Material handicaps on the other hand, present little spectator interest as far as I'm concerned. Once you've seen one, it feels like you've seen them all.
With an equal starting position, you're giving Komodo an opportunity to shine at real chess, which never gets old. More people can probably relate to that, I would imagine.

CL

I agree you.

We want to see real chess vs GM (2600+, 2700+ elo) with same time odds (1-4 core cpu, 8-12 move opening book and more contempt)

Yes, most of the matches will end 4-0.

You are right. Maybe a match vs. a 2600+ GM with same conditions at Erenburg match except time control (90'+30'' vs 10''+0.01" ) is more exciting

Maybe more exciting and much more balanced, but it is of no practical use to Komodo users to determine how well Komodo plays with about 1/8 of a second per move; no one uses it that way. I think I like the takeback idea better if we are going to play from the normal start position.

[Uri Blass]

Back to those odds already? Too bad.

I'm guessing you would have preferred another match like the last one but with larger time odds? I didn't like this so much because making Komodo play much faster than 3' + 1" seems pointless, since hardly anyone uses it with just one or two seconds thinking time I imagine. Or did you have another idea in mind?

A faster TC for Komodo, a longer TC for the GM or a stronger human opponent. Any of those variations, or a viable combination, would give us a more even match.

There's also the option of using a popular phone, instead of a laptop.

U

We might try this sort of match if we get another top ten opponent. Increasing the human's time further isn't very practical for chess.com tv coverage, and I'm told that single core of laptop already falls within the range of cellphones, so using a cellphone wouldn't be much different. We don't want matches where the winner is not in doubt.

Against a 2600 GM, a fair match would be against a mid-range smartphone at 30''+0.3''. Who wants to see such a match?

I have another idea: takebacks. Is it too insulting to GM to allow him taking back? This would solve one prominent issue with humans, that a GM plays generally a very sound chess, but marred by blunders and large inaccuracies. I used "Andscacs Randomizer" engine, which can play a specified percentage of random moves in the total of moves, the rest being regular Andscacs, to compute the value of a blunder defined here as random move. If taken back, this improves the performance of the player by the resulting ELO points. Here are the results, I computed them several weeks ago:

Code: Select all

BLUNDERS (RANDOM MOVES)

Random: 0%   ELO:2700
100%:95%
Score of R2 vs R1: 334 - 36 - 30  [0.873] 400
ELO difference: 334.10 +/- 47.24
Finished match
PlyCount: Total = 400  Range: 3-283  Average = 105.47  StdDev = 36.14
Av.nr.blunders = 2.63675   
Av.ELO per blunder = 126.7
ELO = 2700
Takebacks: 4.7

Random: 5%   ELO:2476
95%:90%
Score of R2 vs R1: 306 - 79 - 15  [0.784] 400
ELO difference: 223.69 +/- 40.49
Finished match
PlyCount: Total = 400  Range: 16-311  Average = 92.0  StdDev = 34.36
Av.nr.blunders = 2.30   
Av.ELO per blunder = 97.3
ELO = 2476
Takebacks: 8.5

Random: 10%   ELO:2324
90%:85%
Score of R2 vs R1: 278 - 113 - 9  [0.706] 400
ELO difference: 152.39 +/- 37.02
Finished match
PlyCount: Total = 400  Range: 17-246  Average = 89.56  StdDev = 34.99
Av.nr.blunders = 2.24  
Av.ELO per blunder: 68.0
ELO = 2324
Takebacks: 14.3

Random: 15%   ELO:2185
85%:80%
Score of R2 vs R1: 269 - 117 - 14  [0.690] 400
ELO difference: 138.99 +/- 36.18
Finished match
PlyCount: Total = 400  Range: 9-282  Average = 90.23  StdDev = 37.57
Av.nr.blunders = 2.26   
Av.ELO per blunder: 61.5
ELO = 2185
Takebacks: 18.1

Random: 20%   ELO:1947
80%:70%
Score of R2 vs R1: 315 - 77 - 8  [0.797] 400
ELO difference: 238.12 +/- 42.06
Finished match
PlyCount: Total = 400  Range: 5-233  Average = 84.18  StdDev = 34.86
Av.nr.blunders = 4.21   
Av.ELO per blunder: 56.6
ELO = 1947
Takebacks: 23.9

Random: 30%   ELO:1784
70%:60%
Score of R2 vs R1: 275 - 100 - 25  [0.719] 400
ELO difference: 162.99 +/- 36.53
Finished match
PlyCount: Total = 400  Range: 5-247  Average = 91.68  StdDev = 43.77
Av.nr.blunders = 4.58
Av.ELO per blunder: 35.6
ELO = 1784
Takebacks: 42.6

A 2700 GM would need to take back thus defined blunders 5 times to reach 3250-3300 FIDE ELO level of Komodo on 24 cores. A 2476 GM would need 8-9 takebacks. A 2324 player about 14 takebacks, A 2185 about 18 takebacks. So on, and a 1784 ELO player would need 43 takebacks. If the blunder (or large inaccuracy) is on average still better than random move, there is a multiplicaive factor on all these numbers of takebacks. One can probably determine pretty accurately what sort of handicap is needed for what strength, say if a ELO 2700 player needs 10 takebacks, then a 2200 player would need 35 takebacks, and a 1900 player needs to take back all moves at least once.

To me it would make for good handicap matches because it eliminates one of the main human weaknesses - blunders an serious inaccuracies. If only GMs are not offended by this proposal, I don't know their tastes.

I don't think GMs would be offended by this idea, it's no more embarassing than playing with material odds. Rather, the problem is that I don't think GMs will have any confidence that takebacks will help much. I don't think your simulation has much to do with the Komodo vs GM situation. Usually GMs make small mistakes, not outright blunders. Sometimes they are apparent after the computer responds, sometimes not.
I suppose any takeback handicap would stipulate that a player may retract his move any time until he makes another move, and if he does so he does not get back any time but Komodo should get back the time of the last move (this may be impractical though). It should probably also stipulate that you cannot take back the same move number more than once. I don't know how to simulate all this, but I think that your numbers are way too optimistic for the human side. In my opinion, even with unlimited takebacks (subject to my above rules) an ordinary GM would have little chance to win a match vs. Komodo. Maybe a top ten player might have a fair chance; even then he might need some other edge like White pieces or short opening book. I'm willing to try the idea if some GM has enough confidence to play knowing he won't make much money if he loses all the games.

I see no reason for the rule that you cannot take back the same move number more than once.

Even with taking back unlimited number of times and taking the same move back unlimited number of times it is not clear that the human can draw or win if the human has limited time to finish the game and the question is what limitation is fair.

For example it may be possible to decide that the computer use 5 seconds per move and the human has 2 hours to finish the game.

If the human play at average speed of 5 seconds per move then
a move by both players can take 10 seconds and the human can practically play 720 moves but with all the take backs it is not clear that it is going to be enough for him even to get a draw.

In order to prevent winning on time in an obvious drawn position it is possible to decide that the human is allowed to play with no take back with 5 seconds per move from the final position that the sides achieved after 2 hours.


Edit:
I have no idea what time control is fair for the human in these conditions.
It is possible that 2 hours is too generous for a GM but in this case it is possible to fix it by less time or by a weaker opponent that is not a GM.

[lkaufman]

Back to those odds already? Too bad.

I'm guessing you would have preferred another match like the last one but with larger time odds? I didn't like this so much because making Komodo play much faster than 3' + 1" seems pointless, since hardly anyone uses it with just one or two seconds thinking time I imagine. Or did you have another idea in mind?

A faster TC for Komodo, a longer TC for the GM or a stronger human opponent. Any of those variations, or a viable combination, would give us a more even match.

There's also the option of using a popular phone, instead of a laptop.

U

We might try this sort of match if we get another top ten opponent. Increasing the human's time further isn't very practical for chess.com tv coverage, and I'm told that single core of laptop already falls within the range of cellphones, so using a cellphone wouldn't be much different. We don't want matches where the winner is not in doubt.

Against a 2600 GM, a fair match would be against a mid-range smartphone at 30''+0.3''. Who wants to see such a match?

I have another idea: takebacks. Is it too insulting to GM to allow him taking back? This would solve one prominent issue with humans, that a GM plays generally a very sound chess, but marred by blunders and large inaccuracies. I used "Andscacs Randomizer" engine, which can play a specified percentage of random moves in the total of moves, the rest being regular Andscacs, to compute the value of a blunder defined here as random move. If taken back, this improves the performance of the player by the resulting ELO points. Here are the results, I computed them several weeks ago:

Code: Select all

BLUNDERS (RANDOM MOVES)

Random: 0%   ELO:2700
100%:95%
Score of R2 vs R1: 334 - 36 - 30  [0.873] 400
ELO difference: 334.10 +/- 47.24
Finished match
PlyCount: Total = 400  Range: 3-283  Average = 105.47  StdDev = 36.14
Av.nr.blunders = 2.63675   
Av.ELO per blunder = 126.7
ELO = 2700
Takebacks: 4.7

Random: 5%   ELO:2476
95%:90%
Score of R2 vs R1: 306 - 79 - 15  [0.784] 400
ELO difference: 223.69 +/- 40.49
Finished match
PlyCount: Total = 400  Range: 16-311  Average = 92.0  StdDev = 34.36
Av.nr.blunders = 2.30   
Av.ELO per blunder = 97.3
ELO = 2476
Takebacks: 8.5

Random: 10%   ELO:2324
90%:85%
Score of R2 vs R1: 278 - 113 - 9  [0.706] 400
ELO difference: 152.39 +/- 37.02
Finished match
PlyCount: Total = 400  Range: 17-246  Average = 89.56  StdDev = 34.99
Av.nr.blunders = 2.24  
Av.ELO per blunder: 68.0
ELO = 2324
Takebacks: 14.3

Random: 15%   ELO:2185
85%:80%
Score of R2 vs R1: 269 - 117 - 14  [0.690] 400
ELO difference: 138.99 +/- 36.18
Finished match
PlyCount: Total = 400  Range: 9-282  Average = 90.23  StdDev = 37.57
Av.nr.blunders = 2.26   
Av.ELO per blunder: 61.5
ELO = 2185
Takebacks: 18.1

Random: 20%   ELO:1947
80%:70%
Score of R2 vs R1: 315 - 77 - 8  [0.797] 400
ELO difference: 238.12 +/- 42.06
Finished match
PlyCount: Total = 400  Range: 5-233  Average = 84.18  StdDev = 34.86
Av.nr.blunders = 4.21   
Av.ELO per blunder: 56.6
ELO = 1947
Takebacks: 23.9

Random: 30%   ELO:1784
70%:60%
Score of R2 vs R1: 275 - 100 - 25  [0.719] 400
ELO difference: 162.99 +/- 36.53
Finished match
PlyCount: Total = 400  Range: 5-247  Average = 91.68  StdDev = 43.77
Av.nr.blunders = 4.58
Av.ELO per blunder: 35.6
ELO = 1784
Takebacks: 42.6

A 2700 GM would need to take back thus defined blunders 5 times to reach 3250-3300 FIDE ELO level of Komodo on 24 cores. A 2476 GM would need 8-9 takebacks. A 2324 player about 14 takebacks, A 2185 about 18 takebacks. So on, and a 1784 ELO player would need 43 takebacks. If the blunder (or large inaccuracy) is on average still better than random move, there is a multiplicaive factor on all these numbers of takebacks. One can probably determine pretty accurately what sort of handicap is needed for what strength, say if a ELO 2700 player needs 10 takebacks, then a 2200 player would need 35 takebacks, and a 1900 player needs to take back all moves at least once.

To me it would make for good handicap matches because it eliminates one of the main human weaknesses - blunders an serious inaccuracies. If only GMs are not offended by this proposal, I don't know their tastes.

I don't think GMs would be offended by this idea, it's no more embarassing than playing with material odds. Rather, the problem is that I don't think GMs will have any confidence that takebacks will help much. I don't think your simulation has much to do with the Komodo vs GM situation. Usually GMs make small mistakes, not outright blunders. Sometimes they are apparent after the computer responds, sometimes not.
I suppose any takeback handicap would stipulate that a player may retract his move any time until he makes another move, and if he does so he does not get back any time but Komodo should get back the time of the last move (this may be impractical though). It should probably also stipulate that you cannot take back the same move number more than once. I don't know how to simulate all this, but I think that your numbers are way too optimistic for the human side. In my opinion, even with unlimited takebacks (subject to my above rules) an ordinary GM would have little chance to win a match vs. Komodo. Maybe a top ten player might have a fair chance; even then he might need some other edge like White pieces or short opening book. I'm willing to try the idea if some GM has enough confidence to play knowing he won't make much money if he loses all the games.

I see no reason for the rule that you cannot take back the same move number more than once.

Even with taking back unlimited number of times and taking the same move back unlimited number of times it is not clear that the human can draw or win if the human has limited time to finish the game and the question is what limitation is fair.

For example it may be possible to decide that the computer use 5 seconds per move and the human has 2 hours to finish the game.

If the human play at average speed of 5 seconds per move then
a move by both players can take 10 seconds and the human can practically play 720 moves but with all the take backs it is not clear that it is going to be enough for him even to get a draw.

In order to prevent winning on time in an obvious drawn position it is possible to decide that the human is allowed to play with no take back with 5 seconds per move from the final position that the sides achieved after 2 hours.


Edit:
I have no idea what time control is fair for the human in these conditions.
It is possible that 2 hours is too generous for a GM but in this case it is possible to fix it by less time or by a weaker opponent that is not a GM.

In theory we could do something like this, but we also have to consider what is acceptable/interesting for spectators. Taking the same move back more than once would probably really annoy the spectators and/or commentator, and likely the GM would be embarrassed to do it. Probably unlimited takebacks (without more than one per move number) with some moderate time odds and/or hardware restriction would make for a fair match with a 2600 GM, but it would take trial and error to determine what those odds would be.

[mehmet karaman]

I think Komodo use 4 seconds per move at Erenburg match

4* 1.000.000 kn = 4.000.000 (kn/move of Komodo at Erenburg match )

With 4.000.000 kn/move Komodo easily beat (3.5-0.5) a 2591 elo chess player (with 3 opening book, black pieces, no tablebases, low contempt)

I guess how many kn/move is needed for Komodo to play a 2600/2700 elo human level ( strong opening book, white/black pieces, tablebase on)

If we know this number ( kn/move) we easily know the human elo of Komodo with strong hardwares