lichess.org
Donate

Add-on to Chess insights: comparing OTB rating with average centipawn loss

Hello,

I see on chess.com that they are creating your feature of analyzing one's game with the "average centipawn loss". They are far away from what you have done and they make you pay for it, so I try to improve your site with a suggestion.

It would be very interesting to know the correlation between OTB rating and "average centipawn loss", for instance 25 "a.c.l." in 100 games correspond to 2200 OTB.... I understand that it is not easy to make this correspondance, so it would have some kind of imprecision, but it would be still very interesting i think.
However, I think it would be easy for you to analyse all games of all GM and IM and draw the correspondance. Of course, it would tell the correspondence for long games, so in blitz games, a coefficient could be applied... but still, analyzing rapid and blitz GM and IM games, you could estimate broadly this coefficient (for bullet, not sure all this has importance of meaning, this kind of chess is just for fun i would say).

It would be of immediate use for a player who want to attain 2200: he knows his goal quantitatively in termes of accuracy of play and he knows at what REAL strength he is playing.
Measuring CPL / CAPS is predicated on two false premises:
1. Ideal human play is "accurate"
2. Amateur players improve by playing more like engines
3. Averages (or sums) are the a useful metric (compared to other possible metrics such as quartile/quintile/decile/etc. ranges)

I have a few remarks:
Amateurs need some way of measuring their opening, middlegame, and endgame mastery. Measuring CPL / CAPS isn't it.

GMs like Tal who play "inaccurate" moves are still immensely strong chess players.

Stockfish doesn't always return the same evaluation for the same position, nor does the UCI protocol require it to. "Accuracy" presumes there exists a correct evaluation for each position!

Human players do not improve by trying to emulate engine play. (If humans and engines played the same way, it would be impossible to detect cheaters.)
Thanks for your very interesting answer.

OK there are exceptions:
- with players like Tal
- with computers who sometimes do not converge "correctly" towards "the" evaluation of the position.

But maybe facts coud indicate that these exceptions can be forgotten, i mean broadly, at 1st order (i speak like a physician here), because:
- Do carlsen and the top 10 have the best "average centipawn loss"? i would think that carlsen is the best player considering "average centipawn loss"
- This non-convergence of computer, how many times does it occur? is it not a scarce phenomenon, let's say in 95% of the cases after 30s the evaluation has converged (for instance, 10% close to the evaluation obtained after 1h)?

I mean, scientifically, if your define an objective tool which correlates with the subjective ELO, your tool is good and this strong result will have a lot of interesting consequences. So my question is simple and easy: does CPL correlate with ELO? this is not a philosophical issue, this is just about statistics.

And in fact, it seems there is one: http://chess-db.com/public/research/qualityofplay.html

And so, a 2200 is about 29 CPL, so I think chess-db has done all the job that i am asking here! (incredible result i would say).
Some doc about the subject :

http://www.chessanalysis.ee/Quality%20of%20play%20in%20chess%20and%20methods%20for%20measuring.pdf
http://www.chessanalysis.ee/a%20study%20on%20chess%20strength.pdf
(both from here : http://www.chessanalysis.ee/chessanalysiseng.htm )

ChessGameReport :
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=234079#pid234079
http://rybkaforum.net/mwf/rybkaattach/68/234068/ReadMe.txt

Chess Game Analyzer :
http://www.talkchess.com/forum/viewtopic.php?p=621764#621764
http://www.talkchess.com/forum/viewtopic.php?p=622079#622079
http://www.talkchess.com/forum/viewtopic.php?p=681976#681976

Chessbase :
en.chessbase.com/post/grandmaster-blunders-a-statistical-analysis
en.chessbase.com/post/computers-choose-who-was-the-strongest-player-

Eval , Win Percentage, and ELO
chessprogramming.wikispaces.com/Pawn+Advantage,+Win+Percentage,+and+ELO
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=15144
First, the rating system is Elo (named after Arpad Elo) or Glicko-2 (named after Glickman), not ELO.

Second, you've ignored or rejected my premise (human players do not improve by attempting to emulate engine play) so of course we're going to arrive at different conclusions.
My conclusion is that there IS a correlation between CPL and Elo rating (by the way, you can write my name in capital letters, or just the first letter... so for me Elo and ELO are completely acceptable). This conclusion is not dependent on your assumption.

Your assumption meets the demand of how to improve one's play from this conclusion.

But I am not sure whether your assumption is true or not. For instance, i have heard/read that professional players tend to be far more tactical due to Fritz or engines, and that a number of "inhuman" opening novelties (like g4 in QG) come directly from computer analysis. Above all, players teach themselves by analyzing and playing against engines, so obviously, they tend to play like them, but you are right, more unconsciously than consciously...

This topic has been archived and can no longer be replied to.