Why is using a logarithmic curve a case of overfitting, but a linear curve is not? To me it seems too arbitrary. I did not want to imply that the combination the engine and hardware is too weak to be used in analysis of top GM games. What I said is that I cannot make trustworthy extrapolations past 3100 ELO mark.perhaps as he says due to overfitting or to the insufficient strength of the "judge" as you point out.
Ok, my bad But the analogy still is valid; just as there are a lot of low-elevation areas besides mountains, so are blunders accompanied by fairly accurate moves. It important that all data be taken into account when analyzing something.You missread my quote above. I said blunder 'wildly', not blunder in wild positions.
It also seems we have different definitions of 'blunder'. For me this must not even be a full pawn equivalent. It can also be a permanent weakness,
like spoiling the pawnshield on Kside or allowing an outpost without need or whatever...
I think I have enough experience to know a thing or two about it. It's not about beliefs, I've seen the data with my own eyes. Just a hobby player with no official rating. But conducting analysis projects like this is even bigger a hobby and I take it more seriously.The frequency is horrible. I have no idea why you don't believe me.
May I ask if you are a chess player too?
There are much too less data points in CCRL at least in 40/40 for programs under 2100 to make any comparisons to Human ratings.
I did consider using 40/4 rating list as a basis, but ended up rejecting the idea, as I feared that it might have too little relevance with respect to modern computer chess.
The methods used in that are by now outdated and not relevant any more.BTW in that link you practically had it all the other way round for lower
CCRL ratings?
How do you know that both tables are way off, if I may ask?I must say both tables are way off in most parts of it.
There is simply not enough data there for doing such a comparison.
Theoretically expanding the range is not impossible. My 2014 database has only 3 games with both players within 1575-1625 range. Perhaps I could add games from 2013 and 2015 to widen the range a little so as to get enough games and also include games from 2013-2015 by players who got TPR of 3000-2900 in tournaments and matches. But that data would be incompatible with the rest of my data. I'm not sure if it would be a good idea.