@e4 said in #10:
> Imagine if the article came to this as one of it's conclusions...Oh wait, it did.
Yes, it pretty much did, but I wanted to point out that high precision also depends on players style (Zukertort vs Capablanca) and very high/low ACPL even in pre-computer or Victorian era does not (necessarily) indicate high/low level of play.
> Imagine if the article came to this as one of it's conclusions...Oh wait, it did.
Yes, it pretty much did, but I wanted to point out that high precision also depends on players style (Zukertort vs Capablanca) and very high/low ACPL even in pre-computer or Victorian era does not (necessarily) indicate high/low level of play.
@RoyalHog said in #13:
> Yes, it pretty much did, but I wanted to point out that high precision also depends on players style (Zukertort vs Capablanca) and very high/low ACPL even in pre-computer or Victorian era does not (necessarily) indicate high/low level of play.
Fair enough, good point - sometimes it feels like people comment without reading the article though :P
> Yes, it pretty much did, but I wanted to point out that high precision also depends on players style (Zukertort vs Capablanca) and very high/low ACPL even in pre-computer or Victorian era does not (necessarily) indicate high/low level of play.
Fair enough, good point - sometimes it feels like people comment without reading the article though :P
Having done scientific data analysis myself, I really do appreciate the effort and thoroughness of this article! It was a very nice read, thank you to all who contributed.
I find it a pity that the first charts only indicates the WC and not the challenger. It takes two players to play a chess game!
On another note, training with engines can obviously help prepare accurate openings, and I guess it can also help develop the endgame theory, but I would imagine the accuracy of the middlegame is entirely up to the human playing it and 'should not' be affected by computers?
On another note, training with engines can obviously help prepare accurate openings, and I guess it can also help develop the endgame theory, but I would imagine the accuracy of the middlegame is entirely up to the human playing it and 'should not' be affected by computers?
Was the default depth of server analysis used? Default one can on rare occasion misunderstand GM's classic games.
For example in this famous Kasparov's game (lichess.org/QB9oj20r#46). His sacrifice on move 24 is considered a -2.8 mistake by server analysis, but when I let it run a bit longer, on depth 22 it already sees the sacrifice as the best move with a rating of -0.5. Then server analysis is stubborn for next 2 half-moves, giving -2.9 and -3 rating, while on slightly deeper analysis it's +0.5. Then the server analysis somewhat correctly evaluates at +1, but on next half-move goes back to -3, then back to +1.5, while the local analysis at depth 30 is giving +6. Another moment is move 28, where local analysis at depth 30 gives +7.5 to Ra7, but Garry moves Qc3 (+0.6), local analysis misses that and doesn't count that as a mistake.
That's probably a very rare type of a game, and maybe not a single game in world championships has any similar miscalculation by medium-low depth engine, but it still would be interesting to see how the results would change if you throw more computational power at this thing.
For example in this famous Kasparov's game (lichess.org/QB9oj20r#46). His sacrifice on move 24 is considered a -2.8 mistake by server analysis, but when I let it run a bit longer, on depth 22 it already sees the sacrifice as the best move with a rating of -0.5. Then server analysis is stubborn for next 2 half-moves, giving -2.9 and -3 rating, while on slightly deeper analysis it's +0.5. Then the server analysis somewhat correctly evaluates at +1, but on next half-move goes back to -3, then back to +1.5, while the local analysis at depth 30 is giving +6. Another moment is move 28, where local analysis at depth 30 gives +7.5 to Ra7, but Garry moves Qc3 (+0.6), local analysis misses that and doesn't count that as a mistake.
That's probably a very rare type of a game, and maybe not a single game in world championships has any similar miscalculation by medium-low depth engine, but it still would be interesting to see how the results would change if you throw more computational power at this thing.
@RoyalHog said in #7:
> The most accurate games will be those that end as a draw right after the opening and are essentially home analysis.
Long simplified games can also produce highly accurate games as small inaccuracies will be avaragede out over many moves moves - or at least that's my understanding on how ACPL is calculated
> The most accurate games will be those that end as a draw right after the opening and are essentially home analysis.
Long simplified games can also produce highly accurate games as small inaccuracies will be avaragede out over many moves moves - or at least that's my understanding on how ACPL is calculated
Fair point. Maybe it would be interesting to consider the maximum centipawn loss in one move over the whole game. Or in between the L1 norm and the L-infinity norm, there is plenty of space.
<Comment deleted by user>
Thanks for the complete collection of world championship games with analysis. This is a treasure!
This topic has been archived and can no longer be replied to.