Maybe something weird. In the game Alekhine Vs Conel Alexander (Nottingham, 1936) the engine analysis detects 0 blunders, 0 innaccuracies and 0 mistakes for both players. And Alekhine won that game in a brillant, tactic way. Usually lichess is enough accurate to detect clear tactical blunders. Someone can explain?
Here you have the pgn
[Event "Casual game"]
[Site "
lichess.org/OJtWyaSc"]
[Date "2014.11.30"]
[White "Anonymous"]
[Black "Anonymous"]
[Result "1-0"]
[WhiteElo "?"]
[BlackElo "?"]
[PlyCount "53"]
[Variant "Standard"]
[TimeControl "-"]
[ECO "E11"]
[Opening "Bogo-Indian Defense, Grünfeld Variation"]
[Annotator "
lichess.org"]
1. d4 Nf6 2. c4 e6 3. Nf3 Bb4+ 4. Nbd2 { Bogo-Indian Defense, Grünfeld Variation } b6 5. g3 Bb7 6. Bg2 O-O 7. O-O Bxd2 8. Qxd2 d6 9. b3 Nbd7 10. Bb2 Rb8 11. Rad1 Ne4 12. Qe3 f5 13. d5 exd5 14. cxd5 Ndf6 15. Nh4 Qd7 16. Bh3 g6 17. f3 Nc5 18. Qg5 Qg7 19. b4 Ncd7 20. e4 Nxe4 21. Qc1 Nef6 22. Bxf5 Kh8 23. Be6 Ba6 24. Rfe1 Ne5 25. f4 Nd3 26. Rxd3 Bxd3 27. g4 { Black resigns } 1-0
Lichess considers an innacuracy starts at 50 centipawns.
Alekhine has an average loss of 11 centipawns per move while Conel Alexander loses an average of 21 centipawns.
So in this game, there are no tactical blunders/mistakes or even inaccuracies but Alekhine manages to slowly grind out a win with slightly better play.
Yeah, there are a couple points to be made.
1) As conor34 pointed out, the average centipawn loss is MUCH more revealing about quality of play than the inaccuracy/mistake/blunder classification.
Against perfect, 0 centipawn loss play, losing 30 centipawns a move would mean being -3.00 after just 10 moves out of book, hardly perfect play, even though no moves would fall into the inaccuracy/mistake/blunder buckets.
2) A few of Alexander's moves really were a bit more than -0.5, but it takes some deeper searching to classify them as such.
The main point is that it's actually remarkably easy to lose a game of chess without any single move being worse than a 50 centipawn loss, especially at lower search depths.
At the risk of beating a dead horse, if I consistently play moves that lose 10 centipawns, and you consistently play moves that lose 20 centipawns, over 30 moves that'll result in your being down a piece or equivalent.
Moral of the story: use the centipawn loss for individual moves to judge the quality of the moves, and average centipawn loss for the quality of a game, not the inaccuracy/mistake/blunder categories.
Categories are nice, but it takes a lot more than avoiding having moves in these categories to play a perfect game :)
Thank you so much for the info, but, if I understand correctly, for example in the Karpov Kramnik in Linares 1994, the average loss for Karpov was 5 centipawns and 6 for Kramnik. After 40 moves, Karpov must have a theoretical +0.40 right?, but lichess points a +1.1 in the position Kramnik resgined. Im skipping something?
(sorry if is a silly question, Im new in this)
No worries...that's a perfectly sensible question.
There are 2 big reasons for discrepancies like that.
The first, less significant reason is White's normal opening advantage. If white starts at, say, +0.3, then after white "earns" 0.4 more in advantage, then you'd actually expect it to be +0.7, not +0.4.
The second, more significant reason is just that looking a little bit deeper into the position can change the evaluation.
So, let's say I force an engine to analyze all the legal moves in a position, and it gets to depth 15 on all of them, and the best move is +0.2 in my opponent's favor, and the move I played is +0.25 in my opponent's favor.
I get charged a 5 centipawn loss. This much is clear. However, we now go one move further into the game to analyze my opponent's reply.
We again get to depth 15 for all the moves, but now with the benefit of a little extra search depth, my opponent's best move is actually +0.5 for him, and not the +0.25 we got when we analyzed this position from one move earlier.
That means that while I only get charged for a 5 centipawn loss, the position actually got worse by 30 centipawns. The first search just wasn't deep enough to realize that.
In the game you mention, you can see a few such points in the game. These moments where the first search is not deep enough introduce some randomness into the relationship between the average error of each side and the final evaluation.
On a practical note, this is why automatic analysis with an engine is usually more stable if it works through the game backwards.
When you do that, the searches for the moves later in the game, which are in one sense "deeper" than the searches earlier in the game, populate the hash table first, so you see fewer of these shallow search-induced evaluation swings.
Hopefully that makes sense :)
