Tuesday 29 October 2019

Ever get the feeling you've been cheated - wrong in fact, wrong in theory?

So how does Chess.com's system work anyway?

I don't really know, and I don't particularly want to speculate, not more than I'm obliged to. It ought to be up to them to explain themselves, not up to me.

But I also don't know
  • whether that system has been assessed independently, and even if so, how thoroughly and how expertly
  • how much it risks (and is understood to risk) catching the wrong people as well as the right ones
  • how much its reliability may vary (and is understood to vary) according to the sample size of games
  • how much it may depend (and is understood to depend) on fallible human inputs, human judgments and so on.
I don't know. But I do know that Chess.com aren't in possession of a foolproof system. Of course they aren't, because there's no such thing as a foolproof system. And I do know that they are wrong in this particular instance. Spectacularly wrong.

What I think, however, is that their method to some degree involves looking at the moves you have played, and seeing how many match with the preferred choice of a computer program. Whether they do anything else, or what precisely their criteria are, who knows. (But how reliable those criteria are - on that, I do have a well-informed opinion.)

One question this raises is - since there is such a thing as theory in chess, when in the game do they start scrutinising? Presumably not on move one. But if not, at what point does the matching begin? If they start too early, when in fact you're still in book (because book use is permitted in these games) isn't that a point where errors can be committed? Because moves which you're finding from a printed source are being marked down as moves you're finding with a program?

Let me give you an example. Let me give you several examples.

When you finish a game on Chess.com, you get a little game report, which includes some basic computer analysis, and a chart that looks like this.

What it means precisely, I couldn't say, but I can guess what Best Move means, and what Book means. And I can guess that 99.3 is a high figure, whatever it means precisely and however they're calculating it. The game it refers to is this one.

[Site "Chess.com"]
[Date "2019.07.01"]
[White "passy234"]
[Black "Justinpatzer"]
[Result "0-1"]
[WhiteElo "2055"]
[BlackElo "2149"]
[EndDate "2019.07.05"]
[Termination "Justinpatzer won by resignation"]
1. Nf3 d5 2. g3 Bg4 3. Bg2 c6 4. c4 e6 5. O-O Nf6 6. Qb3 Qb6 7. Qc2 Nbd7 8. cxd5 exd5 9. d3 Qc5 10. Qb3 Qb6 11. Qc2 Qc5 12. Qd1 Bd6 13. Nc3 O-O 14. Be3 Qa5 15. a3 Rfe8 16. b4 Qd8 17. Rc1 a5 18. Qb3 Qe7 19. Rb1 axb4 20. axb4 Ne5 21. Nxe5 Bxe5 22. Rfe1 d4 0-1

So we've got a twenty-two move minature, in which Black plays five moves of theory, and then turns over White in short order with an extremely high Accuracy rate. Which is pretty suspicious, isn't it?

Except it isn't. Because this, which suggests that theory ends after five moves on each side

and which would mean that the players were playing their own moves from this position

is wrong. Very wrong.

In fact Black was playing published theory until move sixteen.

Specifically, he was following Petrosian v Vovhannisyan, Lake Sevan 2015, which you can see below (to move 14, but as there was a repetition, we had played two more moves apiece) as it appears on page 202

of Delchev and Semkov, Attacking the English/Reti, Chess Stars, 2016

which I have on my bookshelves.

Which is how I came to be in the position below, after Black's 16...Qd8, before I had to play any moves of my own.

White then varied with 17 Rc1. So my original contribution consisted of five moves - five very ordinary moves - and then, after a simple blunder by White

a very obvious pawn fork to win the game.

Suddenly the game looks very different, doesn't it? Suddenly it's perfectly normal, unexceptional. Suddenly there's nothing odd about it at all.

Here's a few more examples. This game, for instance.

[Site "Chess.com"]
[Date "2018.12.25"]
[White "Dave_1969"]
[Black "Justinpatzer"]
[Result "1/2-1/2"]
[WhiteElo "2015"]
[BlackElo "2096"]
[EndDate "2019.01.08"]
[Termination "Game drawn by agreement"]
1. d4 Nf6 2. c4 e6 3. Nc3 Bb4 4. e3 O-O 5. Ne2 d5 6. a3 Bd6 7. c5 Be7 8. b4 b6 9. Nf4 c6 10. Be2 Nbd7 11. Nd3 a5 12. Bd2 Ba6 13. O-O Qc7 14. f4 axb4 15. axb4 Bc4 16. Qc2 Rfb8 17. Rfb1 Qb7 18. Bf3 Ne8 19. Ne5 Nxe5 20. fxe5 Ra6 21. Rxa6 Qxa6 22. Qd1 Nc7 23. Ra1 Qb7 24. e4 b5 25. exd5 cxd5 1/2-1/2

According to Chess.com, we have seven moves of theory for White, and six for Black

which is a funny thing to say anyway, since Black's seventh is forced. But leaving that aside, it would get us this position,  where Black is apparently starting from scratch.

But he's not starting from scratch. he's not even halfway there. Because in my theory here goes as far as move twenty, as per Guliev-Arkhipov, Dubai 1999. You may have to hunt about a bit there to locate the score, or you can find it, as I did, on page 174

of Emms, The Nimzo-Indian move by move, Everyman, 2011

which I have on my bookshelves and which brought me to this position without yet having had to play any of my own moves.

After 21 Rxa6 I once again played five moves - in my view even less distinguished moves than in the previous example - and I then asked for a draw.

(Why would Black, by the way, be using computer assistance in order to play a dull draw in twenty-five moves? Nobody knows. But none of this makes sense.)

Then there's this game, another draw as it happens.

[Site "Chess.com"]
[Date "2019.01.08"]
[White "rembooooo"]
[Black "Justinpatzer"]
[Result "1/2-1/2"]
[WhiteElo "2076"]
[BlackElo "2111"]
[EndDate "2019.01.22"]
[Termination "Game drawn by agreement"]
1. e4 e5 2. Nf3 Nc6 3. Bc4 Bc5 4. O-O Nf6 5. c3 Nxe4 6. d4 exd4 7. cxd4 d5 8. dxc5 dxc4 9. Qe2 Qd3 10. Re1 f5 11. Nc3 O-O 12. Nxe4 fxe4 13. Qxe4 Bf5 14. Qf4 Be6 15. Ne5 Qd5 16. Qg3 Nxe5 17. Qxe5 Qxe5 18. Rxe5 Rfe8 19. f4 Rad8 20. Be3 Bd5 21. Bd4 Bc6 22. Rxe8+ Bxe8 23. Bc3 Bc6 24. Re1 Kf7 25. Re5 Rd1+ 26. Kf2 Rc1 27. g4 Rh1 28. Kg3 Rf1 29. Rf5+ Kg8 30. Rg5 g6 31. f5 Rf3+ 32. Kh4 Kf7 33. fxg6+ hxg6 34. Re5 Rf2 35. h3 Rf3 36. Re1 Bd5 37. a4 c6 38. a5 a6 39. Re2 Rd3 40. Rf2+ Rf3 41. Re2 Rd3 42. Rf2+ Rf3 1/2-1/2

in which according to Chess.com

we played five moves of theory for White, and four for Black, reaching this position.

In fact Black at least was following all fourteen moves of a recommended line on page 379

of Bologan, Bologan's Black Weapons (etc), New In Chess, 2014

so original play only started here.

Again, why Black would user computer assistance to play forty-two dull moves for a draw is a question for Chess.com. It's not a question they seem keen on answering.

One more, for now.

[Site "Chess.com"]
[Date "2019.07.07"]
[White "MRValero"]
[Black "Justinpatzer"]
[Result "1-0"]
[WhiteElo "2182"]
[BlackElo "2167"]
[EndDate "2019.07.26"]
[Termination "MRValero won by resignation"]
1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. Qe2 Bc5 5. c3 O-O 6. O-O Re8 7. d3 a6 8. Ba4 h6 9. Be3 Bxe3 10. fxe3 d6 11. Nbd2 b5 12. Bb3 Be6 13. Bc2 Qe7 14. h3 d5 15. exd5 Nxd5 16. Qf2 g6 17. e4 Nf4 18. Nxe5 Nxh3+ 19. gxh3 Nxe5 20. Qg3 Qg5 21. Qxg5 hxg5 22. d4 Nc4 23. Nxc4 Bxc4 24. Rf2 Be6 25. Kh2 c6 26. Rg2 Kg7 27. Rxg5 Rh8 28. Rg3 Rh4 29. Rf1 a5 30. a3 b4 31. axb4 axb4 32. Rf2 b3 33. Bd3 Ra1 34. Rd2 Rc1 35. Re3 g5 36. Rg2 Kh6 37. Rg1 Rxg1 38. Kxg1 g4 39. hxg4 Kg5 40. Kg2 Rxg4+ 41. Rg3 Kf4 42. Rxg4+ Bxg4 43. Kf2 f6 44. c4 c5 45. dxc5 Bd7 46. Be2 Kxe4 47. c6 Bc8 48. Bd1 Kd4 49. Bxb3 Kc5 50. c7 Kd6 51. Bc2 Kxc7 52. Ke3 Bh3 53. b4 Kc6 54. Bd3 Kd6 55. Kf4 Kc6 56. Ke3 Kd6 57. Kd4 1-0

You may remember some fragments from this game in the previous posting (and as I asked then, why would Black be using a computer to lose a long game with lots of errors in it?) but in this one, theory supposedly takes us to move five

which is this position.

But in fact, I was following, by transposition, Meijers v Crouan, Sautron 2009. No link, sorry, but here it is on page 99

of Lysyj and Ovetchkin, The Berlin Defence, Chess Stars, 2012

so I was still in theory up to here.

I've not been through all my games for this exercise, by the way, wishing to bore neither myself nor the reader. But even so, I'm aware that there are more examples than just these four.

I don't, of course, know for sure - or really at all - how Chess.com come to their conclusions, nor what role may be played by the Accuracy or Book figures.

I don't know, because they do not say. But I know that some of what they do say is a manifest nonsense.

And what happens if you try to draw conclusions from nonsensical data?

[For the record, I referred the first two of these queries to Chess.com, and they ignored them. I've added the other two to this piece, because why not.]


Avital Pilpel said...

A friend of mine, a master, just won a blitz game on chess.com with "98% accuracy". As he showed me, the only "blunder" chess.com thinks he made was the horrible move 1.Nf3 ... everything else was "pefect". The chess.com accuracy algorithm is nonsense.

Avital Pilpel said...

*Perfect*, not "pefect", of course.

Ammerbucher said...

I suppose this is an urban legend.
1. Nf3 surely was not a "blunder" but a "book move", also at chess.com. (Btw, also the category "perfect" does not exist there, but I would not stick to that wording ..).
Maybe you can ask your "friend" for a screenshot of the analysis, then I would believe it ...