correspondence chess in the age of NNUE

[Ferdy]

I tried to simulate with latest stockfish on the Arno Nickel rule for 3/4 points for stalemate and 1 piece ahead vs bare king, still the draw ratio is high. The major problem is the repetition draw result.

One idea is for tournament like round-robin, the players that played into a repetition can be assigned a score of 0. Then also apply the Arno Nickel rule. It would at least discourage players from making repetition quickly. Opening preparation is somewhat compromised.

Yes, repetition has to be addressed to make a major reduction in the draw percentage. Arno Nickel and I are now in agreement that in addition to the stalemate and minor vs. king rule, repetition should also be scored as 1/4 for the repeater and 3/4 for the opponent. Based on some tests with Komodo I estimate that the three rules together will roughly double the win percentage in high level engine vs engine play. Of course at long time limits and many threads even doubling the win percentage will still leave most of the games drawn.

This one "repetition should also be scored as 1/4 for the repeater and 3/4 for the opponent" is good. Have you worked out on what happens if there is a check? How to identify the repeater?

[lkaufman]

I tried to simulate with latest stockfish on the Arno Nickel rule for 3/4 points for stalemate and 1 piece ahead vs bare king, still the draw ratio is high. The major problem is the repetition draw result.

One idea is for tournament like round-robin, the players that played into a repetition can be assigned a score of 0. Then also apply the Arno Nickel rule. It would at least discourage players from making repetition quickly. Opening preparation is somewhat compromised.

Yes, repetition has to be addressed to make a major reduction in the draw percentage. Arno Nickel and I are now in agreement that in addition to the stalemate and minor vs. king rule, repetition should also be scored as 1/4 for the repeater and 3/4 for the opponent. Based on some tests with Komodo I estimate that the three rules together will roughly double the win percentage in high level engine vs engine play. Of course at long time limits and many threads even doubling the win percentage will still leave most of the games drawn.

This one "repetition should also be scored as 1/4 for the repeater and 3/4 for the opponent" is good. Have you worked out on what happens if there is a check? How to identify the repeater?

Check doesn't matter. The player who would first be eligible to claim a draw by repetition is the repeater, and gets 1/4 point. In practice I think that this will almost always mean that a player giving perpetual check will get just the 1/4 point. It's a clean rule; the guy who actually causes the draw takes the hit.

[MikeB]

One can always take up genealogy...

[peter]

Check doesn't matter. The player who would first be eligible to claim a draw by repetition is the repeater, and gets 1/4 point. In practice I think that this will almost always mean that a player giving perpetual check will get just the 1/4 point. It's a clean rule; the guy who actually causes the draw takes the hit.

That might work for human players, say corr., neglecting the possibility for those to simply make draw per mutual consent instead of repeating.
But what would stop engine- engine- games according to this new scoring then? Engines would simply not claim the rep., neither play it, would they?

I mean, to avoid the repetition of the one and the same position wouldn't mean to make the drawn game end other then drawn later on by any other rule, e.g. lack of material or stalemate. I don't think the latter was of as little meaning as it is now, when the many drawn rook endings are not played till the very end with tbs because of their stalemate endings. I guess avoiding these endgames by corr.- players and by engines programmed for the new stalemate- rule and the different evaluated tbs yet would change more than simply avoiding repetition. Arno Nickel described many of those now theoretically drawn endgames leading to stalemate positions in his open letters.

Not quite understanding your repetition- scoring to do as much as that regards

[lkaufman]

Check doesn't matter. The player who would first be eligible to claim a draw by repetition is the repeater, and gets 1/4 point. In practice I think that this will almost always mean that a player giving perpetual check will get just the 1/4 point. It's a clean rule; the guy who actually causes the draw takes the hit.

That might work for human players, say corr., neglecting the possibility for those to simply make draw per mutual consent instead of repeating.
But what would stop engine- engine- games according to this new scoring then? Engines would simply not claim the rep., neither play it, would they?

I mean, to avoid the repetition of the one and the same position wouldn't mean to make the drawn game end other then drawn later on by any other rule, e.g. lack of material or stalemate. I don't think the latter was of as little meaning as it is now, when the many drawn rook endings are not played till the very end with tbs because of their stalemate endings. I guess avoiding these endgames by corr.- players and by engines programmed for the new stalemate- rule and the different evaluated tbs yet would change more than simply avoiding repetition. Arno Nickel described many of those now theoretically drawn endgames leading to stalemate positions in his open letters.

Not quite understanding your repetition- scoring to do as much as that regards

I suppose the actual wording should read something like: If a move is played that repeats the position for the third time, either player may call it a repetition, with the repeating side getting 1/4 point, the other side 3/4 point. For engine vs. engine play I suppose it would be automatically scored as such. The main point isn't to avoid repetitions played to avoid a no draws by agreement rule, it is that testing has shown that prohibiiting (or penalising) repetitions reduces draws far more than the stalemate + bare king rules do. There are many drawn endings where if you aren't allowed to repeat, you will lose (many even material King and pawn endings would be in this category). I don't claim that this rule avoids all draws or even most draws; I do claim that it roughly doubles the percentage of decisive games when the level of play is high enough that that percentage is a fairly small number. If you want to eliminate all or even most draws we would need more rules.

[Ozymandias]

This one "repetition should also be scored as 1/4 for the repeater and 3/4 for the opponent" is good. Have you worked out on what happens if there is a check? How to identify the repeater?

Check doesn't matter. The player who would first be eligible to claim a draw by repetition is the repeater, and gets 1/4 point. In practice I think that this will almost always mean that a player giving perpetual check will get just the 1/4 point. It's a clean rule; the guy who actually causes the draw takes the hit.

I haven't looked that up, but just taking the two threefolds we've seen so far in the TCEC final, we have the side giving the check being the last one to move and the opposite case too:

[pgn][Event "TCEC Season 20 - Superfinal"]
[Site "https://tcec-chess.com"]
[Date "2021.01.19"]
[Round "22.1"]
[White "LCZero 0.27.0d-Tilps-dje-magic_JH.94-100"]
[Black "Stockfish 20210113"]
[Result "1/2-1/2"]
[WhiteElo "3586"]
[BlackElo "3601"]
[ECO "B20"]
[Opening "Sicilian"]
[Variation "wing gambit, Marshall variation"]
[TimeControl "7200+10"]
[PlyCount "86"]
[GameDuration "03:43:01"]
[GameEndTime "2021-01-19T22:34:10.251 UTC"]
[GameStartTime "2021-01-19T18:51:08.869 UTC"]
[TerminationDetails "3-Fold_repetition"]

1. e4 c5 2. b4 cxb4 3. a3 e5 4. Bb2 Nc6 5. Nf3 Qb6 6. axb4 Qxb4 7. Qc1
Qxe4+ 8. Be2 Qg6 9. Nc3 Be7 10. Nb5 Rb8 11. O-O d6 12. d4 a6 13. Nc7+ Kf8
14. d5 Nd8 15. Qe3 Bd7 16. Qb6 Rc8 17. Rfe1 Bh3 18. Bf1 h5 19. Ra3 Rh6 20.
Rc3 h4 21. Ree3 Bd7 22. Rc4 h3 23. g3 f5 24. Rec3 f4 25. Ne6+ Bxe6 26. Rxc8
Bxc8 27. Rxc8 fxg3 28. fxg3 Kf7 29. Qc7 Rh5 30. Rxd8 Qe4 31. Nh4 Qe3+ 32.
Kh1 Qe1 33. Kg1 Rxh4 34. Rxg8 Qe3+ 35. Kh1 Qe4+ 36. Kg1 Qe3+ 37. Kh1 Qe4+
38. Kg1 Rg4 39. Bxh3 Qe3+ 40. Kg2 Qe2+ 41. Kg1 Qe3+ 42. Kh1 Qe4+ 43. Kg1
Qe3+ 1/2-1/2

[/pgn]
[pgn]
[Event "TCEC Season 20 - Superfinal"]
[Site "https://tcec-chess.com"]
[Date "2021.01.23"]
[Round "47.1"]
[White "Stockfish 20210113"]
[Black "LCZero 0.27.0d-Tilps-dje-magic_JH.94-100"]
[Result "1/2-1/2"]
[WhiteElo "3601"]
[BlackElo "3586"]
[ECO "B12"]
[Opening "Caro-Kann"]
[Variation "advance variation"]
[TimeControl "7200+10"]
[PlyCount "134"]
[GameDuration "03:55:39"]
[GameEndTime "2021-01-23T14:19:24.216 UTC"]
[GameStartTime "2021-01-23T10:23:44.869 UTC"]
[TerminationDetails "3-Fold_repetition"]

1. e4 c6 2. d4 d5 3. e5 Bf5 4. h4 h6 5. g4 Bd7 6. h5 e6 7. f4 c5 8. c3 Qb6
9. Nf3 Nc6 10. Be2 f6 11. Nh4 Nge7 12. dxc5 Qxc5 13. b4 Qb6 14. b5 Na5 15.
Qd4 fxe5 16. fxe5 Rg8 17. Nd2 Qc7 18. O-O Rc8 19. Qf4 Kd8 20. Ndf3 Nc4 21.
Bd3 g5 22. hxg6 Bg7 23. Re1 Qa5 24. Qd4 Bxb5 25. Rb1 Rc7 26. g5 h5 27. Qf2
Rf8 28. Be3 b6 29. Qe2 Be8 30. Rbc1 Bd7 31. Bd4 Qa3 32. Rb1 Rf4 33. Rf1 Be8
34. Kh1 Rg4 35. Rb3 Qa4 36. Rb4 Qa5 37. Rbb1 Kc8 38. Rf2 Kb8 39. Rg1 Nc6
40. Ng2 Qa4 41. Rgf1 Rc8 42. Re1 Rc7 43. Bb1 Rc8 44. Bd3 Kb7 45. Rb1 Rc7
46. Ne3 N4xe5 47. Bxe5 Nxe5 48. Nxe5 Rh4+ 49. Kg2 Bxe5 50. Rb4 Rxb4 51.
cxb4 Bg7 52. Rf6 Qc6 53. Nd1 Bxf6 54. gxf6 Qc1 55. Ne3 a6 56. g7 Bf7 57.
Bg6 Bg8 58. Qf3 Qa1 59. Qf4 e5 60. Qf5 Qxa2+ 61. Kg3 d4 62. Qxe5 dxe3 63.
Be4+ Kb8 64. Qe8+ Rc8 65. Qe5+ Rc7 66. Qe8+ Rc8 67. Qe5+ Rc7 1/2-1/2

[/pgn]

But I find that secondary, the important thing is for the player with an advantage to be able to avoid the repetition. I don't think the rule prevents that, unless said player missed something, in which case the penalty will come in the form of a score loss.

[Peter Berger]

I think many answers miss the main point which is (IMHO) that the major task of the humans in correspondence chess these days is to choose between lines that have either been suggested by computers or that at least have been computer-approved.

The question that suggests itself is whether humans are per se more compentent to do this than a computer program.

To me this seems highly unlikely - it's just that no one has made +serious+ efforts to come up with a computer program to take over this role of judge.
There were some attempts like automated backsolving or the triple-brain approach but these approaches have mostly been pretty simplistic and haven't been worked on as seriously as has been worked on the single engines themselves.

It is my impression that the chess understanding of current strong correspondence players is in fact way behind that of current strong chess engines unless in very specific situations, so that the task of judgement could in principle be taken care of by a computer program.

Why is this relevant? If I am right all suggestions of changes of chess rules are pointless, because the engines ( or sometimes the judge) will just have to be adapted a little, and then they will take over again.

The question is whether this still is a somehow interesting task to accomplish for a programmer, as correpondence chess is pretty much dead anyway. But of course this could still be kind of interesting for the preparation of opening lines e.g.

[Cornfed]

The question is whether this still is a somehow interesting task to accomplish for a programmer, as correpondence chess is pretty much dead anyway. But of course this could still be kind of interesting for the preparation of opening lines e.g.

Yes, but then lets not consider it a 'game of chess'...but an opening prep competition.
There are some too invested in what they are doing to allow themselves to think otherwise of course.

[jefk]

In reply to Peter Berger, no, the suggestions for rule changes
are not pointless, certainly not for Fide/ICCf correspondence
chess where comps are allowed. Even today engines are only a decision
support tool, true with NNue and so on they have improved, but
correspondence chess at higher level also needs human understanding,
and the ability to use different tools; opening theory is especially
important in such slow games, as wel as the ability to improve
on existing theory. In addition, in middle game strategy its a
matter of deep analysis, not simply running one engine for a night.
This may be especially true for the transition between middle
game and endgame.

Think of centaur (computer assisted), comp/human hybrid chess
but then for slower times.

For the rest such a discussion belongs better in the other
thread i noticed now about correspondence chess, ie.
what's the 'use' of correspondence chess. Well some Otb
players may find it interesting for example to gradually
improve their opening knowledge, other players may
like the scientific approach, rather than the tactical fight
of faster Otb games. whatever.

So some rule modification to address the drawing problem
in top correspondence chess certainly is relevant and i'm
going to address in a next message.

[jefk]

In response to Larry K:

Based on some tests with Komodo I estimate that the three rules together will roughly double the win percentage in high level engine vs engine play.

Mostly wins for White, i expect, because with those rules you mainly increase the advantage for White

Probably c2 is the best pawn to remove for this Armageddon play. But note that if you do this then you don't need any of the above rule-changes; it's one or the other.

Sure it are independent options, indeed, but because of the white advantage there's a need to equalize the b/w chances in case you reduce the drawing margin. You (LK) mentioned Armageddon chess, well in this variant there also are two independent, separate rules, namely:
1) draw is win for black, but also:
2) white gets 1 min time (in a blitz 6 or 5 mins game with zero inc).

Of course there are several means to to reduce the advantage for White,
(besides the new Nickel/Kaufman rules:

1) let two opponent play both White *and* Black in the same tournament
(in a Swiss tourn, the amount of W/B games is almost the same but
it's still a little bit of a gamble because you play different opponents).
NB with such a method the nr of people in one correspondence tourn
probably should be reduced, whereas it's now on average around 11 players,
it may be a better idea to reduce it to around 7 players or so (playing
each other both with Black and White!)

and in addition one of the two options;

1A) reduce the playing time for White (like in Armageddon) eg. 1 day on average max for White and 3 days (as normal) for Black
1B) remove the c2 pawn for White
1C) something else (?)

See what i mean ?

Personally i would find such new rules (incl the Nickel/Kaufman rules)
worth a try (besides the old rules, and then we'll see which
rules will become most popular For example)

For the rest i like the idea of removing the c2 pawn (*) better than
FRC aka chess960, besides that in the latter setups White also
has the advantage of being the first-mover (or maybe b2 or so, it
could use some further calibration to be sure, but i like the idea;
although classical opening theory won't apply, you get normal
looking games, and some good old-fashioned pattern recognition still
would most likely help).