Balanced opening lines or advantageous opening lines?

[Chessqueen]

There is a place for both but I prefer balanced because there is already too many non balanced games.
It is obvious after lc0 beat berserk with 1.g4 in TCEC that balanced opening lines or even no opening book is not always a draw even between top engines.

I test now stockfish15 at x nodes per move against berserk13 at 2x nodes per move(with contempt 0,10,20,30,40,50 when I started with x=1000 and double x after every match.

it is still not obvious if contempt of berserk change the playing strenght of it and it is still not clear if the berserk team is better or worse than stockfish15 in these conditions.

a win with black from the tournament.
White used 1024000 nodes per move.

Here is another game that after 10 moves White has a huge advantage 0.90, therefore either Stockfish or Berseerk will win with the White side
[pgn][Event "TCEC Season 26 - Division P"]
[Site "https://tcec-chess.com"]
[Date "2024.05.03"]
[Round "34.1"]
[White "Stockfish dev-20240413-c55ae376"]
[Black "Berserk 13"]
[Result "*"]
[BlackElo "3540"]
[ECO "D15"]
[Opening "Slav"]
[Variation "Chameleon, 5.c5"]
[WhiteElo "3645"]
[TimeControl "3600+6"]
[Termination "unterminated"]
[PlyCount "22"]
[WhiteType "program"]
[BlackType "program"]

{WhiteEngineOptions: 0=[object Object]; 1=[object Object]; 2=[object
Object]; 3=[object Object]; 4=[object Object]; 5=[object Object]; 6=[object
Object]; 7=[object Object];, BlackEngineOptions: 0=[object Object];
1=[object Object]; 2=[object Object]; 3=[object Object]; 4=[object Object];
5=[object Object];} 1. d4 {book} d5 {book} 2. c4 {book} c6 {book} 3. Nc3
{book} Nf6 {book} 4. Nf3 {book} a6 {book} 5. c5 {book} Bg4 {book} 6. Ne5
{book} Be6 {book} 7. Qb3 {book} Ra7 {book} 8. e3 {book} g6 {book} 9. Bd2
{book} Bg7 {book} 10. f3 {R50=50, Rd=-9, Rr=-1000, d=43, h=14.3, mt=139819,
n=3620807453, ph=0.0, pv=10. f3 O-O 11. Na4 Nfd7 12. Nd3 Bf5 13. O-O-O e5
14. Nxe5 Nxe5 15. dxe5 Bxe5 16. h3 Bg7 17. g4 Be6 18. f4 f5 19. gxf5 Bxf5
20. Bg2 Re8 21. Rhe1 Be6 22. Qc2 Nd7 23. e4 d4 24. Bf1 Ra8 25. Bd3 Kh8 26.
Rg1 b5 27. cxb6 c5 28. f5 Bg8 29. b3 c4 30. bxc4 Ne5 31. Bg5 Qd7 32. c5
Nxd3+ 33. Rxd3 Rac8 34. f6 Bf8 35. Kb1 Rxe4 36. Ka1 Qd5 37. h4 Rce8 38. Rc1
Qc6, s=25896205, sd=74, tb=17246, tl=3466181, wv=0.98} Nfd7 {R50=49, Rd=-9,
Rr=-1000, d=44, h=46.6, mt=132377, n=9582657181, ph=0.0, pv=10...Nfd7 11.
Nxd7 Bxd7 12. O-O-O O-O 13. Na4 e5 14. dxe5 Bxe5 15. g4 Bg7 16. h3 Be6 17.
f4 Nd7 18. f5 gxf5 19. Bd3 d4 20. Qc2 dxe3 21. Bxe3 Re8 22. Bxf5 Bxf5 23.
Qxf5 Rxe3 24. Rxd7 Qe8 25. Nb6 Re5 26. Qd3 Re1+ 27. Rxe1 Qxe1+ 28. Qd1
Qxd1+ 29. Rxd1 Bf8 30. Na4 Ra8 31. Rd7 Rb8 32. Kc2 a5 33. b3 Re8 34. Rxb7
Re2+ 35. Kb1 Re1+ 36. Kb2 Bg7+ 37. Ka3 Be5, s=72390233, sd=71, tb=3221,
tl=3473623, wv=0.70} 11. Nxd7 {d=41, h=7.3, mt=59393, n=1864475158, ph=0.0,
pv=[object Object], s=31393755, sd=79, tb=16735, tl=3412788, wv=0.99} Bxd7
{d=42, h=30.1, mt=74631, n=5969381968, ph=100.0, pv=[object Object],
s=79988502, sd=64, tb=6262, tl=3404992, wv=0.57} *[/pgn]

[Chessqueen]

I think that TCEC is all about entertainment value now, which is fine for those who enjoy it.
It also demonstrates how to convert an advantageous opening into a win, which some might find useful.

However, the results have no substantial meaning if you're wanting to know the reality of engine comparisons.
Imagine giving advantageous opening lines to Carlsen and the other top players and stipulating that they must be played for entertainment value.

Talking about avantageous Opening, Lets clarify how much of an advantage todays top engines need to secure a win. For example I took this 10 moves Opening and let Stockfish play the White Side and it won easy, but when I reverse the engines and let Berserk play the White side it could NOT convert after having a slight 0.90 of an advantage for the White Side.

[pgn][Event "TCEC Season 26 - Division P"]
[Site "https://tcec-chess.com"]
[Date "2024.05.03"]
[Round "34.1"]
[White "Berserk 13"]
[Black "Stockfish dev-20240413-c55ae376"]
[Result "1/2-1/2"]
[BlackElo "3645"]
[ECO "D15"]
[Opening "Slav"]
[Variation "Chameleon, 5.c5"]
[WhiteElo "3610"]
[TimeControl "3600+6"]
[Termination "normal"]
[PlyCount "174"]
[WhiteType "program"]
[BlackType "program"]

1. d4 d5 2. c4 c6 3. Nc3 Nf6 4. Nf3 a6 5. c5 Bg4 6. Ne5 Be6 7. Qb3 Ra7 8.
e3 g6 9. Bd2 Bg7 10. f3 Nfd7 11. Nxd7 Bxd7 12. O-O-O O-O 13. Na4 e5 14.
dxe5 Bxe5 15. g4 Bg7 16. h3 a5 17. Bc3 Bxc3 18. Qxc3 Qe7 19. h4 Be6 20. h5
Nd7 21. Rg1 Qf6 22. f4 Qxc3+ 23. bxc3 f5 24. gxf5 Bxf5 25. c4 Re8 26. Kd2
Re7 27. cxd5 cxd5 28. Bb5 Kf7 29. Bxd7 Rxd7 30. Kc3 Re7 31. Rge1 Ra8 32.
Nb6 Rae8 33. hxg6+ hxg6 34. Nxd5 Re6 35. Rh1 Kg7 36. Rb1 Bxb1 37. Rxb1
Rxe3+ 38. Nxe3 Rxe3+ 39. Kc4 Re4+ 40. Kd5 Re7 41. Kd6 Kf7 42. Rb5 Ke8 43.
Rxa5 Rd7+ 44. Ke6 Re7+ 45. Kd6 Rd7+ 46. Ke6 Re7+ 47. Kd5 Rf7 48. Rb5 Kd8
49. Ke4 Kc7 50. Rb6 Rf5 51. Rxg6 Rxc5 52. Rg5 Rc4+ 53. Kf3 Rc3+ 54. Ke4
Rc4+ 55. Kf5 Rc2 56. a3 Rc3 57. a4 Rc4 58. a5 Ra4 59. Kg4 Kd6 60. Rf5 Kd7
61. Rc5 Kd6 62. Rg5 Kc7 63. Kf5 Rxa5+ 64. Ke6 Ra6+ 65. Ke7 Ra4 66. f5 Re4+
67. Kf8 Re5 68. Rg7+ Kc6 69. f6 b5 70. f7 Kc5 71. Rg1 Rf5 72. Ke7 Rf2 73.
Rc1+ Kd5 74. Rb1 Kc4 75. f8=Q Rxf8 76. Kxf8 b4 77. Ke7 b3 78. Ke6 Kb4 79.
Rc1 b2 80. Rc7 Kb3 81. Rb7+ Kc2 82. Rc7+ Kd2 83. Rb7 Kc1 84. Rc7+ Kd2 85.
Rb7 Kc1 86. Kd7 b1=Q 87. Rxb1+ Kxb1 {Insufficient material} *[/pgn]

[Viz]

The only problem with balanced lines is that they are literally impossible to use for development process.
Gl developing anything when your "LTC" (which is mere bullet, mind you) has 96% draw ratio. It's literally impossible to get anything statistically signifficant there, and let's not even talk about higher TCs where it increases even more.
Sometimes it looks really bizarre, like there - https://tests.stockfishchess.org/tests/ ... ce609c99c6
180+1.8 8 threads is not even that big of a TC, like 10 seconds/move at best at 8 threads, can be considered rapid with a single core and bullet with some AMD EPYC machines.
Yet to score 10 elo advantage (almost) stockfish 16 vs stockfish 15 needed to win 12 times more games than it lost. And this is 50 elo in uho book and like a year of development which also included a lot of hyperscaling patches (which perform better at longer TCs than at shorter ones). More or less sf 15 got dominated (expected) and it still was 9,5 elo difference.

[Graham Banks]

To be fair, whatever happens in the development process or with those who run ultra-blitz games to churn out thousands of games in a very short time, are not those that I was really referring to in my original post.

I was more interested in the view of those who enjoy watching engine v engine games, such as on their own computer or in sites such as TCEC, CCC or the broadcasts that I run.

However, all opinions and reasons given are interesting regardless.

[Uri Blass]

The only problem with balanced lines is that they are literally impossible to use for development process.
Gl developing anything when your "LTC" (which is mere bullet, mind you) has 96% draw ratio. It's literally impossible to get anything statistically signifficant there, and let's not even talk about higher TCs where it increases even more.
Sometimes it looks really bizarre, like there - https://tests.stockfishchess.org/tests/ ... ce609c99c6
180+1.8 8 threads is not even that big of a TC, like 10 seconds/move at best at 8 threads, can be considered rapid with a single core and bullet with some AMD EPYC machines.
Yet to score 10 elo advantage (almost) stockfish 16 vs stockfish 15 needed to win 12 times more games than it lost. And this is 50 elo in uho book and like a year of development which also included a lot of hyperscaling patches (which perform better at longer TCs than at shorter ones). More or less sf 15 got dominated (expected) and it still was 9,5 elo difference.

9.45 ± 0.6 is clearly a significant result.

Having many draws mean small rating difference but it also means smaller confidence interval for the rating difference.
I see no reason that it is impossible to test with 96% draws ratio and get significant results.
Maybe you are going to need more time for it but it is possible that you find some improvements that are 0 elo change with biased conditions.

Imagine that some improvement only improve the ability of the engine to find good moves in the opening in equal positions and not when it already has evaluation of +1 or when it is equal later in the game.

With a biased book you will never see the improvement because in the opening you are not going to get equal positions so the change make no improvement and when you get equal position later in the game it is not going to be an opening postion so again the change also make no improvement.

[Viz]

The only problem with balanced lines is that they are literally impossible to use for development process.
Gl developing anything when your "LTC" (which is mere bullet, mind you) has 96% draw ratio. It's literally impossible to get anything statistically signifficant there, and let's not even talk about higher TCs where it increases even more.
Sometimes it looks really bizarre, like there - https://tests.stockfishchess.org/tests/ ... ce609c99c6
180+1.8 8 threads is not even that big of a TC, like 10 seconds/move at best at 8 threads, can be considered rapid with a single core and bullet with some AMD EPYC machines.
Yet to score 10 elo advantage (almost) stockfish 16 vs stockfish 15 needed to win 12 times more games than it lost. And this is 50 elo in uho book and like a year of development which also included a lot of hyperscaling patches (which perform better at longer TCs than at shorter ones). More or less sf 15 got dominated (expected) and it still was 9,5 elo difference.

9.45 ± 0.6 is clearly a significant result.

Having many draws mean small rating difference but it also means smaller confidence interval for the rating difference.
I see no reason that it is impossible to test with 96% draws ratio and get significant results.
Maybe you are going to need more time for it but it is possible that you find some improvements that are 0 elo change with biased conditions.

Imagine that some improvement only improve the ability of the engine to find good moves in the opening in equal positions and not when it already has evaluation of +1 or when it is equal later in the game.

With a biased book you will never see the improvement because in the opening you are not going to get equal positions so the change make no improvement and when you get equal position later in the game it is not going to be an opening postion so again the change also make no improvement.

Well, you should gather info about how SPRT math works more or less.
With extremely high draw rates you have infinite time before SPRT converges to anything because it just works like this.
And 9 for 1 year of development is a joke.
Not to mention that this effect of "you gain in UHO but don't gain in normal book" is almost imaginary. After switching to UHO book we had like +2 elo in UHO and -0,5-1 elo in normal book in bullet - and then they grew up more or less in sync.
So this effect is almost non-existent, more or less within error bars.

[Uri Blass]

The only problem with balanced lines is that they are literally impossible to use for development process.
Gl developing anything when your "LTC" (which is mere bullet, mind you) has 96% draw ratio. It's literally impossible to get anything statistically signifficant there, and let's not even talk about higher TCs where it increases even more.
Sometimes it looks really bizarre, like there - https://tests.stockfishchess.org/tests/ ... ce609c99c6
180+1.8 8 threads is not even that big of a TC, like 10 seconds/move at best at 8 threads, can be considered rapid with a single core and bullet with some AMD EPYC machines.
Yet to score 10 elo advantage (almost) stockfish 16 vs stockfish 15 needed to win 12 times more games than it lost. And this is 50 elo in uho book and like a year of development which also included a lot of hyperscaling patches (which perform better at longer TCs than at shorter ones). More or less sf 15 got dominated (expected) and it still was 9,5 elo difference.

9.45 ± 0.6 is clearly a significant result.

Having many draws mean small rating difference but it also means smaller confidence interval for the rating difference.
I see no reason that it is impossible to test with 96% draws ratio and get significant results.
Maybe you are going to need more time for it but it is possible that you find some improvements that are 0 elo change with biased conditions.

Imagine that some improvement only improve the ability of the engine to find good moves in the opening in equal positions and not when it already has evaluation of +1 or when it is equal later in the game.

With a biased book you will never see the improvement because in the opening you are not going to get equal positions so the change make no improvement and when you get equal position later in the game it is not going to be an opening postion so again the change also make no improvement.

Well, you should gather info about how SPRT math works more or less.
With extremely high draw rates you have infinite time before SPRT converges to anything because it just works like this.
And 9 for 1 year of development is a joke.
Not to mention that this effect of "you gain in UHO but don't gain in normal book" is almost imaginary. After switching to UHO book we had like +2 elo in UHO and -0,5-1 elo in normal book in bullet - and then they grew up more or less in sync.
So this effect is almost non-existent, more or less within error bars.

My point is not that there are many cases when you gain in UHO and not gain with normal book but that there may be patches that fail with UHO(not because they are regression but because they have 0 change with UHO) but can pass with a non biased book.

You do not implement them and miss some possible improvements.

This may not be the best example because maybe it can pass also without balanced mode but
I think specifically about choosing the right draw line and if engines have rules to prefer one 0.00 and not another 0.00(when you can expect more 0.00 with a balanced book).

I saw an example of a mistake because of not knowing what 0.00 is better in the following game.

https://tcec-chess.com/#div=gm&game=18&season=26

Berserk blundered by 44...g6 that it evaluated to be a draw.
Nobody told Berserk that not all the draws are the same and it is better not to lose a pawn and I guess this idea has better chances to pass without biased books.

[swami]

Balanced opening book, but not c4 c5 Nf3 Nf6 g3 g6 Bg2 Bg7 type.

Something dynamic and different, but balanced.

[Michel]

Whether or not balanced books are better cannot be decided theoretically but it can be verified experimentally. I wrote something in the Fishtest FAQ about this https://github.com/official-stockfish/f ... ning-books

Such testing shows that unbalanced books are better.