Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / Selective search is bad for technical endgames?
- By rocket (****) Date 2021-03-15 19:20 Edited 2021-03-15 19:29
Technical endgames=no positional factors left on the board. Everything is down to number crunching.

Believe it or not, my testing suggests that engines low on selectivity performs better in technical endgames, including defending rook +knight vs rook.

I have seen selective search engines(Rybka for example) repeatedly lose such positions by blundering and trapping itself into mate (in blitz at least).  I have never seen it from brute force engines like oldie Fritz. They defend as if it was as tablebase.

Anybody care to do some testing with Houdinis tactical mode compared to default?
- - By rocket (****) Date 2021-03-15 19:37
To explain for those of you not familiar with these concepts.

Selective search is more perceptible to short term oversights (tactical shots).. Brute Force search is more perceptible to long term oversights (strategical errors).

All engines use a combination of both, some more of one over the other.

Brute Force searches can also miss tactical shots but they are in that case either misevaluated completely or very deep.. Otherwise it's like punching a brick wall trying to outcalculate it.

Greek sacrifices and stuff of that sort can be missed by selective search under 60 seconds ponder... Brute Force searches will spot them either instantly or within 10-15 seconds.

As a rough average, brute force engines make less errors but the ones they make are bad enough to lose games regardless of how well they play outside of the error.
Parent - - By MrKris (***) Date 2021-03-16 04:45
Computer chess history site:
https://en.wikipedia.org/wiki/Computer_chess
1912 – Leonardo Torres y Quevedo builds El Ajedrecista, a machine that could play King and Rook versus King endgames.
...
2006 – The world champion, Vladimir Kramnik, is defeated 4–2 by Deep Fritz.
2009 – Pocket Fritz 4 running on a smartphone, wins Copa Mercosur, an International Master level tournament, 9½/10 earning a performance rating of 2900.[18]
...


- - - - - - - -

This is most interesting to me.  Sorry if it does not relate to your subject.

The first chess-like program ran on a computer:
https://en.wikipedia.org/wiki/Los_Alamos_chess
Los Alamos chess (or anti-clerical chess[1]) is a chess variant played on a 6×6 board without bishops. This was the first chess-like game played by a computer program. This program was written at Los Alamos Scientific Laboratory by Paul Stein and Mark Wells for the MANIAC I computer[2] in 1956. The reduction of the board size and the number of pieces from standard chess was due to the very limited capacity of computers at the time.
- diagram below

The bishops and regular 8x8 board was soon added but incorrectly - without color-square complex consideration.
The were added, and 65 years later treaded the same, erroneously like any other 64-square chessboard piece instead of the 32-square checkerboard pieces like they are: a light-squared checkerboard for any light-squared bishops and a dark-square checkerboard for any dark-squared bishops.

28.Bh4 - Stockfish's first evaluation below 6.00.
28...Kg6 - Black Diamond's first 0.00 evaluation, then for the rest of the game (except a few brief anomalies).
[Event "G/15'+3'' 1GBh no Syzygy"]
[Site "AMD PhII X6 - 6CPU"]
[Date "2021.03.15"]
[Round "?"]
[White "Stockfish_21021123_x64"]
[Black "Black-Diamond-v13-Eval-x86-64"]
[Result "1/2-1/2"]
[BlackElo "2000"]
[Time "18:59:04"]
[WhiteElo "2000"]
[TimeControl "900+3"]
[SetUp "1"]
[FEN "4k3/8/6b1/6B1/5P2/4P3/3BK3/8 w - - 0 1"]
[Termination "time forfeit"]
[PlyCount "148"]
[WhiteType "program"]
[BlackType "program"]

1. Kf3 {+6.82/54 13} Kd7 {-2.15/22 99} 2. e4 {+6.66/49 9} Ke6 {-2.17/11
141} 3. Ke3 {+6.58/52 25} Bh7 {-1.94/12 94} 4. Kd4 {+6.58/51 8} Bg6
{-1.96/11 110} 5. Bc3 {+6.58/51 10} Bh7 {-2.22/10 54} 6. Bb4 {+6.58/52 16}
Bg6 {-2.02/10 23} 7. Bc3 {+6.58/51 9} Bh7 {-1.87/9 19} 8. Bb4 {+6.58/52 13}
Bg6 {-1.78/9 85} 9. Kd3 {+6.58/53 39} Bh7 {-1.58/8 72} 10. Ke3 {+6.58/54
56} Bg6 {-1.47/14 61} 11. Bc3 {+6.58/51 10} Bh7 {-1.65/10 19} 12. Bgf6
{+6.58/52 28} Bg6 {-1.47/13 41} 13. Bg5 {+6.58/143 31} Bh7 {-1.71/9 42} 14.
Bh4 {+6.58/245 22} Bg6 {-1.57/8 34} 15. Kd4 {+6.58/41 18} Bh7 {-1.78/10 24}
16. Bg5 {+6.58/245 10} Bg6 {-1.82/8 13} 17. Bh4 {+6.58/55 16} Bh7 {-1.61/9
7} 18. Bd2 {+6.58/28 10} Bg6 {-1.46/8 3} 19. Bb4 {+6.58/46 13} Bh7 {-1.55/8
2} 20. Bc3 {+6.58/59 181} Bg6 {-2.02/11 2} 21. Bd2 {+6.58/53 18} Bh7
{-1.57/9 3} 22. Ke3 {+6.58/29 9} Bg6 {-1.42/11 2} 23. Kd4 {+6.58/54 46} Bh7
{-1.42/9 2} 24. f5+ {+6.58/27 59} Bxf5 {-1.01/15 3} 25. exf5+ {+6.58/22 11}
Kxf5 {-1.12/14 3} 26. Bhe1 {+6.56/29 115} Kf6 {-0.86/13 2} 27. Kd5
{+6.40/23 18} Kf7 {-0.25/11 2} 28. Bh4 {+4.29/38 77} Kg6 {0.00/11 3} 29.
Bhg5 {+4.50/34 11} Kf5 {0.00/14 2} 30. Bdf4 {+5.53/41 54} Kg6 {-0.31/19 3}
31. Ke6 {+3.99/28 13} Kh5 {-1.75/16 3} 32. Be7 {+3.73/24 1} Kg4 {-0.26/17
2} 33. Bb8 {+3.90/31 6} Kf3 {-0.17/17 2} 34. Bb4 {+3.62/26 3} Ke3 {-0.06/20
2} 35. Bc5+ {+3.52/28 0} Ke2 {-0.13/19 3} 36. Bf4 {+3.82/30 5} Kf1 {0.00/20
1} 37. Kf5 {+3.95/27 3} Ke1 {0.00/22 2} 38. Ke4 {+2.97/31 3} Kf1 {0.00/23
3} 39. Ke3 {+3.00/27 2} Kg2 {0.00/28 4} 40. Bd4 {+3.16/28 3} Kg1 {-1.15/25
2} 41. Kf3+ {+3.32/25 1} Kf1 {0.00/27 2} 42. Bc3 {+3.07/25 0} Kg1 {0.00/24
0} 43. Ke2 {+2.60/30 6} Kg2 {0.00/26 4} 44. Bd4 {+2.64/25 1} Kh3 {0.00/25
1} 45. Kf3 {+1.82/31 3} Kh4 {0.00/24 1} 46. Kg2 {+1.85/33 2} Kg4 {0.00/26
3} 47. Bde5 {+1.57/31 4} Kh5 {0.00/26 4} 48. Bf6 {+1.21/30 1} Kg6 {0.00/25
1} 49. Bd4 {+1.52/33 4} Kf5 {0.00/28 7} 50. Kf3 {+1.04/29 0} Ke6 {0.00/24
1} 51. Ke4 {+1.02/29 1} Kd7 {-0.05/25 4} 52. Kd5 {+0.95/29 6} Ke7 {0.00/15
3} 53. Bc5+ {+0.87/26 1} Kf7 {0.00/22 0} 54. Be5 {+0.68/28 1} Kg6 {0.00/25
3} 55. Ke6 {+0.51/28 0} Kh5 {0.00/23 1} 56. Kf5 {+0.42/26 1} Kh4 {0.00/25
1} 57. Be7+ {+0.08/30 9} Kh3 {0.00/22 1} 58. Kf4 {+0.05/25 0} Kg2 {0.00/23
4} 59. Bh4 {0.00/36 1} Kf1 {0.00/26 3} 60. Kf3 {0.00/38 0} Kg1 {0.00/23 2}
61. Be1 {0.00/45 0} Kf1 {0.00/23 6} 62. Bd2 {0.00/45 5} Kg1 {0.00/23 3} 63.
Bc1 {0.00/47 0} Kf1 {0.00/22 2} 64. Bc3 {0.00/46 1} Kg1 {0.00/20 0} 65.
B1b2 {0.00/48 1} Kf1 {0.00/21 1} 66. Ba1 {0.00/49 1} Kg1 {0.00/19 0} 67.
Bab2 {0.00/50 1} Kf1 {0.00/20 3} 68. Ba1 {0.00/53 1} Kg1 {0.00/19 0} 69.
Bf6 {0.00/48 1} Kf1 {0.00/21 2} 70. Bh4 {0.00/47 1} Kg1 {0.00/21 0} 71. Bg3
{0.00/51 2} Kf1 {0.00/27 2} 72. Bh4 {0.00/51 2} Kg1 {0.00/26 0} 73. Bd4+
{0.00/47 1} Kf1 {0.00/31 2} 74. Be1 {0.00/51 2} Kxe1 {0.00/27 0
Insufficient material} 75. ... {White forfeits on time} 1/2-1/2
Above after 24.f5?! Bxf5 25.e(e4)xf5 Kxf5 26.Bh4-e1 +6.59/29 115 seconds
without Syzygy Stockfish does not know that two same-color bishops do not win:
8/8/8/5k2/3K4/8/3B4/4B3 b - - 1 26
Parent - - By rocket (****) Date 2021-03-16 11:13
Is Black Diamond a tactical mode of Stockfish?
Parent - - By MrKris (***) Date 2021-03-16 15:47
The author has said Black Diamond somewhere its a fork of Crystal https://en.wikipedia.org/wiki/Computer_chess , which is from Stockfish.

He describes his '13' release, somewhat, here:  http://talkchess.com/forum3/viewtopic.php?f=2&t=76657 .
"BD is currently the only one with 'Fortress Detection'." (Presumably similar to Crystal and the others closer to Stockfish.)

All of his have 'Pro Analysis' and other settings that look similar to the Tactical idea. (He does not say what 'Defensive' is.)
Parent - - By rocket (****) Date 2021-03-16 16:14
Interesting. How does it affect elo compared to SF default ?
Parent - - By MrKris (***) Date 2021-03-16 21:27
I do see them rated anywhere.
I put some results here:

http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=33852

(They also have the ShashChess, https://github.com/amchess/ShashChess , settings: Tall, Capablanca & Petrosian = close to without.)

Black Diamond's (fortress detect), and the others if using Pro will be wider but less deep searches so mainly to see if Stockfish missed something.

My wild guess is they are between Dragon 1.0 and Stockfish (BD lowest/slightly below).
My feeling is they are a good opponent for Stockfish: less draws than Sf vs. Sf, and good strength.
Parent - - By rocket (****) Date 2021-03-16 21:40
It's a good initiative. Stockfish default is completely blind to blockades, which is something I would expect from the man-machine era,... not today.
Parent - By MrKris (***) Date 2021-03-17 05:16

> It's a good initiative.


Excellent. It must be difficult without a large team of contributors and previously unheard of testing hardware.

> Stockfish default is completely blind to blockades, which is something I would expect from the man-machine era,... not today.


I think adding chess (real) improvements ended with "the man-machine era" when major flaws would get world wide attention.
Stockfish 'today" -although it is working!- is quasi-chess in my view.
The latest patch is a very good example https://abrok.eu/stockfish/

Author: Topologist
Date: Wed Mar 10 12:32:53 2021 +0100
Timestamp: 1615375973

Change advanced pawn push threshold

A pawn push is now considered to be "advanced" if the relative destination
rank is > 6 (previously it was > 5). This affects the search heuristic.
... LTC:
LLR: 2.94 (-2.94,2.94) {0.25,1.25}
Total: 107840 W: 4198 L: 3892 D: 99750 Elo +0.99 ...
 

Quasi-chess: 107,840 games, the patch won 306 more games than it lost, 4198 'steps' forward and 3892 back.
(real) Chess: "advanced" depends on the rest of the position. An extreme example is a pawn on a2 can be "advanced" if for example it is passed and difficult to stop.
Parent - - By rocket (****) Date 2021-03-16 21:42
Btw, the neural network programs are much weaker in endgames, and they are more selective than Stockfish right? So that supports my suspicion.
Parent - By MrKris (***) Date 2021-03-17 06:06

> and they are more selective than Stockfish right?


Probably, though I do not understand Lc0 as much as I would like yet.

> the neural network programs are much weaker in endgames


The only thing is I am beginning to wonder if Lc0 looses about 1/3 of the time in the book to early middle-game, 1/3 in the middle-game and 1/3 in the ending, making a ? every so often:

Game 1         | Game 2         |  Game 3         |  Game 4         |
---------------------?---------------------?---------------------?---

:grin:

However, I think most would agree with your

> So that supports my suspicion.

- By rocket (****) Date 2021-03-17 12:52
Also Deep junior is notoriously poor endgame players. A selectively search engine
Up Topic The Rybka Lounge / Computer Chess / Selective search is bad for technical endgames?

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill