Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / A Lc0 B sac., 8.d5N, crushes the Stockfish-family
- By MrKris (***) Date 2020-05-28 22:44
Also see: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=33529
r1bqk1nr/pp1pnppp/1bp5/1B1P4/4P3/5N2/PP3PPP/RNBQ1RK1 b kq - 0 8

G/24 mins. + 8 secs.  2700X 16thrds | RTX 2060 , Sy.6
Lc0 won all 6 with white and drew all 6 with black.
All 6 of the Sf's only drew with white and lost with black:
                                Elo +/-       Games      Score   Draws
0 Lc0.24.1_SV384x30-t60-3010    191 144   9.0/12 W+6 B=6 75.0%   50.0%
1 Crystal-24may2020            -191 nan   0.5/2   B0 W=  25.0%   50.0%
2 ShashChess-11_Tal            -191 nan   0.5/2   B0 W=  25.0%   50.0%
3 Honey-25may2020              -191 nan   0.5/2   B0 W=  25.0%   50.0%
4 Blue-Honey_noCs              -191 nan   0.5/2   B0 W=  25.0%   50.0%
5 Bluefish_PA8                 -191 nan   0.5/2   B0 W=  25.0%   50.0%
6 Black-Diamond_DPA            -191 nan   0.5/2   B0 W=  25.0%   50.0%

1.e4 e5 2.Nf3 Nc6 3.Bb5 Bc5 4.c3 Bb6 5.d4 exd4 6.cxd4 Nce7 7.O-O c6 8.d5 {Book}
8...cxb5 9.d6: +6 =0 -0 (6/6: 100%)
Lc0 as white won 6 of 6:   All 6 Sf's only drew with white, below.
     9       10       11       12       13       14       15       16    
-------------------------------------------------------------------------------
1  d6       a3       e5       Nbd2     Ne4      Re1      Qd3      Bd2       2:
   a5[1]    Ra6      b4       Nc6      Kf8      Qe8      h6       Qe6[2]   100%

2  ...      e5       a3       Nbd2     Ne4      h4       Re1      Ra3       1:
   b4       a5       Ra6      Nc6      h6       Kf8      ba3      Qe8[3]   100%

3  ...      e5*      Nc3      Re1      Bg5      b4       a4       Nd5       1:
   Nc6      Kf8      Qe8      Bd8      Ba5      Nb4      Nc6      h6[4]    100%

4  ...      ...      ...      ...      Nd5      a4       Ng5      Bg5       1:
   ...      ...      ...      Qe6      Bd8      ba4      Bg5      Qf5[5]   100%

5  ...      ...      ...      ...      Nd5      a3       b4       Qd3       1:
   ...      ...      ...      h6       Ba5      Bd8      a6       b6[6]    100%

[1]  9...Ng6 10.Nc3 h6 11.e5 b4 12.Ne4 Kf8 13.a3 bxa3 14.Rxa3 Qe8 15.h4 Nxe5 16.Re1 Ng4 Lc0 1-0 Bf (45).

1.e4 e5 2.Nf3 Nc6 3.Bb5 Bc5 4.c3 Bb6 5.d4 exd4 6.cxd4 Nce7 7.O-O c6 8.d5 {Book}
8...cxb5 9.d6 Nc6 10.Nc3* h6 11.e5 b4: +0 =6 -0 (3/6: 50%)
Lc0 as black drew 6 of 6:   All 6 Sf's lost with black, above.
    12       13       14       15       16       17       18       19    
-------------------------------------------------------------------------------
1  Ne4      Re1      a3       Qb3      Re2      Bd2      Rae1     h4        1:
   Kf8      Qe8      b3       Ba5      Bd8      b6       a5       Bb7[1]   50%

2  Nd5      de7      Nb6      Be3      Qd2      a4       Rfd1     Nd2       1:
   Nge7!    Ne7      Qb6      Qb5      a5       Qd5      Qd2      d5[2]    50%

3  ...      ...      ...      ...      Re1      Rc1      Bc5      a4        1:
   ...      ...      ...      ...      OO       Nf5      Re8      ba3[3]   50%

4  ...      ...      Nb4      Bf4      Nd5      Bh6      Ne7      Bg5       3:
   ...      ...      OO       Bc5      a5       d6       Qe7      Qc7[4]   50%

[4]  20.exd6 Qxd6 1/2-1/2: SC-Lc0, Cr-Lc0 and BD-Lc0.

Attachment: LcBsac-Sfs.pgn - Games (32k)
- By MrKris (***) Date 2020-05-28 22:51
All diagrams after 20... black's move.
- Lc0's 6 of 6 white wins.
- Lc0's 6 of 6 black draws.
- By MrKris (***) Date 2020-05-30 21:30

- Komodo 13.3 is the first and only CPU engine to defeat Lc0-t60-3010 in this opening!
- Lc0-Ls14 is the first and only NN do so!
- Lc0-B.2-178 is the first only engine to draw with black vs. Lc0-3010!


G/24 mins. + 8 secs.  2700X 16thrds | RTX 2060 , Sy.6
                                Elo   +/-   Games   Score   Draws
0 Lc0.24.1_SV384x30-t60-3010    102   149   9.0/14  64.3%   42.9% wh +6 =1, bk =5 -2
1 Komodo-13.3_c0                  0   nan   1.0/2   50.0%    0.0% w1 b0
2 Lc0.24.1_320x24-J13B.2-178      0     0   1.0/2   50.0%  100.0% w= b=
3 lc0.24.1_256x20_Ls14            0   nan   1.0/2   50.0%    0.0% w1 b0
4 Houdini 6 Std--only8threads  -191   nan   0.5/2   25.0%   50.0% w= b0
5 Komodo-13.2.5                -191   nan   0.5/2   25.0%   50.0% w= b0
6 Lc0.24.1_384x30-t40-2036SV   -191   nan   0.5/2   25.0%   50.0% w= b0
7 Lc0.24.1_256x20-t40-1541SV   -191   nan   0.5/2   25.0%   50.0% w= b0


White's move 13 coincidentaly was critical in Komodo 13.3 vs Lc0:
13.Qd3 (!) only game with it shown below.
13.Qd5 in 1 game: Lc0-3010 1/2-1/2 Lc0-B.2-178.
13.axb4 in 2 games: Lc0-2036SV 1/2-1/2 Lc0-3010,
and Ko13.2.5 1/2-1/2 Lc0-3010. [Event "G/24'+8'' Sy.6"]
[Site "2700X Sy50off | 2060"]
[Date "2020.05.29"]
[Round "1"]
[White "Komodo-13.3_c0"]
[Black "Lc0.24.1_SV384x30-t60-3010"]
[Result "1-0"]
[BlackElo "2400"]
[ECO "C64"]
[Opening "Spanish"]
[Variation "Classical, Charousek Variation"]
[WhiteElo "2400"]
[TimeControl "1440+8"]
[Termination "normal"]
[PlyCount "112"]
[WhiteType "human"]
[BlackType "human"]

1. e4 {book} e5 {book} 2. Nf3 {book} Nc6 {book} 3. Bb5 {book} Bc5 {book} 4.
c3 {book} Bb6 {book} 5. d4 {book} exd4 {book} 6. cxd4 {book} Nce7 {book} 7.
O-O {book} c6 {book} 8. d5 {book} cxb5 {-1.03/14 0 21s} 9. d6 {+0.76/32 0
42s} Nc6 {-1.03/14 0 23s} 10. e5 {+0.60/31 0 39s} Ba5 {-1.06/16 0 30s --
played 4 times, this 1-0 and 3 draws.} (10. .. Kf8 {-- the only other move
played, 3 times, all 3 1-0.}) 11. a3 {+0.71/31 0 36s} b4 {-1.05/18 0 21s}
12. Bg5 {+0.68/32 0 35s} Qb6 {-1.10/19 0 25s} 13. Qd3 {+0.63/29 0 39s} Qa6
{-0.83/16 0 74s} (13. .. h6 {-- Lc0-B.2-178 @ 10 mins. -0.15 += to =} 14.
Be3 Qd8 15. Nd4 b6 16. Qb3 Nge7 17. axb4 Bxb4 18. dxe7 Qxe7 19. Nc3 Nxd4
20. Bxd4 Bb7 21. Nd5 Bxd5 22. Qxd5 O-O 23. Bxb6 a6 {etc.}) 14. Qe4
{+0.75/32 0 44s} b3 {-0.88/19 0 16s} 15. Nbd2 {+1.28/29 0 30s} b5 {-0.65/17
0 47s} 16. Rfe1 {+1.01/30 0 37s} Bb7 {-0.02/20 0 47s} 17. Qg4 {+0.61/33 0
39s} h5 {-0.08/22 0 20s} 18. Qg3 {+0.81/34 0 36s} Bxd2 {-0.06/21 0 0.45s}
19. Bxd2 {+0.63/34 0 46s} Kf8 {-0.07/18 0 50s} 20. Ng5 {+0.71/33 0 36s} Nd8
{-0.47/21 0 64s} 21. Rac1 {+0.85/34 0 33s} Bc6 {-0.40/24 0 29s} 22. Qf4
{+0.66/34 0 39s} Nh6 {-0.55/27 0 41s} 23. e6 {+0.95/35 0 37s} dxe6
{-0.53/29 0 23s} 24. d7 {+0.95/36 0 46s} Kg8 {-0.54/30 0 28s} 25. Bb4
{+0.93/34 0 27s} Bb7 {-0.53/16 0 0.28s} 26. Be7 {+0.87/36 0 35s} b4
{-0.53/22 0 0.51s} 27. Qxb4 {+1.80/34 0 27s} Nf5 {-1.88/17 0 113s} 28. Rc7
{+2.39/33 0 27s} Bd5 {-2.45/12 0 166s} 29. Bxd8 {+3.29/38 0 30s} Rxd8
{-2.59/13 0 0.31s} 30. Rc8 {+3.38/36 0 22s} Qxc8 {-2.90/12 0 10.0s} 31.
dxc8=Q {+3.44/36 0 29s} Rxc8 {-2.98/11 0 21s} 32. Qa4 {+3.65/35 0 27s} Rf8
{-3.05/11 0 103s} 33. Qd7 {+3.50/35 0 34s} Rh6 {-3.23/10 0 51s} 34. Rc1
{+5.66/36 0 36s} Rg6 {-3.26/11 0 4.5s} 35. Rc8 {+5.77/34 0 24s} Rxg5
{-3.27/11 0 22s} 36. Rxf8+ {+5.77/34 0 29s} Kh7 {-3.35/10 0 19s} 37. g3
{+6.08/34 0 24s} Nh6 {-3.37/8 0 49s} 38. Rh8+ {+6.10/35 0 25s} Kg6 {-3.62/9
0 100s} 39. f4 {+6.41/33 0 24s} Rg4 {-3.65/7 0 48s} 40. Qxa7 {+7.16/31 0
25s} h4 {-3.67/7 0 31s} 41. Qe3 {+8.15/32 0 29s} hxg3 {-3.48/8 0 35s} 42.
h3 {+11.31/34 0 56s} Nf5 {-3.81/7 0 31s} 43. Qe5 {+12.18/33 0 24s} Rh4
{-4.26/8 0 53s} 44. Rxh4 {+12.78/32 0 52s} Nxh4 {-4.45/8 0 18s} 45. Qg5+
{+12.46/28 0 43s} Kh7 {0s} 46. Qxh4+ {+13.10/28 0 28s} Kg8 {-5.37/7 0 79s}
47. Qxg3 {+14.28/29 0 28s} f5 {-5.42/6 0 36s} 48. a4 {+15.72/26 0 18s} Kf7
{-5.18/6 0 28s} 49. a5 {+250.00/23 0 13s} Ke7 {-5.11/6 0 27s} 50. Qxg7+
{+250.00/22 0 3.6s} Ke8 {-5.26/6 0 25s} 51. a6 {+M17/26 0 4.2s} Bf3
{-5.35/6 0 24s} 52. a7 {+M15/28 0 4.4s} Ba8 {-5.55/5 0 21s} 53. Qg8+
{+M11/60 0 2.6s} Kd7 {-6.07/7 0 23s} 54. Qxa8 {+M9/99 0 0.39s} Kd6 {-9.35/7
0 19s} 55. Qd8+ {+M7/99 0 0.076s} Kc5 {-128.00/2 0 19s} 56. a8=Q {+M5/99 0
0.006s} e5 {-128.00/2 0 18s, White wins by adjudication} 1-0
- - By rocket (****) Date 2020-05-31 02:11
Does it blow your mind?:) Moves like that work when the other side is behind in development. I have no idea whether it wins by force or not.
Parent - - By MrKris (***) Date 2020-06-01 10:20

> Does it blow your mind?:)


Good question (concise :) ). Yes, in several ways:

1) White is a B down, compensation black's locked in B on c8, and black's locked in R on a8 = white in effect a rook up so should win - but why can Lc0 SV3010 draw almost 100% of the time with black?

If Stockfish/CPU engines are so tactical and NN's not then:

2) Why do the CPU's draw instead of win as white vs. the NN's, with the NN's almost 100% as white vs. the CPU's?

3) Why do the CPU's almost always lose as black vs. the NN's instead using their higher depth and supposed tactics to see what is coming and defend as the NN's can with black against them?

Can the CPU authors/contributors answer these?
Parent - By rocket (****) Date 2020-06-01 11:18
I think you are viewing it too black and white.
It's not possible to determine the objective nature of the position (win, loss or draw) It is too early in the game and very double edged.

Stockfish default is not tactically optimised, there is a tactical mode you could try which get's lower depths but solves more positions.
Parent - By FirebrandX (**) Date 2020-06-02 09:04 Edited 2020-06-02 09:13
Latest dev build of SF gives 8...cxb5 9.d6 h6 10.e5 a5 with a flatline evaluation after several hours of analysis on a Ryzen 7 2700x. I don't believe this wins by force at all, but is a matter of horizon effect at low ply or fast time controls. In other words, this line would be a draw on ICCF. So the reason you see Lc0 winning most of these games (if not all) is because the combination of low ply, fast time control, and early opening position just puts NNs in the driver's seat to scare up all sorts of inaccuracies that compound into a win. However, given enough time, SF will hold the draw just as well. It's just a matter of this being a perfect storm of all the right conditions for NNs to play well and AB engines to be left with not enough time/depth to fend off and hold.
- - By rocket (****) Date 2020-06-01 11:27
You could also argue that the compensation is positional, hence why positional engines perform better. Tactical strength= solutions within a limited game tree.

This position is certainly not a "limited game tree"
Parent - - By MrKris (***) Date 2020-06-02 01:53

> This position is certainly not a "limited game tree"


Good point!

8.d5 definitely could be called a "positional sacrifice", "thickets of variations" as in, I think, Kotov's 'Think Like a Grandmaster'; a lack of forced, or 'only', moves at low plies from the root.

> Tactical strength= solutions within a limited game tree.


So Stockfish/CPU engines are the real 'zero knowledge' zero beyond captures, recaptures, material count, mates, 'count-to-50' ending skill, 'count-to-3 repetition, maybe pawn promotion.
Below position:
Ply 1:       3 minutes:
Lc0 +0.06    +0.03 7 plies
Sf +8.95     +9.11 97 plies


[Event "Happy 64th Birthday to the A/B Algorithm!"]
[Site "2700X 14threads | 2060"]
[Date "2020.06.01"]
[Round ""]
[White "Stockfish_20052607"]
[Black "Lc0.24.1-SVlgt60-3010"]
[Result ""]
[BlackElo "2400"]
[Time "18:16:25"]
[WhiteElo "2400"]
[TimeControl "1440+8"]
[SetUp "1"]
[FEN "6k1/8/6b1/8/p1pB4/PpP1B3/1P3B2/K7 w - - 2 1"]
[Termination "unterminated"]
[PlyCount "0"]
[WhiteType "human"]
[BlackType "human"]

{Lc0 depth 7/11, min:secs 03:08, nodes 4,903k, nps 27k, eval. +0.03} 1.Ba7

(1.Bg3 {Stockfish depth 97/98, 03:08, nodes 3,909,788k, nps 20,775k, eval. +9.11} Be4 2.Bb6 Kf7 3.Bd6 Kg6
4.Be5 Bf5 5.Bd2 Be4 6.Bdf4 Kf7 7.Bbd4 Kg6 8.Bf6 Kf7 9.Bh4 Bf5 10.Bhg3 Kg8
11.Bgf2 Kf7 12.Bf2e3 Be4 13.Bd6 Bf5 14.B4e5 Kg6 15.Bb4 Be4 16.Bd2 Kf7
17.Bbd6 Ke8 18.Be3 Kf7 19.B5f4 Kg7 20.Bb4 Kg6 21.Bd4 Kf7 22.Bg3 Kg6 23.Bde5 Kf7
24.Bef4 Kg6 25.Bh4 Kg7 26.Be5+ Kf7 27.Bc5 Kg6 28.Bcd4 Bf5 29.Be7 Be4 30.Bf4 Kf7
31.Bef6 Kg6 32.B6g5 Kf7 33.Bfe5 Kg6 34.Be7 Bf5 35.Bb4 Be4 36.Bf6 Kf7 37.Bd8 Kg6
38.Be3 Kg7 39.Bdg5 Kg6 40.Bef4 Kg7 41.Bbd6 Kf7 42.Bc7 Kg8 43.Bf6 Kf7 44.Bd4 Kg6
45.Bde3 Kg7 46.Ba5 Bd3 47.Bab6 Kf7 48.Bd2 Be4 49.Bg5 Kg7 )

Kh7 2.Bf4 Kg8 3.Bfb6 Kh7 4.Bfb8 Kg8 *
FEN: 6k1/8/6b1/8/p1pB4/PpP1B3/1P3B2/K7 w - - 0 1 

Lc0.24.1-SVlgt60-3010:
  1/2  00:04            17  2k  +0.06  1.Bb6 Bf5
  2/3  00:04           191  2k   -0.00  1.Ba7 Be8 2.Kb1
  3/4  00:04           448  3k  +0.02  1.Be5 Kh7 2.Ba7 Kg8
...
  7/11  03:04        4,786k  27k  +0.03  1.Ba7 Kh7 2.Bf4 Kg8 3.Bfb8 Kh8 4.Bfd4+ Kh7
  7/11  03:08        4,903k  27k  +0.03  1.Ba7 Kh7 2.Bf4 Kg8 3.Bfb6 Kh7 4.Bfb8 Kg8


Stockfish_20052607: Found 510 tablebases
  1/1  00:00            50  50k  +8.95  1.Bg3
  2/2  00:00           338  338k  +8.99  1.Bg3 Bc2
  3/3  00:00           744  744k  +8.95  1.Bg3 Be4 2.Bge5
  4/4  00:00            1k  1,060k  +9.29  1.Bg3 Bd3 2.Bge5
  5/5  00:00           23k  11,667k  +9.30  1.Bg3 Bd3 2.Bg5 Be4 3.Bge5
  6/6  00:00           42k  10,443k  +9.19  1.Bg3 Kf7 2.Bd6 Be4 3.Bef4 Kg6
  7/9  00:00          111k  12,307k  +9.15  1.Bg3 Kf7 2.Bgf4 Be4 3.Bd2 Bf5 4.Bd6 Bg6
  8/8  00:00          112k  12,412k  +9.15  1.Bg3 Kf7 2.Bgf4 Be4 3.Bd2 Bd3 4.Bfe5 Bf5
  9/10  00:00          115k  12,831k  +9.15  1.Bg3 Kf7 2.Bgf4 Be4 3.Bb8 Kg8 4.Bd6 Kf7 5.Bef4 Kg6
 10/10  00:00          117k  12,994k  +9.15  1.Bg3 Kf7 2.Bgf4 Be4 3.Bb8 Kg8 4.Bd6 Kf7 5.Bef4 Kg6
 11/14  00:00          312k  15,607k  +9.11  1.Bg3 Kf7 2.Bgf4 Be4 3.Bb8 Bd3 4.Bd6 Bg6 5.Bb6 Be4 6.Bdc7 Bc2
...
 96/98  03:08    3,909,788k  20,775k  +9.11  1.Bg3 Be4 2.Bb6 Kf7 3.Bef4 Kg6 4.Bd6 Bf5 5.Bdc7 Be4 6.Bge5 Bf5 7.Bed4 Be4 8.Bf4 Kf7 9.Bfe5 Kf8 10.Bbc5+ Kg8 11.Bf4 Kf7 12.Bfe3 Kg6 13.Bb4 Kf7 14.Bdc5 Kg7 15.Bf4 Kg8 16.Bf2 Kf7 17.B2g3 Kg6 18.Bbd6 Kf7 19.Bde5 Bd3 20.Bd4 Be4 21.Bg5 Bg6 22.B3f4 Be4 23.Bde3 Kg8 24.Bg3 Kf8 25.Bd6+ Kg7 26.Bd4+ Kf7 27.Bb6 Bf5 28.Be5 Be4 29.Bbe3 Kg6 30.B5d4 Kf7 31.Bgf4 Bd3 32.Bfe5 Be4 33.Bc5 Kg8 34.Be3d4 Kh7 35.Bcd6 Kg6 36.Bf4 Bc2 37.Bd6e5 Be4 38.Bc7 Kf7 39.Bde5 Kg6 40.Ba5 Bc2 41.Bb4 Be4 42.Bd4 Kf7 43.Bfd6 Kg6 44.Bg3 Kf7 45.Bde5 Kg6 46.Bef4 Kh5 47.Bf2 Kg6 48.B2e3 Kg7 49.Be5+ Kg6
 97/98  03:08    3,909,788k  20,775k  +9.11  1.Bg3 Be4 2.Bb6 Kf7 3.Bd6 Kg6 4.Be5 Bf5 5.Bd2 Be4 6.Bdf4 Kf7 7.Bbd4 Kg6 8.Bf6 Kf7 9.Bh4 Bf5 10.Bhg3 Kg8 11.Bgf2 Kf7 12.Bf2e3 Be4 13.Bd6 Bf5 14.B4e5 Kg6 15.Bb4 Be4 16.Bd2 Kf7 17.Bbd6 Ke8 18.Be3 Kf7 19.B5f4 Kg7 20.Bb4 Kg6 21.Bd4 Kf7 22.Bg3 Kg6 23.Bde5 Kf7 24.Bef4 Kg6 25.Bh4 Kg7 26.Be5+ Kf7 27.Bc5 Kg6 28.Bcd4 Bf5 29.Be7 Be4 30.Bf4 Kf7 31.Bef6 Kg6 32.B6g5 Kf7 33.Bfe5 Kg6 34.Be7 Bf5 35.Bb4 Be4 36.Bf6 Kf7 37.Bd8 Kg6 38.Be3 Kg7 39.Bdg5 Kg6 40.Bef4 Kg7 41.Bbd6 Kf7 42.Bc7 Kg8 43.Bf6 Kf7 44.Bd4 Kg6 45.Bde3 Kg7 46.Ba5 Bd3 47.Bab6 Kf7 48.Bd2 Be4 49.Bg5 Kg7
Parent - By rocket (****) Date 2020-06-02 11:02
Nah, Stockfish has a lot of knowledge as well. The retro engine I post games from, Fritz 2, is a zero knowledge engine by your definition, however. If everything is developed, it will very often play a null move (no progression, no loss type of move) if it can't spot a move which wins material or gains a new permanent square.  Now when you get a knowledge engine however, there still crops up null moves when it can't differentiate between it's knowledge. This results in randomness, which is even worse than null.moves, and is when an active, random, bad move is played
Parent - - By rocket (****) Date 2020-06-02 11:05
Nah, Stockfish has a lot of knowledge as well. The retro engine I post games from, Fritz 2, is a zero knowledge engine by your definition, however. If everything is developed, it will very often play a null move (no progression, no loss type of move) if it can't spot a move which wins material or gains a new permanent square. 

Now when you get a knowledge engine however, there still crops up null moves when it can't differentiate between its knowledge. This results in randomness, which is even worse than regular null moves and is when an active, random, bad move is played
Parent - - By MrKris (***) Date 2020-06-02 20:52

> Nah, Stockfish has a lot of knowledge as well.


But Stockkfish/A-B engines do not have the chess goal of winning, so can not recognize when it is impossible.

It is just a big pile of additions to the base A-B appropriate to the VIC-16 style earliest legacy computing machines. (+/- 'knowledge' added/subtacted blindly to evey position whether or not they apply to that position.)

Six plus decades of hyped-up pseudo-achievement: the same quasi-chess find the highest evaluation instead of chess.

For example it never believes its own search (below its always the same material, pruning 'give aways', so always the same evaluation).
- The only reason for that is that the original A-B did not because it could only go a few ply deep on legacy hardware, so progress or not on that would be very inaccurate.

https://abrok.eu/stockfish/
"Author: Stefan Geschwentner
Date: Mon Jun 1 17:35:02 2020 +0200
Timestamp: 1591025702

Give bonus for bishops that are alligned with enemy kingring.

Inspired by the succesful patch "Give bonus for rooks that are alligned with enemy kingring" from Vizvezdenec,
this idea has been reused for bishops. Here, we only consider attacks that are not blocked by any pawn.
Also we have a 50% higher bonus than for the rooks.
"

The black bishop always covers b1 and never releases the white king: "knowledge", or maybe just material loss pruning, not to do that but no chess knowledge to determine its chess effect: [Event ""]
[Site ""]
[Date "2020.06.02"]
[Round "?"]
[White ""]
[Black "Stockfish_20060117"]
[Result ""]
[BlackElo "2400"]
[Time "11:36:15"]
[WhiteElo "2000"]
[TimeControl "0+3600"]
[SetUp "1"]
[FEN "6k1/8/6b1/8/p1pB4/PpP1B3/1P6/K7 w - - 0 1"]
[Termination "unterminated"]
[PlyCount "97"]
[WhiteType "human"]
[BlackType "human"]

1. Bg5 {+2.36/97 513 , PV:} Kf7 2. Bd8 Ke6 3. Ba5 Be4 4. Bb4 Bg6 5. Bdc5
Bd3 6. Bf2 Be4 7. Be3 Kd5 8. Bd4 Ke6 9. Bbc5 Bg6 10. Bf2 Be4 11. Bcd4 Kd7
12. Be1 Ke6 13. Bg3 Kd5 14. Be3 Ke6 15. Bb8 Bf5 16. Bc7 Be4 17. Bef4 Bd3
18. Bg3 Be4 19. Be1 Bd3 20. Bb6 Be4 21. Bd4 Kd5 22. Bg3 Ke6 23. Bge5 Bg6
24. Bb8 Be4 25. Bf2 Bd3 26. Bf4 Be4 27. Bg5 Bg6 28. Bg3 Be4 29. B3f4 Kd5
30. Bf6 Ke6 31. Bg7 Kd5 32. Be3 Ke6 33. Bf8 Bd3 34. Bd4 Bf5 35. Bh6 Bd3 36.
Bd2 Be4 37. Bc5 Bd3 38. Bf4 Be4 39. Bf2 Bg6 40. Bb6 Be4 41. Bh6 Bd3 42. Bg5
Bg6 43. Bh4 Bf5 44. Bd4 Bd3 45. Bg7 Kd5 46. Bf8 Be4 47. Bb4 Ke6 48. Bg5 Bg6
49. Bf4 *
8.5 minutes, +2.36 evaluation, 8.2 billion nodes, 29,722 Syzygy 6 TBhits
This is not a weird case: TCEC SuperFinal about 70% draws, very roughly half of that failures that require the 50-move rule because of substantively erroneous evaluations.

It still (a problem I have been aware of for years) fails if analysis reaches the 50-move rule: note below it reaches ply 99 in only 16 seconds then goes backward to 97 at 1/3 the former nps for 8 more minutes.
- Possibly no one wants to start the debate if Stockfish should 'think' the 50-move rule is on or off.
FEN: 6k1/8/6b1/6B1/p1pB4/PpP5/1P6/K7 b - - 1 1 
Stockfish_20060117: Found 510 tablebases
  1/1  00:01            1k  660k  +3.68  1.Bf4
  2/2  00:01            2k  917k  +4.05  1.Bf4 Bd3
  3/3  00:01            2k  1,100k  +3.84  1.Bf4 Bd3 2.Bfe5
  4/4  00:01            3k  1,326k  +3.93  1.Bf4 Bd3 2.Bfe5 Be4
  5/5  00:01            6k  3,170k  +3.33  1.Bc1 Be4 2.Bd2 Kf7 3.Bg5
  6/6  00:01            9k  4,328k  +3.03  1.Bf4 Kf7 2.Bd6 Ke6 3.Bb4 Be4
  7/9  00:01           23k  7,565k  +2.70  1.Bf4 Kf7 2.Bd6 Ke6 3.Bc7 Kd5 4.Bcb6 Ke6
  8/10  00:01           65k  12,959k  +2.40  1.Bf4 Kf7 2.Bh2 Ke6 3.Bb8 Kd5 4.Bc7 Ke6 5.Ba5 Ke7
...
 98/99  00:16      652,056k  41,103k  +2.36  1.Bf4 Kf7 2.Bc5 Ke6 3.Bb4 Bd3 4.Be3 Be4 5.Bd4 Bg6 6.Bbc5 Be4 7.Bg1 Kd5 8.Bcd4 Ke6 9.Bge3 Bd3 10.Bc5 Be4 11.Bed4 Kd5 12.Bf8 Kc6 13.Be7 Kd5 14.Bg5 Ke6 15.Bgf6 Bc2 16.Bg7 Bd3 17.Bf8 Be4 18.Bdg7 Bd3 19.Bb4 Bg6 20.Bh6 Be4 21.Bhf8 Bd3 22.Bbc5 Be4 23.Bfe7 Bg6 24.Bb4 Bd3 25.Bh4 Be4 26.Bbe7 Bg6 27.Bhf6 Be4 28.Bg5 Bd3 29.Bd8 Kd7 30.Bdf6 Ke6 31.Be7 Bg6 32.Bb4 Be4 33.Bh4 Bg6 34.Bd8 Be4 35.Bc5 Kd5 36.Bcb6 Ke6 37.Bh4 Bd3 38.Bd4 Be4 39.Bg5 Kd5 40.Bde3 Bd3 41.Bgf4 Be4 42.Bg3 Bd3 43.Bd4 Kc6 44.Bh2 Kd5 45.Bb8 Kc6 46.Be3 Be4 47.Be5 Kd5 48.B3d4 Ke6 49.Bb8 Bd3
 99/99  08:33    8,175,300k  15,961k  +2.36  1.Bf4 Kf7 2.Bc5 Ke6 3.Bb4 Bd3 4.Be3 Be4 5.Bd4 Bg6 6.Bbc5 Be4 7.Bg1 Kd5 8.Bcd4 Ke6 9.Bge3 Bd3 10.Bc5 Be4 11.Bed4 Kd5 12.Bf8 Kc6 13.Be7 Kd5 14.Bg5 Ke6 15.Bgf6 Bc2 16.Bg7 Bd3 17.Bf8 Be4 18.Bdg7 Bd3 19.Bb4 Bg6 20.Bh6 Be4 21.Bhf8 Bd3 22.Bbc5 Be4 23.Bfe7 Bg6 24.Bb4 Bd3 25.Bh4 Be4 26.Bbe7 Bg6 27.Bhf6 Be4 28.Bg5 Bd3 29.Bd8 Kd7 30.Bdf6 Ke6 31.Be7 Bg6 32.Bb4 Be4 33.Bh4 Bg6 34.Bd8 Be4 35.Bc5 Kd5 36.Bcb6 Ke6 37.Bh4 Bd3 38.Bd4 Be4 39.Bg5 Kd5 40.Bde3 Bd3 41.Bgf4 Be4 42.Bg3 Bd3 43.Bd4 Kc6 44.Bh2 Kd5 45.Bb8 Kc6 46.Be3 Be4 47.Be5 Kd5 48.B3d4 Ke6 49.Bb8 Bd3
 97/97  08:33    8,175,300k  15,961k  +2.36  1.Bg5 Kf7 2.Bd8 Ke6 3.Ba5 Be4 4.Bb4 Bg6 5.Bdc5 Bd3 6.Bf2 Be4 7.Be3 Kd5 8.Bd4 Ke6 9.Bbc5 Bg6 10.Bf2 Be4 11.Bcd4 Kd7 12.Be1 Ke6 13.Bg3 Kd5 14.Be3 Ke6 15.Bb8 Bf5 16.Bc7 Be4 17.Bef4 Bd3 18.Bg3 Be4 19.Be1 Bd3 20.Bb6 Be4 21.Bd4 Kd5 22.Bg3 Ke6 23.Bge5 Bg6 24.Bb8 Be4 25.Bf2 Bd3 26.Bf4 Be4 27.Bg5 Bg6 28.Bg3 Be4 29.B3f4 Kd5 30.Bf6 Ke6 31.Bg7 Kd5 32.Be3 Ke6 33.Bf8 Bd3 34.Bd4 Bf5 35.Bh6 Bd3 36.Bd2 Be4 37.Bc5 Bd3 38.Bf4 Be4 39.Bf2 Bg6 40.Bb6 Be4 41.Bh6 Bd3 42.Bg5 Bg6 43.Bh4 Bf5 44.Bd4 Bd3 45.Bg7 Kd5 46.Bf8 Be4 47.Bb4 Ke6 48.Bg5 Bg6 49.Bf4
Parent - - By FirebrandX (**) Date 2020-06-02 21:14
And yet AB engines are still very clearly better at tactics than NNs (as demonstrated in the eval charts of the recent TCEC season finale), so a case can (and should) still be made that you shouldn't pigeon-hole yourself into thinking only one type of engine is 'correct'. Both have their issues, and a good analyst will always get a 2nd opinion.
Parent - - By rocket (****) Date 2020-06-02 21:56
How are they clearly better at tactics? Alpha Zero made tactical shots against Stockfish that took days for Houdini Tactical to solve.
Parent - - By user923005 (****) Date 2020-06-03 00:06
They are clearly better at tactics because they solve a much larger percentage of tactical positions both in test sets and in games.

I also guess that some moves you guess were tactical were in fact strategic (you can usually tell by looking at the eval where the key move is made.  If there is a huge leap it was largely tactical, if it is a gradual rise it was probably strategic).  The NN engines do not see very far ahead, as a function of how they search.  That is why the tend to lose advantages or even get clobbered in the endgame.

It is (of course) possible for a NN engine to see a tactical move that an AB engine does not see.  And LC0 is not terrible in tactics, just worse than the best AB engines.  Similarly, the best NN engines are not terrible in endgames, just a lot worse than the AB engines.
Parent - - By rocket (****) Date 2020-06-03 00:37

>The NN engines do not see very far ahead, as a function of how they search.  That is why the tend to lose advantages or even get clobbered in the endgame.<


They don't? I thought they employed self-play to the end of the game for each line they select to examine?
Parent - - By user923005 (****) Date 2020-06-03 01:00
If they went to the end of the game, that is 12,000 plies, potentially.
If you would like to know which search technique is stronger, look at Komodo.
You can play Komodo in MCTS mode or AB mode.  It is a lot stronger in AB mode.

NN engines don't use MCTS because  they want to.  They use it because they have to.  A recursive algorithm like AB searching does now work well on GPUs.

Eventually, MCTS may be refined to the point it equals AB.  But it is not there today.

Some games (like GO) are so complicated with such an enormous branching factor that MCTS is just as good or better.  But chess is not one of those games.
Parent - - By rocket (****) Date 2020-06-03 01:30

>If they went to the end of the game, that is 12,000 plies, potentially.<


But how did Alpha Zero determine which move to make if it was not based on win/loss from each self play position?
Parent - - By user923005 (****) Date 2020-06-03 04:53
After a game was completed, during the learning phase, it then knew the outcome.
Parent - By rocket (****) Date 2020-06-03 11:56
I don't get it. What does the training games have to do with the unique positions arrising against Stockfish?
Parent - - By Sesse (****) Date 2020-06-04 20:27
Some MCTS types self-play random moves to finish the game; this was long a popular tactic in e.g. Go engines. (Obviously, you have to do it a bunch of times and hope that the average converges to some reasonable truth about how winning the position is for one side.) NN engines have something much better; the NN evaluation. So no playouts needed.
Parent - - By rocket (****) Date 2020-06-04 20:33
So how is NN evaluation different from type A or B chess engines?
Parent - - By Sesse (****) Date 2020-06-04 22:14
The question is ill-posed; type A or B is about the search, not the evaluation.
Parent - By rocket (****) Date 2020-06-05 14:20
I mean engines that use those types.
Parent - - By rocket (****) Date 2020-06-03 00:59 Edited 2020-06-03 01:02 Upvotes 1
Fair enough but positional moves can be long-term as well.

Strategical moves have to do wich path you want to take the position into.
Positional moves have to do with how you want to place your pieces and your pawn structure.

Weak strategical moves can still be good positionally. Example, blocking up a winning position, leading to huge space advantage but no winning line. An F in strategy but A looking at it purely positionally.  Or even doubling someones pawn structure but losing to a kingside attack with open lines for the bishops.

All weak positional moves are however weak strategical moves.
Parent - - By user923005 (****) Date 2020-06-03 04:54
That is an interesting thought, I have never thought about it that way before.
Parent - By rocket (****) Date 2020-06-03 11:58

>That is an interesting thought, I have never thought about it that way before.<


Take the bayonett attack of the Kings Inidan. After black replies a5.. If white goes b5, he will lock down the side that he's supposed to attack, and thus commits a strategical mistake. But it is not a positional mistake. His own pieces and structure is just as good as before

A lot of engines play b5....
Parent - - By rocket (****) Date 2020-06-02 21:55 Upvotes 1

>It is just a big pile of additions to the base A-B appropriate to the VIC-16 style earliest legacy computing machines. (+/- 'knowledge' added/subtacted blindly to evey position whether or not they apply to that position.<


Very true and this is where the Monte Carlo approach is vastly superior for positional play because it actually tests out moves rather than dogmatically prune certain lines out.
Parent - - By user923005 (****) Date 2020-06-03 00:08
The NN engines are not using a random search, it is guided.  And the NN engines also prune based in previous searches.  I guess that some of the AB concepts like null move pruning may not be performed, since that is more or less an AB idea, but the NN pruning is very much like LMR at least conceptually.
Parent - By rocket (****) Date 2020-06-03 01:45

>The NN engines are not using a random search, it is guided<


I never claimed that it was random.
Parent - By user923005 (****) Date 2020-06-03 00:27
Re: "But Stockkfish/A-B engines do not have the chess goal of winning, so can not recognize when it is impossible"

AB engines definitely have the chess goal of winning.
If you let an AB engine think long enough (and it may be so long that  the machine would disintegrate before it could finish) and if the game is finite (chess is finite) then if a win is possible an AB engine will find the win.

Alpha-Beta pruning is totally sound, which means that if pure mini-max examination of the entire search tree could find a win, then AB searching will also find a win.  The search result is provably identical to examination of every single node in the tree (which is a bit miraculous and also probably why AB searching is so popular).

Some techniques (such as null move pruning) can hide a win.  But an AB searcher will eventually find that win too, if you let it run long enough, because search reductions only lessen the search depth.  Eventually, on some subsequent ply, the depth searched will be great enough to overcome the reductions (although the time may be  so long as to make it impractical).

Most of the time when you see a really stupid evaluation from an AB engine due to a long, undisturbed pawn wall it seems that the engine will never solve the fact  that the game is drawn.  But it will, once 100 plies with no capture occur.  Suddenly the engine will know that the score is zero.

People imagine that MCTS searching is better than AB.  Clearly it's not.  Why do I say that?  Because Komodo Chess has two versions - one with AB and  one with MCTS search.  Everything else is the same.  Same evaluation, same everything but the search.  And the AB version is stronger, a lot stronger.

The NN engines win because of better evaluation.  In fact, the evaluation is absurdly better.  Good enough to make up for inferior search depths.  Imagine that you can look ahead flawlessly 7 moves.  And you play against a player who can look flawlessly ahead 2  moves.  Imagine further that the 2 move thinker was just as good at winning as the 7 move thinker.  Clearly, the 2 move thinker is making better use of those two moves as far as understanding structure and eventual consequences (even though he cannot see the consequences yet).
Parent - By MrKris (***) Date 2020-06-02 20:58
I found this, just in case anyone is interested.

https://www.coursera.org/learn/game-theory-1
(see also: https://www.coursera.org/learn/game-theory-1#syllabus )
Browse > Social Sciences > Economics
Game Theory
4.6 stars 2,958 ratings
Matthew O. Jackson +2 more instructors 

Offered By
Stanford University
The University of British Columbia

Enroll for Free
267,065 already enrolled

100% online
Start instantly and learn at your own schedule.
Flexible deadlines
Reset deadlines in accordance to your schedule.
Beginner Level
Approx. 12 hours to complete
English Subtitles: English, Chinese (Simplified)
Parent - By MrKris (***) Date 2020-06-02 05:36
G/56'+8" , RTX 2060
Stockfish's 28.Qf3 maybe could have kept more winning chances at least.
[Event "?"]
[Site ""]
[Date "2020.06.01"]
[Round ""]
[White "Lc0.24.1_SV384x30-t60-3010"]
[Black "Lc0.24.1-SV384x30-t60-3010.2"]
[Result "1/2-1/2"]
[BlackElo "2400"]
[ECO "C64"]
[Opening "Spanish"]
[Variation "Classical, Charousek Variation"]
[WhiteElo "2400"]
[TimeControl "3360+8"]
[Termination "normal"]
[PlyCount "113"]
[WhiteType "human"]
[BlackType "human"]

1. e4 {book} e5 {book} 2. Nf3 {book} Nc6 {book} 3. Bb5 {book} Bc5 {book} 4.
c3 {book} Bb6 {book} 5. d4 {book} exd4 {book} 6. cxd4 {book} Nce7 {book} 7.
O-O {book} c6 {book} 8. d5 {book} cxb5 {-1.01/15 0 40s} 9. d6 {+1.12/15 0
37s} Nc6 {-0.98/16 0 46s} 10. e5 {+1.00/18 0 69s} Ba5 {-0.95/17 0 47s} 11.
a3 {+1.07/16 0 75s} b4 {-0.97/17 0 61s} 12. Bg5 {+1.09/20 0 42s} Qb6
{-1.00/20 0 29s} 13. Qd5 {+1.00/18 0 93s} h6 {-0.94/18 0 120s} 14. Be3
{+0.99/19 0 34s} Qd8 {-0.96/19 0 42s} 15. Rc1 {+0.99/17 0 34s} Kf8
{-1.03/19 0 83s} 16. h3 {+1.13/20 0 120s} b3 {-0.86/19 0 244s} 17. a4
{+0.95/16 0 108s} g5 {-0.81/22 0 0.26s} 18. Na3 {+0.83/17 0 100s} Rh7
{-0.81/21 0 36s} 19. Nc4 {+0.97/25 0 102s} b6 {-0.86/25 0 101s} 20. Nxa5
{+0.97/20 0 1.3s} bxa5 {-0.95/23 0 33s} 21. Nd4 {+1.23/27 0 135s} Bb7
{-1.16/26 0 111s} 22. Nxc6 {+1.23/27 0 39s} Bxc6 {-1.16/28 0 39s --A N
down, next white gets more positional for a R vs. pawn down.
Othwise the advantage would switch to black, Stockfish 1 min. 23.Qxb3? Rg7 +0.91.}
[FEN "r2q1kn1/p2p1p1r/2bP3p/p2QP1p1/P7/1p2B2P/1P3PP1/R1R3K1 w - - 0 23"]
23. Rxc6 {+1.23/26 0 30s --Stockfish agrees at 2 minutes: 23.Rxc6 +1.46/47} dxc6 {-1.17/27
0 43s} 24. Qxc6 {+1.23/17 0 0.24s} Rg7 {-1.15/25 0 43s} 25. Qd5 {+1.22/25 0
97s} Rg6 {-1.22/25 0 76s} 26. Rc1 {+1.22/26 0 8.8s} Kg7 {-1.19/27 0 42s}
27. Rc7 {+1.19/28 0 232s} Re6 {-1.25/19 0 45s} 28. Bd4 {+1.19/20 0 0.95s}
(28. Qf3 {best, Stockfish +1.50 1 min., for example:} Qf8 29. Bd4 a6 30.
Qd5 Rc8 {?} 31. Qc4 Rxc7 32. Qxc7 {Sf +4.07 30 secs.}) 28. .. Ne7 {-1.07/31
0 174s} 29. Qf3 {+1.10/25 0 83s --Stockfish +0.70 1 min.} Rc8 {-1.07/22 0
1.3s} 30. Rxa7 {+1.07/24 0 61s} Kg8 {-1.04/23 0 72s} 31. Bc3 {+0.97/25 0
310s} Qb6 {-0.91/23 0 181s} 32. Ra8 {+0.97/27 0 0.34s} Qd8 {-0.91/25 0 31s}
33. Ra7 {+0.93/23 0 127s} Qb6 {-0.86/24 0 98s} 34. Ra8 {+0.92/25 0 0.30s}
Qd8 {-0.85/23 0 46s} 35. dxe7 {+0.90/22 0 29s} Rxe7 {-0.87/22 0 0.61s} 36.
Ra6 {+0.91/20 0 0.40s} Re6 {-0.84/21 0 90s} 37. Rxa5 {+0.87/19 0 78s} Qe8
{-0.81/19 0 67s} 38. Qd3 {+0.78/19 0 389s} Rd8 {-0.80/19 0 97s} 39. Qf3
{+0.75/19 0 132s} Rc8 {-0.75/19 0 158s} 40. Qd3 {+0.76/18 0 0.62s} Rd8
{-0.71/19 0 85s} 41. Qe4 {+0.88/23 0 149s} Rc6 {-0.85/25 0 180s} 42. Qb4
{+0.89/24 0 0.32s} Rb8 {-0.83/25 0 27s} 43. Rb5 {+0.88/23 0 0.32s} Rxb5
{-0.83/23 0 0.37s} 44. axb5 {+0.85/26 0 48s} Rb6 {-0.77/25 0 66s} 45. Qd4
{+0.85/19 0 16s} Qe6 {-0.74/23 0 49s} 46. Bb4 {+0.79/21 0 64s} Kg7
{-0.60/22 0 139s} 47. Be7 {+0.64/22 0 140s} Rb8 {-0.58/24 0 27s} 48. Bf6+
{+0.54/24 0 172s} Kg8 {-0.50/21 0 103s} 49. h4 {+0.55/21 0 0.32s} Qc8
{-0.48/20 0 41s} 50. hxg5 {+0.49/19 0 93s} hxg5 {-0.44/19 0 71s} 51. g3
{+0.48/19 0 36s} Qf5 {-0.37/18 0 79s} 52. b6 {+0.47/20 0 0.26s} g4
{-0.34/17 0 47s} 53. Qd6 {+0.38/17 0 80s} Rc8 {-0.26/18 0 79s} 54. Qd1
{+0.25/18 0 88s} Kh7 {-0.14/18 0 78s} 55. b7 {+0.20/15 0 0.60s} Rb8
{-0.12/18 0 29s} 56. Qd5 {+0.14/17 0 71s} Kg8 {-0.08/17 0 54s} 57. Qa5
{+0.10/16 0 63s, Draw by adjudication} 1/2-1/2
- By rocket (****) Date 2020-06-03 01:15 Upvotes 1
There is one exception when it comes to piece play that is strategical and not positional move, and that's remaneuvering.

The Breyer Ruy Lopez line where black goes Nc6, Nb8, Nd7 is a strategical move. The move has  to do with how black wants to position his piece, when there is no clear dichotomy of good and bad (it stands positionally sound on c6 as well).
- By rocket (****) Date 2020-06-03 13:20
For all you knowledge advocates. Here's how Rybka 3 with all those new evaluation terms added played with no opening book. The best measure of the evaluation heuristic since search can't compensate.

It's ugly.. very ugly. I guarantee you that a retro engine with no evaluation heuristic beyond piece tables and doubled pawns penalty will play vastly better than this positionally.

[Event "DESKTOP-SFGJC3V, Blitz:5'+3""]
[Site "DESKTOP-SFGJC3V"]
[Date "2020.05.31"]
[Round "1"]
[White "Rybka 3"]
[Black "Deep Rybka 4.1 x64"]
[Result "0-1"]
[ECO "A13"]
[Annotator "0.22;0.06"]
[PlyCount "30"]
[TimeControl "300+3"]

{Intel(R) Celeron(R) CPU  N3050  @ 1.60GHz  W=11.1 plies; 33kN/s  B=13.7 plies;
47kN/s} 1. Nf3 {0.22/12 11 Both last book move} d5 {0.06/14 18} 2. c4 {0.29/12
12 (d4)} e6 {0.07/13 15 (c6)} 3. Nc3 {0.27/11 8} d4 {-0.17/13 16 (Nf6)} 4. Nb5
{0.11/10 9} Nc6 {-0.33/13 6} 5. e3 {-0.09/11 15} e5 {-0.33/14 16} 6. exd4 {-0.
14/11 11} exd4 {-0.33/14 7} 7. Qe2+ {-0.17/11 13 (Bd3)} Be7 {-0.67/13 13} 8.
Qd3 {-0.34/11 9} Nb4 {-0.74/13 8 (Bc5)} 9. Qb3 {-0.56/10 11} d3 {-0.71/14 14}
10. Nbd4 {-0.56/11 5} Bg4 {-0.70/14 14} 11. a3 {-0.65/11 8} Bxf3 {-0.70/14 8}
12. gxf3 {-0.77/12 12} Qxd4 {-0.70/15 13} 13. axb4 {-0.77/11 0} O-O-O {-0.82/
15 13} 14. Qc3 {-0.58/13 65} Qb6 {-1.03/13 9 (Qf4)} 15. Bxd3 {-0.46/10 6} Bxb4
{-1.15/13 12 adjud.} 0-1

- By rocket (****) Date 2020-06-03 13:23
What do you consider the most important evaluation improvements in Rybka 3 compared to the previous version?

Larry: The contempt and related terms are very important against humans, even if they only add 3-5 Elo points against computers. New endgame knowledge is worth several points, revised definition of game phase helped, pawn structure knowledge is dramatically improved, king safety is much better understood, special terms about piece placement relative to other pieces helped, more sophisticated definition of mobility gained some points, defense of the king was much improved, proper use of the king in the endgame was emphasized. The list is almost endless.


Yet when the search is of no help, it plays like crap:roll:

Rybka 3 gains over Rybka 2.32a was clearly 95% search, not eval.
- By rocket (****) Date 2020-06-03 13:35
Vas did good work rewriting the evaluation function of Rybka 4


You can only start testing when your software is (nearly) ready. Rybka 4 wasn't ready in April. Vas rewrote the evaluation function completely. This took time. And all rumors about a delay because of the world championship are total bogus.



>

Up Topic The Rybka Lounge / Computer Chess / A Lc0 B sac., 8.d5N, crushes the Stockfish-family

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill