Announcement: Open 10x8 WB engine tournament

[hgm]

If the piece values given on those pages are those that you use in Smirf, then the Smirf piece values are definitely not OK, as the derivation given in the link has no connection with reality. The 'value' of a piece is supposed to be a measure for its contribution to winning the game. Not for the number of squares or directions it can cover, or some abstract combination thereof. What do I care if piece X covers twice as many squares as piece Y? If playing with piece Y against a piece X would win 90% of the time, I would prefer to have a Y anytime, even if X covers 100 times as many squares...

Fact is that A beats R+N+P more often than not, between equally strong opponents. So equating A to slightly over R+P, as your theory does, is basically equivalent to setting a Knight to zero. With as a consequence that Smirf often voluntarily gives piece odds to the opponent...

Another good example is this game from the 4th round:

Code: Select all

[Event "Computer Chess Game"]
[Site "SCHAAK_PC"]
[Date "2008.04.03"]
[Round "4.3"]
[White "Smirf 1.73f-X"]
[Black "BigLion 15mar"]
[Result "0-1"]
[TimeControl "3300+5"]
[Variant "capablanca"]
[FEN "rnbqckabnr/pppppppppp/10/10/10/10/PPPPPPPPPP/RNBQCKABNR w KQkq - 0 1"]
[SetUp "1"]
[Number "29"]

{--------------
r n b q c k a b n r
p p p p p p p p p p
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
P P P P P P P P P P
R N B Q C K A B N R
white to play
--------------}
1. Af3 Af6 2. Ag5 Ag4 3. Cf3 Cf6 4. Cg1 Cg8 5. Af3 Af6 6. Ae1 Ae8 
{ Engines play from here: }
7. Nh3 {+0.32/11 2:07} d5 {+0.12/7 1:00} 8. g3 {+0.31/11 2:08}
Nj6 {+0.36/7 59} 9. Nc3 {+0.25/10 1:16} Ni4 {+0.34/6 7} 10.
Af3 {+0.68/12 2:28} d4 {+0.26/7 57} 11. Ni5 {+0.86/11 1:41}
Ch6 {+0.21/7 56} 12. h3 {+0.54/11 2:20} dxc3 {+0.80/7 55} 13.
d4 {+0.49/10 1:13} Cj6 {+0.80/7 54} 14. hxi4 {+0.66/10 57} cxb2 {+0.65/6 8}
15. Bxb2 {+0.34/11 1:04} h6 {+0.52/6 16} 16. Ah5 {+0.19/12 2:04}
Cxh5 {+2.04/7 29} 17. Nh7+ {+0.08/11 1:09} Kg8 18. ixh5 {+0.07/10 31}
Kxh7 {+2.15/7 38} 19. c3 {-0.10/10 1:56} f5 {+2.16/7 52} 20.
Bf3 {+0.06/11 1:10} Af7 {+2.16/6 11} 21. Ch3 {-0.02/11 1:50}
e5 {+2.29/7 50} 22. dxe5 {+0.24/11 1:47} Nd7 {+2.28/7 49} 23.
Bd5 {+0.41/9 1:21} Axe5 {+2.32/7 42} 24. O-O {-0.07/10 1:40}
Nb6 {+2.40/7 49} 25. Bf3 {-0.40/11 1:32} Qxd1 {+3.24/9 1:32} 26.
Rhxd1 {-0.50/12 1:01} Nc4 {+3.47/7 6} 27. Bc1 {-0.95/13 1:28}
Axc3 {+3.52/7 7} 28. Bf4 {-0.54/12 1:24} c6 {+3.66/7 28} 29.
Rac1 {-0.36/11 1:21} Axa2 {+4.34/8 1:36} 30. e4 {-1.04/11 1:19}
Nb6 {+4.25/7 39} 31. Rd8 {-1.11/10 1:11} Af7 {+4.34/6 31} 32.
Rf8 {-1.42/11 1:03} Ae6 {+4.42/7 8} 33. Bd6 {-1.98/11 51}
Bd7 {+4.26/8 1:31} 34. Rxa8 {-1.85/11 52} Nxa8 {+4.40/7 8} 35.
Rd1 {-2.02/10 24} Ag5 {+4.54/8 43} 36. Cxg5+ {-1.69/13 44} hxg5 {+4.38/8 4}
37. Bb8 {-1.73/12 18} Be6 {+4.09/10 1:20} 38. exf5 {-1.37/13 17}
Bxf5 {+3.65/9 24} 39. g4 {-1.14/13 15} Be6 {+3.64/8 8} 40.
Rd6 {-1.27/11 17} Bd5 {+3.44/8 24} 41. Bxd5 {-1.55/13 15} cxd5 {+3.16/8 0}
42. Bxa7 {-2.14/14 22} i6 {+3.27/9 10} 43. Rxd5 {-2.26/14 31}
Bi7 {+3.28/9 25} 44. Bd4 {-2.45/13 28} ixh5 {+3.35/9 33} 45.
gxh5 {-2.38/13 24} Rc8 {+3.34/8 24} 46. Kh2 {-2.39/12 9} Rc4 {+3.45/8 27}
47. Rd7 {-2.49/12 15} Bh6 {+3.51/9 45} 48. i3 {-3.24/12 21} b5 {+3.55/9 38}
49. Be5 {-2.85/12 18} g4 {+3.47/8 26} 50. Ki2 {-3.03/12 9}
Rc5 {+3.89/9 1:08} 51. Re7 {-3.37/13 20} b4 {+3.84/8 15} 52.
Bg3 {-3.44/12 12} Rxh5 {+4.56/7 0} 53. Rb7 {-3.78/11 3} Bd2 {+4.49/7 3} 54.
Rb8 {-3.86/11 9} Ra5 {+4.66/8 20} 55. Bd6 {-4.45/11 9} Ra2 {+4.54/9 1:04}
56. Kj3 {-4.09/12 5} Be1 {+4.60/9 1:00} 57. Rj8 {-4.53/13 16}
Ki7 {+4.56/9 58} 58. Bf4+ {-4.22/13 15} Ki6 {+4.70/9 55} 59.
Ri8+ {-4.37/14 15} Kh5 {+4.66/9 24} 60. Rg8 {-4.11/12 11} Bc3 {+4.83/8 28}
61. i4+ {-4.07/12 3} Kg6 {+5.06/9 51} 62. Bg3 {-4.66/11 3} b3 {+7.51/9 21}
63. Rb8 {-8.18/13 9} b2 {+7.88/8 0} 64. Rb3 {-8.22/12 8} Ra6 {+8.51/9 32}
65. Rxc3 {-8.42/11 7} b1=Q {+9.03/8 17} 66. Rc4 {-8.63/11 7}
Qh1+ {+9.03/7 28} 67. Kj4 {-8.89/12 11} Qj1 {+9.04/7 25} 68.
Bh4 {-9.51/11 5} Qg1+ {+9.20/7 26} 69. Bi3 {-11.09/14 10} Qg2 {+9.48/8 23}
70. Rc5 {-11.03/11 4} Qxf2 {+10.52/8 16} 71. Rb5 {-12.09/12 9}
g3 {+10.81/7 18} 72. Bh4 {-12.27/11 5} Qh2+ {+11.59/7 13} 73.
Bi3 {-14.56/14 8} Qh6+ {+12.68/8 39} 74. Kj3 {-15.48/13 8}
Qh1+ {+12.83/8 38} 75. Kj4 {-15.36/3 0} g2 {+12.87/8 36} 76.
Rh5 {-15.76/12 8} g1=Q {+12.88/7 19} 77. Bxg1 {-27.69/13 0}
Qxg1+ {+13.38/6} 78. Kj5 {-24.20/9 0} Ra3 {+15.03/8 33} 79. j4 {-55.50/8 0}
Rj3 {+20.34/6 2} 80. Rh6+ gxh6 {+309.87/1} 81. Kj6 {-24.43/2 0}
Qxj4+ {+309.89/1} 82. Ki7 Qj6+ {+309.91/1} 83. Ki8 {-21.65/2 0}
Qi6+ {+309.93/1} 84. Kj8 Qg8+ {+309.95/1} 85. Ki7 Qh7+ {+309.97/1} 86.
Kj8 {-22.37/2 0} Qh8# {+309.99/1}
{Xboard adjudication: Checkmate} 0-1

With 16. Ah5? Smirf traps the black Chancellor such that black is forced to trade it for the Archbishop. But since Smirf does this when its Knight is attacked by a Pawn, he loses the Knight in the process, for no other compensation that a bit of reduced black King safety (6. ..., CxA 7. Nh7+, Kg8 8. PxC, KxN). A Knight that could have been simply withdrawn. After that, even outseraching the opponent by 4-5 ply cannot avert the loss.

Which "other weaknesses" do cause such disastrous trades as A vs B+N against Fairy-Max and A+N vs C against BigLion, according to you, if it is not equating A to R+P+1/4?

Let me put it another way: If you let Smirf play 100 games against itself from the position

rnbqck1bnr/pppppppppp/10/10/10/10/PPPP1P1PPP/1NBQCKABNR w Kkq 0 1

(i.e. white has Archbishop, against a Rook plus two Pawns), what would be white's score, roughly? I am sure that with the current piece values Smirf would blunder away its A for two minors or R+P on many occasions, but in most cases it will not get an opportunity to do that. (As it is self-play, the opponent will stupidly avoid such trades!) I predict you that white will score 75% or better! Have you ever tried this?

[hgm]

To SMIRF's values derivations see at: http://www.chessbox.de/Compu/schachansatz1_e.html

SMIRF's weaknesses are caused by other things - its average piece values are ok.

Reinhard.

BTW, you could add the piece values as they they follow from empirical win probabilities to your overview page, if you want:

P=1, N=3.5, B=4 (+0.5 for pair), R=5.5, A=10.25, C=10.5, Q=11

(on 10x8).

And on an 8x8 board K is about half a Pawn less valuable than N, while Nightrider is Rook plus half a Pawn.

The values look strange, because in 10x8 Chess the Pawn is worth less than in normal Chess. To make comparison with other sets easier, it is better to normalize of Queen, rather than Pawn. Then you get (in cP) approximately:

P=85, N=300, B=350 (+40 for pair), R=475, A=875, C=900, Q=950

[George Tsavdaris]

With 16. Ah5? Smirf traps the black Chancellor such that black is forced to trade it for the Archbishop. But since Smirf does this when its Knight is attacked by a Pawn, he loses the Knight in the process, for no other compensation that a bit of reduced black King safety.

What a great example!
White just loses a Knight for nothing!
Smirf thinks white is OK since it evaluates the capture of opponent's Chancellor to be worth much more from its loss of Archbishop, so it allows its Knight to be lost.

I'm not optimistic that Reinhard will consider changing this, but i advice him just to create a Smirf with the values you give or i give and play against the current Smirf with the current values.

The aforementioned values are given in the following picture as also the comparisons between each person's predictions.
I have given these predictions 2 years before in the Brainking.com forum, and they were based in the "safe check" method but with multiple combinations of pieces.
And now you, with a complete different method(more logical and practical) find more or less the same correlations between the pieces.

Here i give the prediction of piece values from Ed Trice, Larry Kaufman, myself, Reinhard, and you.
It's amazing how mine and your predictions agree so much!
We only disagree on RP vs BN where i give it slightly better for RP, while you slightly better for BN.

[hgm]

Well, to see if we really disagree, we should define very carefully what the piece values mean. In my case they are early opening values. This means that the Rooks value does not include an open-file bonus. Near the end-game essentially every file will be open, and this bonus will get integrated into the Rook value automatically.

Second point of attention should be the value of the Bishop pair. In comparing B+N vs R it is important whether the B is part of a pair or not.

I did find that N+N heavily beats R+P (from the opening). An alternative system to account for this, however, would be to make the Pawnvalue dependent on the fact if you have a numerical Piece majority or not. As I always measure with 19 or 20 Pawns on the board, I cannot detect such a dependence. I could very well imagine that R+P on the average beats N+N in the late middle-game.

It would be interesting to measure N+N+5P vs R+6P end-games. Perhaps I will do so. (I already tried this for A+5P vs R+N+6P, confirming that the Archbishop value did not decrease towards the end-game. As is already suggested by the fact that KAKBN is mostly won.)

If I normalize to Q=950, we get:

Code: Select all

pc H.G. George
P   86  112
N  302  285
B  345  345
R  475  529
A  885  889
C  907  925
Q  950  950

Note my values are not more accurate than 10-20 cP, so I tend to round them. The main difference is in the Rook value, which might be largely explained by the open-file effect. My Pawn is one devoid of any bonuses (g2/g7). Your Knight is a bit lighter thn mine.

[Matthias Gemuh]

I will surely update BigLion80.

Matthias.

Hi Harm,
I hope you will update BigLion80 now or soon.
I uploaded a stronger version "long" ago, but silently.
Best,
Matthias.

[hgm]

Sorry, I did not know that, so I have been playing the old version so far. And it doesn't do that bad at all, with wins against Smirf Bonus and TSCP-G. Is the upgraded version much stronger?

As I didn't make rules for this, I would like to know if there are any objections from your remaining opponents this cycle (TJchess and Smirf Donation) if I would decide to replace it now. They would be disadvantaged a little compared to their competitors if they had to play a stronger BigLion.

In any case I can replace it before the next cycle starts.

[smrf]

It would make a lot of sense to normalise the piece values having 3.0 as the value for a knight. Because implementation of position depending pawn values is done very differently, it makes a lot of sense not to use pawns as compare unit.

Reinhard.

[hgm]

I prefer to normalize on Queen, because the value of a Queen is the least dependent of all pieces on its position on the board. At least, I think so. Knights and Bishops have comparatively large centralization values, and Bishops have the pair complication. Rooks have open-file and 6th-rank bonuses. But anything is better than Pawns...

In Joker I have no piece-square tables for Queen at all.

[Matthias Gemuh]

Sorry, I did not know that, so I have been playing the old version so far. And it doesn't do that bad at all, with wins against Smirf Bonus and TSCP-G. Is the upgraded version much stronger?

As I didn't make rules for this, I would like to know if there are any objections from your remaining opponents this cycle (TJchess and Smirf Donation) if I would decide to replace it now. They would be disadvantaged a little compared to their competitors if they had to play a stronger BigLion.

In any case I can replace it before the next cycle starts.

The current thing scores about 66% against older and matches eval of BigLion (8x8) more closely.

Matthias.

[smrf]

No veto from SMIRF against replacements, if SMIRF then also would be replaced by MS-173g.

Reinhard.