My review of Strelka Code

[Dann Corbit]

I get this from a standing start:
[d]r7/r6p/2p1p3/1p3pk1/1P1P4/p1P3PP/R2K4/R7 w - -

E:\programme\winboard\strelka>strelka
new
post
setboard r7/r6p/2p1p3/1p3pk1/1P1P4/p1P3PP/R2K4/R7 w - -
go
1 -624 0 4 a2a3 a7a3
1 -83 0 7 d4d5 c6d5
1 76 0 19 h3h4 g5g4
1 97 1 26 d2e3
1 119 1 27 d2d3
2 118 1 117 d2d3 f5f4 g3f4 g5f4
3 133 3 480 d2d3 h7h5 a1e1
4 116 3 1870 d2d3 g5f6 a1e1 a7d7
4 145 3 3030 d2c2 g5f6 c2b3 a7g7
5 140 6 5104 d2c2 g5f6 c2b3 a7g7 a1g1
6 126 6 7780 d2c2 g5f6 c2b3 a7g7 a1g1 f5f4
7 126 7 17126 d2c2 g5f6 c2b3 a7g7 a1g1 h7h5 g1g2
8 113 7 27958 d2c2 g5f6 c2b3 a7g7 a1g1 h7h5 g1g2 f5f4
9 112 12 67314 d2c2 g5f6 c2b3 a7g7 a1g1 h7h5 h3h4 f5f4 g1f1
10 111 15 131405 d2c2 g5f6 c2b3 a7g7 a1g1 h7h5 h3h4 a8a7 g1g2 f5f4 g3f4 g7g2 a2g2
11 114 35 423305 d2c2 h7h5 a1e1 a7e7 h3h4 g5g4 e1e3 f5f4 g3f4 g4f4 c2d3 f4g4 e3e4 g4f5 a2f2 f5g6
12 103 82 1151527 d2c2 h7h5 a1e1 a7e7 c2b3 h5h4 e1e3 h4g3 e3g3 g5h4 g3f3 e7g7 f3e3
13 98 139 2057132 d2c2 h7h5 a1e1 a7e7 h3h4 g5g4 e1e3 e6e5 d4e5 f5f4 g3f4 g4f4 e3d3 f4e5 c2b3
13 106 289 4405880 c3c4 h7h5 a1e1 a7d7 d2c3 g5f6 e1f1 d7g7 f1f3 a8g8 g3g4 h5g4 h3g4 g7a7
14 57 885 13792316 c3c4 g5h5 d2e3 b5c4 a1c1 a7a4 c1c4 a8b8 c4c6 b8b4 c6e6 b4b3 e3f2 a4d4 e6e2
14 95 925 14427556 d2c2 h7h5 a1e1 a7e7 h3h4 g5g4 e1e3 e6e5 d4e5 f5f4 g3f4 g4f4 e3d3 f4e5 c2b3 e5f4
15 95 1125 17613014 d2c2 h7h5 a1e1 a7e7 h3h4 g5g4 e1e3 e6e5 d4e5 f5f4 g3f4 g4f4 e3d3 e7e5 c2b3 e5e1 d3d4 f4f5
16 80 1542 24348254 d2c2 a7g7 a1f1 h7h5 c2b3 h5h4 g3h4 g5h4 f1e1 g7e7 e1e3 f5f4 e3f3 e7f7 c3c4 h4g5 a2g2 g5f5 h3h4 b5c4 b3c4
17 73 3134 49647160 d2c2 a7g7 a1f1 h7h5 c2b3 h5h4 g3h4 g5h4 f1f2 g7g1 f2e2 g1b1 b3c2 b1g1 e2e6 g1g2 c2b3 g2a2 b3a2 f5f4 e6c6 h4h3
17 80 5032 79678814 a1e1 g5f6 d2e3 a7g7 e3f4 h7h5 e1a1 g7a7 a1f1 a7g7 f1h1 g7a7 h1b1 f6e7 b1b3 e7f6 f4e3
18 79 6970 110335082 a1e1 g5f6 d2e3 a7g7 e3f4 h7h5 e1a1 g7a7 a1f1 a7g7 f1h1 g7a7 h1g1 f6e7 f4e3 e7f6 g1d1 a7g7
19 66 10415 165685028 a1e1 g5f6 d2c2 a7g7 e1e3 a8g8 g3g4 g7a7 e3f3 h7h5 g4f5 e6f5 c2b3 g8g1 a2a3 a7a3 b3a3 f6g5 a3b3 f5f4 f3f2 g5f5
20 67 20620 326094310 a1e1 g5f6 d2c2 a7g7 e1e3 a8g8 g3g4 g7a7 e3f3 h7h5 g4h5 g8g2 c2b3 g2a2 b3a2 f6g5 f3g3 g5h5 g3e3 a7e7 a2a3 f5f4 e3e1

Bryan thanks very much for the new compile!

Dann don't you think something strange is happening especially to the scores of your compilation version?

For the first plydepths PV and nodecounts are still the same, then they begin to differ slightly; 4;7;19;26;27;117;480;1870, and with the regular binary 4;7;19;26;27;117;477;1444
but the scores are roughly +1.4 to +0.8 pawns in your version, -1.4 pawns or less in the regular version, no matter what the plydepth is the difference is more than two pawns, and I don't think the score is inverted in your version, besides it is White to move? Output without game history or hash contents:


r7/r6p/2p1p3/1p3pk1/1P1P4/p1P3PP/R2K4/R7 w - -

Engine: Strelka_10b (64 MB)
gemaakt door Juri Osipov

1.00 0:00 -6.70 1.Txa3 Txa3 2.Txa3 Txa3 (4) 4

1.00 0:00 -4.08 1.d5 cxd5 (7) 7

1.00 0:00 -2.49 1.h4+ Kg4 (19) 19

1.00 0:00 -2.27 1.Ke3 (26) 26

1.00 0:00 -2.06 1.Kd3 (27) 27

2.00 0:00 -2.06 1.Kd3 f4 2.gxf4+ Kxf4 (117) 117

3.00 0:00 -1.85 1.Kd3 h5 2.Te1 (477) 477

4.00 0:00 -2.02 1.Kd3 Kf6 2.Te1 Td7 (1.444) 1444

5.00 0:00 -1.91 1.Kd3 Kf6 2.Kc2 Tg7 3.Txa3 Txa3
4.Txa3 Txg3 (4.954) 4954

6.00 0:00 -1.91 1.Kd3 Kf6 2.Kc2 Tg7 3.Txa3 Txa3
4.Txa3 Txg3 (8.274) 8274

6.00 0:00 -1.90 1.c4 Kf6 2.Kc3 Tg7 3.g4 fxg4 4.hxg4 (18.650) 1865

6.00 0:00 -1.57 1.Kc2 Kf6 2.Kb3 Tg8 3.g4 fxg4 4.Tf1+ Kg5
5.hxg4 Kxg4 6.Txa3 Txa3+ 7.Kxa3 (25.236) 2523

7.00 0:00 -1.46 1.Kc2 Kf6 2.Kb3 Tg8 3.g4 fxg4 4.hxg4 Txg4
5.Tf1+ Kg5 6.Txa3 Txa3+ 7.Kxa3 (32.938) 1097

8.00 0:00 -1.46 1.Kc2 Kf6 2.Kb3 Tg8 3.g4 fxg4 4.hxg4 Txg4
5.Tf1+ Kg5 6.Txa3 Txa3+ 7.Kxa3 (52.058) 1301

9.00 0:00 -1.46 1.Kc2 Kf6 2.Kb3 Tg8 3.g4 fxg4 4.Tf1+ Ke7
5.hxg4 Txg4 6.Txa3 Txa3+ 7.Kxa3 (81.071) 1158

10.00 0:00 -1.62 1.Kc2 Kf6 2.Kb3 h5 3.Tg1 Tg7 4.Tgg2 Tag8
5.Txa3 Txg3 (137.397) 1249

11.00 0:00 -1.54 1.Kc2 Kf6 2.Kb3 h5 3.Te1 Tg7 4.Te3 h4
5.gxh4 f4 6.Te1 (228.032) 1266

12.00 0:00 -1.57 1.Kc2 Kf6 2.Te1 Tg7 3.Te3 Tag8 4.g4 fxg4
5.hxg4 Ta7 6.Kb3 Txg4 7.Tf3+ Ke7
8.Txa3 Txa3+ 9.Kxa3 (545.388) 1298

13.00 0:01 -1.47 1.Kc2 Kf6 2.g4 fxg4 3.hxg4 e5 4.Tf1+ Kg5
5.dxe5 Te7 6.Te1 Kxg4 7.Kb3 Kf5
8.Tf2+ Ke6 (1.381.429) 1291

14.00 0:02 -1.31 1.Kc2 h5 2.Kb3 h4 3.gxh4+ Kxh4 4.Te1 Te7
5.Te3 f4 6.Tf3 Kg5 7.Tg2+ Kf5 8.Tg4 e5
9.dxe5 (2.614.476) 1300

15.00 0:04 -1.30 1.Kc2 h5 2.Kb3 Ta6 3.Tf1 T8a7 4.Tfa1 Ta4
5.h4+ Kg4 6.Te1 Kxg3 7.Txe6 f4
8.Tg6+ Kh3 9.Te2 Kxh4 10.Txc6 (5.314.452) 1305

16.00 0:06 -1.27 1.Kc2 h5 2.Kb3 Ta6 3.Tf1 h4 4.gxh4+ Kxh4
5.Tg1 Tf8 6.Tg6 f4 7.Ta1 f3 8.Tag1 Tf5
9.Th6+ Th5 10.Tf6 a2 11.Tf4+ Kxh3
12.Txf3+ Kh2 (7.868.624) 1309

17.00 0:10 -1.43 1.Kc2 h5 2.Kb3 Kf6 3.Te1 Ke7 4.Te3 Kd6
5.c4 bxc4+ 6.Kxc4 Ta6 7.Te1 Ta4
8.Kb3 T4a7 9.Te3 Kd5 10.Te5+ Kd6 (13.343.207) 1294

18.00 0:17 -1.46 1.Kc2 h5 2.Kb3 Kf6 3.Te1 Ke7 4.Te3 Kd6
5.c4 Ta6 6.Te5 bxc4+ 7.Kxc4 Ta4
8.Kb3 T4a7 9.Te3 Ta6 10.Te1 Tg8 (22.771.386) 1292

19.00 0:27 -1.45 1.Kc2 h5 2.Kb3 Kf6 3.Te1 Ke7 4.Te3 Kd6
5.c4 Ta6 6.Te5 bxc4+ 7.Kxc4 Ta4
8.Kb3 T4a7 9.Kc4 Ta6 10.Te3 Ta4
11.Te1 (35.292.166) 1297

20.00 0:48 -1.43 1.Kc2 h5 2.Kb3 Kf6 3.Te1 Ke7 4.Te3 Kd6
5.c4 Ta6 6.Te5 bxc4+ 7.Kxc4 T6a7
8.Te3 Ta4 9.Kb3 T4a6 10.Kc4 T6a7
11.Kb3 Kd5 12.Te5+ Kd6 (62.816.225) 1297

21.00 1:30 -1.43 1.Kc2 h5 2.Kb3 Kf6 3.Te1 Ke7 4.Te3 Kd6
5.c4 Ta6 6.Te5 bxc4+ 7.Kxc4 T6a7
8.Te3 Ta4 9.Kb3 T4a6 10.Kc4 T6a7
11.Kb3 Kd5 12.Te5+ Kd6 (117.226.681) 1288

21.00 2:31 -1.43 1.Kc2 h5 2.Kb3 Kf6 3.Te1 Ke7 4.Te3 Kd6
5.c4 Ta6 6.Te5 bxc4+ 7.Kxc4 T6a7
8.Te3 Ta4 9.Kb3 T4a6 10.Kc4 T6a7
11.Kb3 Kd5 12.Te5+ Kd6 (194.902.892) 1288

beste zet: Kd2-c2 tijd: 2:31.250 min n/s: 1.288.784 nodes: 194.902.892

Regards, Eelco

Yes, it appears that my build is broken. Probably there is some undefined behavior somewhere, or I did an optimization that is unsafe for this program (e.g. assume no alias).

I also noticed that Strelka was wildly optimistic about his position while he was playing Rybka. It appears that Brian Hoffman's build does not suffer from this defect.

[Werner]

Yes, it appears that my build is broken. Probably there is some undefined behavior somewhere, or I did an optimization that is unsafe for this program (e.g. assume no alias).

I also noticed that Strelka was wildly optimistic about his position while he was playing Rybka. It appears that Brian Hoffman's build does not suffer from this defect.

Hi Dann,
I can confirm this. With Bryan´s build I now get the same results as with the original engine - but a little bit faster!
Thanks Bryan for the very fast compilation. In this position it was around 60%!!