Probcut
Posted: Fri May 24, 2013 7:15 am
I'm trying an experiment with Probcut http://chessprogramming.wikispaces.com/ProbCut on Stockfish. I took 55000 positions from games, and tried depth 15 searches, recording the scores at each move. Then did a least squares fit, predicting how well the scores of the low depth search predicts the high depth searches.
SF uses a depth reduction of 4 for probcut, and the fitted values are below. I used a simple formula, instead of the static reduction that SF currently uses. Should be interesting to see how it goes!
https://github.com/glinscott/Stockfish/ ... ...46b0441
The code (a giant mess!) is here https://gist.github.com/glinscott/5641353.
Here is the output for all depth pairs up to 15.
SF uses a depth reduction of 4 for probcut, and the fitted values are below. I used a simple formula, instead of the static reduction that SF currently uses. Should be interesting to see how it goes!
https://github.com/glinscott/Stockfish/ ... ...46b0441
Code: Select all
lo hi slope off stdev
1 5 - 0.9 -11.3 62.2
2 6 - 0.9 -3.2 57.8
3 7 - 0.9 -5.4 56.3
4 8 - 0.9 -1.0 62.4
5 9 - 1.0 -3.6 44.3
6 10 - 1.0 -1.1 64.0
7 11 - 1.0 -1.1 83.7
8 12 - 1.1 -0.8 110.9
9 13 - 1.2 -0.1 136.0
10 14 - 1.2 -0.1 166.3
11 15 - 1.2 0.3 177.6
Here is the output for all depth pairs up to 15.
Code: Select all
lo hi slope off stdev
1 2 - 0.9 -11.1 34.8
1 3 - 0.9 -8.6 42.6
1 4 - 0.9 -13.1 61.0
1 5 - 0.9 -11.3 62.2
1 6 - 0.9 -13.5 65.6
1 7 - 0.9 -13.6 69.1
1 8 - 0.9 -14.2 72.3
1 9 - 0.9 -14.8 76.1
1 10 - 0.9 -14.8 92.5
1 11 - 0.9 -15.1 112.4
1 12 - 0.9 -15.8 138.8
1 13 - 1.0 -15.9 170.2
1 14 - 1.0 -16.5 201.5
1 15 - 1.0 -16.7 223.8
2 3 - 1.0 2.3 31.4
2 4 - 0.9 -2.6 51.6
2 5 - 0.9 -0.8 54.7
2 6 - 0.9 -3.2 57.8
2 7 - 0.9 -3.3 62.0
2 8 - 0.9 -3.9 65.3
2 9 - 0.9 -4.4 69.3
2 10 - 0.9 -4.3 86.8
2 11 - 1.0 -4.5 107.5
2 12 - 1.0 -4.9 134.5
2 13 - 1.0 -4.6 166.3
2 14 - 1.1 -4.8 198.0
2 15 - 1.1 -4.7 220.3
3 4 - 1.0 -4.7 45.3
3 5 - 1.0 -3.0 47.6
3 6 - 1.0 -5.3 52.2
3 7 - 0.9 -5.4 56.3
3 8 - 1.0 -6.0 59.9
3 9 - 1.0 -6.6 63.9
3 10 - 1.0 -6.5 82.5
3 11 - 1.0 -6.7 103.6
3 12 - 1.0 -7.2 131.1
3 13 - 1.1 -7.1 162.7
3 14 - 1.1 -7.4 194.6
3 15 - 1.1 -7.4 217.2
4 5 - 0.9 2.1 51.6
4 6 - 0.9 -0.3 55.0
4 7 - 0.9 -0.4 59.3
4 8 - 0.9 -1.0 62.4
4 9 - 0.9 -1.5 66.2
4 10 - 0.9 -1.4 84.1
4 11 - 0.9 -1.5 104.7
4 12 - 1.0 -1.8 131.9
4 13 - 1.0 -1.6 162.8
4 14 - 1.1 -1.6 194.8
4 15 - 1.1 -1.4 217.2
5 6 - 1.0 -2.3 26.1
5 7 - 1.0 -2.5 34.0
5 8 - 1.0 -3.1 39.2
5 9 - 1.0 -3.6 44.3
5 10 - 1.0 -3.5 67.9
5 11 - 1.0 -3.7 91.6
5 12 - 1.1 -4.1 121.0
5 13 - 1.1 -3.9 153.3
5 14 - 1.1 -4.0 186.4
5 15 - 1.2 -4.0 208.5
6 7 - 1.0 -0.1 26.3
6 8 - 1.0 -0.7 32.6
6 9 - 1.0 -1.2 38.8
6 10 - 1.0 -1.1 64.0
6 11 - 1.0 -1.2 88.4
6 12 - 1.1 -1.5 118.2
6 13 - 1.1 -1.2 150.7
6 14 - 1.2 -1.3 184.0
6 15 - 1.2 -1.1 206.2
7 8 - 1.0 -0.5 22.9
7 9 - 1.0 -1.1 31.2
7 10 - 1.0 -1.0 59.3
7 11 - 1.0 -1.1 83.7
7 12 - 1.1 -1.4 114.3
7 13 - 1.1 -1.2 146.1
7 14 - 1.2 -1.2 179.7
7 15 - 1.2 -1.1 201.9
8 9 - 1.0 -0.5 22.4
8 10 - 1.0 -0.4 54.1
8 11 - 1.1 -0.5 79.7
8 12 - 1.1 -0.8 110.9
8 13 - 1.2 -0.6 142.1
8 14 - 1.2 -0.6 176.1
8 15 - 1.3 -0.5 198.2
9 10 - 1.0 0.1 46.8
9 11 - 1.1 -0.0 74.1
9 12 - 1.1 -0.3 105.8
9 13 - 1.2 -0.1 136.0
9 14 - 1.2 -0.1 171.3
9 15 - 1.3 0.1 193.4
10 11 - 1.0 -0.1 55.1
10 12 - 1.1 -0.5 90.2
10 13 - 1.2 -0.2 123.7
10 14 - 1.2 -0.1 166.3
10 15 - 1.2 0.1 188.6
11 12 - 1.0 -0.3 70.3
11 13 - 1.1 -0.1 109.3
11 14 - 1.1 0.0 154.5
11 15 - 1.2 0.3 177.6
12 13 - 1.1 0.3 80.8
12 14 - 1.1 0.4 131.5
12 15 - 1.1 0.6 156.0
13 14 - 1.0 0.0 98.8
13 15 - 1.1 0.2 126.1
14 15 - 1.0 0.2 75.6