Author: Robert Hyatt
Date: 07:34:33 10/17/99
Go up one level in this thread
On October 16, 1999 at 15:36:06, Ratko V Tomic wrote:
>> I thought that might be the case after I did this calculation on my own
>> program's results and got 5.1--I thought I'd introduced a bug or something
>> at first. Then I decided it was blowing up due to excessive extensions.
>> Anyway, that calculation was on a tactical possition which (elsewhere in
>> this thread) Bob says will give larger nunbers.
>
>I guess it depends what kind of tactical position. Something that unfolds at the
>knife's edges, as it were, with only one, at most two, decisevely winning and
>the rest decisevely losing moves, would have very efficient cutoffs and the
>clear cut best move (or couple) from itereation to iteration. Position with lots
>of pieces but without any urgent threats and not cluttered or blocked, positions
>where many viable long term manouvers exist, would be the most expensive since
>several good alternatives differ a little and for fragile reasons which will
>vary from iteration to iteration. So the sorting will fluctuate for random
>reasons and cutoffs will be less efficient.
>
>> I know I've measured my program doing BF's of 3 or so, but on most of
>> the test positions I've just tried I get about 4.6...
>
>That's exactly the middle of the interval 4.5-4.7 I measured on Crafty 16.6
>engine (with root width 38).
Here is some real data taken from long games played on ICC (level 0 30 30,
30 minutes on the clock to start with, 30 seconds added after each move):
(4) 8-> 7.77 -0.18 10. ... Bf6 11. Nbd2 Nc5 12. Bc2 Qd7
13. b4 Ne6 14. Qe2
(3) 9 10.72 -0.15 10. ... Bf6 11. Na3 Nc5 12. Bd5 Bxf3
13. gxf3 Qd7 14. b4 Na6 15. b5
(2) 9-> 14.34 -0.15 10. ... Bf6 11. Na3 Nc5 12. Bd5 Bxf3
13. gxf3 Qd7 14. b4 Na6 15. b5
10 22.02 -0.21 10. ... Bf6 11. Nbd2 Bg6 12. Qe2 Qe7
13. Nf1 Na5 14. Bc2 Nc6 15. Ng3
10 23.37 -0.20 10. ... Bg6 11. Nbd2 Na5 12. Bc2 f5
13. b4 fxe4 14. Nxe4 Nc6 15. Bb3+ Bf7
16. Qe2 Bxb3 17. axb3
(3) 10-> 30.19 -0.20 10. ... Bg6 11. Nbd2 Na5 12. Bc2 f5
13. b4 fxe4 14. Nxe4 Nc6 15. Bb3+ Bf7
16. Qe2 Bxb3 17. axb3
(2) 11 37.45 -0.16 10. ... Bg6 11. Nbd2 Na5 12. Ba4 c5
13. Qe2 f5 14. exf5 Bxf5 15. Ne4 Nf6
16. Neg5
11-> 46.52 -0.16 10. ... Bg6 11. Nbd2 Na5 12. Ba4 c5
13. Qe2 f5 14. exf5 Bxf5 15. Ne4 Nf6
16. Neg5
12 1:09 -0.13 10. ... Bg6 11. Nbd2 Na5 12. Ba4 c5
13. a3 Nb6 14. Bc2 Nc6 15. Qe2 d5 16.
(2) 12-> 1:33 -0.13 10. ... Bg6 11. Nbd2 Na5 12. Ba4 c5
13. a3 Nb6 14. Bc2 Nc6 15. Qe2 d5 16.
Rad1
based on time:
N N+1 time ratio
7 8 5.49 3.66
8 9 6.57 1.20
9 10 15.85 2.41
10 11 16.33 1.03
11 12 52.48 3.21
those are calculated by first determining how long it took for each
iteration by subtracting the end time for successive iteration. Then
simple division.
nss depth time score variation (1)
6-> 0.26 -0.04 22. ... Bd8 23. Qb4 Re6 24. Nf4 Rf6
25. Nxg6 Rxg6
7 0.50 -0.13 22. ... Bd8 23. Qb4 h6 24. Bd4 Qe6
25. Qc3 Bg5
(3) 7-> 1.22 -0.13 22. ... Bd8 23. Qb4 h6 24. Bd4 Qe6
25. Qc3 Bg5
(2) 8 3.44 -0.14 22. ... Bd8 23. Bf4 Re6 24. Nc4 Nxc4
25. Qxc4 Qe8 26. f3 Bh4
8-> 6.31 -0.14 22. ... Bd8 23. Bf4 Re6 24. Nc4 Nxc4
25. Qxc4 Qe8 26. f3 Bh4
9 10.06 -0.19 22. ... Bd8 23. Qb4 Nb7 24. Rc6 Nc5
25. Bf4 Re6 26. Qd4 Rc8 27. Rxc8 Qxc8
9-> 14.38 -0.19 22. ... Bd8 23. Qb4 Nb7 24. Rc6 Nc5
25. Bf4 Re6 26. Qd4 Rc8 27. Rxc8 Qxc8
10 19.99 -0.11 22. ... Bd8 23. Qb4 Nb7 24. Rc6 Qe6
25. Bf4 f5 26. Nc7 Bxc7 27. Rxc7 Nc5
10-> 25.02 -0.11 22. ... Bd8 23. Qb4 Nb7 24. Rc6 Qe6
25. Bf4 f5 26. Nc7 Bxc7 27. Rxc7 Nc5
11 36.94 -0.14 22. ... Bd8 23. Qb4 Nb7 24. Rc6 Re5
25. Bf4 Bxe4 26. Nxe4 Rxd5 27. Bxd6
Nxd6 28. Nxd6
11-> 49.43 -0.14 22. ... Bd8 23. Qb4 Nb7 24. Rc6 Re5
25. Bf4 Bxe4 26. Nxe4 Rxd5 27. Bxd6
Nxd6 28. Nxd6
N N+1 time ratio
7 8 5.09 5.3
8 9 8.08 1.58
9 10 10.64 1.32
10 11 24.41 2.29
As you can see, at two random places it is well below 3 most of the
time. the 5.3 above is pretty high, and might be a time quantizaton
error as the previous iteration claimed to take .96 seconds... that can be
off by up to .2 in unix.
The main thing is to not count cumulative times, just the time iteration to
iteration.
Also note that the EBF is lower in games than in test positions. Because the
hash table carries move ordering from search to search, which avoids some bad
EBF values you can see when you just set a position and say search this, even
though you know nothing about what has happened so far...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.