Author: Dann Corbit
Date: 11:30:05 04/28/04
Go up one level in this thread
On April 27, 2004 at 23:19:53, Dann Corbit wrote: >On April 27, 2004 at 22:10:56, Joel wrote: > >>Hey Dann, >> >>When you say tune, do you mean the evaluation function only? >> >>If this is the case, don't even think you could say you improved your tactical >>strength if you just tuned it for a few (or maybe many testsuites). What I think >>would be likely is that you would be mucking your eval terms up so they solved >>the positions the search struggled with, with the dominating woodcount along >>with the search doing well in the rest of the positions. >> >>Is there someplace you suggest (your ftp perhaps?) where I can find a bit more >>information about this parabolic fit technique you use? I would be interested in >>trying some experiments of my own with it. > >ftp://cap.connx.com/pub/chess-engines/new-approach/beocurve.zip > >Take a look at curve.c, pers.c, and parser.c > >The test positions are defined in testdat.h I added 10,000 quiet positions, and did some curves for some pruning constants. It seems that prune_soften is not a parabolic fit (or the data varies wildly to the degree that a parabola cannot be fitted with any confidence). When it says "at 4" that means at a depth of 4 plies. prune_ratio=51 at 4; stddev=3.415650 prune_factor=373 at 4; stddev=0.000000 prune_constant=599 at 4; stddev=0.000000 prune_soften=4999 at 4; {badfit} prune_ratio=52 at 5; stddev=6.750661 prune_factor=305 at 5; stddev=0.000000 prune_constant=1 at 5; stddev=0.000000 prune_soften=1667 at 5; {badfit} prune_ratio=51 at 6; stddev=108.080933 prune_factor=456 at 6; stddev=0.243975 prune_constant=512 at 6; stddev=0.000000 prune_soften=4999 at 6; {badfit}
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.