Author: J. Wesley Cleveland
Date: 15:29:45 12/07/01
Go up one level in this thread
On December 07, 2001 at 14:57:52, Frank Phillips wrote: >On December 06, 2001 at 21:13:35, Dan Newman wrote: > >>On December 06, 2001 at 17:29:43, Severi Salminen wrote: >> >>>Hi >>> >>>How many collisions crafty gets on average using 32-bit keys? So how many nodes >>>Crafty searchs on average to get 1 collision? I'm now using plain Visual C++ 6.0 >>>rand() with no hamming distance tests and I get about 80 collisions out of >>>10'000'000 evaluations from initial position. I'd like to know if that is more >>>or less than Robert and David were seeing. Funny thing was that first I searched >>>about 5'000'000 nodes with no collisions, then I saw 40 collisions in a short >>>time, then again no collisions and finally 40 more in a short time. >>> >>>Severi >> >>I did this test on my program, Shrike, and got 62 collisions out of >>66 million pawn hash probes with a 32-bit hash code, so I get about >>1 collision per million probes. This was on the WAC test suite and >>so may vary in actual games... >> >>-Dan. > >I store the lower 32bits as the hash key and use (some of) the top bits as the >index. This is effectively more than 32bits, but I feel I now feel the need to >check it is enought. > >How did you test for clashes: did you store the board along with the hash entry >and check that the position on the board and stored board were the same - or is >there a smarter way? Simplest way is to store all 64 bits of hash. If the lower 32 bits match, and the upper 32 don't, you have a clash.
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.