Author: Dan Newman
Date: 04:10:35 11/14/01
Go up one level in this thread
On November 14, 2001 at 06:08:46, Gian-Carlo Pascutto wrote: >On November 14, 2001 at 05:23:08, Dan Newman wrote: > >>I think several of us here have done the experiment. ISTR that Bob did >>this some time ago (perhaps with Cray Blitz) and I vaguely recall that >>Vincent (or was it Ed?) also has measured the (high) error rate that you >>get with too few bits. I'm sure others have, too. >> >>When I measured it several years back I got about one collision (error) >>per second with a 32-bit hash code. (I can't remember the probe rate, >>but it was probably about 100k/s at the time.) >> >>Anyway, 32 bits seems too small, but I don't really know if the errors >>would have any measurable effect on engine strength. I've always just >>assumed that any such error is a bad thing, and 64 bits virtually >>eliminates them... > >This is _exactly_ my point. Bob and others, and now you, have been >claiming for a long time that 32 bits is not enough because they >see a lot of collisions with it. > >Why would this necessarily mean 32 bits hashing is not enough? > >-- >GCP Well, I guess I never thought about it too hard. A lot of things in chess programs never get tested and are just put in because it seems right to do so...or because everybody else does it... I guess there would be one of two extremes of result: 1) the effect on strength is unmeasurable (via games or test suites) or 2) it is measurable. In the case that it is measurably better you surely want to use the 64 bits since putting this in your program doesn't require much thought. In the case that it isn't easily measured, paranoia tells me that it will cause me to lose a game here and there by producing a bad score in exactly the wrong place. So either way I've got to use 64 bits :). (I guess I left out the third case where 32 bits is faster and so does better than 64.) I decided to try an experiment to see if I got different results on a test suite using just 32 bits. Part way into the test, Shrike crashed. So it looks like I've probably got a bug in my hash table move validity checker. Looks like 64 bits spares my program from such failures--so it has at least some utility :). -Dan.
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.