Author: Sune Fischer
Date: 05:32:49 12/04/01
Go up one level in this thread
On December 04, 2001 at 07:33:08, David Rasmussen wrote: >On December 04, 2001 at 04:30:03, Sune Fischer wrote: > >>> >>>In the exact same test, Crafty gets similar results. 115 collisions in 3 >>>minutes, but at a search-rate of 137kn/s compared to my 450kn/s. So at this >>>rate, Crafty would get about 3-400. >>> >>>/David >> >>Are you sure this is a pawn hash table only? >>I can believe these numbers if they are for the standard hash, but it doensn't >>sound right for the pawns. >> > >Of course I'm sure. It's pretty difficult to mess that up, since the trans/ref >table and the pawn hashtable in Crafty used very differently. > >What I have done in the test is exactly this: >In EvaluatePawns(), whenever I get a hit in the beginning of the function, >Crafty would usually return with the values from the table, but I don't. I just >remember that there was a hit, and the score that came with it. And now, I am >only talking about the final score value, not the pawn structure properties that >Crafty also stores in the table. If I checked on those too, there might be even >more collisions. >At the end of the function, the score has been calculated, and I check if there >was a hit at the beginning. If there was, the score (and other data, which I >don't check) from the pawn hashtable, better be the same, or else we have a >collision. We might have even more, but we definately have these. I just count >these collisions, and print them out. Very simple, and it proves without doubt >that "damaging" collisions do happen, and fairly often. They might not be >damaging after all, since there are relatively few of them overall, but still, >the rate of collisions are very much like what I see with my program. > >/David Hmm, how many different pawn positions do you get? If you get millions it makes sense, but not if you only have a few thousand. The odds of a collision between a few thousand when you have 4 billion slots is not that great (albeit not neglible), it is the famous "birthday problem". I'm not clear on what you mean by: "And now, I am only talking about the final score value, not the pawn structure properties that Crafty also stores in the table." I only pawn hash the pawn structure scores (like double, passers, connected, isolated etc..) which is why I get so few different positions. In more than 999 times of 1000 I get a hit. -S.
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.