Author: David Rasmussen
Date: 12:20:05 12/05/01
Go up one level in this thread
On December 05, 2001 at 14:11:25, Robert Hyatt wrote: >> >>You showed me _some_ math. Whether it describes the situation or not remains to >>be shown. > > >Some points to ponder. > >Assuming you search 1M nodes per second, you really only have to do a >full evaluation on 10% of them, typically. The rest can be eliminated >by 'lazy eval' testing since if you are two queens down, positional scores >likely won't help. > >This means that for 3 minutes of search, you will fully evaluate maybe 20M >positions. Now, for those 20M positions, how many will contain _different_ >pawn positions that you have never seen before? > >I believe that number is very small, simply because of the testing I have >done in the past. > > I _know_ the number is "small". I have done testing myself, you know :) I didn't say anything that contradicts that. I am just saying I am not convinced that it describes the situation thoroughly (i.e. that there will be no collisions), as I find collisions. is >>testing exactly the invariant that makes (or should make) the pawn hashtable >>work, namely that a hit implies data about the correct position, not another one >>with same hash signature (a collision). The experiment doesn't even alter the >>way the program runs at all, it just makes it run a bit slower. So take your >>pick: > > >One easy test in Crafty is to save the "hashed" stuff and _still_ call >EvaluatePawns() to see if the computed stuff differs. I have done this That is exactly what I have done. Have I really described my experiment so badly that no one understands? :) >often in debugging and I do not ever see any disagreement here. Which >doesn't necessarily mean there are no "collisions" but that if there are, >the two different positions produce both the same signature _and_ the same >score (and all the other bitmaps that are hashed for other evaluation >terms). > Sure, that's _exactly_ what I've been saying. Although I _do_ find collisions. "Many" collisions. > > > >> >>1) Crafty has pawn hash collisions, but they don't matter because they are so >>few (why?) >> >>2) Crafty has pawn hash collisions, and it makes it does matter performancewise >>somehow > >(3) Crafty doesn't any pawn hash collisions to speak of... > > It does on the machines I've tried, using the experiment you describe. /David
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.