Author: Uri Blass
Date: 00:58:41 04/12/03
Go up one level in this thread
On April 11, 2003 at 22:00:21, Peter McKenzie wrote: >[D] 6k1/7p/4Q1p1/3pP1P1/p6q/Pp1r2R1/1P5P/7K b - - 0 1 > >This position was posted a while ago, and it was noted that some programs (eg. >Hiarcs, Warp) still report a very big score for white after a very long search. >On the other hand, other programs like Crafty, Yace, and movei seem to see the >draw (or close to it) within 10-12 ply. > >I decided to investigate this a bit further, starting with the hypothesis that >the main differentiator here is the effectiveness of a program's main hash >table. I modified Warp's replacement scheme with a little hack, telling it >never to replace a draw score with a non-draw score. > >With this change, Warp can see the draw at depth 11 while previously it was >still showing +4.8 at depth 15! > >While its unlikely this little hack will help overall strength, it does indicate >that its the hashing scheme which is critical here. I use a single hash table, >with 4 probes. Its basically an always replace strategy, where I replace the >record with the lowest draft (of the records probed). My replacement strategy is not very important because I use hash tables only for order of moves(I have explanations for the reasons that I do it). > >Interested in how others get on with this position, and if the hack gives >similar results if your program has trouble with it. > >cheers, >Peter I think that movei can see scores that are close to draw thanks to extensions. It is also a problem because there are positions that it cannot go very deep because of check extensions but I think that I have an idea how to do less check extensions without doing less check extensions in the relevant lines of this position. Movei still does not use hash tables for pruning. The problem is that the score is dependent not only on the leaf position but also on other factors. Today the influence of the other factors is small but I have ideas that I still did not test to make it bigger in the future. The fact that the score is dependent on other factors does not mean that it is impossible to use hash tables for pruning(there are scores like mate score when I am sure about the score so it is possible to use them for pruning in that case) but it is not simple to do it. I decided that in the near future I prefer to spend time on other things and not on investigating how to use hash tables for pruning in a productive way. I believe that there is a lot of things that I can do to improve movei that are not about using hash tables for pruning. I may consider increasing my small scores in some cases and changing my small scores in other cases but I prefer before doing it to improve the static part of my evaluation. Uri
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.