Author: Uri Blass
Date: 07:41:18 03/10/01
Go up one level in this thread
On March 10, 2001 at 10:05:04, Robert Hyatt wrote: >On March 10, 2001 at 03:34:19, Uri Blass wrote: > >>On March 10, 2001 at 02:55:29, Dann Corbit wrote: >> >>>On March 10, 2001 at 00:26:57, Michael Fuhrmann wrote: >>> >>>>[D] 8/8/7P/5kP1/5P2/r7/5K2/8 b - - 0 114 >>>> >>>> >>>>Crafty scores this as -3.26 for white on my computer, yet it seems to be a draw. >>>>I don't have all the 5-piece tbs. Is that the problem? >>> >>>I have the complete 5 piece tablebase file set. It does not seem to help much. >>>However, it is clear that it is a draw. The output of the chase below is rather >>>amusing. We might call it "I'll get that f4 pawn if it's the last thing I do!" >>> >>>Past the last actual search depth in plies, for each new search, it always ends >>>in Kxf4. But each time we go one more ply (when we have actually examined the >>>consequences of that horizon capture) it gets pushed out. It is clear that this >>>is a draw. >> >>It seems that the problem is that Crafty does not look at the tablebases in the >>last ply. >>Maybe probing the tablebases everywhere can help Crafty to get more realistic >>evaluation because it is not going to have wrong evaluation for KR vs KPP >>endgames and evaluation for KR vs KPPP endgames is going to be closer to draw. >> >>Uri > > >It isn't quite the "last ply". It is the first ply of the q-search, and I >don't probe there as it is too expensive. Which means the goal of the search >becomes "push the capture off to the q-search where I won't realize it is a >draw." I have seen this in rare positions before... :) I understand the reason that you do it in the first ply of the q-search(in this case it was also the last ply). I do not know if changing it in part of the cases can be an improvement. Some possible ideas: 1)If the main line leads to tablebases position and if the difference between the evaluation of this position with tablebases and without tablebases is big enough then probe everywhere in the next iteration. 2)compare evaluations with tablebases and without tablebases for 1 out of 100 positions when you probe the tablebases. If you get a big difference in significant part of the cases then probe the tablebases everywhere. 3)Start the game with probing the tablebases also in the qsearch. If the number of tablebases hits in the last iteration is not big enough then continue probing the tablebases everywhere. Stop to probe the tablebases in the qsearch only when you get enough tablebases hits so you are afraid from a significant reduction in speed. 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.