Author: James T. Walker
Date: 07:04:39 12/02/01
Go up one level in this thread
On December 01, 2001 at 21:57:05, Mike S. wrote: >On December 01, 2001 at 20:10:09, James T. Walker wrote: > >>On December 01, 2001 at 18:47:10, Russell Reagan wrote: >> >>>On December 01, 2001 at 16:35:31, Mike S. wrote: >>> >>>>(...) >>>>I'm aware most people will say "use all 5 men tbs", but large experiments have >>>>been made showing that the slowdown effect and other effects from using the tbs, >>>>often result in *worse* performances in practical positions, than without the >>>>tbs. Misjudgement of continuations which keep the game out of tbs material, is >>>>one of these effects, which i.e. can make the win unecessarily easy for the >>>>opponent (which may - without tbs - fail to win if the best defense in played). >>>>(...) > >>>(...) I don't see how this >>>slows down anything, unless you're talking about incorporating probing the >>>tablebases into your search and evaluation of many positions. I hope someone >>>will explain this to me, because I've never implemented tablebases into my own >>>program and I plan to eventually, so I'm sure this will be something I'll need >>>to know eventually. > >Yes, I meant using them in the search (much) before a tbs situation is actually >reached in the game. > >Btw. in conjunction to that, I mentioned in earlier discussions M-Chess Pro's >ability to configure *two types* of endgame table directories: One which is >accessed during the search, and another one which is only accessed when the >endgame material is on the board already. M-Chess didn't use the Nalimovs yet, >but I think because of the doubts expressed in my 1st posting, it would make >sense now too. Because now if I leave most 5 piece tbs away, they are always >left away... I don't recall any replies though - people are either not >interested in too much detail regarding the tbs, or underestimate the possible >downside when using everything "as is". > >>Most all programs are probing the tablebases at the leafs. This slows some down >>to about half the normal K/nps. Some like Shredder slow down to 10% of normal >>(In some positions) but the ply depth is greater because of the information >>obtained from the tablebases. >>By the way, I had a position today where one program with tablebases had one >>queen and announced mate on itself. The other program had two queens but >>without tablebases only managed a draw. > >This is a drastic example. If resigning was on, the program with the single >queen has wasted half a point... but that doesn't mean here, it played moves >worse than without tablebases. > >From the experiment I mentioned, this was the position (NET#10) with the most >unexpected results. Seven different programs played the position twice against >themselves, but only one side each used the 5 piece tbs. With one exception, >White *with* the tablebases lost always while *without* the tbs, White could >most always draw. > >[D]3R4/1p6/2b5/2P1k2p/p3p2P/P6r/1P2KB2/8 b - - 0 1 >"No TBS" vs. "TBS": 9.5 - 4.5 > >So, what (sub-)set of the tablebases is most recommendable? > >I'm sure programmers will have an opinion built on their own comprehensive >testing. > >Furthermore, what do you thing about the configuration idea M-Chess Pro did use >(differents directories for dynamic/static access)? Is there any other program, >or GUI, which uses this idea? > >Regards, >M.Scheidl Hello Mike, I'm interested in your experiment. What time control did you use for this test. Also can you tell me the names of the programs you used. I find your results very hard to believe and would like to do some test myself along this line. Regards, Jim
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.