Author: Heinz van Kempen
Date: 13:31:44 12/08/05
Go up one level in this thread
On December 08, 2005 at 14:14:46, Dann Corbit wrote: >On December 08, 2005 at 04:03:26, Heinz van Kempen wrote: > >>On December 08, 2005 at 03:50:37, Jouni Uski wrote: >> >>>3 Fruit 2.2 2768 16 16 1385 67.8 % 2638 27.8 % >>>4 Fruit 2.2.1 2764 15 15 1414 59.1 % 2700 34.5 % >>> >>>EGTBs weaken engines play? >>> >>>Jouni >> >>Hi Jouni, >> >>this is frm the Blitz rating list. In the 40/40 list the have the same rating >>according to BayesELO. >> >>There is another example: compare Loop List 600 EGTB and Loop List 600 noEGTB. >>Of course more games needed here. > >Only 16 Elo difference (within experimental uncertainty). I think it is safe to >say that EGTB files (in general) do not increase strength of play. The cost of >a probe seems to be about equal to the value of the knowledge gained. I have >never seen any EGTB experiment where the EGTB version showed a clear strength >increase over the same program playing without them. I guess that a weak >program might generate an improvement. But generally, EGTB files are added at a >later date in development. After all, KRk is going to be won with the rook, >with or without EGTB files. The only real question is whether or not the path >will be optimal. > >Of course, a really fast disk subsystem or smarter probing could change the >balance. > >Bitbase experiments do seem to have measurable strength increases. > >>http://kd.lab.nig.ac.jp/chess/cegt/rating-table-all-300-shifted.shtml >> >>Best Regards >>Heinz Hi Dann and Uri, we wil soon have more games for Loop List 600 EGTB and will also have the opportunity to test Rybka with EGTB support. So maybe after this there will be more clues about this interesting subject if there is improvement for engines using EGTB, or only for some, and if for which ones and why. Maybe this is also partly dependent on a more conservative approach like for Fruit 2.2.1 with default tablebases depth=8 or using it earlier when speed loss will be a disadvantage that must be (over-)compensated by effective hits. Best Regards Heinz
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.