Author: Uri Blass
Date: 14:33:44 12/22/04
Go up one level in this thread
On December 22, 2004 at 16:50:40, Ulrich Tuerke wrote: >On December 22, 2004 at 16:27:13, Uri Blass wrote: > >>On December 22, 2004 at 16:19:05, Ulrich Tuerke wrote: >> >>>On December 21, 2004 at 05:11:13, Uri Blass wrote: >>> >>>>On December 21, 2004 at 03:53:49, Joachim Rang wrote: >>>> >>>>>On December 20, 2004 at 16:32:24, Jouni Uski wrote: >>>>> >>>>>>Shredder 8 reaches quite stunning averaga search dept. E.g. with only 15 >>>>>>second/move average with Pentium 2400 and 128MB hash it averages 15,2 ply! For >>>>>>comparison Fritz Bilbao and Gandalf 6 reach both same 12,7 ply. No wonder >>>>>>Shredder beats them 13 - 7 and 14 - 6. Of course 2,5 ply is a lot difference! >>>>>>Is there any amateur or pro engine, which can equal Shredder in depth - Junior's >>>>>>half plys can be forgotten. >>>>>> >>>>>>Jouni >>>>> >>>>>Here is a 1-minute search on my Athlon XP@1540 MHz, with Shredder 7.04 default >>>>>and Fruit X with aggressive delta, futility and history pruning. Both engines >>>>>were analysing together under Arena. Compare the depth! >>>> >>>>I am surprised to read that fruit2 has history pruning. >>>>I understood that Fabien does not plan to implement history based pruning and >>>>here is the source of my misunderstanding: >>>> >>>> >>>>The reason is the following link: >>>>http://f11.parsimony.net/forum16635/messages/68164.htm >>>> >>>>Fabien claimed in that link: >>>> >>>>"Forward pruning in the main search is a separate topic that I don't intend to >>>>address now. If I ever do, I expect it will require years of work (same for a >>>>proper move ordering)." >>>> >>>>Note that I consider history based pruning as forward pruning in the main >>>>search. >>> >>>Doesn't the history heuristic simply mean that moves which had often produced >>>cuts in the previous iteration will be tried as killers ? >> >>No >> >>It means reducing depth of moves that almost always caused fail low based on >>statistics. > >Hm ..., with the intention to avoid production of these nasty fail lows again by >reducing their search depth ? No The intention is not to avoid production of these fail lows that I am almost sure that I will get but to produce them faster. Note that if I get unexpected fail high I research again to the original depth. As far as I know this idea is used by Terra and Gothmog and smarthink. This idea was original idea of me and I thought not to share it but after sergei markof (author of smarthink) already talked about it there is no reason to hide it. I do not share the exact conditions that I decide to reduce depth but one of the conditions is that based on history experience the move failed low in many cases that it was tried and almost never failed high. history tables give you only the frequency of fail high so you need to build different tables. 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.