Author: martin fierz
Date: 06:31:36 06/17/03
Go up one level in this thread
On June 17, 2003 at 08:25:32, Kim Roper Jensen wrote: >Hi > >After reading the posts regarding hashtables, where Robert Hyatt and others show >that the most significant improvements is in the 24M-96M area I was thinking >about the whole hashtable thingie... > >Reading some papers(of course i forgot which, i will try to locate again) it >seems that stuffing more of the PV in the hashtable will improve the efficiency >another notch(article states 3% in middlegame and 12% in the endgame) > >As i recall they stuffed 2.5 ply PV in the hashtable. > >So what i am really asking is this : > >1. Have anyone ever tried to put longer lines of PV in the hashtable ? > >2. Could there also be a "better" number for how many ply's that should be >inserted ? > >I was just thinking instead of using a big hashtable, maybe it would be better >making a bigger entry with PV in it instead. Mainly because the efficiency dont >really skyrocket when we get over the 24M-96M size. > >Of course im on work doing some really boring administrative tasks, so i may >just be daydreaming..... > >Any thoughts or comments ?? > >Regards >Kim hi kim, i always assumed that the optimal hashtable size depends linearly on the search time you're going to do. obviously, there is no sense in having a huge hashtable if you don't even fill it to close to 100%. your statement about 48 or 96 or whatever MB only makes sense in the context of bob's original post, which also states how long he's searching or how big his typical search trees are in this experiment. on saving a longer pv, i don't quite see how that will help - generally you use the hash move for move ordering, so what you will do with your longer pv? - pass it on as parameter to the child node and use it there for move ordering? sounds awfully complicated. the paper mentions a 3% gain for the 5-ply-pv, and a corresponding 1% gain for a doubled hashtable size; so i guess the potential of this technique is rather small for the pains you would go through with it.... cheers martin
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.