Author: Bas Hamstra
Date: 06:59:44 09/28/99
Go up one level in this thread
On September 28, 1999 at 09:25:58, Robert Hyatt wrote: >On September 28, 1999 at 08:13:47, Bas Hamstra wrote: > >>Hello, >> >>A while ago it was posted here that movesorting near the root was much more >>important than near the leafs. >> >>- Is that true? >>- If so, how much more important? > > >Think about this for a minute. If you get the wrong move first near the >leaves, >how much work does it take to search the wrong move and then get the right one >to get that cutoff? Compare this to positions near the root. So yes, it is >_exponentially_ more important to get the move ordering right nearer to the > root than nearer to the leaves. >>For instance Crafty has a FH FirstMove/FH rate of 94%. Now most nodes are near >>the leafs, on average 94%, so near the root it is probably a good bit better. >>Say 98%. Many moves come from hash, there. >> >>Now what if it went from 98% to 99%? What would be the impact on total nodes? >>Any ideas? > >again, that 1% is an exponential improvement in the search, not linear... >That last (missing) 6% would make the tree 2x smaller if I could get it... > >> >>If sorting near the root is more important, isn't it an idea to try to improve >>sorting for the first couple of plies (say 3) by Enhanced Transp Cutoffs? >> >>Some reported rootsorting by failsoftvalue gave good results. Why not the first >>3 plies try something similar? >> > >this doesn't work. for positions where you have to search _all_ moves, >ordering >is totally unimportant. And even with failsoft, you do _not_ discover the best >move. Alpha/beta simply won't do it. ETC is certainly worth a try. When I >tried it it made the tree smaller, but it was no faster due to the extra >overhead it entails. Trying it near the root is something I didn't try, so you >ought to give that a test.. > > >> >>Regards, >>Bas Hamstra.
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.