Author: Rudolf Huber
Date: 09:47:34 05/31/01
Go up one level in this thread
On May 31, 2001 at 12:26:19, Vincent Diepeveen wrote: >On May 31, 2001 at 08:11:16, Dan Andersson wrote: > >>You are comaring apples with oranges, re-search in MTD is never full width. Thus >>even a thusand re-searches (clearly an extreme case) may be less expensive than >>one full width re-search (not bloody likely, but still possible). You never seem >>to get that point. > >Please annotate in each research how many NEW nodes you searched >and how many of those you had to evaluate. That would be cool >info. > >It's too expensive for my poor evaluation. Even though i store >evaluation both in transposition table AND special eval table, >note i do 8 probes which is pretty something compared to most >which do 2 probes or so. > >But if i do researches i search loads of new nodes and loads of new >evaluations are needed. > >Please post here how many NEW nodes you search! > >If you measure that you will get scared to death! > >Also search fluctuations of 0.3 pawn are very normal first >few plies in diep. Of course above 10 ply fluctuations get less, >but i do not even get to 10 ply with MTD... > >If i have bound 0.100 now which fails high >and suppose i need to go to 0.400 >(but of course i do not know that i get to 0.400). > >If i want to profit optimal from hashtable with MTD then my >next bound should be 0.101 right? > >So you advice me to do 300 researches? > >That's *pretty* expensive. No. Lets say you search 20 ply deep. With 300 researches times 20 ply you visit 6000 nodes. Is that much? Is is really not an issue. I also can not follow you millipawn argument. I myself use centipawns but I tried also milli and dezipawns. There is a only small difference in efficency (5-10 percent). What really helps however is full pawn evaluation. With this I search the starting position to 30 ply full width over night. > >>Regards Dan Andersson > >Vincent Rudolf
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.