Author: Dave Gomboc
Date: 11:48:55 08/08/00
Go up one level in this thread
On August 08, 2000 at 08:21:26, Vincent Diepeveen wrote: >On August 07, 2000 at 20:04:18, Dave Gomboc wrote: > >>On August 06, 2000 at 20:26:06, Vincent Diepeveen wrote: >> >>>Actually MTD is great for testsets. MTD is having a huge problem if >>>you start playing games with it however and the vice versa happens. >> >>Completely unjustified, as always. >> >>>If i'm at 0.600 now with PVS at iteration is 8, and the chess prog >>>starts smelling trouble, then suppose we fail low to 0.300, now >>>aren't that 300 researches with MTD? >>> >>> THREEHUNDERD RESEARCHES? >> >>Vincent, if it takes you 300 researches to fail low by 0.3 pawns with mtd(), >>your code is seriously broken. >> >>Dave > >Using MTD it's easy to research using only 1 point: > bound,bound+1 > >if bound a bit lower then > newbound = bound-1; > >then search at newbound,newbound+1 > >Now that's something you can do quick. > >As soon as you go jump in scores in a binary way > > newbound = bound-2; > then search at bound-3,bound-2 > > no if fails again low then > bound-5,bound-4 > >Then you lose a big advantage of MTD, namely that >you search the same space with about the same score. > >Now you suddenly need to see entire new trees, and that >all because you don't use but prefer to get a lot of >overhead to get there. Better works: > -infinite > >So any number of researches is a bigger overhead as using PVS, >and that in positions which are CRUCIAL to your game! I repeat: >>Vincent, if it takes you 300 researches to fail low by 0.3 pawns with mtd(), >>your code is seriously broken. 1. A program which exhibited the above behaviour contains evaluation terms that specify accuracy without corresponding precision being available. 2. A linear adjustment of the score window after multiple researches is a beginner's mistake. Dave
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.