Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: MTD(f)

Author: Tony Werten

Date: 02:21:17 07/28/04

Go up one level in this thread


On July 28, 2004 at 01:11:44, Stuart Cracraft wrote:

>>
>>Play with the granularity of the search.  If a pawn is worth 64 centipawns you
>>will do far less searches than 1000 centipawns.  But then you lose resolution.
>>So the trick is to find a happy medium, I would guess.
>
>Done. I made a pawn 64 centipawns instead of 1000 millipawns.
>and got a tree reduction for the 30-position test of about 5.589%
>so it's on the outer fringes of what I've read in Schaffer's
>writings, there being 5-15% reduction. I wonder what the other
>tricks are. I cringe at going to fail-soft and having to give up
>PVS/NEGASCOUT.

There are still some improvements

- Fail soft. It doesn't have to break your pvs-negascout since it works in there
too ( but gives less ) It is aimed at reducing the number of bestmove searches.

- ETC. I never seen anything coming from it in PVS, but heard it works well in
MTD.

- Don't resolve bestmove. If after 11 searches, you still have a faillow, you
know the bestmove score dropped with >= 2 pawns. Rather than resolving the exact
score, you can take the score, set a BESTMOVE_STILL_SUCKS flag and continue
searching. The idea is that you will find a better move anyway. If not you can
still resolve the bestmove score ( or just go 1 ply deeper anyway) but you're in
trouble then anyway.

-  Search deeper :) Tree size gains aren't static. What gives a 5% at fe 10 ply,
will most likely give a 7% on 11 ply and a 10% on 12 ply. Unfortunately not so
easy to do.

Tony


>
>Googling around for MTD(f) is remarkably sparse as far as amount of practical
>examples are concerned.
>
>Here are the results.
>
>MTDF 64 centipawns to a pawn
>**** 90% 27/30 25.81 6492995 216433/1/251608 0/0/303344/0/0/0
>
>MTDF 1000 millipawns to a pawn
>**** 90% 27/30 25.69 6640618 221354/1/258440 0/0/307973/0/0/0
>
>PVSNEGASCOUT 1000 millipawns to a pawn
>**** 90% 27/30 25.56 6877388 229246/1/269110 0/0/316245/0/0/0
>
>Tree size reduced from 6,877,388 nodes to 6,492,995 nodes.
>
>Stuart



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.