Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Beating MTD(n,f)

Author: Vincent Diepeveen

Date: 06:06:40 06/06/01

Go up one level in this thread


On June 06, 2001 at 06:25:40, Gian-Carlo Pascutto wrote:

>On June 06, 2001 at 05:05:28, Dan Andersson wrote:
>
>>>Behaviour of MTD is a horror in normal positions when
>>>score slowly drops. Also with pawn being 1000 units MTD is hard to use,
>>>as well as most commercials which give big scores for some patterns there
>>>MTD is not usable.
>>>
>>
>>This is simply not true.
>
>Actually I think Vincent is right here, although I would
>really like an objective measure, because this is just
>guesswork based on how we understand the search.
>
>The catch is that most strong MTD'f implmementations use convergence
>accelerators, and those will make the difference between centipawn
>and millipawn scoring void.

Noop it doesn't make it void.

To get from 100 to 200 is harder as it is to get from 10 to 20.

That's the diff between 0.100 and 0.200 in millipawns and
0.10 and 0.20 in centipawns.

If i use alfa=beta-1
and in the root try to get to a truebound, then the number of
extra nodes is huge. I did this at a time when pawn=200 for diep.

In diep last 2 years pawn=1000 everywhere now.

I get big differences in node counts using MTD,
let's formulate it this way. If MTD would work great,
why am i not using it?

Why is no other commercial program using it?

The only program using it is SOS, Postmodernist, cilkchess.

Those programs all share that they have
  a) very little chessknowledge
  b) pawn is worth not 1000 but more like 100 at most
  c) the knowledge the programs have is not scored
     as high as in other programs.

Best regards,
Vincent

>--
>GCP



This page took 0.32 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.