Computer Chess Club Archives




Subject: Re: Beating MTD(n,f)

Author: Robert Hyatt

Date: 14:16:21 06/07/01

Go up one level in this thread

On June 07, 2001 at 06:52:07, Tony Werten wrote:

>On June 06, 2001 at 10:32:14, Gian-Carlo Pascutto wrote:
>>On June 06, 2001 at 09:06:40, Vincent Diepeveen wrote:
>>>Noop it doesn't make it void.
>>>To get from 100 to 200 is harder as it is to get from 10 to 20.
>>If they mean the same thing, not at all.
>>Note that I said you should be taking convergence acceletrators
>>into account. Every competetive implementation of MTD(n,f) has
>>them and the papers suggest them so that is a valid requirement.
>>If you increase the window forecefull from 100 to 200 when you
>>are using millipawns it's the exact same thing as increasing it
>>from 10 to 20 when you are using centipawns.
>>If you rely purely on the fail-soft and do not adjust the bounds
>>with the accelerators, I think you are right, but I really would
>>like to see evidence of it. When forcing the bounds it's just the
>>same thing however, and that is what every good MTD(n,f) implementation
>>is doing.
>>You are basically saying that the difference between 0.1 and 0.2
>>is different from that between 0.1 and 0.2. No it's not :)
>I think what Vincent is saying is that there's a difference between 0.1 and 0.2
>and 0.10 and 0.20 The first 2 are following each other while the second 2 have 9
>numbers between them.
>Suppose the current score is 0.10 Now I find a move with score 0.12 MTD will
>fail and I have to research. If I had scored 0.1 then the new move would have
>also been 0.1 and MTD wouldn't have to research.
>Of course is this not only true for MTD, but for all minimum window searches,
>but in MTD the first (best) move is also done with mws and this is most of the
>time the one that has the fluctuating scores.

First, I came to the conclusion years ago that "millipawns" offers resolution
that is unusable.  Even units of 1/100ths of a pawn are very tiny.

Second, there are solutions to the problems he mentioned.  A well-behaved search
can do ok with mtd...  I just personally happen to not like the results I got
when I tested it a couple of years ago when it became "the rage".  It is OK.
That is all I will say.  I am not convinced that it is significantly better (or
worse) than PVS.  The extra iterations can be costly.  There are lots of other
isses that require rewriting major pieces of code for what might be no real
benefit.  I did try it.  I don't use it.  Others should _always_ try something
before (a) giving up on it or (b) commenting on it.  It is just like bitmaps.
YMMV depending on how much time you put into learning to use/understand them


This page took 0.03 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.