Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: MTD(f)

Author: Ricardo Gibert

Date: 16:38:53 01/07/01

Go up one level in this thread


On January 07, 2001 at 18:30:04, James Swafford wrote:

>>
>>The catch here is that it keeps failing in the same direction most of the
>>time, so one bound will often be enough.
>
>Why does it keep failing in the same direction most of the time?
>
>My understanding is that you're basically playing a game of "hi-lo."
>In other words, take a guess.  If it's too high, take a lower guess.
>Too low, make a higher guess.  Eventually you'll zero in on the
>correct score.  If it kept failing in the same direction, you'd
>continue to make lower and lower (or higher and higher) guesses
>until you hit the correct score.  Doesn't sound right, but maybe
>I'm misunderstanding you...

MTD(f) is not based on a type of bisection search, but is rather a type of
sequential search. What you suggest has been tried and been found less
effective. That may seem surprising, but the justification is the initial
estimate of the root positions value is assumed to be already very close. With
optimal move ordering, you only do one research and trap the correct value right
away between 2 bounds for an exact result. With bisection, you must do a lot of
researches before the correct value is found.

>
>
>>If you use convergence-accelerating heuristics, dual bounds will probably
>>be better 'cos you will be failing in both directions then.
>
>Maybe that's what I'm thinking about. (?)
>
>--
>James
>
>
>>
>>--
>>GCP



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.