Computer Chess Club Archives




Subject: Re: Question for the MTD(f) experts

Author: Robert Hyatt

Date: 10:15:29 04/15/04

Go up one level in this thread

On April 15, 2004 at 05:11:29, Fabien Letouzey wrote:

>On April 14, 2004 at 13:33:38, Robert Hyatt wrote:
>>On April 14, 2004 at 03:30:22, Dann Corbit wrote:
>>>I decided to toss an MTD(f) search into TSCP, and I've got something wrong, but
>>>I can't quite see what it is.
>>There is a lot more to do.
>>1.  you need to modify the hash table to store 2 bounds, not 1.
>>2.  the search must be fail-soft.  TSCP isn't.
>>3.  PV has to be yanked from the hash table and that makes it flakey at times as
>>has been discussed many times.  There is another way to get the PV, but it is a
>>special case solution only for mtd...
>>4.  the convergence has to be accelerated.  IE on a fail high searching v and
>>v+1 won't cut it.
>>IE this is not a "drop in and go" it is a "rewrite and test" thing to do. :)
>Could you give me a pointer on the special way to get the PV (not with hash
>tables)?  I'm interested in the preconditions.
>I think 5) rounding the eval to the nearest multiple of whatever is also common.
>Maybe it's a very small gain but I think most MTD(f) engines do this, unless
>their eval has large grain to start with.

Not my idea, but it goes like this.

On the next-to-last search, you will fail low or high.  You adjust the bound and
search again and on the last search you fail in the opposite direction.  You cam
catch this condition and use it to recognize the path that leads to the same
position for both searches.

I don't remember who came up with this, memory is suggesting Brian Richardson,
but I'll let him speak for himself if I an correct...  He and I were playing
with mtd(f) at the same time, he kept it, I didn't.  :)  Which certainly means
he knows more about it than I do...

This page took 0.05 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.