Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Couple of chess programming questions

Author: José Carlos

Date: 14:46:52 09/10/02

Go up one level in this thread


On September 10, 2002 at 15:19:10, Uri Blass wrote:

>On September 10, 2002 at 14:21:33, Omid David wrote:
>
>>On September 10, 2002 at 14:09:47, Dann Corbit wrote:
>>
>>>On September 10, 2002 at 13:59:26, Omid David wrote:
>>>[snip]
>>>>Has anyone compared NegaScout vs. MTD(f) on a set of tactical test suites? It
>>>>seems to be interesting to test the behavior of MTD(f) in real complicated
>>>>positions.
>>>
>>>I think that the quiet positions will be a lot more interesting.  That is where
>>>the games are really won or lost anyway.
>>>
>>>Most chess moves in a game are quiet moves.  How often do you see the eval shoot
>>>up a full pawn?  Before the game is clearly won, maybe once or maybe not at all.
>>> The major meat of the game is manuvering into a position where you can perform
>>>that tactical shot.
>>
>>Finding the good quiet move has to do largly with the evaluation function. In
>>such positions, searching one ply deeper is usually less important than having a
>>better eval function, and thus, a better understanding of the position. So,
>>while quiet positions are good means for general testing of programs, they
>>aren't of any great use for pure comparison of different algorithms against each
>>other.
>
>
>I do not agree.
>
>I think that finding good quiet moves is often a question of searching deeper.
>programs can translate by search positional advantage that they do not
>understand to positional advantage that they understand.
>
>They may not see winning material but only winning a small positional advantage
>thanks to search.
>
>>
>>But in tactical positions, the determining factor is the search depth, so the
>>algorithm which results in a smaller search tree will fair better. As Vincent
>>pointed out, there is some evidence that MTD(f) behaves worse than NegaScout in
>>complicated positions.
>
>The program does not have to use the same algorithm in all the game.
>It can decide based on the changes in the evaluation if the position is
>quiet or not quiet and decide to use MTD only in part of the cases.

  I thought about that a while back and got the conclusion that it's not a good
idea:
  - you need different hash table structure for PVS than for MTD(f).
  - both searches will give different results so, in the moment you switch from
PVS to MTD(f) or viceversa you need to clear the hash tables and you might find
the new algorithm could not understand what the old one "was trying to do",
which would cause a bad fail low. Or the other way, a fail high on a bad move
the other algorithm had discovered to be bad.
  So the search would be very unstable.

  José C.


>Tactical positions are important but quiet positions are also important.
>
>I do not say that MTD is a good idea but my reason to dislike it is different
>
>I do not like losing information and I understood from posts that at least
>Fritz7 lost the correct pv thanks to mtd.
>
>Uri



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.