Author: Christophe Theron
Date: 11:15:08 08/25/00
Go up one level in this thread
On August 25, 2000 at 11:58:10, Ernst A. Heinz wrote:
>Hi Christophe,
>
>>On August 25, 2000 at 11:01:45, Ernst A. Heinz wrote:
>>
>>>Hi Christophe,
>>>
>>>>Yes, but have you seen my own reply to your answer?
>>>>
>>>>It's not to be picky actually. I think there is a serious problem with the trick
>>>>when you have a QSearch, and I wanted to have your opinion.
>>>
>>>Yes, still as I said before -- personally, I do not think that
>>>the quiescence search is the _real_ problem. A variable-depth
>>>capture-only quiescence without any kind of alpha-beta related
>>>stuff (e.g. futility pruning and lazy evaluation) would be a
>>>perfectly fine black-box routine for scoring horizon nodes that
>>>should work together with the negascout "trick" _without_ losing
>>>theoretical correctness. The "trick" only fails with variable
>>>depths in the _main_ search layer because that is where the
>>>true fixed-depth 2-ply research always yields the same result as
>>>the original zero-window search.
>>>
>>>Have you read Dave Gomboc's answer where he tells about Jonathan
>>>Schaeffer's experience with the negascout "trick". As soon as he
>>>added extensions in the _main_ search, he ran into problems. If,
>>>as I assume, he already had a quiescence search before, this
>>>would provide empirical evidence and support for my argument.
>>
>>I do not disagree with you about extensions, but I disagree about SQearch, even
>>it is very simple. Have you read my other post in which I explain what the
>>problem is, in my opinion?
>
>Yes, I have read it -- that is why I hypothetically assume
>a variable-depth quiescence search above which is _independent_
>of alpha and beta.
Are you talking about a QSearch that would not use alphabeta cutoffs?
In this case you are right, but it's a very hypothetical case, as using such a
QSearch does not make sense.
For a practical chess program, the important thing is that:
1) A QSearch kills the NS trick
2) Extensions kill the NS trick
Disabling extensions in the last plies does not work if the program uses a
QSearch based on an alphabeta search.
Using a static evaluation function taking captures into account does not work if
the program uses extensions in the last plies.
> Thereby I meant to show that variable-depth
>search in the scoring of horizon nodes does not _necessarily_
>kill the negascout "trick", i.e., it is not any variable-depth
>search below horizon nodes which kills the "trick" but
>variable depths in the main search.
>
>What else have I overlooked in your other post?
Nothing else I think.
Christophe
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.