Author: Christophe Theron
Date: 10:11:56 10/07/01
Go up one level in this thread
On October 07, 2001 at 02:36:52, Uri Blass wrote:
>On October 06, 2001 at 21:23:38, Christophe Theron wrote:
>
>>I was thinking about some linear equivalence between depth and "knowledge"
>>(evaluation), very much like e=mc^2.
>>
>>But this is too far stretched at this time.
>>
>>One basic fact that supports my point is, and I think Bob described the
>>phenomenon himself some time ago, that improving the knowledge (evaluation) of a
>>program is especially needed when the program cannot reach high depths.
>>
>>He talked about this about the older version of Cray Blitz of whatever was
>>before Cray Blitz.
>>
>>I have noticed the same thing, and this is the basis of my belief that there are
>>dimishing returns from improved knowledge.
>>
>>This goes against the already old urban legend saying that "more knowledged
>>programs" (programs with slower evaluations) will be superior on faster
>>computers (which is clearly denied year after year).
>>
>>
>>
>> Christophe
>
>I think that it depends on the knowledge and there is knowledge that cannot be
>replaced practically by search depth because you need to search more than 30
>plies forward.
>
>Suppose that a program that has no knowledge about king safety plays against
>program with knowledge about kingsafety.
>
>At small depth it is not going to be important because tactical mistakes are
>going to dominate and searching 1 ply deeper than the opponent may be more
>important.
>
>At big depthes it is going to be more important because the program with more
>knowledge is going to have enough depth to play for a sound king attack without
>tactical mistakes and 1 ply is not going to help the program without knowledge
>about king safety because there are moves that in order to see that they are bad
>(not for king safety reasons) you need to search at least 30 plies forward.
>
>Uri
...And on the other hand, if you can search 30 plies forward, your knowledge
about king safety is probably going to be useless.
Because it does not matter if you know how to build a king attack or not, your
opponent can see deep enough anyway to see the consequences just by calculating.
I don't think it is possible to "solve" this problem by thinking from a theoric
point of view, because you can find extreme examples to support both sides.
Chess is a statistical game (in the real world), so what matters is how good or
bad knowledge performs when you play real games.
From experience it seems to me that knowledge (in evaluation) is more important
on slow hardware (when you cannot search deep). But I'm open to the idea that
some sort of knowledge (in evaluation) performs better for deeper searches...
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.