Author: Vincent Diepeveen
Date: 21:24:45 10/22/03
Go up one level in this thread
On October 22, 2003 at 21:34:31, Anthony Cozzie wrote:
>On October 22, 2003 at 21:00:52, Dave Gomboc wrote:
>
>>On October 22, 2003 at 14:40:23, Uri Blass wrote:
>>
>>>On October 22, 2003 at 13:13:37, Vincent Diepeveen wrote:
>>>
>>>>On October 22, 2003 at 06:30:13, Sergei S. Markoff wrote:
>>>>
>>>>>Hello All!
>>>>>
>>>>>DS - is a term for using some features of classical evaluation that consists of
>>>>>two parts - material and positional.
>>>>>There are a lot of positions in that for one side material evaluation is >0 but
>>>>>positional evaluation is <0 or vice versa. The root of big part of mistakes made
>>>>>by modern engines is underestimating of positional eval because the positional
>>>>>evaluation is constructed of several "atomic" factors. The _sum_ of this factors
>>>>>frequently isn't good positional evaluation (anyway there are a lot of
>>>>>"palliative" methods to avoid this problem like evaluation the relationship
>>>>>between several factors). We can't fully trust positional evaluation and that's
>>>>>why most of modern programs using a small values for a lot of factors.
>>>>
>>>>In 1990 your statement would have been true.
>>>>
>>>>However in 2003, i know very little modern programs with small values for the
>>>>positional factors. Perhaps diep is one of them in some sense, yet the quantity
>>>>makes the total positional score overrule any material reality.
>>>>
>>>>>The idea of DS is to use disagreement between positional and material
>>>>>evaluation. There are a lot of ways how to use it. For example we can check
>>>>>nodes in which sum_eval < alpha, but positional eval is large (for example we
>>>>>sacrificed a pawn for attack e.t.c.). For this nodes we can:
>>>>>1. Rebuild quiescence to include checks e.t.c.
>>>>>2. Extend search
>>>>>3. Change eval for the case of losing pawn or quality (trade bishop or knight
>>>>>for rook) for big passed pawn / king attack eval.
>>>>>4. Do assymetric eval.
>>>>>5. Something else?
>>>>>
>>>>>Do you have some ideas in this area?
>>>>
>>>>In case you forgot, the evaluation can just return 1 score and that's a total
>>>>score it can't return 2 scores for either positional or tactical matters.
>>>
>>>I do not think that sergei forgot something.
>>>He is a good programmer and smarthink is one of the best free engines.
>>>
>>>The fact that the evaluation can return only one score does not mean that the
>>>program cannot compute more than one score to get decisions because decisions
>>>are not only about evaluation but also about which lines to extend.
>>>
>>>Uri
>>
>>There's no law that says a score must be scalar.
>>
>>Dave
>
>
>I just had a vision of a program that returned tensors in its eval. Hmm.
>
>I'm with Sergei on this one though - Zappa already keeps track of the various
>"parts" of the eval, so I can tell if the position has a high kingsafety score
>for white or some such.
>
>anthony
Sure, but that's not the point.
A year or 5-7 ago i had the idea that i should seperate some scores. So i
returned 2 values to the search algorithm.
My score was roughly said (but more interesting labels):
struct Score { int value1; int value2; };
The point is HOW to use this for search.
I figured out after not a too long time that i had to use 1 score to combine the
2 scores to 1 score anyway or you can never give a cutoff.
So my viewpoint is that sooner or later you must trigger a decision and use the
alfabeta features of the search.
So why not do that directly and just return 1 score.
Extensions have nothing to do with this.
Best regards,
Vincent
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.