Author: Bob Durrett
Date: 09:45:19 01/10/04
Go up one level in this thread
On January 10, 2004 at 11:59:35, Ed Trice wrote:
>Hello Reinhard and Michel,
>
>Just so you know, the values on the website are just approximate values that I
>offered. Gothic Vortex does something very complicated in the form of a
>statistical "strategic" adjustment to the values as a function of a very
>granular evaluation of the game stage.
>
>Here is a martix you should compose and test your weights:
>
>Knight + Rook vs. Chancellor
>Knight + Bishop vs. Chancellor
>
>Chancellor + Pawn vs. Queen
>Archibishop + Pawn vs. Chancellor
>
>Queen vs. Archibishop + Knight
>Queen vs. Archbishop + Bishop
>
>Chancellor + Bishop vs. Archbishop + Rook
>
>Rook vs. Knight + Knight
>
>Rook + Pawn vs. Archbishop
>
>I have many more of these in the evaluation function that STATISTICALLY
>estimates which group should be preferred at many different game stages.
>
>IF, and that is a BIG IF, the weights are tuned well, Vortex has 'strategic
>vision' equivalent to extending its search horizon.
>
>If the weights are off, I would expect catastrophic results, such as wildly
>fluctuating scores, or many researches, etc.
>
>>On January 10, 2004 at 06:17:35, Michel Langeveld wrote:
>>
>>>This is the current piece value array of TSCP:
>>>
>>>/* the values of the pieces */
>>>int piece_value[BOARD_MAX_PIECE_TYPES] = {
>>> 100, //pawn
>>> 250, //knight
>>> 300, //bishop
>>> 475, //rook
>>> 650, //archbishop
>>> 825, //chancellor
>>> 875, //queen
>>> 000 //king
>>>};
>>
>>Well, I think Ed Trice also like to make some runs with those values.
>>
>>>What do I have to file in for your values?
>>
>>Pawn = 100
>>Knight = 306
>>Bishop = 360
>>Rook = 543
>>Archbishop = 665
>>Chancellor = 849
>>Queen = 903
>>-----------------
>>King = 372 (usage depending on your program)
>>
>>>Then I will play two games.
>>>One with white against the old version and one with black against the old
>>>version and tell you what versions what version won the most :-)
>>
>>Though I think two games would not be that representative I will thank you very
>>much for such experiments, because I myself are still far from being able to
>>test this.
>>
>>Regards, Reinhard.
The technical issue of how piece values can be properly estimated is very
interesting to me.
In "ordinary" chess, the amount of human experience is measured in the millions
of games and so there is plenty of data available to estimate piece values for
human vs human games.
For a new variant of chess where a new piece is to be used, there will not
initially be the extremely large database from which to draw piece valuation
estimates and such large databases may be a long time in coming.
This begs the following question: "What would be a practical way to develop
information which could be used to get better piece valuations?
Having a large amount of data provides two benefits: First, it makes
statistical evaluation feasible. Secondly, it provides many examples which could
be studied individually to improve our understanding of this topic.
Engine versus engine experiments may be a practical solution. The time limits
might be blitz or faster and still give useful data. [Slow time limits provide
smaller databases in a given amount of time but may give better data.]
The difficulty might be in deciding how to analyze the data to glean the desired
"piece valuations." Generally, piece valuations depend on a number of things
such as whether in opening, middlegame, endgame among many other things.
Incidentally, my guess is that the overarching strategic concepts of "ordinary
chess" would still apply to chess variants as long as the variant is reasonably
close to the original. What "reasonably" might be is unclear.
Bob D.
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.