Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Piece Values in Chess Programs (Larry Kaufman)

Author: Vincent Diepeveen

Date: 05:35:07 05/06/01

Go up one level in this thread


On May 05, 2001 at 12:01:08, Dana Turnmire wrote:

>  Here is an interesting article found in a 1989 CCR article.

Exactly the biggest misunderstanding in chessbooks is that 2 rooks
are stronger as a queen. Even in most endgames a queen wins easily
against 2 rooks.

However grandmasters only get the exceptions on the board, so if a
GM gets 2 rooks versus a queen then *usually* the 2 rooks are stronger.

My piece values are about this:
  pawn    1000
  knight  3625
  bishop  3675
  rook    5800
  queen   11750

However there are always things wrong in piece values.

In the above values an obvious problem is that a queen is stronger
as 3 pieces which is NOT true.

Also 2 pieces in the above example are very close to rook+pawn.

However i have special code for that to fix it :)

>  "Most elementary chess textbooks assign relative values to the pieces, based
>on pawn=1, as follows: N=3, B=3, (or 3+), R=5, Q=9 (or 9.5 or 10).  Most chess
>computers use these numbers in their programs; in fact they play a critical
>role.  But there are serious problems arising from relying on these numbers."
>
>The most glaring problem is the exchange of two minor pieces for a rook and
>pawn.  Any tournament player should know that the minor pieces are nearly always
>superior, except in simple endgames.  In the middle game they are fully equal
>(or even superior) to a rook and two pawns, as I learned the hard way in the
>1972 U.S. Championship against fellow Senior Master Greg Defotis.  Yet many
>chess computers, especially Fidelity's will give up the two pieces for a rook
>and one pawn at the drop of a hat, and nearly always go on to lose.  I discussed
>this problem with Fidelity's programmers last year, but as the Excel68000 makes
>this losing exchange with alarming frequency it is clear that the problem has
>not been corrected.
>
>  Other piece value problems are improper bishop or knight exchanges (a
>Turbostar flaw), and unsound sacrifices of a knight for two pawns and meager
>positional compensation (Mephisto Amsterdam).  Fidelity machines tend to vive up
>their queens a bit cheaply, while Novag machines (especially the Super
>Constellation) sacrifice the exchange too readily.
>
>  It is my opinion that many of these problems relate to the fact that the
>accepted piece value tables were derived from endgame theory, and are not
>accurate for the middle game, in which pawns are more expendable and minor
>pieces more valuable.  To prove this I ran a series of blitz games, using the
>autoplay feature on the Mephisto Dallas, in which I removed a white knight and
>three black pawns (not rook pawns), varying the choice of knight and pawns,
>before starting the games.  Black won 8-0!  I raised the compensation to four
>pawns, and black still won 3-1 (at 5 pawns they split two games).  So it seems
>clear that at least in the early stages a piece is worth at least four pawns,
>unless king safety or center domination is involved.
>
>  Apparently, Mephisto reached the same conclusion.  In a major departure from
>the Amsterdam program as well as all others, piece values were changed for the
>Mephisto Dallas program to knight=4, bishop=4+, rook=6, and queen=11.  Suddenly,
>two minor pieces equal rook and two pawns (as they should), the exchange retains
>its standard two pawn value, and the unsound piece sacrifices of the Amsterdam
>dissapear.  The major drawback is a tendency to underestimate sacrifices of a
>minor piece for two king-protecting pawns, but this may be dealt with in the
>future by heuristics.  The endgame is not harmed because passed pawns receive
>sufficient bonuses in that phase to reduce the effective value of minor pieces
>to around three pawns.  My only criticism is that the queen should probably rate
>a tad higher or the rook a hair lower, since queen and pawn are usually superior
>to two rooks.  I predict that the Mephisto Dallas piece values will soon be
>copied by other programmers, and I recommend them to those human players who
>feel the need for numbers.



This page took 0.01 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.