This is a temporary, read-only recovery of the chessprogrammingwiki while a longer-term plan is worked out. Editing is not possible right now, but will be again soon.
Chess Programming Wiki All pages Other namespaces

Evaluation Overlap

Home * Organizations * ICGA * Investigations * Evaluation Overlap

by Mark Watkins

Here might be a historical perspective. From page 62 of Hayes and Levy, World Computer Chess, Stockholm 1974 1.

Contents
  1. Numerical Evaluation of Positions
    1. CHAOS
    2. The Ostrich
    3. Chess 4.0
    4. Kaissa
  2. Analyse
    1. Center Control
    2. Levy Development
    3. Queen Bonus
    4. Rook Bonus
    5. Bishop Bonus
    6. Knight Mobility
    7. King's mobility
    8. Attack/Defence
  3. Final Definition
  4. See also
  5. References

Numerical Evaluation of Positions

In the section devoted to notes on the competing programs, the respective evaluation functions for CHAOS and The Ostrich are discussed in some detail. CHAOS uses nineteen features and The Ostrich only thirteen, but the latter plans to increase this number in later incarnations. They have approximately eleven features in common: 'approximately', because the concepts are divided differently. The overlap here is not surprising since all scoring functions aim to embody 'reliable' chess heuristics: the weightings may differ, but the features themselves are similar. One of the eleven concepts shared by CHAOS and The Ostrich is material: this is the dominant factor no only in these two programs but also in the majority of chess playing programs. Exceptions are Freedom, which stresses mobility, and Papa, which concentrates on 'entropy' (basically a mobility measure). Other shared concepts include mobility, control of the centre, castling, king safety and assorted terms concerned with pawn structure: the latter include bonus points for occupancy of the centre, advancement, passed pawns, a bonus for doubling opponents' pawns with a penalty for doubling one's own, and a penalty for blocking development of one's own pieces. The delicate matter of handling pawn structures has not been mastered satisfactorily by any current program [...] For instance, all seem to place disproportionate faith in the concept of doubling the opponent's pawns (see records of games 1 and 4). CHAOS also considers the number of threatened pieces, pins and discovered checks, king end-game position, and capturing and mobility potential.

The Ostrich has an interesting term, concerned with tempi, which aims to penalize time-wasting moves such as taking two moves to reach a square that could be reached in one, or repeating a move (see game 24, where the Swiss program Tell causes a draw by repetition). [...]

CHAOS

4. Evaluation function. [...] Nineteen factors are taken into account, and their weights modified according to the stage reached in the game.

The evaluation function has been based as far as possible on general principles, avoiding special cases. For testing purposes the weights of these factors were set differently for each side while the program played against itself. This yielded valuable information about best settings.

The Ostrich

The static evaluation function. This consists of thirteen subroutines each corresponding to a basic chess heuristic:

Chess 4.0

[...] In addition to material score in Chess, terms are added which express in a primitive way, notions of mobility (number of squares attacked), pawn structure (passed, isolated, doubled, backward, etc.), piece placement (e.g. rooks on seventh rank), and king safety (king in castled position, adequate pawn cover). [...]

Kaissa

One of the papers which is available (Adelson-Velskii et al. 1970 2) describes the program as it was in the late 1960s and from the evidence available it is likely that Kaissa has many of the same features. [...] This paper has notes on the evaluation function; the present function is likely to be similar, if not identical. Bonus points are given for:

Penalties are incurred by

Analyse

Another thing to discuss could be Chapter 6 of Hartmann's Notions of Evaluation Functions tested against Grandmaster Games in Advances in Computer Chess 5 (1989) 3.

Here is a brief synopsis. First there are three notions of mobility: Legal Moves, Pseudo Moves, and de Groot Moves. These should be self-explanatory.

Center Control

With B as a bonus for a given square, this is B*[AT+2*OC] where AT is #attackers and OC is #occupants.

Levy Development

This is a rather complicated construct, that doesn't seem of much direct relevant here.

Queen Bonus

CHESS 4.5 uses 0.8*AT-0.8*DI where AT is #squares attacked that are not attacked by enemy pawns or minor, and DI is the minimum of the rank/file differences to the enemy king. ANALYSE rescaled this to AT-DI (so as to make everything integral).

Rook Bonus

Chess 4.5 used 1.6*AT-1.6*DI+8*DR+8*op+3*SO+22*SR where AT is #squares attacked, DI as above, DR is #rooks in file or rank w/o intervening pieces, SR is #rooks on 7th rank, OP is #rooks on open files, SO is #rooks on semi-open file (defined as File in which there is no [own] Pawn, and at least one enemy Pawn that is no defended by a Pawn and that cannot move one single square w/o being attacked by a Pawn). ANALYSE rescaled to make it integral.

Bishop Bonus

Jaap Herman helped with this, as PS+EM+OM+DL+OB+OW, where PS #pseudomoves, EM is sum of enemy material on diagonals of the Bishop (unclear if it includes x-rays, even for blocked pawns?) where P=0,NB=3,R=5,Q=9,K=10, OM is sum of own material forwards on diagonal of the Bishop, P=0,NB=3,R=5,Q=9,K=0, DL is length of diagonals [#pseudomoves on empty board] minus 7, OB is pawn obstruction of enemy pawns (-25,-15,-10,-5,+1 according to specific characteristics), OW is own pawns, every pawn on diagonal in front of bishop is -5 if rank < 4 and +1 otherwise, and every pawn on diagonal behind the bishop is +1.

Knight Mobility

Hartmann discusses having this, and then points out that it doesn't seem to contribute one way or the other, so he concludes "It is therefore no use to have a separate mobility component for the knight, detached from the Pseudo Mobility."

King's mobility

Again Hartmann points out that this might be useless, but it seems OK in the endgame.

Attack/Defence

Final Definition

MAT+PS+DIF+0.5*CEN+LD+2*RB+QB+2*BB+0.5*DEF+OFF

MAT is material, PS is #pseudo-legal moves, DIF is #different attacked squares, CEN is centre control, LD is Levy development, RB is rook bonus, QB is queen bonus, BB is bishop bonus, DEF is #attacked own squares, OFF is #attacked opp squares.

See also

References

Up one level


  1. Jean E. Hayes, David Levy (1976). The world computer chess championship, Stockholm 1974. University Press (Edinburgh) ISBN 0852242859↩︎

  2. Georgy Adelson-Velsky, Vladimir Arlazarov, Alexander Bitman, Alexander Zhivotovsky and Anatoly Uskov (1970). Programming a Computer to Play Chess. Russian Mathematical Surveys, Vol. 25, pp. 221-262↩︎

  3. Dap Hartmann (1989). Notions of Evaluation Functions Tested against Grandmaster Games. Advances in Computer Chess 5↩︎

What links here

Contributors: GerdIsenberg.