Previous: Checkers
Up: Preliminary Results
Next: New Games
Previous Page: Checkers
Next Page: New Games

Chess

In chess, we are playing METAGAMER against GnuChess, a very strong publicly available chess program, winner of the C Language division of the 1992 Uniform Platform Chess Competition. GnuChess is vastly superior to METAGAMER (at chess, of course!), unless it is handicapped severely in time and moderately in material.

The preliminary results from the experiments are that METAGAMER is around even to GnuChess on its easiest level, when given a handicap of one knight. For purposes of comparison, a version of METAGAMER with only a standard hand-encoded material evaluation function (queen=9, rook=5, bishop=3.25, knight=3, and pawn=1) [\protect\citenameAbramson, 1990][\protect\citenameBotvinnik, 1970] played against METAGAMER with all its advisors and against the version of GnuChess used above. The result was that the material program lost every game at knight's handicap against GnuChess, and lost every game at even material against METAGAMER with all its advisors. This showed that METAGAMER 's performance was not due to its search abilities, but rather to the knowledge in its evaluation function.

On observation of METAGAMER's play of chess, we have seen the program develop its pieces quickly, place them on active central squares, put pressure on enemy pieces, make favourable exchanges while avoiding bad ones, and restrict the freedom of its opponent. In all, it is clear that METAGAMER's knowledge gives it a reasonable positional sense and enables it to achieve longer-term strategic goals while searching only one or two-ply deep. This is actually quite impressive, given that none of the knowledge encoded in METAGAMER's advisors or static analyser makes reference to any properties specific to the game of chess- METAGAMER worked out its own set of material values for each of the pieces (see Section 3), and its own concept of the value of each piece on each square. On the other hand, the most obvious immediate limitation of METAGAMER revealed in these games is a weakness in tactics caused in part by an inability to search more deeply within the time constraints, in part by a lack of quiescence search, and also by the reliance on full-width tree-search. These are all important areas for future work.