Robotics Reference
In-Depth Information
from the outset. De Grey's concept is an example of Artificial Intelligence
par excellence
—programs that determine by themselves, with no outside
help, which aspects of a game should be codified into the features of an
evaluation function. When applied to Chess such programs recognize
the importance of material as an evaluation feature, counting the num-
ber of Chess pieces of each type. When applied to the game of Reversi,
de Grey's program produced features which measure different aspects of
positions that are correlated with mobility, recognized by strong players
as being one of the most important features of the game. These methods
do not generate the weights for the features—instead they serve as the
input to systems that learn their weights either from the experience of
playing games or by observing experts playing the game.
Another of the early systems for automatically devising evaluation
functions was Tom Fawcett's ZENITH, developed as part of his PhD
research at the University of Massachusetts at Amherst. When tested on
the game of Reversi, ZENITH was able to create most of the features that
are generally well-accepted by human expert players. It also designed a
useful feature that no-one had previously thought of !
Apparently inspired by de Grey's work, another Cambridge student,
Barney Pell, developed de Grey's remarkable concept into the somewhat
different idea of the “Metagame”, on which he wrote
The idea is to write programs which take as input the rules of a
set of
new
games within a pre-specified class, generated by a pro-
gram which is publicly available. The programs compete against
each other in many matches on each new game, and they can then
be evaluated based on their overall performance and improvement
through experience. [3]
The principal difference between de Grey's original concept and Pell's
work is that, while de Grey's program would automatically devise its
own features for its evaluation function, and then learn the appropri-
ate weightings for those features based on its experience when playing
games, Pell's approach was to endow his program with “advisors”,
35
each
of which encapsulated “a piece of advice about why some aspect of a po-
sition may be favourable or unfavourable to one of the players” Then,
based on a collection of such advisors, Pell's program would analyze the
35
For example, one advisor was called dynamic-mobility—the number of squares to which a
piece can move directly from its current square on the current board position. Another advisor,
static-mobility, counted the number of squares to which a piece could move on an otherwise empty
board.
