Robotics Reference
In-Depth Information
Both methods of achieving this, knowledge engineering and supervised
training, have been found to be extremely difficult challenges that are
loaded with pitfalls. Although there has been some success in using these
techniques, there has always been a huge gap between the positional judg-
ment of the strongest human Go players and the judgement encapsulated
in the heuristic evaluation functions in Go programs.
Playing Metagames—Programs that Learn to Play from
the Rules
In 1985 Aubrey de Grey investigated a new concept in games program-
ming as his BSc project in Computer Science at Cambridge University.
De Grey's idea was to write a program that would learn to play a game
despite being given no more than the rules.
De Grey split the task of automating the whole process into the au-
tomatic generation of the features for an evaluation function and the
automatic discovery of the best weightings for those features. He de-
vised a method for generating the features, and this was implemented
and refined, to the stage where it could produce a suitable feature list for
a few games. 34 De Grey also devised an original method for optimiz-
ing the weightings in the evaluation function, based on processing the
game trees to a depth of only two-ply. His method consisted of refining
or rejecting various constraints on the values of the weightings, so that
the constraints became progressively tighter, resulting in each weighting
having a smaller range and therefore becoming more accurate.
Unfortunately for the world of games programming, Cambridge un-
dergraduate project reports are not normally published, so de Grey's
work has not accrued the credit it deserves. But within a few years the
same idea was being investigated by other researchers, who developed
programs that could generate evaluation features automatically from the
rules of the games rather than have the features available to the program
player or a to computer programmer that if you crowd your stones together (bad shape) you lose out
by having them control less territory, while if you spread them wide apart (also bad shape) you make
it easier for your opponent to adopt a divide-and-conquer strategy. So good shape is somewhere in
the middle ground between the two. But this is totally inadequate as a definition or specification.
Francis Roads explains the problem of defining good shape thus: “A game of go is a living organism.
You can never fully understand one part or aspect of it without taking the whole into account—and
the whole is usually beyond human understanding.” [2]
34 The games de Grey chose were Checkers (Draughts), Reversi (Othello TM ), Qubic (4 x 4 x 4
Tic-Tac-Toe), the “L” Game, Chinese Checkers and Kensington.
Search WWH ::




Custom Search