対局経験を利用した相手の手の予測システム
大島 安簡、Grimbergen, R. (2008).
in: The 13th Game Programming Workshop in Japan (GPW2008),
pp. 80--86, Kanagawa, Japan. (in Japanese)
Abstract
The general assumption when performing search in game-playing research is that the opponent will play
the best move. Even when there is experience against the same opponent, this is still being assumed.
However, human players use their experience against the same opponent to make a prediction about which
move the opponent will play. In this research, we will propose a searching system with a more human-like
approach, trying to predict the moves of the opponent. In order to make such prediction, a three-layer
neural network will be used as memory, which is being trained by the games played against the same
opponent. Furthermore, this network was used to predict the move in the case of unknown patterns, as well
as guiding a search algorithm. As a result, it was shown that the system could predict moves in unknown
patterns, but that improvements are needed to actually lead to improved playing strength.