A Neural Network for Evaluating King Danger in Shogi

Grimbergen, R. (2002)

in: The 7th Game Programming Workshop (GPW2002), pp. 36--43, Kanagawa, Japan.

Abstract

Neural networks are a promising tool for automatically tuning parts of game programs. For neural networks to be useful, it must be clear which features need to be weighted. There also must be a way to decide automatically if the output of the neural network is correct or not. King danger in shogi meets both these requirements. In this paper a simple neural network is given that uses 161 features as input units. This network is trained with 500 positive and 500 negative examples of king danger. The learning curves for this neural network are not ideal, but for large training sets the learning behaviour is good, giving a prediction accuracy of over 90%. Furthermore, a self-play experiment shows that it is likely that the resulting network can be used in a normal evaluation function.