人間と対戦可能なカーリングロボットに関する研究

第7報 戦術決定のための評価関数の重み学習法

抄録

This paper deals with the optimization of the weight in evaluation function. The evaluation function is expressed by sum of product of the feature values based on stone distribution and each weights. There are two aims in this study. One is optimization of weight for evaluation function, and the other is to determine the appropriateness of each feature value which selected for curling robot. After a several feature values are proposed, supervised learning is performed using the averaged perceptron algorithm. Teacher data is selected delivery parameter that show the "good" results from the game record of digital curling. Learning results showed a high reproducibility with the training set, but showed low reproducibility for testing set. As a result, optimization of weights do not proceed, and determination of the appropriateness of the feature value selection is failure. The results suggest that learning should perform after delivery parameters are classified by purpose such as "take", "draw","freeze".