|[Oshita Lab.][Research Theme]||[Japanese]|
In this paper, we propose a method for automatically learning motion rules for autonomous characters from control logs. We ask a few players to control their characters in a gaming application. Based on the control logs obtained, motion rules are constructed for each character. We use a support vector machine (SVM) to learn the motion rules. However, the SVM cannot be applied directly to our problem, as the necessary variables and their values vary according to the motion. To solve this problem, we introduce a layered mechanism and adaptive parameters to the SVM. As an experiment, our method has been applied in simulating football players.