Symbolizing the learning environment for acquiring game bots

Abstract

The purpose of this research is to symbolize learning objects in game bots. By symbolizing the learning target using object detection, we can extract knowledge about common scenes on different screens by reducing the amount of information in the screens. We used YOLO, one of the object detection methods, to perform object detection on Atari titles. Fifty screenshots for each were prepared as training and test images, and the detection rate was examined through training. The average detection rate was about 80%, so it is necessary to increase the number and type of training images.