Abstract
Car racing game is a game of computer programs in IEEE 2007 Car racing Competition, where two car agents compete with each other for taking waypoints in a two-dimensional real-number plane. We have proposed a method that a car agent gradually learns its actions through several stages with fuzzy Q-learning from the initial state that it does not know how it acts. In the previous research, we introduced four stages where waypoints only appear forward, left and right, backward, and in a small range. We got a good result. However we could not analyze completely how four stages effect on the result. In this paper, we propose improved stages and their combinations based on our research result, and investigate the relation between scores and stages to get high scores. As a result, we get higher scores when we combine stages where a car agent learns actions such as turn quickly after it gets score and put on the brake after it turns. We design stages for a car agent to learn mainly these actions, and can get the best results.
