Abstract
Car racing game is a game of computer programs in IEEE CEC 2007 Car Racing Competition, where two car agents compete with each other for passing through waypoints in a two-dimensional real-member plane. In a competitive game, passing through the current waypoint so as to pass the next one easily is more important than passing the current one quickly when our agent has a great advantage for the opponent. We defined it as an action to pass through the current waypoint in the front direction for the next one, but we have several other cases, e.g., in the back direction for the next one. Thus, these actions are considered as an action with the minimum steps from the current waypoint to the next one. We propose an improved reward condition that the present steps are smaller than the past steps. We show simulation results of its application to fuzzy Q-learning with two Q-tables.
