Host: Japan Society for Fuzzy Theory and Intelligent Informatics (SOFT)
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-number plane. Generally it is important for the car agent to pass through the current waypoint with smaller steps, while it is more important to pass through the current one so as to pass easily through the next one when the car agent can much more easily pass through the current one than the opponent. To learn such two different actions, we have applied a fuzzy Q-learning to a car racing game in a single trial, where the Q-table is updated by two kinds of rewards, and car agent's actions are strongly affected by frequently given reward. We, therefore, propose a fuzzy Q-learning with two Q-tables according to two kinds of rewards. This selects an action with two Q-tables and updates two Q-tables according to two kinds of rewards. We show a simulation result in this method.
