サンプリング実時間Q-MDP法 : 不完全な観測に基づき実時間行動する自律サッカーロボットへの適用 Sampling Real-Time Q-MDP Value Method : Application on an Autonomous Soccer Robot under Uncertainty

The sampling real-time Q-MDP value method, which is a novel real-time decision making method under uncertain information, is proposed in this paper. The proposed method is implemented and evaluated on an autonomous goalkeeper robot for RoboCup four legged robot league. In the original Q-MDP value method, a proper action is chosen by calculations of expected values with the following data: 1) a planning result that is obtained with an assumption that information is perfectly known, and 2) a probability distribution that represents uncertainty of information in the planning space. The real-time Q-MDP method has utilized particle filters for real-time calculation of the values. The proposed method is an improved method of the real-time Q-MDP method. An proper action is chosen only a part of the particles in the proposed method. Faster or more proper decision making can be possible by sophisticated omission of calculations.