Host: Japan Society for Fuzzy Theory and Intelligent Informatics (SOFT)
In our laboratory, we have already proposed the method based on the Multi-Layered Fuzzy Behavior Control, which describes behavior decision fuzzy rules for each subtask after dividing complex tasks and acquires the behavior strategy, and the Fuzzy State Division-Type Reinforcement Learning that is able to input the continuous value for the input state. And we have also proposed the Differential Reinforcement-type Shaping Q-Learning applied a general idea of "differential reinforcement" to reinforce the special behavior step by step such as in real animal training. Therefore, in this research, by adopting the concept of shaping to the fuzzy state division type reinforcement learning we propose a method incorporated an automatic Shaping system to make decrease the burden of the trainer which occurs when he gives a Shaping reward. Furthermore, we also propose the method to let it learn the behavior strategy according to the situation for the behavior acquisition of the soccer robot by obtaining the weighted value of attack and defense for the robot by using the fuzzy rule.
