In daily interaction with others, we can behave cooperatively by estimating the intention of others and predicting their behavior. It is important to understand the mechanism of such smooth cooperative behavior for both comprehending human social cognitive ability and designing social artifacts that can interact with humans. To clarify effective functions that constitute the mechanism, we constructed a computational model of cooperative behavior in which an action decision process is dynamically controlled based on an estimation of intention of other. The model estimates the other's intention from his⁄her behavior by reusing the knowledge of one's own action decisions. However, in a condition in which agents mutually estimate the intentions of others, the cooperation performance is unsatisfactory. We addressed the problem by introducing “level” of estimation. Our model combines three types of action decision strategies: an action strategy based on the estimation of other's intention, one based on the estimation of one's own intention by other, and an action strategy that doesn't use the intention of other. To analyze the model, we simulated tasks in which two hunters cooperatively chased two preys and demonstrated that this model achieved smooth cooperation. The result suggests the effectiveness of our model.
