Interactive learning of social agents based on confidence degree

In Multi-Agent Systems (MAS), plural autonomous agents are able to acquire cooperative behavior by learning process. It is useful for agents to learn experience by sharing policies of others. However, in the normal Reinforcement Learning (RL), heterogeneous agents are difficult to acquire useful experience from others directly, because they have different characteristics and their actions would bring different effects. It is necessary for heterogeneous social agents to obtain suitable experience during the interaction with others. In our proposal, each agent learns not only the experience to adapt itself to environment, but also the suitable experience from other agents. When agents are learning by trial and error, the confidence to others is formed by learning process. The confidence is useful to select the suitable experience from various experience, and will be refreshed after each trial learning. Interactive relationship among social agents will be shown based on the confidence degree. After interactive learning processes, social agents will acquire the better cooperative performance than using normal RL method.