Emergence of Individuality and Sociality by Reinforcement Learning in Multi-Agent Systems

In this paper, a new viewpoint, "individuality" and "sociality", is introduced for analyzing the multi-agent system's behavior. A hypothesis is set up that the both behavioral characters emerge to generate different actions from the other agent to increase each agent's bene t. "Individuality" is de ned as the di erence of actions between agents based on the di erence of its internal information processing. While "sociality (rule)" is de ned as the di erence of actions based on the di erence of its sensory inputs. Next, a model is proposed in which "individuality" and "sociality" are obtained by reinforcement learning. It is also mentioned that there exist some factors like asymmetry of the environment, which in uence the di erentiation into one of the two characters. Finally through some simulations of con ict avoidance problems among passengers getting on and o a train, it is examined that the di erentiation is adaptive to some of the above factors appropriately, and the rule that the passengers getting o have a priority to go.