Emergence and Differentiation Model of Individuality and Socility by Reinforcement Learning

Abstract

In this paper, a concept of individuality and sociality is introduced as a method to avoid conflicts of individual interests in multi-agent systems. It is considered that each agent has its individuality when the conflicts are resolved by making its own mapping from the sensory input to the action output. On the other hand, each agent has sociality when the conflicts are avoided by some common input-output mapping, which is commonly called rules. A conflict avoidance task in which passengers are getting on and off a train are taken as an example, and the emergence processes of both behavioral characters are explained. Furthermore, it is shown that the differentiation of the agent into one of them is adaptively realized by reinforcement learning based on local rewards according to the asymmetry of environment, number of agents, identification of the other agents, or physical ability of agents.