A Social View to Multiagent Reinforcement Learning

摘要

This paper presents a study on Multiagent Reinforcement Learning (RL) for cooperating work from a social view to solve the problems of individual agent's incomplete world model, conflict of individual interests and inconstant reinforcement from environment. Through a modeling of four software robot agents' cooperative work to balance a ball on a plate, two social RL approaches-observing reinforcement and vicarious reinforcement-are applied to individual RL agents in a multiagent learning environment. The comparisons between social RL and independent RL are provided in several aspects, and we show that social RL accelerates the convergence of agents' learning efficiency in an organization and also contributes to the learning efficiency of unskilled agents, who are newcomers and join to an existing organization.