On the Rationality of Profit Sharing in Multi-Agent Reinforcement Learning

• Authors: Kazuteru Miyazaki;Shigenobu Kobayashi
• Venue: ICCIMA '01 Proceedings of the Fourth International Conference on Computational Intelligence and Multimedia Applications
• Year: 2001

Reinforcement learning is a kind of machine learning.It aims to adapt an agent to anUnknown environment according to rewards.Traditionally, from the vertical point of view,many reinforcement learning systems assume that environment has Markovian properties.However it is important to treat non-Markovian environments in multi-agent reinforcement Learning systems.In this paper, we use Profit Sharing (PS) as a reinforcement learningSystem and discuss the rationality of PS in multi-agent environments.Especially, we Classify non-Markovian environments and discuss how to share a reward among Reinforcement learning agents.Through cranes control problem, we confirm the effectiveness of PS in multi-agent environments.