On the rationality of profit sharing in multi-agent reinforcement learning

Reinforcement learning is a kind of machine learning. It aims to adapt an agent to an unknown environment according to rewards. Traditionally, from theoretical point of view, many reinforcement learning systems assume that the 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 learning system 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.