A Cooperative Learning Method for Multi-Agent System with Different Input Resolutions

Abstract

Multi-Agent Reinforcement Learning controls some agents to learn group action with cooperation each other. For example, AGVs in warehouse as the agents cooperate with others and put on and off the supplies to organize them. Though Multi-Agent Reinforcement Learning seems to make advantage to apply multi-robot and more domains, this method has some problems, in particular, it cannot consider the sensor resolution in real world problem. This paper addresses this problem as hetero informational problem, and discuss how to solve the problem by the topology and learning of the neural network of the deep reinforcement learning. Concretely, This paper employed Asynchronous Advantageous Actor-Critic (A3C) with some kinds of neural networks to discuss through two experimental cases, single and multi agent domains. This paper compared performance of agents with different number of hidden layers of neural networks in the single agent domain, and investigate the performance on the environment whose agents have different resolution each other in the multi-agent domain.