A Novel Multi-Agent Deep Reinforcement Learning Approach

Abstract

Borrowing the power of deep neural networks, deep reinforcement learning achieved big success in games, and it becomes a popular method to solve the sequential decision-making problems. However, the success is still restricted to single agent training environment. Multi-agent reinforcement learning still is a challenge problem. Although some multi-agent deep reinforcement learning methods have been proposed, they can only perform well when the number of agents is very limited. In this paper, by analyzing the dynamic changing observation space and action space of multi-agent environment, we propose a novel multi-agent deep RL method that compress the joint observation space and action space as the time goes on. The proposed method is potential for a large number of agents cooperative or competitive tasks