An Improved Deep Reinforcement Learning-Based Multi-Agent Cooperative Game Approach

EasyChair Preprint 11210

Abstract

Multi-agent collaborative games based on deep reinforcement learning have been one of the hot topics in the field of artificial intelligence in recent years. Building on existing research, this paper selects the on-policy Multi-Agent Proximal Policy Optimization (MAPPO) algorithm to explore its performance in multi-agent collaborative games, providing new insights for further research. Using the Hanabi game environment, this paper implements the MAPPO algorithm with an appropriate action space to maximize collaborative efficiency and competitiveness. Experimental results demonstrate that the MAPPO algorithm performs well in collaborative gaming scenarios. Compared to the off-policy Value-Decomposition Networks (VDN) algorithm [1], it improves the decision efficiency and outcomes of intelligent agents. This study highlights the feasibility and advantages of the MAPPO algorithm in multi-agent collaborative games. Furthermore, this experiment delves into the application of the MAPPO algorithm in multi-agent collaborative games, offering valuable insights for enhancing reinforcement learning algorithms and their practical applications. This study also poses new questions and provides guidance and inspiration for future researchers.

Keyphrases: MAPPO, Reinforcement Learning, multi-agent collaborative games

@booklet{EasyChair:11210,
  author    = {Zhongqi Zhao and Chuang Zhang and Haoran Xu and Jiawei Kou and Hui Cheng},
  title     = {An Improved Deep Reinforcement Learning-Based Multi-Agent Cooperative Game Approach},
  howpublished = {EasyChair Preprint 11210},
  year      = {EasyChair, 2023}}