Deep Reinforcement Learning for Multi-Agent Systems: A Review of Challenges, Solutions and Applications

  • 2018-12-31 13:39:54
  • Thanh Thi Nguyen, Ngoc Duy Nguyen, Saeid Nahavandi
  • 7

Abstract

Reinforcement learning (RL) algorithms have been around for decades and beenemployed to solve various sequential decision-making problems. These algorithmshowever have faced great challenges when dealing with high-dimensionalenvironments. The recent development of deep learning has enabled RL methods todrive optimal policies for sophisticated and capable agents, which can performefficiently in these challenging environments. This paper addresses animportant aspect of deep RL related to situations that demand multiple agentsto communicate and cooperate to solve complex tasks. A survey of differentapproaches to problems related to multi-agent deep RL (MADRL) is presented,including non-stationarity, partial observability, continuous state and actionspaces, multi-agent training schemes, multi-agent transfer learning. The meritsand demerits of the reviewed methods will be analyzed and discussed, with theircorresponding applications explored. It is envisaged that this review providesinsights about various MADRL methods and can lead to future development of morerobust and highly useful multi-agent learning methods for solving real-worldproblems.

 

Introduction (beta)

None

 

Conclusion (beta)

None