Abstract Multiple agents become increasingly required in various fields for both physical robots and software agents, such as, robot soccer, search and rescue robots, automated driving, auctions and electronic commerce agents, and so on. In multiagent domains, agents interact with others, and coadapt with others and act on the best choice available. Each agent's choice of policy depends on the others' joint policy which also aims to achieve the best available performance. Since all the agents are evolving, the environment is no longer stationary, and this brings in a difficult learning problem that violates the basic stationary assumption of traditional techniques for behavior learning. This paper reviews seminal research in the multiagent learning field from a reinforcement learning perspective. Under the framework of stochastic games, model-based, model-free, and no-regret learning technique have achieved notable success in the field, while more research is still required to meet the theoretical fundation for non-stationary learning as well as apply in practice. Open issues are discussed from scalability, dynamics, communication and evaluation criteria aspects.

Download paper in PDF format.

The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.