Decentralized Mean Field Games

Multiagent reinforcement learning algorithms have not been widely adopted in large scale environments with many agents as they often poorly the number of agents. Using mean field theory to aggregate has proposed a solution this problem. However, almost all previous methods area make strong assumption centralized system where environment learn same policy and are effectively indistinguishable from each other. In paper, we relax about propose new known Decentralized Mean Field Games, agent can be quite different others. All independent policies decentralized fashion, based on their local observations. We define theoretical concept for provide fixed point guarantee Q-learning algorithm system. A practical consequence our approach is that address `chicken-and-egg' problem empirical algorithms. Further, actor-critic use give stronger performances compared common baselines area. setting, do need clones other fully fashion. Hence, first time, show application competitive environments, large-scale continuous action space heterogeneous Importantly, also apply method ride-sharing using real-world dataset. problem, which more than existing training methods.