Multi-Level Credit Assignment for Cooperative Multi-Agent Reinforcement Learning

Multi-agent reinforcement learning (MARL) has become more and popular over recent decades, the need for high-level cooperation is increasing every day because of complexity real-world environment. However, multi-agent credit assignment problem that serves as main obstacle to coordination still not addressed properly. Though lots methods have been proposed, none them thought perform assignments across multi-levels. In this paper, we aim propose an approach a better scheme by First, hierarchical model consists manager level worker level. The incorporates dilated Gated Recurrent Unit (GRU) focus on plans uses GRU execute primitive actions conditioned plans. Then, one centralized critic designed each learn level’s scheme. To end, construct novel MARL algorithm, named MLCA, which can achieve multi-level assignment. We also conduct experiments three classical challenging tasks demonstrate performance proposed algorithm against baseline methods. results show our method gains great improvement all maps require cooperation.