MA-MIX: Value Function Decomposition for Cooperative Multiagent Reinforcement Learning Based on Multi-Head Attention Mechanism

Multi-Agent Deep Reinforcement Learning (MADRL) is a research field that combines deep learning and multi-agent reinforcement learning. In complex tasks, a single agent often finds it difficult to complete the task independently, thus requiring cooperation and communication between agents. However, communication between agents remains a key issue in multi-agent cooperative reinforcement learning. To address this issue, we propose a new method called Multi-Head Attention Mixing Network (MA-MIX), which aims to solve key challenges in multi-agent systems. MA-MIX is based on the multi-head attention mechanism and innovatively applied to agent networks, effectively solving the problem of information exchange and cooperation in multi-agent systems. We compared MA-MIX with traditional QMIX algorithms and other baseline algorithms. The experimental results show that MA-MIX has superior performance under the StarCraft Multi-Agent Challenge (SMAC) environment.