Analysis of Collaborative Policy Learning Mechanisms in the Multi-Agent PPO-Comm Framework for Game

Effective collaboration among agents in Multi-Agent Systems (MAS) is a crucial challenge for accomplishing complex tasks. However, existing methods often suffer from insufficient collaboration efficiency in dynamic environments, and their communication mechanisms are vulnerable to noise interference. To address these issues, this paper proposes a communication mechanism framework named PPO-Comm, based on Multi-Agent Proximal Policy Optimization (PPO), aiming to enhance communication efficiency and collaborative performance among agents through collaborative policy learning. The specific methods include: 1) Designing a Graph Neural Network (GNN)-based communication module where agents achieve efficient information sharing through message passing; 2) Introducing an Attention Mechanism to enhance the ability to extract key information; 3) Combining a multi-agent credit assignment mechanism to optimize agent behavior through a weighted reward function; 4) Adopting a distributed training architecture to improve model training efficiency and convergence speed. Experiments are conducted on platforms such as StarCraft II and MPE. The results show that PPO-Comm achieves a task success rate of $\mathbf{9 5 \%}$ in collaborative tasks, reduces communication costs to $\mathbf{4. 5 M B}$, and demonstrates stronger robustness in dynamic environments. The proposed method effectively solves communication and efficiency problems in multi-agent collaboration, providing new insights for the design of collaborative strategies in MAS.