Multi-agent Collaborative Decision-making Mechanism Combining Reinforcement Learning and Graph Attention Network
Xinyue Zhang · 2025
Multi-agent collaborative decision-making faces three core problems: inefficient modeling of relationships between agents in partially observable environments, poor scalability of traditional reinforcement learning algorithms, and inaccurate global reward allocation mechanisms. This study proposes a collaborative framework that integrates graph attention networks (GAT) and multi-agent deep deterministic policy gradients (MADDPG), aiming to build a relationship-aware agent decisionmaking mechanism: first, GAT is used to dynamically learn the adjacency matrix, and the agent interaction relationship is encoded through a two-layer attention mechanism; then a joint reward function is designed, and the centralized critic network is combined to evaluate the collaborative utility; the Centralized Training architecture is used to verify it in the StarCraft II multiagent challenge environment. Experiments show that in the 15agent pursuit task, the success rate of this framework reaches 95.7%; in the heterogeneous agent formation task, the path planning efficiency is improved by 36.4 seconds. This method explicitly models the agent dependencies through a graph structure, solving the problem of traditional methods in complex collaboration.