Graph-Based Contribution Evaluation for Multi-agent Systems via Graph Neural Networks

Assessing subsystem contributions in large-scale intelligent architectures is essential for understanding internal cooperation mechanisms and guiding system optimization. This paper presents a contribution evaluation method that integrates multi-agent reinforcement learning and graph neural networks. By constructing dynamic graphs based on agent interactions, the framework models structural dependencies and quantifies functional contributions of individual nodes. The approach is validated in a StarCraft II simulation scenario, where coordinated strategies are learned via Hierarchical Attention-based Proximal Policy Optimization (HAPPO) and evaluated through graph-based representations. Results show that the method effectively identifies core components and reveals the structural roles of subsystems within the intelligent system.