A graph convolutional attention network with Q-net on multi-agent systems
Yupei Li, Tianqi Shen, Zhiyuan Liu · 2nd International Conference on Artificial Intelligence, Automation, and High-Performance Computing (AIAHPC 2022) · 2022
In recent years, GNNs have becoming one of the hottest topics of deep learning due to their powerful ability in modeling relational data and their wide applications in real world. Besides, reinforcement learning is widely applied in many fields. However, the tradition way to train the reinforcement learning is not that accurate enough. It can be assumed that it is wise to combine GNN and reinforcement learning, making good use of the modeling capability of GNNs, to enhance the reinforcement learning performance. In this paper, the most ordinary GNN is compared within the existing works to see whether it is possible to combine GNN and reinforcement learning. PettingZoo is chosen on behalf of the multi-agent system, and a Graph Convolutional Attention Network with Q-net (GCANQ) is designed and validated, which shows the efficient computational capability of GNN. Experiments show that GNN performs strikingly well on a common multiagent system and in a small scale with some overfit as well. Moreover, the speed of the training of GCANQ is acceptable for the complex calculation needed.