Graph attention networks based multi-agent path finding via temporal-spatial information aggregation.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Multi-Agent Path Finding (MAPF) algorithms face challenges with partial observability, hindering optimal strategy determination.
• Effective information fusion is critical to expand agent observation ranges and improve MAPF system performance.

Purpose of the Study:

• To develop a robust information fusion method for MAPF that addresses partial observability by integrating temporal and spatial dimensions.
• To enhance the accuracy and time efficiency of MAPF algorithms in complex, dynamic environments.

Main Methods:

• A novel fusion approach utilizing Graph Attention Networks (GAT) for inter-agent communication and scaled dot-product aggregation.
• Integration of Gated Recurrent Unit (GRU) and Convolutional Neural Network (CNN) for historical observation encoding.
• Decoding aggregated agent information into final action strategies to solve partial observability.

Main Results:

• The proposed method significantly improves accuracy and time efficiency, outperforming existing GNN and GAT methods by up to 73%.
• Performance gains are more substantial in larger maps, demonstrating the algorithm's scalability.
• Achieved accuracy and time efficiency improvements of 24.5%, 47%, 37.5%, and 73% over GNN and GAT.

Conclusions:

• The developed information fusion technique effectively mitigates partial observability in MAPF.
• The GAT-based approach offers a scalable and efficient solution for complex multi-agent navigation tasks.