GR-MAPPO Algorithm for Perimeter Defense Problem in Multi-Agent Systems.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Multi-agent perimeter defense is crucial for unmanned swarms.
• Existing deep reinforcement learning methods struggle with coordination, temporal information, and scalability under communication constraints.
• Dynamic variations in swarm size pose generalization challenges.

Purpose of the Study:

• To propose a novel multi-agent reinforcement learning (MARL) strategy for perimeter defense in unmanned swarms.
• To enhance coordination and temporal information processing under local communication constraints.
• To improve generalization capability across dynamic swarm sizes.

Main Methods:

• Implemented a GraphSAGE-based spatial aggregation module for enhanced inter-agent information exchange.
• Utilized a GRU-based temporal encoding module to process historical observations for improved anticipation.
• Employed an inductive node-level aggregation mechanism to ensure scalability and adaptability to varying swarm sizes.

Main Results:

• The proposed GR-MAPPO strategy significantly improved capture performance under limited communication.
• Demonstrated enhanced coordination and anticipatory capabilities.
• Showcased superior performance retention during cross-scale transfers across different swarm scales.

Conclusions:

• The integrated spatial and temporal modeling approach effectively addresses limitations of existing MARL methods in swarm defense.
• The proposed method offers a scalable and adaptable solution for dynamic multi-agent perimeter defense scenarios.
• This research advances the cooperative defense capabilities of unmanned swarms in complex operational environments.