Hierarchical Attention Master-Slave for heterogeneous multi-agent reinforcement learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Current multi-agent reinforcement learning (MARL) often uses homogeneous agents, limiting complex task coordination.
• Real-world complex tasks require diverse agents that leverage each other's strengths.

Purpose of the Study:

• To develop a novel MARL approach for optimizing decision-making and communication in heterogeneous agent systems.
• To address the limitations of homogeneous agents in complex, multi-agent scenarios.

Main Methods:

• Proposed Hierarchical Attention Master-Slave (HAMS) MARL framework.
• HAMS utilizes hierarchical attention for inter- and intra-cluster weight balancing.
• Master-Slave architecture enables independent agent reasoning and guidance.

Main Results:

• Achieved over 80% win rates across all StarCraft II micromanagement tasks.
• Exceeded 90% win rate on the largest map.
• Demonstrated up to a 47% improvement in win rate compared to existing algorithms.

Conclusions:

• HAMS effectively fuses information among heterogeneous agents, optimizing decisions while minimizing communication overhead.
• The proposed method offers a novel approach for heterogeneous multi-agent policy optimization.
• HAMS significantly outperforms state-of-the-art methods in complex MARL tasks.