Global and Coalition Cognition Graph Modeling for Interpretable Multiagent Reinforcement Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Multiagent Systems

Background:

• Multiagent reinforcement learning (MARL) excels in complex tasks but lacks transparency in agent decision-making.
• Current MARL methods often produce opaque neural network policies, hindering human understanding of agent behavior.

Purpose of the Study:

• To introduce a novel framework, global and coalition cognition graph modeling (GC²GM), for interpretable MARL.
• To equip agents with explicit cognitive representations for coalition-level reasoning and global state awareness.

Main Methods:

• Developed an information exchange graph network (IEGN) for agents to reason about teammate intentions and behaviors.
• Implemented mutual-information regularization to align agents' local perspectives with global environmental states.
• Designed GC²GM within the centralized training with decentralized execution (CTDE) paradigm for practical application.

Main Results:

• GC²GM demonstrated strong performance across various challenging MARL benchmarks.
• The framework achieved a significant improvement in the interpretability of agent behaviors.
• GC²GM successfully balanced task performance with enhanced explainability.

Conclusions:

• GC²GM offers a promising solution for creating more transparent and understandable MARL systems.
• The modular design allows seamless integration into existing MARL frameworks.
• This work paves the way for more trustworthy and interpretable multiagent AI.