Robust Multi-Agent Communication With Graph Information Bottleneck Optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Multi-agent reinforcement learning (MARL) benefits from communication learning for improved action coordination.
• Graph neural networks (GNNs) offer a framework for MARL communication, modeling agents and channels as graph nodes and edges.
• Existing GNN-based MARL communication is vulnerable to adversarial attacks and noise, a challenge largely unaddressed.

Purpose of the Study:

• To develop a robust communication learning mechanism for MARL systems.
• To enhance the resilience of GNN-based communication against perturbations.
• To optimize the effectiveness and robustness of MARL communication.

Main Methods:

• Introduced a robust communication learning mechanism employing graph information bottleneck optimization.
• Developed two information-theoretic regularizers to learn minimal sufficient message representations.
• Maximized mutual information between message representation and action selection.
• Minimized mutual information between agent features and message representation.
• Integrated the communication mechanism into a MARL framework with value decomposition methods.

Main Results:

• The proposed method demonstrates superior robustness compared to state-of-the-art GNN-based MARL techniques.
• Experimental results confirm enhanced efficiency of the developed communication learning mechanism.
• The approach effectively mitigates the impact of noise and adversarial perturbations on MARL communication.

Conclusions:

• The graph information bottleneck optimization provides an effective strategy for robust MARL communication.
• The proposed regularizers enable learning of efficient and resilient communication protocols.
• This work advances the field of robust MARL by addressing critical vulnerabilities in GNN-based communication.