Efficient multi-agent communication via entity-aware causal network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Multi-agent reinforcement learning (MARL) cooperative tasks heavily rely on effective communication strategies.
• Existing MARL communication methods often use predefined agent orders or identifiers, limiting scalability and efficiency.
• These methods overlook that communication target selection is primarily state-dependent, not order-dependent.

Purpose of the Study:

• To introduce the Entity-Aware Causal (EAC) framework for more efficient and scalable MARL communication.
• To address the limitations of predefined communication targets in existing MARL approaches.
• To enhance communication efficiency through entity-aware target selection and causal inference.

Main Methods:

• Designed an entity-aware hypernetwork to identify communication targets based on individual agent states.
• Employed a masked-attention mechanism for scalable and sparse communication topologies.
• Proposed a causal inference beliefs mechanism to reduce redundant messages and strengthen communication belief.

Main Results:

• The EAC framework demonstrated superior performance compared to baseline MARL algorithms across SMAC, SMAC_v2, GRF, and MPE benchmarks.
• The algorithm showed robustness across diverse network topologies and varying sparsity levels.
• EAC enables scalable and sparse communication by focusing on entity states rather than fixed identifiers.

Conclusions:

• The EAC framework offers a novel entity-centric approach to MARL communication, improving efficiency and scalability.
• State-based communication target selection and causal inference are key to overcoming limitations of traditional MARL communication.
• EAC provides a robust and high-performing solution for complex cooperative MARL tasks.