CCKS framework improves multi-agent RL with consensus-based knowledge sharing

In decentralized multi-agent reinforcement learning (MARL), action-advising helps agents learn from experienced teachers, but existing methods often blindly follow advice without evaluating compatibility, leading to excessive advising and instability. Researchers from Renmin University and Macau University present CCKS (Consensus-based Communication and Knowledge Sharing), a framework that allows agents to adopt recommendations based on consensus-derived constraints. The key innovation is using contrastive learning during training to build consensus models from each agent's local observations. During action selection, agents score and choose actions by weighing both shared knowledge and the consensus model, enabling them to balance exploration with smart adoption of teacher advice.

CCKS is designed as a plug-and-play module that integrates with existing Decentralized Training and Decentralized Execution (DTDE) algorithms without architectural changes. Experiments on the Google Research Football environment and the StarCraft II Multi-Agent Challenge (SMAC) showed that adding CCKS significantly improved cooperation efficiency, learning speed, and overall performance compared to current DTDE baselines. The work addresses a critical bottleneck in scalable multi-agent systems, where agents must coordinate without a central controller. By making action-advising more selective and context-aware, CCKS paves the way for more robust decentralized AI in applications like robotics, autonomous driving, and game AI.