Bottom-Up Reputation Promotes Cooperation with Multi-Agent Reinforcement Learning

Reputation serves as a crucial mechanism for fostering cooperation in multi-agent systems, as agents are more likely to collaborate with those who have established trustworthiness. However, existing reinforcement learning approaches often rely on predefined social norms to assign reputations, leaving open the question of how independent agents form a consensus on reputation in a decentralized manner.

Our paper introduces Learning with Reputation Reward (LR2), a bottom-up reputation learning framework that promotes cooperative behavior through reputation-driven reward shaping. LR2 allows agents to assign and utilize reputations autonomously without the need for centralized control or predefined norms.

Key Contributions:

• Dilemma Policy & Evaluation Policy: LR2 agents operate with two policies—one for making cooperative decisions while considering neighborhood impact, and another for assigning reputations to influence others strategically.
• Decentralized & Adaptive Mechanism: Agents learn reputation-based interactions solely through local observations and experience-driven evaluations, without external intervention.
• Enhanced Cooperation Stability: By fostering strategy clustering in structured populations, LR2 not only stabilizes cooperation but also enhances it over time.

Our experimental results demonstrate that LR2 improves cooperation across various spatial social dilemma settings. The findings highlight how bottom-up reputation assignments reshape agent behaviors, promoting robust and sustained cooperation.

If you’re interested, please read the full paper here.

The implementation code is available in the associated GitHub repository.

Thank you for your interest in our research! I look forward to discussing your insights and feedback!

Recommended citation: T. Ren, X. Yao, Y. Li, and X.-J. Zeng, “Bottom-Up Reputation Promotes Cooperation with Multi-Agent Reinforcement Learning,” in Autonomous Agents and Multiagent Systems (AMMAS), 2025.

Recommended citation: T. Ren, X. Yao, Y. Li, and X.-J. Zeng, "Bottom-Up Reputation Promotes Cooperation with Multi-Agent Reinforcement Learning," Autonomous Agents and Multiagent Systems (AMMAS), 2025.
Download Paper