Enhancing Cooperation through Selective Interaction and Long-term Experiences in Multi-Agent Reinforcement Learning

Why focus on interaction rules? They’re fundamental in society, yet aligning everyone to common standards from the start is challenging.

Our paper introduces a novel training framework that enables reinforcement learning agents to develop cooperative behaviors and strategic interactions in social dilemmas, without relying on predefined social norms or external incentives. The approach combines deep reinforcement learning with complex game theoretical models to study cooperative dynamics:

• Selective Interaction: Introduces mechanisms for agents to selectively interact based on past experiences, enhancing cooperative outcomes.
• Integration of Long-term Experiences: Employs reinforcement learning to adapt strategies over time, factoring in the history of interactions to optimize future decisions.

Our framework employs two distinct Q-networks per agent, enabling interactions with up to four neighbours. Results show agents distinguish between cooperative neighbours and free-riders. This enhances network reciprocity and fosters strategy clusters in networked populations.

If you are insterested, please read the full paper here.

The Implementation code can be found in the associated GitHub repository.

Poster Presentation of This Work

Thank you for your interest in our research! I’m eager to discuss your comments further!

Recommended citation: T. Ren and X. -J. Zeng, “Enhancing Cooperation through Selective Interaction and Long-term Experiences in Multi-Agent Reinforcement Learning,” in International Joint Conference on Artificial Intelligence (IJCAI). Jeju, Korea (2024).

Recommended citation: T. Ren and X.J. Zeng, "Enhancing Cooperation through Selective Interaction and Long-term Experiences in Multi-Agent Reinforcement Learning," International Joint Conference on Artificial Intelligence (IJCAI). Jeju, Korea (2024).
Download Paper