Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Safe Multiagent Coordination via Entropic Exploration

Created by
  • Haebom

Author

Ayhan Alp Aydeniz, Enrico Marchesini, Robert Loftin, Christopher Amato, Kagan Tumer

Outline

This paper proposes a method that leverages constraints on the entire team, rather than individual agents, to address safety issues in multi-agent reinforcement learning. Existing safe reinforcement learning algorithms constrain agent behavior to limit exploration, which is crucial for discovering effective cooperative behaviors. In this paper, we present Entropy Search (E2C), a method for constrained multi-agent reinforcement learning. E2C encourages exploration by maximizing observation entropy, facilitating the learning of safe and effective cooperative behaviors. Extensive experimental results demonstrate that E2C performs equally or better than existing unconstrained and constrained baseline models, reducing unsafe behaviors by up to 50%.

Takeaways, Limitations

•
Takeaways:
◦
We demonstrate that leveraging constraints on the entire team can effectively address the safety problem in multi-agent reinforcement learning.
◦
We experimentally demonstrate that an observation entropy maximization-based search strategy (E2C) is effective for safe and effective cooperative behavior learning.
◦
We confirm that the proposed method can simultaneously improve safety and performance compared to existing methods.
•
Limitations:
◦
There is a possibility that the effectiveness of the proposed method may be limited to certain environments.
◦
Further research is needed on generalization performance in more complex and diverse multi-agent environments.
◦
Further research may be needed on the design and optimization of team constraints.
👍
Made with Slashpage