State-Blocked Coordination Improves Zero-Shot Agent Cooperation
Researchers introduce State-Blocked Coordination (SBC), a framework enhancing zero-shot coordination (ZSC) in multi-agent systems. ZSC aims for agents to cooperate with unseen partners without prior interaction, crucial for human-AI collaboration. Existing methods focus on increasing partner diversity but often fail to generalize. SBC generates virtual environments via state blocking, exposing agents to diverse suboptimal partner policies. The method achieves superior performance across benchmarks, including strong generalization to human partners.
Key facts
- Zero-shot coordination (ZSC) enables agents to cooperate with independently trained partners without prior interaction.
- Existing approaches emphasize increasing partner diversity during training.
- State-Blocked Coordination (SBC) is a framework that improves ZSC.
- SBC generates a family of virtual environments through state blocking.
- SBC allows agents to experience a wide range of suboptimal partner policies.
- SBC demonstrates superior performance in zero-shot coordination across multiple benchmarks.
- SBC shows strong generalization to human partners.
- The paper is from Computer Science > Machine Learning on arXiv.
Entities
Institutions
- arXiv