[공지사항]을 빙자한 안부와 근황 
Show more

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.

Causal Knowledge Transfer for Multi-Agent Reinforcement Learning in Dynamic Environments

Created by
  • Haebom

Author

Kathrin Korte, Christian Medeiros Adriano, Sona Ghahremani, Holger Giese

Outline

In this paper, we present a causal knowledge transfer framework to address the knowledge transfer problem of multi-agent reinforcement learning (MARL) in unpredictable environments. Effective knowledge transfer between agents in unpredictable environments with changing goals is a challenging task. This study enables agents to learn and share concise causal representations of paths in the environment. When environmental changes such as new obstacles occur, conflicts between agents are modeled as causal interventions, and these are implemented as recovery action sequences (macros) to bypass obstacles and increase the probability of goal achievement. These recovery action macros are transferred online from other agents without retraining, and applied as lookup model queries using local context information (conflicts).

Takeaways, Limitations

•
Takeaways:
◦
When adapting to an unfamiliar environment, we show that agents with heterogeneous objectives can reduce the performance gap between random exploration and fully retrained policies by about half.
◦
We reveal that the effectiveness of causal knowledge transfer depends on the complexity of the environment and the interactions between the heterogeneous goals of the agent.
◦
We present a new way to transfer knowledge online without retraining, increasing efficiency.
•
Limitations:
◦
The performance of the proposed method depends on the complexity of the environment and the heterogeneity of the agent's goals, and may not guarantee consistent performance in all situations.
◦
The performance of the lookup model can have a significant impact on the performance of the entire system, and further research is needed on the design and training of the lookup model.
◦
Additional experiments and analysis are needed to determine generalization performance across different types of environments and agents.
👍
Made with Slashpage