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 Reinforcement Learning in Black-Box Environments via Adaptive Shielding

Created by
  • Haebom

Author

Daniel Bethell, Simos Gerasimou, Radu Calinescu, Calum Imrie

Outline

This paper presents ADVICE (Adaptive Shielding with a Contrastive Autoencoder), a novel postprocessing technique for safe exploration of reinforcement learning (RL) agents. It focuses on reducing safety risks that arise when training RL agents in black-box environments without prior knowledge. ADVICE distinguishes between safe and unsafe features of state-action pairs, thereby protecting the agent from performing actions likely to lead to unsafe outcomes. Experimental results demonstrate that it reduces safety violations by approximately 50% compared to existing safe RL exploration techniques, while achieving competitive rewards.

Takeaways, Limitations

•
Takeaways:
◦
We present an effective postprocessing technique for safe navigation of RL agents in black-box environments.
◦
Achieve competitive performance while significantly reducing safety violations.
◦
A novel approach to distinguishing the safe/unsafe characteristics of state-action pairs is presented.
•
Limitations:
◦
The performance of ADVICE may depend on the performance of the contrastive autoencoder.
◦
Further research is needed on generalization performance across different environments and tasks.
◦
Consideration should be given to additional safety concerns that may arise in real-world applications.
👍
Made with Slashpage