Reinforcement Learning by Guided Safe Exploration

Reinforcement Learning by Guided Safe Exploration

Yang, Q. (TU Delft Algorithmics)
Simão, T. D. (TU Delft Algorithmics)
Jansen, Nils (Radboud Universiteit Nijmegen)
Tindemans, Simon H. (TU Delft Intelligent Electrical Power Grids)
Spaan, M.T.J. (TU Delft Algorithmics)

Gal, Kobi (editor)
Gal, Kobi (editor)
Nowe, Ann (editor)
Nalepa, Grzegorz J. (editor)
Fairstein, Roy (editor)
Radulescu, Roxana (editor)

2023

Safety is critical to broadening the application of reinforcement learning (RL). Often, we train RL agents in a controlled environment, such as a laboratory, before deploying them in the real world. However, the real-world target task might be unknown prior to deployment. Reward-free RL trains an agent without the reward to adapt quickly once the reward is revealed. We consider the constrained reward-free setting, where an agent (the guide) learns to explore safely without the reward signal. This agent is trained in a controlled environment, which allows unsafe interactions and still provides the safety signal. After the target task is revealed, safety violations are not allowed anymore. Thus, the guide is leveraged to compose a safe behaviour policy. Drawing from transfer learning, we also regularize a target policy (the student) towards the guide while the student is unreliable and gradually eliminate the influence of the guide as training progresses. The empirical analysis shows that this method can achieve safe transfer learning and helps the student solve the target task faster.

http://resolver.tudelft.nl/uuid:65bcb73a-e4f6-4e35-9ab5-975461fd2466

https://doi.org/10.3233/FAIA230598

9781643684369

ECAI 2023 - 26th European Conference on Artificial Intelligence, including 12th Conference on Prestigious Applications of Intelligent Systems, PAIS 2023 - Proceedings

26th European Conference on Artificial Intelligence, 2023-09-30 → 2023-10-04, Kraków, Poland

Frontiers in Artificial Intelligence and Applications, 0922-6389, 372

Institutional Repository

conference paper

© 2023 Q. Yang, T. D. Simão, Nils Jansen, Simon H. Tindemans, M.T.J. Spaan