Home WakeSpace Scholarship › Electronic Theses and Dissertations

Task Agnostic Safe Reinforcement Learning

Electronic Theses and Dissertations

Item Files

Item Details

Task Agnostic Safe Reinforcement Learning
Rahman, Md Asifur
Reinforcement learning (RL) has the potential to be applied in many real-world scenarios due to its ability to learn through exploring the problem domain. However, this same learning capability also makes RL prone to being unsafe. For example, a robot may compromise its structural integrity while blindly aiming to accomplish task objectives in situations where ensuring its own safety is a top priority. Research has provided important insights on how to enhance the safety of RL, but many prior studies have portrayed exploration as being in conflict with safety. This is because an RL agent can transition to uncharted unsafe regions of the environment, resulting in safety violations. This issue is addressed either through joint optimization of task and safety or by implementing constraints for safe exploration. This study introduces a novel safety approach called Task Agnostic Safe Reinforcement Learning (TAS-RL) that does not restrain RL's natural exploratory task learning. Rather, it acquires the unsafe policy model of behaviors that are responsible for most safety violations by utilizing exploration and later derives a safety policy with respect to the unsafe policy. The proposed TAS-RL has been tested against state-of-the-art safe-RL baselines in terms of its safety performance and robustness while accounting for uncertainty. TAS-RL consistently outperforms the baselines by achieving an average safety performance of over 75% in the continuous action space with 10 times more variations in the testing environment dynamics. Using a standalone safety policy independent of conflicting objectives, TAS-RL also paves the way for interpretable safety behavior analysis, as demonstrated through a user study. Moreover, this thesis presents a novel study to investigate the robustness and interpretability of safe RL methods under deliberate perturbations.
Behavior Exclusion
Reinforcement Learning
Safe Reinforcement Learning
Safety Learning
Task Agnostic Safety
Alqahtani, Sarra Dr. (advisor)
Pauca, Pau ́l Dr. (committee member)
Fulp, Errin Dr. (committee member)
Han-Vanbastelaer, Xueyuan Michael Dr. (committee member)
2024-02-13T09:36:05Z (accessioned)
2024-02-13T09:36:05Z (available)
2023 (issued)
Computer Science (discipline)
http://hdl.handle.net/10339/102904 (uri)
en (iso)
Wake Forest University

Usage Statistics