Target-Oriented Soft-Robust Inverse Reinforcement Learning

26 Sept 2024 (modified: 13 Nov 2024)ICLR 2025 Conference Withdrawn SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Imitation Learning, Inverse Reinforcement Learning, Soft-Robust Optimization, Robust Optimization, Optimization Algorithm
Abstract:

In imitation learning, when the learning agent is at a state that is outside the demonstration of the expert, it could be difficult for her to choose an action. To overcome this challenge, inverse reinforcement learning (IRL) learns a parameterized reward function based on which we can generalize the expert's behavior to those states that are unseen in the demonstration. However, on the one hand, there could be multiple reward functions that can explain the expert's behavior, leading to reward ambiguity in IRL. On the other hand, though we often consider the transition kernel of the expert to be known to the agent, sometimes the transition kernel of the agent is different from the expert's and is unknown, leading to transition kernel ambiguity in IRL. Drawing on the notion of soft-robust optimization, we build a target-oriented soft-robust IRL (SRIRL) model where the performance of the output policy strikes a flexible balance between risk aversion and expected return maximization towards reward uncertainty in IRL. Moreover, by employing the robust satisficing framework, our SRIRL is also robust to transition kernel ambiguity in IRL. In our target-oriented SRIRL, we keep a target for the performance of the output policy that balances expected return and risk, and we minimize the constraint violation incurred by the difference between the ambiguous transition kernel and the empirical one. We derive tractable reformulation for SRIRL, and we design tailored first-order methods for SRIRL. Numerical results showcase the soft robustness towards reward uncertainty and the robustness against transition kernel ambiguity of SRIRL, as well as the stronger scalability of our first-order methods compared to a state-of-the-art commercial solver.

Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6233
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview