A Probabilistic Generative Method for Safe Physical System Control Problems
Keywords: safe PDE control, physical systems, generative models, conformal prediction, fine-tuning
TL;DR: To solve safe PDE control problems, we propose Safe Conformal Physical system control (SafeConPhy), which iteratively improves model safety with a provable safety bound.
Abstract: Controlling complex physical systems is a crucial task in science and engineering, often requiring the balance of control objectives and safety constraints. Recently, diffusion models have demonstrated a strong ability to model high-dimensional state spaces, giving them an advantage over recent deep learning and reinforcement learning-based methods in complex control tasks. However, they do not inherently address safety concerns. In contrast, while safe reinforcement learning methods consider safety, they typically fail to provide guarantees for satisfying safety constraints.
To address these limitations, we propose Safe Conformal Physical system control (SafeConPhy), which optimizes the diffusion model with a provable safety bound iteratively to satisfy the safety constraint.
We pre-train a diffusion model on the training set. Given the calibration set and the specific control targets, we derive a provable safety bound using conformal prediction. After iteratively enhancing the safety of the diffusion model with the progressively updated bound, the model's output can be certified as safe with a user-defined probability.
We evaluate our algorithm on two control tasks: 1D Burgers' equation and 2D incompressible fluid. Our results show that our algorithm satisfies safety constraints, and outperforms prior control methods and safe offline RL algorithms.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
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: 1241
Loading