Bayesian Policy Distillation via Offline RL for Lightweight and Fast Inference
Keywords: neural network compression, reinforcement learning, robot learning
Abstract: High-performance deep reinforcement learning faces tremendous challenges when implemented on cost-effective low-end embedded systems due to its heavy computational burden. To address this issue, we propose a policy distillation method called Bayesian Policy Distillation (BPD), which effectively retrains small-sized neural networks through an offline reinforcement learning approach. BPD exploits Bayesian neural networks to distill already designed high-performance policy networks by adopting value optimizing, behavior cloning, and sparsity-inducing strategies. Simulation results reveal that the proposed BPD successfully compresses the policy networks, making them lighter and achieving faster inference time. Furthermore, the proposed approach is demonstrated with a real inverted pendulum system and reduced the inference time and memory size by 78 \% and 98 \%, respectively.
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.
Supplementary Material: zip
Submission Number: 1812
Loading