Entropy regularisation has proven effective in reinforcement learning (RL) for encouraging exploration. Recent work demonstrating the equivalence between entropy regularised RL and approximate probabilistic inference suggests the potential for improving existing methods by generalising the inference procedure. We develop the Rényi regularised RL framework by using Rényi variational inference to learn a stochastic policy. We present theoretical results for policy evaluation and improvement within this new framework. Additionally, we propose two novel algorithms, $\alpha$-SAC and $\alpha$-SQL, for large-scale RL tasks. We show that these algorithms attain higher returns on games from the Atari suite relative to an entropy-regularised benchmark, SAC-Discrete.
Keywords: Reinforcement Learning, Variational Inference, Rényi divergence
TL;DR: We generalise soft-actor critic using Rényi divergences.
Abstract:
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: 10959
Loading