A Connection between One-Step RL and Critic Regularization in Reinforcement Learning

Benjamin Eysenbach, Matthieu Geist, Sergey Levine, Ruslan Salakhutdinov
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:9485-9507, 2023.

Abstract

As with any machine learning problem with limited data, effective offline RL algorithms require careful regularization to avoid overfitting. One class of methods, known as one-step RL, perform just one step of policy improvement. These methods, which include advantage-weighted regression and conditional behavioral cloning, are thus simple and stable, but can have limited asymptotic performance. A second class of methods, known as critic regularization, perform many steps of policy improvement with a regularized objective. These methods typically require more compute but have appealing lower-bound guarantees. In this paper, we draw a connection between these methods: applying a multi-step critic regularization method with a regularization coefficient of 1 yields the same policy as one-step RL. While our theoretical results require assumptions (e.g., deterministic dynamics), our experiments nevertheless show that our analysis makes accurate, testable predictions about practical offline RL methods (CQL and one-step RL) with commonly-used hyperparameters.

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-eysenbach23a, title = {A Connection between One-Step {RL} and Critic Regularization in Reinforcement Learning}, author = {Eysenbach, Benjamin and Geist, Matthieu and Levine, Sergey and Salakhutdinov, Ruslan}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {9485--9507}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/eysenbach23a/eysenbach23a.pdf}, url = {https://proceedings.mlr.press/v202/eysenbach23a.html}, abstract = {As with any machine learning problem with limited data, effective offline RL algorithms require careful regularization to avoid overfitting. One class of methods, known as one-step RL, perform just one step of policy improvement. These methods, which include advantage-weighted regression and conditional behavioral cloning, are thus simple and stable, but can have limited asymptotic performance. A second class of methods, known as critic regularization, perform many steps of policy improvement with a regularized objective. These methods typically require more compute but have appealing lower-bound guarantees. In this paper, we draw a connection between these methods: applying a multi-step critic regularization method with a regularization coefficient of 1 yields the same policy as one-step RL. While our theoretical results require assumptions (e.g., deterministic dynamics), our experiments nevertheless show that our analysis makes accurate, testable predictions about practical offline RL methods (CQL and one-step RL) with commonly-used hyperparameters.} }
Endnote
%0 Conference Paper %T A Connection between One-Step RL and Critic Regularization in Reinforcement Learning %A Benjamin Eysenbach %A Matthieu Geist %A Sergey Levine %A Ruslan Salakhutdinov %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-eysenbach23a %I PMLR %P 9485--9507 %U https://proceedings.mlr.press/v202/eysenbach23a.html %V 202 %X As with any machine learning problem with limited data, effective offline RL algorithms require careful regularization to avoid overfitting. One class of methods, known as one-step RL, perform just one step of policy improvement. These methods, which include advantage-weighted regression and conditional behavioral cloning, are thus simple and stable, but can have limited asymptotic performance. A second class of methods, known as critic regularization, perform many steps of policy improvement with a regularized objective. These methods typically require more compute but have appealing lower-bound guarantees. In this paper, we draw a connection between these methods: applying a multi-step critic regularization method with a regularization coefficient of 1 yields the same policy as one-step RL. While our theoretical results require assumptions (e.g., deterministic dynamics), our experiments nevertheless show that our analysis makes accurate, testable predictions about practical offline RL methods (CQL and one-step RL) with commonly-used hyperparameters.
APA
Eysenbach, B., Geist, M., Levine, S. & Salakhutdinov, R.. (2023). A Connection between One-Step RL and Critic Regularization in Reinforcement Learning. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:9485-9507 Available from https://proceedings.mlr.press/v202/eysenbach23a.html.

Related Material