DoMo-AC: Doubly Multi-step Off-policy Actor-Critic Algorithm

Yunhao Tang, Tadashi Kozuno, Mark Rowland, Anna Harutyunyan, Remi Munos, Bernardo Avila Pires, Michal Valko
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:33657-33673, 2023.

Abstract

Multi-step learning applies lookahead over multiple time steps and has proved valuable in policy evaluation settings. However, in the optimal control case, the impact of multi-step learning has been relatively limited despite a number of prior efforts. Fundamentally, this might be because multi-step policy improvements require operations that cannot be approximated by stochastic samples, hence hindering the widespread adoption of such methods in practice. To address such limitations, we introduce doubly multi-step off-policy VI (DoMo-VI), a novel oracle algorithm that combines multi-step policy improvements and policy evaluations. DoMo-VI enjoys guaranteed convergence speed-up to the optimal policy and is applicable in general off-policy learning settings. We then propose doubly multi-step off-policy actor-critic (DoMo-AC), a practical instantiation of the DoMo-VI algorithm. DoMo-AC introduces a bias-variance trade-off that ensures improved policy gradient estimates. When combined with the IMPALA architecture, DoMo-AC has showed improvements over the baseline algorithm on Atari-57 game benchmarks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-tang23e, title = {{D}o{M}o-{AC}: Doubly Multi-step Off-policy Actor-Critic Algorithm}, author = {Tang, Yunhao and Kozuno, Tadashi and Rowland, Mark and Harutyunyan, Anna and Munos, Remi and Avila Pires, Bernardo and Valko, Michal}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {33657--33673}, 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/tang23e/tang23e.pdf}, url = {https://proceedings.mlr.press/v202/tang23e.html}, abstract = {Multi-step learning applies lookahead over multiple time steps and has proved valuable in policy evaluation settings. However, in the optimal control case, the impact of multi-step learning has been relatively limited despite a number of prior efforts. Fundamentally, this might be because multi-step policy improvements require operations that cannot be approximated by stochastic samples, hence hindering the widespread adoption of such methods in practice. To address such limitations, we introduce doubly multi-step off-policy VI (DoMo-VI), a novel oracle algorithm that combines multi-step policy improvements and policy evaluations. DoMo-VI enjoys guaranteed convergence speed-up to the optimal policy and is applicable in general off-policy learning settings. We then propose doubly multi-step off-policy actor-critic (DoMo-AC), a practical instantiation of the DoMo-VI algorithm. DoMo-AC introduces a bias-variance trade-off that ensures improved policy gradient estimates. When combined with the IMPALA architecture, DoMo-AC has showed improvements over the baseline algorithm on Atari-57 game benchmarks.} }
Endnote
%0 Conference Paper %T DoMo-AC: Doubly Multi-step Off-policy Actor-Critic Algorithm %A Yunhao Tang %A Tadashi Kozuno %A Mark Rowland %A Anna Harutyunyan %A Remi Munos %A Bernardo Avila Pires %A Michal Valko %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-tang23e %I PMLR %P 33657--33673 %U https://proceedings.mlr.press/v202/tang23e.html %V 202 %X Multi-step learning applies lookahead over multiple time steps and has proved valuable in policy evaluation settings. However, in the optimal control case, the impact of multi-step learning has been relatively limited despite a number of prior efforts. Fundamentally, this might be because multi-step policy improvements require operations that cannot be approximated by stochastic samples, hence hindering the widespread adoption of such methods in practice. To address such limitations, we introduce doubly multi-step off-policy VI (DoMo-VI), a novel oracle algorithm that combines multi-step policy improvements and policy evaluations. DoMo-VI enjoys guaranteed convergence speed-up to the optimal policy and is applicable in general off-policy learning settings. We then propose doubly multi-step off-policy actor-critic (DoMo-AC), a practical instantiation of the DoMo-VI algorithm. DoMo-AC introduces a bias-variance trade-off that ensures improved policy gradient estimates. When combined with the IMPALA architecture, DoMo-AC has showed improvements over the baseline algorithm on Atari-57 game benchmarks.
APA
Tang, Y., Kozuno, T., Rowland, M., Harutyunyan, A., Munos, R., Avila Pires, B. & Valko, M.. (2023). DoMo-AC: Doubly Multi-step Off-policy Actor-Critic Algorithm. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:33657-33673 Available from https://proceedings.mlr.press/v202/tang23e.html.

Related Material