Representations for Stable Off-Policy Reinforcement Learning

Dibya Ghosh, Marc G. Bellemare
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:3556-3565, 2020.

Abstract

Reinforcement learning with function approximation can be unstable and even divergent, especially when combined with off-policy learning and Bellman updates. In deep reinforcement learning, these issues have been dealt with empirically by adapting and regularizing the representation, in particular with auxiliary tasks. This suggests that representation learning may provide a means to guarantee stability. In this paper, we formally show that there are indeed nontrivial state representations under which the canonical SARSA algorithm is stable, even when learning off-policy. We analyze representation learning schemes that are based on the transition matrix of a policy, such as proto-value functions, along three axes: approximation error, stability, and ease of estimation. In the most general case of a defective transition matrix, we show that a Schur basis provides convergence guarantees, but is difficult to estimate from samples. For a fixed reward function, we find that an orthogonal basis of the corresponding Krylov subspace is an even better choice. We conclude by empirically demonstrating that these stable representations can be learned using stochastic gradient descent, opening the door to improved techniques for representation learning with deep networks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v119-ghosh20b, title = {Representations for Stable Off-Policy Reinforcement Learning}, author = {Ghosh, Dibya and Bellemare, Marc G.}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {3556--3565}, year = {2020}, editor = {III, Hal Daumé and Singh, Aarti}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/ghosh20b/ghosh20b.pdf}, url = {https://proceedings.mlr.press/v119/ghosh20b.html}, abstract = {Reinforcement learning with function approximation can be unstable and even divergent, especially when combined with off-policy learning and Bellman updates. In deep reinforcement learning, these issues have been dealt with empirically by adapting and regularizing the representation, in particular with auxiliary tasks. This suggests that representation learning may provide a means to guarantee stability. In this paper, we formally show that there are indeed nontrivial state representations under which the canonical SARSA algorithm is stable, even when learning off-policy. We analyze representation learning schemes that are based on the transition matrix of a policy, such as proto-value functions, along three axes: approximation error, stability, and ease of estimation. In the most general case of a defective transition matrix, we show that a Schur basis provides convergence guarantees, but is difficult to estimate from samples. For a fixed reward function, we find that an orthogonal basis of the corresponding Krylov subspace is an even better choice. We conclude by empirically demonstrating that these stable representations can be learned using stochastic gradient descent, opening the door to improved techniques for representation learning with deep networks.} }
Endnote
%0 Conference Paper %T Representations for Stable Off-Policy Reinforcement Learning %A Dibya Ghosh %A Marc G. Bellemare %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-ghosh20b %I PMLR %P 3556--3565 %U https://proceedings.mlr.press/v119/ghosh20b.html %V 119 %X Reinforcement learning with function approximation can be unstable and even divergent, especially when combined with off-policy learning and Bellman updates. In deep reinforcement learning, these issues have been dealt with empirically by adapting and regularizing the representation, in particular with auxiliary tasks. This suggests that representation learning may provide a means to guarantee stability. In this paper, we formally show that there are indeed nontrivial state representations under which the canonical SARSA algorithm is stable, even when learning off-policy. We analyze representation learning schemes that are based on the transition matrix of a policy, such as proto-value functions, along three axes: approximation error, stability, and ease of estimation. In the most general case of a defective transition matrix, we show that a Schur basis provides convergence guarantees, but is difficult to estimate from samples. For a fixed reward function, we find that an orthogonal basis of the corresponding Krylov subspace is an even better choice. We conclude by empirically demonstrating that these stable representations can be learned using stochastic gradient descent, opening the door to improved techniques for representation learning with deep networks.
APA
Ghosh, D. & Bellemare, M.G.. (2020). Representations for Stable Off-Policy Reinforcement Learning. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:3556-3565 Available from https://proceedings.mlr.press/v119/ghosh20b.html.

Related Material