Decoupling Representation Learning from Reinforcement Learning

Adam Stooke, Kimin Lee, Pieter Abbeel, Michael Laskin
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:9870-9879, 2021.

Abstract

In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at \url{https://github.com/astooke/rlpyt/tree/master/rlpyt/ul}.

Cite this Paper


BibTeX
@InProceedings{pmlr-v139-stooke21a, title = {Decoupling Representation Learning from Reinforcement Learning}, author = {Stooke, Adam and Lee, Kimin and Abbeel, Pieter and Laskin, Michael}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {9870--9879}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/stooke21a/stooke21a.pdf}, url = {https://proceedings.mlr.press/v139/stooke21a.html}, abstract = {In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at \url{https://github.com/astooke/rlpyt/tree/master/rlpyt/ul}.} }
Endnote
%0 Conference Paper %T Decoupling Representation Learning from Reinforcement Learning %A Adam Stooke %A Kimin Lee %A Pieter Abbeel %A Michael Laskin %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-stooke21a %I PMLR %P 9870--9879 %U https://proceedings.mlr.press/v139/stooke21a.html %V 139 %X In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at \url{https://github.com/astooke/rlpyt/tree/master/rlpyt/ul}.
APA
Stooke, A., Lee, K., Abbeel, P. & Laskin, M.. (2021). Decoupling Representation Learning from Reinforcement Learning. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:9870-9879 Available from https://proceedings.mlr.press/v139/stooke21a.html.

Related Material