Pretraining Representations for Data-Efficient Reinforcement Learning
Abstract
Data efficiency is a key challenge for deep reinforcement learning. We address this problem by using unlabeled data to pretrain an encoder which is then finetuned on a small amount of task-specific data. To encourage learning representations which capture diverse aspects of the underlying MDP, we employ a combination of latent dynamics modelling and unsupervised goal-conditioned RL. When limited to 100k steps of interaction on Atari games (equivalent to two hours of human experience), our approach significantly surpasses prior work combining offline representation pretraining with task-specific finetuning, and compares favourably with other pretraining methods that require orders of magnitude more data. Our approach shows particular promise when combined with larger models as well as more diverse, task-aligned observational data -- approaching human-level performance and data-efficiency on Atari in our best setting.
Cite
Text
Schwarzer et al. "Pretraining Representations for Data-Efficient Reinforcement Learning." Neural Information Processing Systems, 2021.Markdown
[Schwarzer et al. "Pretraining Representations for Data-Efficient Reinforcement Learning." Neural Information Processing Systems, 2021.](https://mlanthology.org/neurips/2021/schwarzer2021neurips-pretraining/)BibTeX
@inproceedings{schwarzer2021neurips-pretraining,
title = {{Pretraining Representations for Data-Efficient Reinforcement Learning}},
author = {Schwarzer, Max and Rajkumar, Nitarshan and Noukhovitch, Michael and Anand, Ankesh and Charlin, Laurent and Hjelm, R Devon and Bachman, Philip and Courville, Aaron C.},
booktitle = {Neural Information Processing Systems},
year = {2021},
url = {https://mlanthology.org/neurips/2021/schwarzer2021neurips-pretraining/}
}