Pretraining Representations for Data-Efficient Reinforcement Learning
This addresses data efficiency for reinforcement learning practitioners, offering an incremental improvement over existing pretraining methods.
The paper tackles data inefficiency in deep reinforcement learning by pretraining an encoder with unlabeled data and finetuning it on limited task-specific data, achieving significant improvements over prior methods with only 100k interaction steps on Atari games and approaching human-level performance in best settings.
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. We provide code associated with this work at https://github.com/mila-iqia/SGI.