Timezone: »

 
CoBERL: Contrastive BERT for Reinforcement Learning
Andrea Banino · Adrià Puigdomenech Badia · Jacob C Walker · Tim Scholtes · Jovana Mitrovic · Charles Blundell

@
in ICML 2021 Workshop on Unsupervised Reinforcement Learning »

Many reinforcement learning (RL) agents require a large amount of experience to solve tasks. We propose Contrastive BERT for RL (CoBERL), an agent that combines a new contrastive loss and a hybrid LSTM-transformer architecture to tackle the challenge of improving data efficiency. CoBERL enables efficient, robust learning from pixels across a wide range of domains. We use bidirectional masked prediction in combination with a generalization of recent contrastive methods to learn better representations for transformers in RL, without the need of hand engineered data augmentations. We find that CoBERL consistently improves performance across the full Atari suite, a set of control tasks and a challenging 3D environment.

Author Information

Andrea Banino (DeepMind)
Adrià Puigdomenech Badia (Deepmind)
Jacob C Walker (Carnegie Mellon University)
Tim Scholtes (DeepMind)
Jovana Mitrovic (DeepMind)
Charles Blundell (DeepMind)

More from the Same Authors