Timezone: »

 
Poster
Stabilizing Transformers for Reinforcement Learning
Emilio Parisotto · Francis Song · Jack Rae · Razvan Pascanu · Caglar Gulcehre · Siddhant Jayakumar · Max Jaderberg · Raphael Lopez Kaufman · Aidan Clark · Seb Noury · Matthew Botvinick · Nicolas Heess · Raia Hadsell

Wed Jul 15 05:00 AM -- 05:45 AM & Wed Jul 15 04:00 PM -- 04:45 PM (PDT) @ None #None
in Poster Session 21 »

Owing to their ability to both effectively integrate information over long time horizons and scale to massive amounts of data, self-attention architectures have recently shown breakthrough success in natural language processing (NLP). Harnessing the transformer’s ability to process long time horizons of information could provide a similar performance boost in partially observable reinforcement learning (RL) domains, but the large-scale transformers used in NLP have yet to be successfully applied to the RL setting. In this work we demonstrate that the standard transformer architecture is difficult to optimize, which was previously observed in the supervised learning setting but becomes especially pronounced with RL objectives. We propose architectural modifications that substantially improve the stability and learning speed of the original Transformer and XL variant. The proposed architecture, the Gated Transformer-XL (GTrXL), surpasses LSTMs on challenging memory environments and achieves state-of-the-art results on the multi-task DMLab-30 benchmark suite, exceeding the performance of an external memory architecture. We show that the GTrXL has stability and performance that consistently matches or exceeds a competitive LSTM baseline, including on more reactive tasks where memory is less critical.

Keywords: Deep Reinforcement Learning · Representation Learning · Time Series and Sequence Models · Transfer and Multitask Learning · Reinforcement Learning - Deep RL

Author Information

Emilio Parisotto (Carnegie Mellon University)
Francis Song (DeepMind)
Jack Rae (DeepMind)
Razvan Pascanu (DeepMind)
Caglar Gulcehre (DeepMind)
Siddhant Jayakumar (DeepMind)
Max Jaderberg (DeepMind)
Raphael Lopez Kaufman (Deepmind)
Aidan Clark (DeepMind)
Seb Noury (DeepMind)
Matthew Botvinick (DeepMind)
Nicolas Heess (DeepMind)
Raia Hadsell (DeepMind)

Raia Hadsell, a senior research scientist at DeepMind, has worked on deep learning and robotics problems for over 10 years. Her early research developed the notion of manifold learning using Siamese networks, which has been used extensively for invariant feature learning. After completing a PhD with Yann LeCun, which featured a self-supervised deep learning vision system for a mobile robot, her research continued at Carnegie Mellon’s Robotics Institute and SRI International, and in early 2014 she joined DeepMind in London to study artificial general intelligence. Her current research focuses on the challenge of continual learning for AI agents and robotic systems. While deep RL algorithms are capable of attaining superhuman performance on single tasks, they cannot transfer that performance to additional tasks, especially if experienced sequentially. She has proposed neural approaches such as policy distillation, progressive nets, and elastic weight consolidation to solve the problem of catastrophic forgetting and improve transfer learning.

More from the Same Authors