Can Transformer Be Applied To Reinforcement Learning?

3 main points
✔️ Applying the transformer to Reinforcement Learning
✔️ GTrXL is proposed as a modified transformer to stabilize the learning process.
✔️ Performance and robustness exceeding LSTM

Stabilizing Transformers for Reinforcement Learning
written by Emilio Parisotto, H. Francis Song, Jack W. Rae, Razvan Pascanu, Caglar Gulcehre, Siddhant M. Jayakumar, Max Jaderberg, Raphael Lopez Kaufman, Aidan Clark, Seb Noury, Matthew M. Botvinick, Nicolas Heess, Raia Hadsell
(Submitted on 13 Oct 2019)
Comments: Accepted to ICML2020
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Paper Official Code COMM Code

Introduction

The transformers proposed at "Attention is all you need" have been very successful in various domains. In particular, they have a large presence in natural language processing, and their performance and growth rate are astonishing, especially in the area of prior learning models such as BERT, and especially in GPT-3, which has recently become a major topic of discussion. And this success is not limited to natural language processing. For example, its power has been demonstrated in the area of image processing, such as DETR for object detection and Image GPT for unsupervised representation learning. So, how many areas can we expect to see transformers applied to?　How versatile is it?
In this article, we present a paper that successfully applied the Transformer to reinforcement learning and brought out its capabilities.

anonymous