Unsupervised State Representation Learning in Atari

Kendrick Tan, Lawrence Huang, Shekar Ramaswamy, Tyler Jiang

Abstract:

In this study, we reproduced the paper "Unsupervised Representation Learning in Atari" as part of the NeurIPS 2019 Reproducibility Challenge.

The original paper presents ST-DIM, or Spatiotemporal Deep Infomax, which is an unsupervised method for learning useful state representations. ST-DIM attempts to learn an encoding that maximizes shared information between sequential game frames. The authors propose that this method allows for more effective learning of meaningful environmental features, such as the game state variables of Atari games.

We review the authors' goals, claims, and results. We also discuss our process, challenges encountered, and results of our work.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=HkelAVBeIr