In brief
What the paper does.
A review of deep reinforcement learning on the Atari benchmark. It argues that the usual human-normalized criteria understate what humans can do, proposes a benchmark based on human world records that asks for both final performance and learning efficiency, reports state-of-the-art methods on the new metrics, and identifies the open challenges that keep agents from superhuman play.
Cite
Cite the paper.
@inproceedings{fan2022atarireview,
title={A Review for Deep Reinforcement Learning in Atari: Benchmarks, Challenges, and Solutions},
author={Fan, Jiajun},
booktitle={AAAI-22 Workshop on Reinforcement Learning in Games},
year={2022}
}