Insights From the NeurIPS 2021 NetHack Challenge
Abstract
In this report, we summarize the takeaways from the first NeurIPS 2021 NetHack Challenge. Participants were tasked with developing a program or agent that can win (i.e., 'ascend' in) the popular dungeon-crawler game of NetHack by interacting with the NetHack Learning Environment (NLE), a scalable, procedurally generated, and challenging Gym environment for reinforcement learning (RL). The challenge showcased community-driven progress in AI with many diverse approaches significantly beating the previously best results on NetHack. Furthermore, it served as a direct comparison between neural (e.g., deep RL) and symbolic AI, as well as hybrid systems, demonstrating that on NetHack symbolic bots currently outperform deep RL by a large margin. Lastly, no agent got close to winning the game, illustrating NetHack's suitability as a long-term benchmark for AI research.
Keywords
Cite
@article{arxiv.2203.11889,
title = {Insights From the NeurIPS 2021 NetHack Challenge},
author = {Eric Hambro and Sharada Mohanty and Dmitrii Babaev and Minwoo Byeon and Dipam Chakraborty and Edward Grefenstette and Minqi Jiang and Daejin Jo and Anssi Kanervisto and Jongmin Kim and Sungwoong Kim and Robert Kirk and Vitaly Kurin and Heinrich Küttler and Taehwon Kwon and Donghoon Lee and Vegard Mella and Nantas Nardelli and Ivan Nazarov and Nikita Ovsov and Jack Parker-Holder and Roberta Raileanu and Karolis Ramanauskas and Tim Rocktäschel and Danielle Rothermel and Mikayel Samvelyan and Dmitry Sorokin and Maciej Sypetkowski and Michał Sypetkowski},
journal= {arXiv preprint arXiv:2203.11889},
year = {2022}
}
Comments
Under review at PMLR for the NeuRIPS 2021 Competition Workshop Track, 10 pages + 10 in appendices