Insights From the NeurIPS 2021 NetHack Challenge

Eric Hambro, Sharada Mohanty, Dmitrii Babaev, Minwoo Byeon, Dipam Chakraborty, Edward Grefenstette, Minqi Jiang, Jo Daejin, Anssi Kanervisto, Jongmin Kim, Sungwoong Kim, Robert Kirk, Vitaly Kurin, Heinrich Küttler, Taehwon Kwon, Donghoon Lee, Vegard Mella, Nantas Nardelli, Ivan Nazarov, Nikita Ovsov, Jack Holder, Roberta Raileanu, Karolis Ramanauskas, Tim Rocktäschel, Danielle Rothermel, Mikayel Samvelyan, Dmitry Sorokin, Maciej Sypetkowski, Michał Sypetkowski
Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track, PMLR 176:41-52, 2022.

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.

Cite this Paper


BibTeX
@InProceedings{pmlr-v176-hambro22a, title = {Insights From the NeurIPS 2021 NetHack Challenge}, author = {Hambro, Eric and Mohanty, Sharada and Babaev, Dmitrii and Byeon, Minwoo and Chakraborty, Dipam and Grefenstette, Edward and Jiang, Minqi and Daejin, Jo and Kanervisto, Anssi and Kim, Jongmin and Kim, Sungwoong and Kirk, Robert and Kurin, Vitaly and K{\"u}ttler, Heinrich and Kwon, Taehwon and Lee, Donghoon and Mella, Vegard and Nardelli, Nantas and Nazarov, Ivan and Ovsov, Nikita and Holder, Jack and Raileanu, Roberta and Ramanauskas, Karolis and Rockt{\"a}schel, Tim and Rothermel, Danielle and Samvelyan, Mikayel and Sorokin, Dmitry and Sypetkowski, Maciej and Sypetkowski, Micha\l{}}, booktitle = {Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track}, pages = {41--52}, year = {2022}, editor = {Kiela, Douwe and Ciccone, Marco and Caputo, Barbara}, volume = {176}, series = {Proceedings of Machine Learning Research}, month = {06--14 Dec}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v176/hambro22a/hambro22a.pdf}, url = {https://proceedings.mlr.press/v176/hambro22a.html}, 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.} }
Endnote
%0 Conference Paper %T Insights From the NeurIPS 2021 NetHack Challenge %A Eric Hambro %A Sharada Mohanty %A Dmitrii Babaev %A Minwoo Byeon %A Dipam Chakraborty %A Edward Grefenstette %A Minqi Jiang %A Jo Daejin %A Anssi Kanervisto %A Jongmin Kim %A Sungwoong Kim %A Robert Kirk %A Vitaly Kurin %A Heinrich Küttler %A Taehwon Kwon %A Donghoon Lee %A Vegard Mella %A Nantas Nardelli %A Ivan Nazarov %A Nikita Ovsov %A Jack Holder %A Roberta Raileanu %A Karolis Ramanauskas %A Tim Rocktäschel %A Danielle Rothermel %A Mikayel Samvelyan %A Dmitry Sorokin %A Maciej Sypetkowski %A Michał Sypetkowski %B Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track %C Proceedings of Machine Learning Research %D 2022 %E Douwe Kiela %E Marco Ciccone %E Barbara Caputo %F pmlr-v176-hambro22a %I PMLR %P 41--52 %U https://proceedings.mlr.press/v176/hambro22a.html %V 176 %X 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.
APA
Hambro, E., Mohanty, S., Babaev, D., Byeon, M., Chakraborty, D., Grefenstette, E., Jiang, M., Daejin, J., Kanervisto, A., Kim, J., Kim, S., Kirk, R., Kurin, V., Küttler, H., Kwon, T., Lee, D., Mella, V., Nardelli, N., Nazarov, I., Ovsov, N., Holder, J., Raileanu, R., Ramanauskas, K., Rocktäschel, T., Rothermel, D., Samvelyan, M., Sorokin, D., Sypetkowski, M. & Sypetkowski, M.. (2022). Insights From the NeurIPS 2021 NetHack Challenge. Proceedings of the NeurIPS 2021 Competitions and Demonstrations Track, in Proceedings of Machine Learning Research 176:41-52 Available from https://proceedings.mlr.press/v176/hambro22a.html.

Related Material