Insights From the NeurIPS 2021 NetHack Challenge

TL;DR

The NeurIPS 2021 NetHack Challenge highlighted advances in AI with diverse approaches, revealed the superiority of symbolic AI over deep RL in NetHack, and underscored the game's value as a long-term benchmark.

Contribution

This report summarizes the challenge outcomes, compares neural and symbolic AI approaches, and emphasizes NetHack's role as a complex benchmark for AI research.

Findings

Symbolic AI outperforms deep RL in NetHack.

No agent nearly achieved winning the game.

The challenge fostered diverse AI approaches and progress.

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.