NetHack is Hard to Hack

TL;DR

This paper investigates why neural policy learning struggles in complex, long-horizon environments like NetHack, and develops a neural agent that significantly improves performance but still lags behind symbolic and human strategies.

Contribution

The paper provides an extensive analysis of neural policy learning in NetHack, introduces a large demonstration dataset, and achieves a new state-of-the-art neural agent performance.

Findings

Neural agents outperform previous policies by 127% offline and 25% online.

Symbolic agents outperform neural approaches by over four times in median score.

Scaling neural models alone is insufficient to match symbolic or human performance.

Abstract

Neural policy learning methods have achieved remarkable results in various control problems, ranging from Atari games to simulated locomotion. However, these methods struggle in long-horizon tasks, especially in open-ended environments with multi-modal observations, such as the popular dungeon-crawler game, NetHack. Intriguingly, the NeurIPS 2021 NetHack Challenge revealed that symbolic agents outperformed neural approaches by over four times in median game score. In this paper, we delve into the reasons behind this performance gap and present an extensive study on neural policy learning for NetHack. To conduct this study, we analyze the winning symbolic agent, extending its codebase to track internal strategy selection in order to generate one of the largest available demonstration datasets. Utilizing this dataset, we examine (i) the advantages of an action hierarchy; (ii) enhancements in neural architecture; and (iii) the integration of reinforcement learning with imitation learning. Our investigations produce a state-of-the-art neural agent that surpasses previous fully neural policies by 127% in offline settings and 25% in online settings on median game score. However, we also demonstrate that mere scaling is insufficient to bridge the performance gap with the best symbolic models or even the top human players.