NetHack is Hard to Hack
This addresses the problem of improving neural agents for complex, multi-modal tasks like NetHack, but it is incremental as it builds on existing methods and does not fully close the performance gap.
The paper tackled the challenge of neural policy learning in long-horizon, open-ended environments like NetHack, where symbolic agents had previously outperformed neural ones by over four times in median game score; the result was a state-of-the-art neural agent that improved over previous neural policies by 127% offline and 25% online on median game score, but scaling alone was insufficient to match symbolic models or top human players.
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.