Facebook Wants To Leverage AI To Beat Notoriously Difficult 1980s Game NetHack
Games are one of the best testbeds for AI as they require problem-solving, forward-thinking, and other skills that normally only humans possess. So far, AI has gotten a handle on games like Go and up to 57 different Atari 2600 titles, but those are not terribly difficult. What if AI was pitted against one of the most notoriously hard games of all time? As it turns out, that is exactly what Facebook wants to do to bring down NetHack.
NetHack is an 80s text-based single-player dungeon exploration game with a focus on “discovering the detail of the dungeon and not simply killing everything in sight,” as the NetHack website explains. If you decide to kill everything in sight, or at least try to, you will end up dead rather quickly. In any case, though the game and website are dated, as evidenced by the simple design and Y2K statement that still resides at the bottom of the site, it is still actively updated and played by many hardcore fans (see masochists) around the world. Moreover, it is still regarded as an incredibly difficult game to master, all things considered.
Thanks to these attributes, the game could make for a great reinforcement learning tool for AI while having an incredibly low computational overhead. As such, Facebook open-sourced its NetHack Learning Environment last year and has launched a competition called NeurIPS 2021 to create an agent that can reliably beat or achieve a high score in the game. As the Facebook blog post states, “making progress in NetHack is making progress toward RL in a wider range of applications.”
As the AI field advances, so too does the need for better and more complex training methods, which will further spur progress. While beating NetHack with AI is only a small step in the grand scheme of things, it is an incredibly interesting development that we cannot wait to see where it ends up. In any case, let us know what you think of this venture and NetHack's difficulty in the comments below.
(NetHack Image Courtesy Of Ars Technica)