NetHack may not be the most well-known video game, but the open-source, ASCII-based dungeon-crawler remains one of the most beloved (and brutal) titles in gaming history. Released in 1987, the RPG has since garnered a dedicated, diehard community of fans who live and digitally die — quite often — by the roguelike game’s difficulty, nuance, and bone-dry sense of humor. It’s one of the rare examples of pure internet joy... and because of this, Facebook must obviously taint it with a machine learning competition that will probably one day lead to humanity’s enslavement by an army of murderous, self-aware ZuckerBots.
Okay, maybe not, but for this year’s NeurIPS conference, Facebook announced an open-call NetHack Challenge utilizing its 2020 NetHack Learning Environment to highlight artificial intelligence/machine learning programs with the best completion rates and high scores.
Entrants are given free range on how they build and train their agents “with or without machine learning, using any external information you’d like, and with any training method and computational budget,” according to the competition’s partner and co-organizer, AIcrowd. Winners will be crowned by the number of successful games, or the best median score if no AI beats NetHack — an outcome that may actually be a procedurally generated possibility, given the game’s notorious difficulty.
Our fingers are crossed for NetHack to do what it does best: frustrate the living hell out of its players, be them human or AI... if only to stave off our impending robot apocalypse by a few more years.
Computational costs — Developers have achieved advancements in reinforcement learning (RL) for years by running their programs through video game simulations like StarCraft II and Minecraft, but NetHack is inadvertently a perfect venue to quickly and efficiently train AI decision-making. In these scenarios, “progress came at substantial computational costs, often requiring running thousands of GPUs in parallel for a single experiment, while also falling short of leading to RL methods that can be transferred to more real-world problems outside of these games,” according to Facebook’s blog posted earlier today.
NetHack’s computing simplicity is key — But given that NetHack was primarily written using the C programming language and uses ASCII artwork instead of pixel-based imagery, AI models can learn extremely quickly within the game’s environment without expending this unnecessary computational power. Instead of those “thousands of GPUs,” agents can train over 1.2 billion steps a day using only a couple GPUs. Because of this, NetHack is a perfect venue to better train AI and ML programs, as well as opens up the playing field (no pun intended) to developers on smaller budgets and hardware capabilities.
A showdown between neural and symbolic methods — Although neural networks are generally the preferred means to build modern reinforcement learning agents, the NetHack Challenge is open to anyone, including bot developers using symbolic methods for a “direct comparison with neural agents.” The NetHack Challenge will run from early June until October 15, with winners announced during the NeurIPS Conference in December.