Facebook Aims to Develop AI Capable of Mastering the 'Hardest Game in the World'

Artificial Intelligence Challenges in Gaming: The NetHack Initiative

Real-time strategy, role-playing, and building games have become essential training grounds for cutting-edge artificial intelligence (AI) in recent years. Digital competitors now routinely surpass human opponents in iconic titles like StarCraft II, Dota 2, Minecraft, and Go. In this context, Facebook is reaching out to the AI community for assistance in tackling NetHack—one of the most challenging games in history. The goal is to enhance computers' ability to simulate game instances more efficiently while consuming fewer resources.

NetHack is a tactical rogue-like dungeon crawler that expects players to fail—often and spectacularly. Developed in the 1980s and still actively updated, the game resets the entire dungeon each time a player dies. To conquer NetHack, players must blend luck, creative problem-solving, and diligent research using resources like the NetHack Wiki to learn from the experiences of past adventurers.

As part of the NeurIPS 2021 NetHack Challenge, Facebook invites researchers to create and train AI systems capable of developing agents that either complete the game or secure the highest possible scores. According to a recent blog post, this endeavor aims to showcase the NetHack Learning Environment as a robust reinforcement learning framework while exploring various AI and machine learning solutions based on neural and symbolic methodologies.

The challenge requires participants to design agents that will navigate multiple games, each featuring randomly assigned character roles and fantasy races. For each evaluation episode, the average number of times an agent completes the game will be calculated, alongside the median end-of-episode score. Entries will be ranked primarily by the average number of wins and secondarily by median score in case of ties.

This competition runs through October 15th, with winners to be announced at NeurIPS in December.