Real time strategy, role playing and building games have become fertile training grounds for cutting edge artificial intelligence systems in recent years with digital competitors easily besting their human opponents in everything from StarCraft II and Dota 2 to Minecraft and Go. Now, Facebook is asking for the AI community’s help in bringing down NetHack — one of the most notoriously difficult titles in gaming history — and maybe help computers learn to simulate instances faster using fewer resources.
NetHack is a tactical curb stomping that passes for a rogue-like dungeon crawler. Originally developed in the 1980s but still actively updated today, the game doesn’t expect you to win — it expects you to die. And die you will. In bunches. And every time the player perishes, the entire dungeon resets in its entirety. The only way to actually best the game lies in your ability to combine luck, outside-the-box problem solving, and old fashioned research skills at the NetHack Wiki to learn from the misfortunes of explorers who have come before you.
As part of the NeurIPS 2021 NetHack Challenge, Facebook is inviting researchers to design, train and release AI systems able to “develop agents that can reliably either beat the game or (in the more likely scenario) achieve as high a score as possible,” according to a Wednesday FB blog post. In doing so, Facebook hopes that not only will this showcase the NetHack Learning Environment as a viable reinforcement learning system but also enable a range of potential AI/ML solutions based in both neural and symbolic methodologies.
“The candidate agents will play a number of games, each with a randomly drawn character role and fantasy race,” the Wednesday post explained. “For a given set of evaluation episodes for an agent, the average number of episodes where the agent completes the game will be computed, along with the median in-game end-of-episode score. Entries will be ranked by average number of wins and, if tied, by median score.”
The competition runs from this month through October 15th with winners announced at NeurIPS in December.