🥁NetHack Challenge at NeurIPS 2021 from Facebook: open-source project for Reinforcement Learning (RL) as a game
Many advances in RL have been achieved through simulation environments in games such as Dota 2, Minecraft, and StarCraft II. But this requires a lot of computation on thousands of GPUs at a time for just one experiment. To reduce the cost of RL modeling, Facebook in 2020 initiated the development of the open-source NetHack Learning Environment project. And in 2021, Facebook announced the NeurIPS 2021 competition as part of the NetHack Challenge in conjunction with AIcrowd, an AI crowdsourcing organization. The competition runs from early June to October 15, 2021, and the winners will be announced on NeurIPS in December.
The NetHack game has actually existed since the 1980s. It is visually straightforward and completely free to play, but more complicated than StarCraft II due to the very confusing interaction of players with their environment and related objects, users have to think outside the box or refer to the NetHack Wiki. And the main difficulty of NetHack is that after the death of a character, the game session of this player ends. Therefore, within this RL environment, researchers hope to find new ways to control agents so that in the future, AI can think creatively in difficult situations, helping people. Because NetHack runs on a terminal, players can quickly simulate gameplay by training billions of agents a day on just 2 GPUs. This is how the NetHack Challenge tests the latest AI techniques in a complex environment without the enormous power of a supercomputer. https://techxplore.com/news/2021-06-facebook-nethack-neurips.html