Facebook wants an AI to overcome, alone, the most difficult video game in the world
Artificial Intelligence (AI) experts have used this technology for various purposes. From being able to predict natural disasters or diagnose diseases more quickly, to pilot fighter jets or develop levels of play. The applications are very broad and now Facebook is making a new proposal, it wants to make a machine capable, autonomously, of overcoming one of the most difficult video games in the world: NetHack.
According to information from the Engadget medium, real-time strategy, role-playing and construction games have become, in recent years, a fertile training ground for cutting-edge artificial intelligence systems. So far the developers of these digital competitors have managed to easily outperform their human opponents in titles ranging from StarCraft II and Dota 2 to Minecraft and Go. But, as in video games, the level of difficulty has to increase.
Facebook is asking for the help of the community of specialists in artificial intelligence to develop a system capable of surpassing the game NetHack, one of the most difficult titles in the history of games.
It should be noted that NetHack is a complicated tactical game that takes the user through dungeons and other scenarios. Although it was originally developed in the 1980s, it has not stopped receiving updates to date. What makes it so difficult is that it is programmed so that you do not win and it takes you through all the necessary circumstances waiting for you to die and, for sure you will.
One of the frustrating elements of the title is that once the player dies, the entire dungeon is fully reset. The only way to advance in the game lies in the ability to combine luck, problem-solving, and investigative skills on the NetHack Wiki to learn from the explorer misfortunes that have come before.
A new training system
Beyond how interesting the idea of testing an AI in this title may sound, the reality is that behind the experiment there is relevant research since specialists consider that the title can help computers learn to simulate many more solutions fast and using fewer resources.
Thus, according to a blog post and, as part of the NeurIPS 2021 NetHack Challenge, Facebook is inviting researchers to design, train and launch artificial intelligence systems capable of developing agents that can beat the game in a reliable, or at least in the most likely scenario, achieve as high a score as possible.
With this initiative, Facebook not only hopes to be able to better understand the NetHack learning environment but also to make an AI achieve what they call viable reinforcement learning, that is, the technology is able to devise a range of possible solutions based on neural methodologies and symbolic.
Facebook said, candidates, will play a series of games, each with a randomly drawn fantasy character and career. He explained that to evaluate the results of the systems, the average number of episodes in which the AI completes the game will be calculated together with the average score at the end of the episode in the title. This is how they will rank success in each case based on the average number of wins and, if tied, based on the average score.
The competition runs from this month through October 15, and the winners are announced on NeurIPS in December. If you want to know more about this initiative, click here.