Google DeepMind releases source code to the Quake III levels its using to train AIs
Alphabet artificial intelligence wing Google DeepMind has been developing and utilising mazes and challenges, designed using video game Quake III Arena’s 17-year-old software, to teach its artificial intelligence programmes how to operate in 3D spaces.
Somewhat reminiscent of the rat-in-a-maze experiment, the ‘DeepMind Lab’ project was used to create enviroments capable of testing AI systems’ “ability to achieve goals in a wide range of environments”.
Tasks such as navigation in mazes, collecting fruit, traversing dangerous passages, laser tag and interaction with bots have been developed to refine the programmes – and as an additional aid to external developers, the group has made the source code available on GitHub.
The levels will look familiar to Quake players, from everything encompassing the aesthetic to how the physics engine operates. The first-person perspective has been deemed ideal in teaching systems how to navigate space in the way that humans and animals would. These findings could in turn be utilised to help people navigate spaces in more sophisticated ways, in addition to numerous more applications.
One of Deepmind’s founders, Shane Legg told Wired: “It's up to [AIs] to learn how to behave in this environment and solve the problems we can give to them.
"We live in a 3D environment, if we want artificial intelligence systems that can perceive the real-world; to move around in that environment; interact with it; and solve problems, this is a virtual proxy to that."
Bloomberg reports that Legg also issued an old-school, high score challenge to other programmers to see which systems perform best on each level.