Video games offer up countless learning experiences for fundamental life skills needed both online and in the real world. From controlling your empire's resource management in Civilization VI to improving your situational awareness whilst driving through the streets of Los Santos in GTA V, a video game can often be a teacher. But what can artificial intelligence, rather than humans, learn from games?
Heaps, if you're Elon Musk's artificial intelligence consortium OpenAI. The group, which is dedicated to open source research into AI, has today released a new platform that broadens its goal of training AI with video games. It's called Universe, and is described as "a software platform for measuring and training an AI's general intelligence across the world's supply of games, websites and other applications."
The goal of Universe is actually fairly succinct: let AI programs play the world's library of video games so that is can use the experiences to master unfamiliar, new and difficult environments without specific training. The goal is to move beyond "narrow AI," programs that are good at very specific tasks, but lack more generalized skills and learning capabilities. One example OpenAI cites is that of Google DeepMind's AlphaGo: it can easily beat you at Go, but can it explain to you the rules of chess?
"Universe allows an AI agent to use a computer like a human does: by looking at screen pixels and operating a virtual keyboard and mouse. We must train AI systems on the full range of tasks we expect them to solve, and Universe lets us train a single agent on any task a human can complete with a computer," reads OpenAI's blog post about Universe.
Universe has already won support from EA, Microsoft Studios, Valve, and a bunch of other game studios. AI researchers can let their AI's loose in games including Red Alert 2, Portal, World of Goo, RimWorld, as well as Atari 2600 games and 1,000 browser-based Flash games.
"Systems with general problem solving ability—something akin to human common sense, allowing an agent to rapidly solve a new hard task—remain out of reach. One apparent challenge is that our agents don't carry their experience along with them to new tasks," said OpenAI. Universe could change this.
The initiative is also letting AI agents complete form-filling tasks using interface elements like buttons, lists, and sliders. Hopefully, reckons OpenAI, in the future AI will carry out more complex tasks such as searching for answers to questions on the internet, managing email accounts, and working on Amazon Mechanical Turk tasks.
One aspect that OpenAI doesn't touch on however is the question of morality. We've already seen how AI can be negatively influenced by humans, and letting an AI loose on video games with specific goals of killing and breaking the law surely may concern some human observers (especially since violent video games are still mildly controversial in real life). Is it really wise to teach AI with video games that involve violence?
OpenAI's communications director Jack Clark told Motherboard in an email that OpenAI doesn't condone violence, and it doesn't plan on training its AIs to be violent, either. "Games like GTA V give us access to rich, dynamic, surprising worlds, and will be useful in developing AIs with better driving/motor-control abilities," he said. Clark pointed Motherboard to previous research that concluded games such as GTA V are able to act as driving simulators for artificial intelligence, as reported in September by Motherboard.
For now though, the group is calling on contributors to send in their demonstrations of AIs in Universe environments, as well as asking developers to grant OpenAI permission to use their games, apps, and websites for the purpose of teaching artificial intelligence.
Get six of our favorite Motherboard stories every day by signing up for our newsletter.