OpenAI, an artificial intelligence research company, wants to let AIs loose in their own universe, where they can learn to play games, use apps and interact with websites.
Universe is the name of OpenAI’s tool for training AIs on, as it puts it, “any task a human can complete with a computer.” Using a VNC (Virtual Network Computing) remote desktop, it allows the AI to control the game or app using a virtual keyboard and mouse, and to see its output by analyzing the pixels displayed on the screen. It’s essentially an interface to the company’s Gym toolkit for developing reinforcement algorithms, a type of machine learning system.
The source code for Universe posted to Github on Monday, and includes interfaces to a thousand online environments, the company said.
Among them are simple browser tasks such as clicking buttons, or copying and pasting items — but there are also more complex environments, such as the video game Grand Theft Auto V.
Training AIs how best to steal cars and shoot bystanders might seem a strange direction to take, given that OpenAI’s mission is to build safe AI, and that one of its goals is to enable a general-purpose robot to wield tools — initially to perform housework.
But playing games is an important test of artificial intelligence: Many games are like small virtual worlds, the company said, and AI learning techniques will have to advance significantly in order to play a diverse range of games quickly and well. Being able to quickly master new, unfamiliar environments in this way is a first step toward general intelligence.
The non-profit company is backed by Elon Musk and Peter Thiel among others, and receives support from Microsoft, Amazon Web Services and Infosys.
OpenAI says it has already struck deals with game publishers including Microsoft Studios, EA and Valve, allowing AIs to learn to play games such as Portal, Wing Commander III, Command & Conquer: Red Alert 2, and Sid Meier’s Alpha Centauri. Less playful AIs will in future also be able to learn to use Wolfram’s Mathematica software using Universe, OpenAI said.