During training, players are first confronted with simple single-player games, such as finding a purple cube or placing a yellow ball on the red floor. They progress to more complex multiplayer games, such as hide-and-seek or flag capture, where teams compete to find and grab an opponent’s flag first. The playground manager has no specific goal, but aims to improve the general abilities of the players over time.
Why is this great? Artificial intelligences like DeepMind’s AlphaZer have beaten the world’s best players in chess and Go. But I can only learn one game at a time. As DeepMind co-founder Shane Legg said when I spoke to him last year, that seems to be the case you have to replace your chess brain for your Go brain every time you want to change the game.
Researchers are now trying to build an AI that can learn multiple tasks at once, which means teaching them general skills that make it easier to adapt.
One exciting trend in this direction is open learning, where artificial intelligence is trained for many different tasks without a specific goal. In many ways, humans and other animals seem to learn through aimless play. But that requires a huge amount of data. XLi generates this data automatically, in the form of an endless series of challenges. It’s similar with POET, an artificial intelligence dojo where bipedal bots learn to manage obstacles in a 2D landscape. XLand’s world, however, is much more complex and detailed.
XLi is also an example AI learns to create itself, or what Jeff Clune, who helped develop the POET and leads the team working on this topic on OpenAI, calls AI (AI-GA) generation algorithms. “This work is pushing the boundaries of AI-GA,” says Clune. “It’s very exciting to see.”