The ever-growing playground teaches AI how to teach multitasking

During training, players first play single-player games, such as finding a purple cube or placing a yellow ball on the red floor, playing more complex multiplayer games like hiding, or catching a flag for teams to compete with. first find and take the opponent’s flag. The playground manager does not have a specific goal, but over time it aims to improve the overall ability of the players.

Why is this cool? DeepMind’s AlphaZero-like AI has won some of the best human players in the world in chess and Go. But at the same time they can learn only one game. As I said with Shane Legg, one of the founders of DeepMind, when I spoke to him last year, your chess brain may need to change in exchange for the brain every time you want to change games.

Researchers are now trying to build IAs that can learn multiple tasks, which teaches general skills that make it easier to adapt to new tasks.

An exciting trend in this direction is open learning, where AIs are trained in a variety of tasks without specific goals. In many ways, humans and other animals learn this way through aimless play. But this requires a large amount of data. XLand automatically generates this data in an endless stream of challenges. It’s similar POETAn AI training dojo where two-legged boats learn to navigate obstacles in a 2D landscape. The world of XLand is much more complex and precise, however.

XLand is also an example AI learns to do itself, or Jeff Clune, who helped develop the POET and leads a group working on this topic In OpenAI, it is called AI Creation Algorithms (AI-GA). “This work pushes the boundaries of AI-GA,” Clun says. “It’s very exciting to watch.”