[ad_1]
In preparation, players meet for the first time with one simple game, such as finding a purple box or placing a yellow ball on the red floor. They advance in more difficult sports such as hiding and seeking or taking up a flag, while teams compete to be the first to find and carry their opponent’s flag. The stadium manager has no real goal but his goal is to promote the full potential of the players over time.
Why is this good? AI like DeepZind’s AlphaZero has beaten the world’s top players in chess and Go. But they can learn one game at a time. As DeepMind researcher Shane Legg pointed out when I spoke to her last year, it’s almost like you have to change your chess brain to get to your brain any time you want to change a game.
Researchers are now trying to create AIs that can learn several tasks at once, which means they teach them skills that contribute to change.
THINK
One of the most exciting aspects of this feature is unlimited learning, where AI is trained for a variety of tasks without any real purpose. In many ways, this is how humans and other animals seem to learn, through a meaningless toy. But this requires more. XLand releases this all of a sudden, like a lot of complexity. It’s the same with POET WRITER, an AI training dojo where two-legged bots learn to deal with obstacles in the 2D area. The world of XLand is very complex and detailed, however.
XLand is an example of AI learning to make itself, or what Jeff Clune, who helped form POET and led the team working on the topic in OpenAI, calls for algorithms for developing AI (AI-GAs). “The project pushes the boundaries of AI-GAs,” says Clune. “It’s wonderful to see.”
[ad_2]
Source link
