AIX is a program being used to try to teach an AI how to learn from things… using minecraft. Microsoft researchers are trying to get an AI to figure out how to go up a hill, using a system of incremental rewards to let it know when it has achieved all or part of its goal (I looked all over but I couldn’t find what they use for the rewards anywhere! Probably just some sort of programming.). The AI has to learn what’s important, climbing the hill, and what’s not, such as whether it’s light or dark.
Of course, they could do this in real life, but it would require them to make an extremely costly machine, which as it tried to learn, would inevitably tumble down the hill. The AI goes into the world with no information given beforehand of what to do, how to move, anything. So it’s going to fall into lava pools or lakes and encounter many other obstacles many times as it tries to learn how to obtain its goal. With Minecraft, it’ll just respawn, whereas in real life they’d have to make an entire new, expensive machine—which is why games are such a good place to do this. Why Minecraft in specific is a good way to teach AIs to learn, is because it offers endless possibilities for things to do. You could build a grand palace with some friends, you could go fight monsters, you could even make your own games inside of it.
Trying to make an AI that learns this sort of way isn’t a completely new thing, just coming up in the last year or so. In 1997 Garry Kasparov lost a chess match to Deep Blue, a supercomputer built by IBM. But Google’s Deepmind (bought by Google for $400m in 2014), named AlphaGo, has triumphed over Lee Sedol in a game of Go (Lee Sedol is one of the best human players of that ancient and notoriously taxing board game). Even though the rules or simpler than chess, it’s much harder to program an AI to play Go than chess, just because of the sheer number of possible moves—“There are [1 * 10171] possible positions—that’s more than the number of atoms in the universe, and more than a googol times larger than chess,” said Demis Hassabis, CEO and co-founder of DeepMind.
AlphaGo learned how to play by watching other human players, and having numerous matches against itself. Finally, it played against other Go-playing AIs, winning all but one of the 500 matches it played. Then in its game against Mr. Lee, it won its first three rounds. Commentators were convinced that it had made serious mistakes, but it kept winning, and they had to admit there had been no mistakes after all— it was using valid strategies that its human masters had overlooked. On the fourth round, however, Mr. Lee changed tactics—playing around the edge while leaving AlphaGo to its own devices in the middle. He won that round. However, in the fifth and final round AlphaGo won again.
I think that it’s pretty cool that they’re trying to make AIs be able to learn from scratch and experience, much like humans do. Some people probably will think that if they manage to do this, the robots will take over the workspace—and while yes, many jobs might have many more robots in them, I think that robots could never completely fill in the creativity humans have, and they should always be other jobs as well. Besides, it would be expensive to get an entire task force of robots, and many companies (Besides the huge ones) wouldn’t be able to afford it. So I’m pretty sure that pertains to self-reliance.
However, this is a long blog post, so I’m sure you won’t blame me when I say that I’m tired and I’m done.
Sources:
AIX:
AlphaGo:
No comments:
Post a Comment