Sunday October 29, 2017
Bots that teach themselves to be the best
Google made headlines around the world when its AI AlphaGo beat the reigning Go champion a year ago.
It achieved that by studying millions of games played by human players and learning what worked best from the outcomes of all those games.
But last week, a new AI, AlphaGo Zero, practiced playing Go by itself for three days, and then beat the original, human-trained AlphaGo.
One hundred games to zero.
This bleeding-edge form of AI, called self-play, has proven to be a surprisingly effective way for AIs to learn certain things, says Greg Brockman, the co-founder and CTO of the non-profit group OpenAI.
OpenAI recently built a similar machine that took on a game more complex than Go, a multiplayer online game called Dota 2. And by playing against itself over and over, the AI quickly taught itself to beat the best in the world — people who play the game professionally.
While chess and Go are certainly difficult games, they aren't complicated in the same way videogames are. Rather than simple grids, they take place in varying settings, and can have many different characters and types of moves. That's why it makes for such a good testing ground for AI self-play.
"You want an environment that is very complicated, that requires real kinds of intelligence," Greg says. "Strategic planning, mechanical aspects, figuring out how different combos work," he adds.
This type of learning can allow machines to adapt to differing situations more quickly, and learn tasks more effectively.
"People have shown that on board games, things like Go, things like chess, are all things that artificial intelligence can solve. But the question of can you take these technologies and push them and surpass human abilities in these very complicated games is something that before now was not really known," Greg says.
OpenAI is a group dedicated to making artificial intelligence research accessible to all, and not merely to those with the resources available to pursue it. It is founded, in part, by Elon Musk, who has said publicly that he's concerned about the future of AI should it become too powerful to be controlled.
"We should be sure that, before they're at that amazing level, that they do the things we want," Greg says. "You need to be certain they they behave in the way the designer intends."
What will be essential for a safe future for AI, according to Greg, is proactive government monitoring. "The more that we have consciousness of what these technologies will be like, the more we will proactively be able to make the right choices."