Tales of the Rampant Coyote

Adventures in Indie Gaming!

Computer beats Go grandmaster for the first time in history

Posted by Rampant Coyote on March 9, 2016

AlphaGoVictory1Wow. It happened.

‘WE MADE HISTORY’: Google’s DeepMind AI just beat a human world champion at Go

It’s only the first of a five-game tournament. But still, this is a huge development. Much bigger, in my opinion, than Deep Blue’s victories over Gary Kasparov 20 years ago. Why? At that point in the 1990s, Chess was already an established, “solved” problem. It had been that way for about four decades, and computers had been beating top players for a couple of decades. The algorithms were progressively improved, and of course the hardware progressively improved, over the years, so that it was really just a matter of time and refinement.

With Go, it was a major leap. First of all, for the last several years, computer Go AI has only been able to play at an amateur level. Then, within months, it made the leap to defeating a professional in a 5-0 sweep, and has now had its first victory against a master. For so long Go was considered the game to beat, the true test of the capability of artificial intelligence. Quite suddenly, that barrier has been shattered.

Secondly, this wasn’t about optimizing a well-understood strategy with better heuristics, but a dramatic practical application of neural networks, machine learning, and other “machine intelligence” systems to learn to play the game. The downside is that there’s no easy way that this skill can be ‘transferred’ back to humans. It doesn’t “understand” the game the way humans do, and what it has learned about playing the game bears no resemblance to human knowledge.

Thirdly – and this ties into the second point – this example of machine learning is not limited to Go. With something like Chess, the basic search and alpha-beta pruning which have general applications, but beyond that it’s all pretty specific to Chess with limited application to other fields. But AlphaGo is an implementation of more general-purpose machine learning. This has applications almost anywhere. In fact, the systems powering AlphaGo were originally used to learn to play old Atari video games.  This represents a very practical breakthrough in artificial intelligence as a whole.

So… I guess I should welcome our new robot overlords or something. Even if AlphaGo ends up losing the tournament, this one victory represents an incredibly big deal on many levels. What’s next? Self-driving cars? Oh, yeah, about that…

Filed Under: Geek Life - Comments: 2 Comments to Read

  • Victor said,

    Finally after a long time of disappointment they have managed to find a reliable tool they can use.

    What they manged is a way to simulate knowledge barely.

    I guess the next step is a way to merge 2 knowledge types to get a third type.

    They are still nowhere near human intelligence and it might be possible that this is a dead end.

    Hopefully they will solve Japanese language translation and voice generation soon.

  • Rampant Coyote said,

    “Simulate knowledge” – good term for it. It’s kind of a weird place… simulated knowledge, real skill.

    I think there are a ton of practical applications… a very broad area… which is awesome. But yeah, in the end, it’s a tool.