Go is a complex game you may have not heard of. So complex, in fact, that a artificial intelligence (AI) has not been able to beat a human professional until now.
That AI’s name is AlphaGo and it comes from the minds of DeepMind, a Google company in London. AlphaGo, despite it’s name, was not engineered specifically to play the game, but rather a general-purpose machine designed to learn. Here, it was asked to learn Go.
“There are more board configurations than atoms in the universe,” Said Demis Hassabis, CEO of Google DeepMind. Hassabis himself is one of the best chess players in the world, was a child prodigy, and dabbled in Go. In fact, his team wanted this project to work just so they could beat him at it.
This was achieved thanks to a process called deep learning, in which software can look at past mistakes (either from its previous iterations or some other source) and “learn” from them to improve.
David Silver, an employee at Google DeepMind, explains how this AI was able to crack a game many thought a computer never would.
The key to AlphaGo is the deep neural networks that we use under the hood. In fact, AlphaGo uses two neural networks. The first is called the “Policy Network”, and the second is called the “Value Network”.
And the way AlphaGo uses [the Policy Network] is to reduce the enormous complexity of the search tree into something more manageable. Instead of having to consider hundreds of different moves, at each step of the search, it just considers, maybe, a hanful of promising moves.
The Value Network we use to reduce the depth of the search. So, instead of having this enormously deep search that has to go all the way down perhaps 300 moves until the end of the game, we have this this modest search that goes to, perhaps, 20 moves and then we evaluate that position without playing all the way until the end of the game.
The search isn’t based on brute force, it’s based on something more akin to imagination.
The “human professional” here was Fan Hui, the European Go champion. In their five matches, AlphaGo cleaned house with five straight victories.
The next test for this invention is Lee Se-dol, who many consider to be the greatest Go player in the world. Place your bets, ladies and gentlemen, who will win?
[Source – Nature.com, Image – C.C 2.0 by Linh Nguyen]