The computer Deep Blue made headlines in 1997 for beating chessmaster Garry Kasparov at his own game, but the ultimate test of an artificial intelligence’s game-playing and strategizing ability is the game Go. The game, which originated in China and is played by placing stones on a board in order to capture space, might share with chess the basic idea of outmaneuvering and taking an opponent’s pieces, but the number of possible moves is exponentially higher – more than a gogol times higher.
Yesterday, Google announced that its system of neural networks beat three-time European Go champion Fan Hui 5 games to 0. This marks the first time a computer program has beaten a professional, human Go player.
Google’s system was called AlphaGo, and it isn’t only one neural network. Instead, it combines an advanced tree search algorithm with two different deep learning neural networks. The two neural networks are separated by tasks – the “policy network” suggests the next move the computer will play, while the “value network” speculates on who will win the game. The two neural networks process the description of the Go board through 12 different network layers, and were trained to recognize and understand 30 million moves played by human players.
Becoming a Go master requires making intuitive moves based on the general layout of the pieces as well as predicting one’s opponent ahead of time. Therefore, AlphaGo used reinforcement learning, supported by the Google Cloud Platform, to strategize through trial-and-error.
The last computer program to make headlines in the difficult world of AI-driven Go was CrazyStone, a Go program that defeated accomplished player Yoshio Ishida in 2013. However, that program had a four-move head start. AlphaGo’s five decisive victories are a bigger sign for the development of complex, game-playing AI.
Google’s next job is to pit AlphaGo against Lee Sodol, the best Go player in the world. Afterward, they can start applying the AI’s game-playing capabilities to wider applications and real-world problems: neural networks this complex might be able to be used to address large-scale phenomena like disease or climate change.
