There’s one game that robots have, at least until now, struggled to beat—an ancient Chinese game called “Go.”
Its looks are deceiving; Go’s simple black-and-white checkers seem like they should be relatively transparent to a computer algorithm. But unlike more knowledge-based games like Jeopardy! and Scrabble, Go has long been regarded as an ultimate test for computing power. Previously, scientists have only been able to develop programs that play at the level of human amateurs. If a machine could outwit a human expert at Go, which involves complex board positions and moves, then artificial intelligence will have taken a huge step forward.
David Silver and Demis Hassabis of Google DeepMind in London have now created such a program, reported in the journal Nature today. Called “AlphaGo,” it builds on the concept of deep learning, which approximatesneural networks of the human brain as a means of absorbing information and acquiring new skills. In other words, Silver and Hassabis trained AlphaGo by repeatedly exposing it to human-against-human games. They also let AlphaGo play against itself to reinforce those lessons. As a result, AlphaGo achieved a 99.8% winning rate against other Go algorithms and defeated the human European Go champion in a recent tournament five separate times.
Go requires more intensive strategizing than most games—so AlphaGo’s victory shows that it can do long-term planning. For that reason, it is considered to be “the pinnacle of game AI research for more than 20 years,” according to Hassabis. “For us, it was an irresistible challenge.”
What’s more, the Google team members say that the European expert Go player felt as though he could intuit what AlphaGo was “thinking” during their match. That’s far from a Turing test , but it does illustrate the extraordinary power of neural networks. AlphaGo’s next challenge will be to defeat Lee Sedol, the top Go player in the world for the last decade. That tournament is scheduled for March.
While AlphaGo has no immediate applications, Silver and Hassabis say that it will spawn technological developments and, of course, many ethical questions down the road.
The news comes just a few days after artificial intelligence pioneer Marvin Minsky passed away at the age of 88.
“I would have loved to hear how surprised he would have been,” Silver said, adding that most of Minsky’s contemporaries hadn’t anticipated such a development would be possible until the year 2025 or 2030.