Today, in the first round of a weeklong match pitting man against machine in the ancient Chinese game of Go, Google’s AlphaGo computer program won the first battle.
Advertisement
Hassabis and Silver fed AlphaGo a collection of 30 million moves from games played by skilled human Go players, until it could correctly predict a player’s next move 57 percent of the time; the previous record was 44 percent.
Dennis Hassabis, founder and CEO of Google DeepMind, says one advantage AlphaGo has over human players is that it never gets exhausted or intimidated.
The Internet company is playing a five-match tournament against Lee Sedol, who Google said has been the top-ranked Go player of the past decade, to show off the capabilities of software developed by its London-based AI subsidiary DeepMind.
He also singled out DeepMind’s Reinforcement Learning Framework technologies for praise and for making the match possible.
It will not be the first time that computers have challenged human minds.
But he doesn’t seem sure: “I’ve been telling people that I was certain that I would be the victor of all five matches”.
Lee was expected to brief reporters later, but commentators following the match said he appeared to have made a bad error in the final stages of the game, which was being watched by hundreds of experts, reporters and VIPs packed into an adjacent press room with a large TV screen. Some computer scientists said it is only a matter of time for AI to dominate the ancient game of Go.
Lee will get his crack at a victory Thursday. “Go is a game primarily about intuition and feel rather than brute calculation, which is what makes it so hard for computers to play well”.
AlphaGo taps neural networks and advanced “tree search” programs. The rank gap between Hui and Sedol is significant, and Sedol’s had the added benefit of being able to study Hui’s games against the program.
This is mainly due to the complexity of Go. But with AlphaGo, “it’s the ability, given the description, and the value of the game state, which action should I take?”
“I have been in countless matches, but this unfamiliar feeling that I am getting is a first”, he said.
The hotly anticipated match-up is being closely followed by tens of millions of Go fans in Asia as well as experts in AI. If AlphaGo wins, the proceeds will be donated to UNICEF, STEM and Go charities, Google DeepMind said.
Lee said he wouldn’t be affected psychologically by the loss. Last October, in a match that was kept secret so the Google team could publish its work in Nature, AlphaGo trounced the European champion, Fan Hui, 5-0.
“AlphaGo made some moves that no human would ever make”.
It worked out which positions were likely to be winning ones, and then – and this is the master-stroke – it played itself repeatedly to improve its own play. “To me it was a relief”, says Frank Lantz, the director of the New York University Game Center in New York City and an enthusiastic Go amateur.
Advertisement
AlphaGo won’t evolve over the course of the five games, as the algorithm has to learn things through training and learning from just a few games does not involve very much data.
