Human-AI Match
On February 22, Google DeepMind CEO Demis Hassabis had a press conference in London, with Lee Se-dol in Seoul, to explain the details of the Go matches between the world champion and the AlphaGo AI system scheduled for March 9, 10, 12, 13 and 15 at the Four Seasons Hotel located in Seoul. “During the matches, the AlphaGo is to be operated on the Google Cloud,” said the CEO, adding, “The server is actually located in the Midwest of the United States and we will link the AI system to the hotel with a high-speed connection.”
In the matches, Aja Huang, who is a member of the British Go Association and Google DeepMind alike, serves as the eye and hand of the AlphaGo in front of the human champion. The games are to be broadcast live worldwide via the YouTube channel of Google DeepMind and those in Korea, China and Japan can watch the matches on TV as well. The winner takes US$1 million.
The game of Go has a huge number of cases, more than that of atoms in the universe, and thus a search tree is insufficient to guarantee a win. In this regard, the AlphaGo analyzes the Go board with 12 processes including millions of links. Specifically, its policy network selects its next move and the value network predicts the winner. The AI system has won 99.8% of its games against the other computer programs and recently beaten Fan Hui, the European champion from China.
Lee Se-dol showed some confidence though. “I watched the game between the AlphaGo and Fan Hui in October last year and came to a conclusion that the former is not comparable to me,” he said, continuing, “I’m checking how the AlphaGo plays every day while trying to maintain a good condition.”