Off the wire
• 1st LD-Writethru: Chinese shares struggle to extend winning streak  • Philippine supreme court junks disqualification case against woman presidential bet  • 21 militants killed in military operation in NW Pakistan  • China Hushen 300 index futures close mixed Tuesday  • 2nd LD-Writethru-China Focus: Chinese exports slump, outlook worrying  • China treasury bond futures close higher Tuesday  • Roundup: S. Korea unveils unilateral sanctions on DPRK  • 27 arrested for toxic waste dumping in north China  • Seoul shares end lower  • 2nd LD Writethru: S.Korea unveils unilateral sanctions on DPRK  
You are here: 　 Home

S.Korean Go champion gets nervous ahead of match with Google's AI

Xinhua, March 8, 2016 Adjust font size:

South Korean Lee Sedol, the world champion of the ancient Chinese board game Go, said Tuesday that he felt slightly nervous ahead of a match with Google's computer program AlphaGo.

Lee told a press conference in Seoul that he seemed to get slightly "nervous" though he still has confidence in his victory in the five-game match scheduled for Wednesday to next Tuesday.

"(My) winning rate does not seem to go as far as 5-0," the 33-year-old said, slightly lowering his confidence compared with his Feb. 22 press conference in which he said AlphaGo could by no means defeat him.

The winner of the match will receive 1 million U.S. dollars in prize. If AlphaGo wins, the prize will be given to charities as donation.

AlphaGo, developed by Google's London-based subsidiary DeepMind, demonstrated a major step forward in artificial intelligence (AI) after defeating European Go champion Fan Hui in October 2015.

It thrilled the public as some experts predicted decades of years needed to see AI program's victory over human professionals at the ancient Chinese board game.

Demis Hassabis, CEO of the DeepMind, held a press conference in Seoul to explain the principle of AlphaGO algorithm, saying AlphaGo became stronger now than in October as it has made many upgrades since then.

AlphaGo combines an advanced tree search with neural networks. The networks take a description of the Go board as an input and process it through 12 different network layers containing millions of neuron-like connections.

The "policy" neural network selects a next move to play, and the "value" network predicts a winning rate in order to mimic humans' intuition. The AlphaGo developer trained neural networks on 30 million moves from games by human experts.

The algorithm learned to discover new strategies for itself, by playing thousands of games between its neural networks and adjusting connections by use of a trial-and-error process known as reinforcement learning.

Hassabis told reporters that intuition is important in the Go game, known as Weiqi in China and Baduk in South Korea, saying the neural network approach for human intuition is at the core of the AlphaGo system. Enditem