Developers admit AI is not perfect

By Park Sae-jin Posted : March 14, 2016, 11:10 Updated : March 14, 2016, 11:10

[Courtesy of Google]


Google's computer program AlphaGo has surprised human Go players in a historic showdown with South Korea's Lee Sedol, but developers admit artificial intelligence is not perfect and can make mistakes.

On Sunday Lee won his first match against AlphaGo, developed by Google's artificial intelligence (AI) company DeepMind, in the ancient board game Go to stop AI's clean sweep in a five-game match.

DeepMind head Demis Hassabis said Sunday's loss was a valuable learning tool and would help identify weaknesses in the programme that his team needed to address.

"It's a real testament to Lee's incredible fighting spirit and he was able to play so brilliantly today after three defeats," Hassabis said, praising Lee as an "incredible" player.

Hassabis admitted that Lee's "fantastic moves" forced AlphaGo to make mistakes.

"AlphaGo started off well, and it was estimated it was ing very well in the middle position, but then because of Lee Se-dol's fantastic play, it was pressured into some mistakes," he said.

"But actually, we are also happy, because this is why we came here, to test AlphaGo to its limits and find out what its weaknesses are and try to improve the program."

Lee, one of the world's top players and a holder of 18 international titles, was surprised at AlphaGo's skills but also said the program was not perfect.

The 33-year-old has already lost the Go tournament after suffering three straight defeats, but he became the first to beat AlphaGo.

The one-million-dollar prize money went to AlphaGo, but Lee will receive $20,000 for a win on top of $150,000 for participating in the series.

In the previous matches, Lee had tried to find AlphaGo's weaknesses, tweaking the game, luring AlphaGo into his traps or changing his style of play to shake AlphaGo's calculations.

In Sunday's match, AlphaGo was ahead of Lee at the beginning, but Lee's aggressive alternating battles pushed AlphaGo into overtime and forced mistakes.

The AI has honed itself by studying old matches and training through simulated games. Google executives say Go offers too many possible moves for a machine to win simply through brute-force calculations.

"AlphaGo is still (a) prototype," Hassabis said.



 
기사 이미지 확대 보기
닫기