AlphaGo最終以4:1戰勝李世石 AlphaGo conquers Korean grandmaster Lee Se dol

Google’s AlphaGo computer system sealed a 4-1 victory over a South Korean Go grandmaster on Tuesday, in a landmark moment for the development of artificial intelligence.

週二，谷歌(Google)的AlphaGo計算機系統以4比1的總比分戰勝了韓國圍棋大師李世石(Lee Se-dol)。這是人工智能發展中的一個里程碑時刻。

Mastery of the east Asian game of Go was long seen as a stern challenge for computers given its huge complexity. AlphaGo’s creators estimate that there are about 250 potential moves at each point of a game, against 35 in chess, yielding a possible number of board configurations of 10 squared by 170.

鑑於圍棋極其複雜，長期以來，精通圍棋一直被視爲計算機面臨的一項嚴峻挑戰。AlphaGo的創建者估計，圍棋的每一步都有250種可能走法（國際象棋只有35種），產生的可能局面數量爲10的170次方。

Lee Se-dol, arguably the best player of the past decade, had expected to win a crushing victory, arguing that AlphaGo lacked the “intuition” needed to beat him. But the program won the first three games in the series, which began last Wednesday, before Mr Lee clawed back a victory on Sunday.

李世石可以說是過去10年最棒的圍棋手，他曾預計自己會取得壓倒性勝利。他認爲，AlphaGo缺少擊敗他所需的“直覺”。但在上週三開始這場對決中，計算機程序贏了前三局，而後李世石在上週日扳回一局。

Tuesday’s final game was one of the closest: AlphaGo recovered from an early error to force Mr Lee into resignation in overtime, with each player having used up the allotted two hours.

週二的最後一局是雙方拼殺得最難解難分的一局：AlphaGo起先出現了一次失誤，但後來挽回了局面，把李世石拖入讀秒，李世石在讀秒階段投子認輸。雙方都用盡了分配給自己的兩小時。

The victory demonstrates the power of the “deep learning” systems employed by AlphaGo’s creators at DeepMind, a London-based start-up acquired by Google two years ago.

AlphaGo的勝利證明了創建者使用的“深度學習”系統的威力。AlphaGo是由谷歌在兩年前收購的倫敦創業型企業DeepMind創建的。

Go’s huge complexity rules out the “brute force” approach of IBM’s Deep Blue chess computer, which beat Garry Kasparov in 1997 by evaluating 200m positions per second. Instead, AlphaGo learned to recognise promising moves by playing huge numbers of Go matches against itself.

圍棋的巨大複雜性超出了IBM深藍(Deep Blue)國際象棋計算機“蠻力”方法的處理能力。1997年，每秒評估2億步的深藍戰勝了加里•卡斯帕羅夫(Garry Kasparov)。AlphaGo則是通過自己跟自己大量對弈，學會了如何推測對手可能的走法。

Demis Hassabis, DeepMind’s chief executive, said the series would enable his team to make further improvements to the system, which had some flaws exposed during the contest — notably when an unorthodox move by Mr Lee in the fourth match prompted AlphaGo to make a series of amateurish blunders.

DeepMind首席執行官傑米斯•哈薩比斯(Demis Hassabis)表示，此次對弈將使他的團隊能對AlphaGo系統進行更多改進。該系統在對弈過程中暴露出了一些缺陷——尤其是在第四局中，李世石走出了反常規的一步，導致AlphaGo出現一連串業餘選手般的失誤。

The system’s log showed that it had assessed the likelihood of Mr Lee’s move at less than one in 10,000, Mr Hassabis tweeted on Tuesday.

週二，哈薩比斯在Twitter上發帖稱，系統日誌顯示，AlphaGo認爲李世石走出那一步的可能性低於萬分之一。

Mr Lee, meanwhile, refused to concede that the era of human supremacy in Go was at an end. “I don’t necessarily think AlphaGo is superior to me — there’s more that a human being can do against artificial intelligence,” he said. “I don’t feel this was a loss for human beings. It showed my weaknesses, not the weaknesses of humanity.”

李世石拒絕承認人類統治圍棋的時代已經終結。“我並不完全認爲AlphaGo比我高明——人類在對抗人工智能方面還能做得更多。”他說，“我不認爲這是人類的失敗。它暴露出了我個人的弱點，而非人類的弱點。”