Alphabet Program Beats the European Human Go Champion
By John Markoff

Artificial intelligence researchers are closing in on a new benchmark for comparing the human mind and a machine. On Wednesday, DeepMind, a research organization that operates under the umbrella of Alphabet, reported that a program combining two separate algorithms had soundly defeated a high-ranking professional Go player in a series of five matches.

The result, which appeared in the Jan. 27 edition of the journal Nature, is further evidence of the power created when a class of A.I. machine learning programs known as “deep neural networks” is combined with immense sets of data.

Go is seen as a good test for artificial intelligence researchers because it is more complex than chess, with a far larger range of possible positions. This makes strategy and reasoning in the game more challenging.

Go is played with round black and white stones, and two players alternately place pieces on a square grid with the goal of occupying the most territory. Until recently, software programs had not been able to do better than beat amateur Go players. In the Nature paper, engineers at DeepMind described a program, AlphaGo, that had achieved a 99.8 percent winning rate against other Go programs. It also swept five games from the European Go champion, Fan Hui.

The match between the AlphaGo program and Fan Hui was in October, and the DeepMind program has continued to train since then, said Demis Hassabis, a researcher who founded DeepMind Technologies, which was acquired by Google in 2014. Google changed its name to Alphabet last year, though the company’s traditional ad-based businesses still operate under the Google label.
“The machine has continued to get better. We haven’t hit any kind of ceiling yet on performance,” he said.

The Alphabet approach relies on the newest so-called deep learning approach combined with a more traditional type of algorithm known as a Monte Carlo, which is designed to exhaustively explore large numbers of possible combinations of moves. The researchers said they had also trained their program using input from expert human Go players.

The research and the game have created a rivalry among the public relations departments of companies like Alphabet, Microsoft and Facebook.

The day before the Alphabet paper was published, Facebook republished an earlier paper the company had posted on the arXiv.org website. At the same time, Facebook issued blog posts from Yann LeCun, one of its artificial intelligence researchers, and one from the company’s chief executive, Mark Zuckerberg.

The statement by Mr. Zuckerberg resulted in a swift response from one Facebook user that may express a deeper human concern than the narrow results of the research: “Why don’t you leave that ancient game alone and let it be without any artificial players? Do we really need an A.I. in everything?” wrote Konstantinos Karakasidis.

Those concerns are not likely to be heeded. In a blog post Wednesday morning, Alphabet stated that, in an effort to reprise the winning IBM Deep Blue chess playing program that defeated the chess champion Garry Kasparov in 1996, Alphabet will match its AlphaGo program against Lee Sedol, the current Go champion, for a five-game match in March.

There will be a $1 million prize for the winner, and Mr. Hassabis said that Alphabet would donate the prize to charity if AlphaGo won. The match will be streamed live on YouTube.

Mr. Hassabis, who is a skilled chess player and has been a professional gamer as well, said that Go was a beautiful game, but that “building an A.I. is also a human endeavor and a kind of ingenious one, too. The reason games are used as a testing ground is that they’re kind of like a microcosm of the real world.”

Google人工智慧AlphaGo 完勝歐洲圍棋棋王

Google開發的人工智慧程式AlphaGo擊敗歐洲三屆圍棋棋王樊麾，是人工智慧一大突破，下一回合將於3月在首爾挑戰南韓棋王李世乭。首爾的人腦和電腦對決勝方可獲100萬美元獎金。

AlphaGo去年10月擊敗樊麾，但這項成就27日才發表於「自然」期刊。

Google的人工智慧部門DeepMind表示，他們的程式與樊麾對弈，五盤皆勝。不過，研發程式的DeepMind部門主管哈薩比斯坦言，AlphaGo的「棋藝」仍只是業餘棋士水平。

發源於中國古代的圍棋已有2500年以上歷史，棋盤縱橫各19條線，棋子黑白兩種，圍住對方棋子便可吃子。圍棋的工具和規則簡潔優雅，棋法卻變化萬千，公認是全世界最複雜的棋盤遊戲。

人工智慧（AI）的先驅，莫過於國際商業機器公司（IBM）開發的超級電腦「深藍」，它曾於1997年擊敗西洋棋世界冠軍卡斯帕洛夫。

然而人工智慧要贏過圍棋人類棋手，難度遠高於西洋棋，因為相較於西洋棋只要將死「國王」，圍棋的下法更複雜。西洋棋每一步平均約有20種走法，而圍棋有近200種，一盤棋中可能出現的變化，比宇宙中原子數量還多；且圍棋棋子的分布並非代表絕對會贏或不利，棋士既可進攻、也可布下看似不利的陷阱，這些並非電腦演算可及。

不過兩年以前，多數圍棋棋士和程式設計師當時還認為，人工智慧要贏過人類，還要花至少十年才能追上，如今人工智慧已能打敗圍棋棋士。

臉書創辦人祖克柏27日發文透露，臉書的AI研究團隊去年也開始研發會下圍棋的程式；過去六個月來，程式已進步到僅0.1秒就能下一子。祖克柏說，他們的研究以搜尋為基礎，把對手所有可能的下法建立模組，且與臉書「電腦願景團隊」建立的系統相容。這個程式由專研人工智慧的陸裔學者田淵棟負責。

哈薩比斯在Google官方部落格中說，正因為圍棋玩法簡單，所以棋局變化更複雜，棋局組合高達10的171次方。

AlphaGo曾挑戰目前已開發的「同儕」人工智慧圍棋程式，對弈五百場，只輸過一次，即使讓子。

原文參照：
http://bits.blogs.nytimes.com/2016/01/27/alphabet-program-beats-the-european-human-go-champion/

2016-01-28．聯合晚報．A6．國際焦點．編譯莊蕙嘉