AlphaGo paper in Nature reveals win over Fan Hui

Background

Go is played by two people who take turns placing black and white stones on a grid, each trying to enclose more territory than the other. The rules fit in a few sentences, but the game is hard to play well: the number of possible positions and move sequences is enormous, so a program cannot simply check everything, and the value of a move usually depends on the judgement of the whole board rather than on one local fight.

On 27 January 2016, DeepMind announced AlphaGo, and timed the announcement to coincide with a paper in Nature dated 28 January 2016 ^[1][2]. The announcement and the paper together made public a result that had until then been kept quiet: a program had beaten a professional Go player.

The match behind closed doors

The paper reported that AlphaGo had beaten European Go champion Fan Hui 5–0 in October 2015 ^[1][2]. Google's blog post described Fan Hui as the reigning three-time European Go champion and said the games were held behind closed doors between 5 and 9 October 2015 ^[2]. According to Google, the 5–0 result was the first time a computer program had beaten a professional Go player ^[2]. The announcement therefore carried two messages at once: a specific scoreline, and a first in the history of the game.

How the program worked

The paper described a combination of policy and value networks with Monte Carlo tree search ^[1]. In outline, a policy network narrows the search by proposing moves that look worth considering, while a value network estimates how promising a position is, so the program does not have to play every line out to the end; the tree search then spends its time on the branches those networks rate most highly. The networks themselves were trained on human games and then improved by self-play, with the program playing against versions of itself ^[1].

The paper also reported a broad test against other Go programs: the distributed version of AlphaGo won all 500 games it played ^[1].

The Lee Sedol match

Even as it announced the Fan Hui result, DeepMind named the next test. Google's blog announced a match for March 2016 in Seoul, South Korea ^[2]. That match was played in Seoul in March 2016, and AlphaGo won it 4–1 ^[1].

Why it matters

Three moments carry most of the story. In October 2015 a professional player was beaten behind closed doors; on 27 January 2016 the result was announced together with a paper in Nature; in March 2016 the program won a public match in Seoul. The order matters. A claim that would once have been hard to believe arrived with a written account of the method and of the games, so readers could examine how the result had been produced instead of taking it on trust.

For the wider field, the interest lay in how the result was reached. The program did not rely on hand-written rules of thumb about Go positions; it learned from human games and from playing itself, and then combined those learned judgements with search ^[1]. That recipe — learned networks guiding a tree search — was the part other researchers could study and argue with. The 500 games against other programs ^[1] also showed that the outcome was not a lucky series of games against a single opponent.

The announcement also set up the next step in public: the Seoul match was named in the same blog post that revealed the Fan Hui result ^[2], and the paper recorded its outcome, a 4–1 win for AlphaGo ^[1].