Deep Learning Takes Computer Go to the Professional Level
Two human vs computer even game matches were played in Europe last October. In one, former German champion Franz-Josef Dickhut disposed of Zen 3-1, apparently reaffirming human superiority. But as reported in the 28 January issue of Nature, in the other match AlphaGo, a new program from Google DeepMind, trounced European Champion Fan Hui 5-0. Fan earned a 2-dan professional ranking in China before emigrating to France, so it appears that go software has reached the level of professional play. The new ingredient responsible for this startling advance is deep learning, a technique also coming into use in fields such as speech recognition and medical diagnosis.
In the first of the five AlphaGo-Fan games, both sides played conservatively and AlphaGo won by 2.5 points. In the rest of the match Fan played aggressively, but AlphaGo outfought him and won four times by resignation. Fan described AlphaGo as “very strong and stable…like a wall.” Game records can be found here.
This March, plans call for AlphaGo to take on a tougher professional opponent, in fact, one of the toughest there is: Korean 9-dan Lee Sedol, winner of numerous world titles since 2002. The outcome of this match is hard to predict, but it is worth noting that the AlphaGo programming team reports that AlphaGo can beat the best rival computer programs with a four-stone handicap. That is something that several other 9-dan pros have had trouble doing in the past few years.
Further information can be found, here, here, here, and elsewhere on the Internet.