AlphaGo's Seoul victory redraws a boundary for machine learning

DeepMind's AlphaGo completed its five-game match against Lee Sedol in Seoul today with a 4-1 victory. Go has long resisted computer mastery because each position offers too many plausible continuations for brute-force calculation alone.

The program combines neural networks trained on expert games and self-play with a search procedure that evaluates promising moves. Its unconventional play repeatedly surprised professional observers, although Lee's inventive victory in game four showed that the contest was not a simple procession.

This result gives machine learning a vivid public demonstration beyond laboratory benchmarks. AlphaGo has mastered a narrow task under fixed rules, not general intelligence, but its methods may help researchers approach other problems where judgment depends on recognizing patterns among enormous numbers of choices. The match also leaves Go players with an unfamiliar new source of ideas to study.