At the end of January, Carnegie Mellon computer scientists achieved a major milestone: their algorithm, Libratus, beat a set of professional poker players in a 120,000-hand tournament. While humans have fallen to computers in a variety of games, notably chess and go, poker is fundamentally different, in that each player has information that’s not available to the rest. A fundamentally different sort of AI is required to deal with this sort of imperfect information.

This week in Science, a different team describe its human-beating poker algorithm, DeepStack. Both teams say their approach isn’t specific to poker, so 2017 may mark the end of human dominance at all imperfect-information games.

Imperfect strategies

A perfect information game is relatively simple: all players can know the full state of the game, often just by looking at the board. They also know the full set of legal rules. So it’s relatively trivial to calculate all the possible moves available given any specific board. With enough computing power, it’s also possible to calculate all possibilities many moves out—enough to effectively bring any game to a conclusion. In the case of a simple game like checkers, this means all possible future moves. For something more complicated like chess, calculations may effectively be limited to 10 moves ahead.