Machines versus champions
From Samuel's checkers to AlphaGo's Go, games measure how far AI has come. Two paths take turns, brute-force search and learning by self-play, until they merge.
Samuel's checkers program (1952)
Samuel's checkers program learns by playing against itself, adjusting its evaluation of each position using the evaluation of later ones.
Temporal differences (August 1988)
Sutton turns that idea into a general method, temporal-difference learning.
TD-Gammon (1992)
Tesauro applies temporal differences to a network that plays backgammon against itself and approaches the level of the best players.
Deep Blue beats Kasparov (May 11, 1997)
Meanwhile, IBM takes the other path. Deep Blue does not learn; it searches with specialized hardware, and it beats Kasparov in 1997.
Monte Carlo tree search (May 2006)
In Go, that kind of search is not enough. Coulom, and Kocsis and Szepesvári, evaluate moves with thousands of random games and a tree that grows selectively.
Watson wins Jeopardy! (February 14, 2011)
IBM pitches Watson as Deep Blue's successor, and it beats the greatest champions of the quiz show Jeopardy!
AlphaGo beats Fan Hui (October 5, 2015)
AlphaGo brings the two paths together. It combines Monte Carlo tree search with networks that evaluate positions and improve through self-play, and it beats a Go professional for the first time.