AlphaGo

    Keywords: Software

Computer go bot developed by Google DeepMind.

Algorithm

Combines MCTS with two neural networks: a policy network and a value network.

Accomplishments

October 2015 - First AI program to defeat a professional player on a 19x19 board without any handicap. This was against the French player Fan Hui (2p) who lost 0-5[1].

March 2016 - Defeated Lee Sedol 4-1 in a 5-game match with a prize purse of one million USD. AlphaGo won the first three games by resignation. Then, in the fourth game, Lee Sedol played a tesuji in the centre that led to victory after it had looked like AlphaGo had been winning in the opening. AlphaGo made several clear mistakes before resigning, showing how its policy networks are susceptible to over-valuing forcing moves if they lead to a higher perceived winning percentage. In a close match, Lee resigned the fifth game, giving a final score of 4-1 for AlphaGo. Lee had been winning early in the game after AlphaGo could not correctly read out a two-stone edge squeeze and made a mistake locally, but later in the game he neglected to play a more aggressive extension in the upper part of the board, instead preferring to play safe and make eye shape. When Lee resigned, AlphaGo was estimated to be winning by about 2.5 points, with 7.5 komi under Chinese scoring.

[1] These are the results of the five formal games with longer time controls. Five informal games with shorter time controls were also played and Fan Hui won two of them by resignation (see Silver et al. 2016; p. 493). The kifu of the formal games seem to be available at page 488 (op. cit.).

AlphaGo vs. Lee Sedol Challenge match 2016

External Links



This is a copy of the living page "AlphaGo" at Sensei's Library.
(OC) 2016 the Authors, published under the OpenContent License V1.0.
[Welcome to Sensei's Library!]
StartingPoints
ReferenceSection
About