AlphaGo

AlphaGo is a computer program that plays the board game Go.^[1][2] It was made by DeepMind Technologies (Google affiliate).^[3] This program became famous due to the victories against professional players.^[4][5]

AlphaGo logo

Many new technologies were used to create AlphaGo, including deep learning,^[6] optimization,^[7] and the Monte Carlo algorithm.^[8]

Powered versions

After the release of AlphaGo, DeepMind Technologies has made powered versions such as the AlphaGo Zero^[9][10] and the AlphaZero.^{[11][12][13][14][15]} AlphaZero is a self-taught program.^[16] This means that it became powerful without human guidance.

Details

The following table is the summary of AlphaGo achievements (including its variants).

Strength of AlphaGo and its variants^[17]

Versions	Hardware	Elo rating	Date	Results
AlphaGo versus Fan Hui	176 GPUs,[18]distributed	3,144[19]	Oct 2015	5:0 against Fan Hui (professional player)
AlphaGo versus Lee Sedol	48 Tensor processing units (TPUs),[18] distributed	3,739[19]	Mar 2016	4:1 against Lee Sedol (former Korean & world champion)
AlphaGo Master	4 TPUs,[18] single machine	4,858[19]	May 2017	60:0 against professional players
AlphaGo Zero (40 block)	4 TPUs,[18] single machine	5,185[19]	Oct 2017	100:0 against AlphaGo version that defeated Lee Sedol 89:11 against AlphaGo Master
AlphaZero (20 block)	4 TPUs, single machine	5,018 [20]	Dec 2017	60:40 against AlphaGo Zero (20 block)

Rivals

After the appearance of AlphaGo, several research groups have created computer Go programs with similar technical viewpoints.

Darkforest

This was made by Facebook.^[21] The source codes are available on GitHub.^[22]

DeepZenGo

This was made in Japan.^[23][24] Nihon Ki-in was also involved in its research and development.

Related pages

• Demis Hassabis