Go is a game which started somewhere 3000 years ago in China. The rules of the game are simple. There are two players where one has black and the other has white marbles. The two players have to place their marbles on full sized 19×19 (biggest size) board. The aim of the player is to surround more region than the antagonist. Though the Go has very simple rules but it is considered more complex than chess.
1. Start with empty board
2. Marbles are placed on intersection of lines not in the squares of board
3. The weaker player take the black marbles and black player plays first
4. Once marbles are placed are not removed
5. If a black or white marble or marbles are surrounded by opposite color marbles will be removed from board as a prisoner
6. The game will end once the players decide that they cannot make any more progress
So the total number of legal positions of a square board of 19×19 are 10^170 which are more than the number of atoms in the universe.
AlphaGo is a computer program that plays the game of Go. AlphaZero was the first Go which has achieved super human status without watching any human game play. The first version of AlphaGo used two neural networks which cooperated to choose its moves, both were convolutional neural networks CNN with twelve layers. CNN are useful for classifying images. They take images as inputs and output class probabilities after been labeled on image dataset. The learn network between inputs and outputs. So in AlphaGo board. The first network is called Policy Network which takes board positions as input and decide best next move as output. These deep minds are made by using the millions of examples of moves made by strong human mind players of AlphaGo. It is fast enough to take one good move but it needs to check thousands of possible moves before making a decision. So they modify the network. So it doesn’t look on entire board of 19×19. Instead a smaller window around the opponent previous move. This make it thousand time faster. The second network is called the Value network. It answers a different question than what move to play next. It actually estimates the chances of each player winning given a board position. In addition to these two networks AlphaGo uses algorithm known as Monte Carlo tree search algorithm. It reads the sequences of future moves effectively.