A harder kind of game

The Atari system struggled when a good move paid off much later. DeepMind chose as its next challenge.

Players place their stones where the board's lines cross. A full board has 19 horizontal lines and 19 vertical lines, making 361 intersections where a stone can go. The spaces between those lines form 18 by 18 squares, but Go counts the intersections, not the squares.

Players try to surround empty territory. They can also capture opposing stones by surrounding them. The rules are short. Choosing a good move is not.

Go gave a computer two problems.

First, there were too many moves to check. A chess player usually chooses among about 35 legal moves. In Go, the number is closer to 250. Looking only four moves ahead creates nearly 4 billion possible paths.

Second, there was no simple way to judge who was ahead. There are no queens or pawns with fixed values. A stone matters because of its place among all the others. Strong players could recognise a promising position without being able to turn that judgement into a tidy rule.

To tackle Go, DeepMind needed both search and learned judgement.