AlphaGO from Google DeepMind in 2016, beat human grandmasters uses Monte Carlo game tree search Science Breakthrough of the year runner-up Silver et al (2016). Nature, 529:484–489.
Go - the game 19x19, black vs white stones images from https://en.wikipedia.org/wiki/Go_(game) Go - the game 19x19, black vs white stones chess: O(bd) ~ 3580, go: O(bd) ~ 250150 Rule 1: Every stone remaining on the board must have at least one open "point" ("liberty") directly adjacent (up, down, left, or right), or must be part of a connected group that has at least one such open point next to it. Stones or groups of stones which lose their last liberty are removed from the board. Strategies: Connection: Keeping one's own stones connected means that fewer groups need to make living shape, and one has fewer groups to defend. Stay alive: The simplest way to stay alive is to establish a foothold in the corner or along one of the sides. At a minimum, a group must have two eyes (separate open points) to be "alive" liberties capture
AlphaGo implementation uses deep networks (13 layers) to represent "value function" and "policy function" these are for Reinforcement Learning learn the value of good moves or positions by whether they lead to wins (discounted rewards) performs Monte Carlo game search explore state space like minimax random "rollouts" simulate probable plays by opponent according to policy function
AlphaGo implementation hardware: 1920 CPUs, 28O GPUs training of networks: phase 1: supervised learning from database of 30 million human moves phase 2: play against self using reinforcement learning
match against Tang Weixing (2016)