A Linear Classifier Outperforms UCT in 9x9 Go

The dominant paradigm in computer Go is Monte-Carlo Tree Search (MCTS). This technique chooses a move by playing a series of simulated games, building a search tree along the way. After many simulated games, the most promising move is played. This paper proposes replacing the search tree with a neural network. Where previous neural network Go research has used the state of the board as input, our network uses the last two moves. In experiments exploring the effects of various parameters, our network outperforms a generic MCTS player that uses the Upper Confidence bounds applied to Trees (UCT) algorithm. A simple linear classifier performs even better.