Influence Function-based Convolutional Neural Network for Move Prediction in Go

For improving the accuracy of the move prediction based on a convolutional neural network in Go, a move prediction method was presented based on board features created by influence functions. First, the influence of a current board was computed by means of an influence function. Second, regions controlled by both players were recognized according to the given threshold and the corresponding feature maps were created. Finally, all features including stone configuration were used for training the convolutional neural network. Results show that the new method combined with influence functions can improve the accuracy of move prediction in Go, and enhance the strength of Go playing.