Sarkozy, Gabor N
Institute for Computer Science and Control, Hungar
Deep networks have been proved to be useful in predicting moves of human Go experts. Combining Upper Confidence bounds applied to Trees (UCT) with a large deep network creates an even more powerful AI in playing Go. Our project introduced a new feature, board patterns at the end of a game, used as inputs to the network. By adding the new feature, our model correctly predicted the experts’ move in 18% of the positions, compared to 6% without this feature. In practice, although the board pattern at the end of the game is invisible to Go players until the end of the game, collecting statistics during each simulation in UCT can approximate the board pattern at the end of game. With the dataset generated by UCT simulation, the network correctly predicted the experts’ move in 9% of positions.
Worcester Polytechnic Institute
Major Qualifying Project
All authors have granted to WPI a nonexclusive royalty-free license to distribute copies of the work, subject to other agreements. Copyright is held by the author or authors, with all rights reserved, unless otherwise noted.