high performance computing on graphics processing units: hgpu.org

The purpose of this paper is to introduce the use of convolutional neural network for prediction of the next appropriate move in the Go game. The paper contains description of the crucial Go game rules, neural networks theory, description of implemented programs and final evaluation of the trained neural networks. The programs were implemented with programming language C++ using Caffe framework for the initialization and management of predesigned convolutional neural network models. The thesis compares various models and ways of neural networks learning. The outcome of the experiments was poor; nevertheless, the analysis revealed the main root for this, which was insufficient hardware power. Changes were proposed, probably leading to successful neural network, predicting appropriate moves in the Go game. Despite the imperfections, the experiments proved convolutional neural networks are applicable for the next step prediction in the Go game if the training process is performed properly.