Move prediction in Gomoku using deep learning

The Gomoku board game is a longstanding challenge for artificial intelligence research. With the development of deep learning, move prediction can help to promote the intelligence of board game agents as proven in AlphaGo. Following this idea, we train deep convolutional neural networks by supervised learning to predict the moves made by expert Gomoku players from RenjuNet dataset. We put forward a number of deep neural networks with different architectures and different hyperparameters to solve this problem. With only the board state as the input, the proposed deep convolutional neural networks are able to recognize some special features of Gomoku and select the most likely next move. The final neural network achieves the accuracy of move prediction of about 42% on the RenjuNet dataset, which reaches the level of expert Gomoku players. In addition, it is promising to generate strong Gomoku agents of human-level with the move prediction as a guide.