Deep Learning Approaches with Attention Mechanisms for Go Strength Imitation and Style Recognition

With the advancement of AI technology, imitating human intelligence and endowing machines with personalities have become crucial research areas. This study aims to enhance the performance of Go AI through techniques such as CNN, ResNet, and attention mechanisms, mimicking different levels of playing strength and imitating machines with unique play styles. This research focuses on the algorithms and model architectures for strength imitation and play style recognition in Go. Factors such as move order, board state, and liberties are considered for strength imitation, employing CNN with residual blocks, and introducing Channel Attention and Spatial Attention mechanisms. For play style recognition, a similar model architecture is used with the addition of Spatial Dropout to prevent overfitting. The results show that incorporating liberties as input yields better performance in strength imitation and play style recognition. The final model achieved approximately 56% accuracy in strength imitation with a Top-5 accuracy of 86% and 82% accuracy in play style recognition.