Improved CNN Model Using Innovative Adaptive-DropMessage for Gomoku Game

Convolutional Neural Networks (CNNs) has been used in many Gomoku AI systems. This paper introduces the innovative Adaptive-Drop Message method within the CNN framework. This method, combined with techniques including attention mechanisms, residual dense blocks, and depth-wise separable convolutions. The Adaptive-Drop Message method dynamically discards portions of feature map nodes based on loss values, finely controlling the activity of neurons. The model does not overly rely on any single feature set while learning critical features on the game board. Experiments were conducted on the A100 GPU platform. Compared to other outstanding Gomoku AI models with equivalent parameter scales and resource consumption, the model proposed performs excellent in terms of winning rates, average winning moves, and average winning time. it is validated that the proposed methods can improve rapid and efficient decision-making.