摘要
目前传统卷积网络在爱恩斯坦棋中的运用已颇显成效,但存在着训练速度慢,在浅层次的卷积中无法关注到全局信息的缺点,通过改进深度学习算法和使用GNN取代卷积神经网络(CNN),发现可以显著提升模型性能。研究方法包括将爱恩斯坦棋的棋盘和移动规则表示为图结构,构建GNN以在较浅层次中捕捉局部与全局特征。同时结合蒙特卡洛树搜索(monte carlo tree search,MCTS),通过神经网络的策略头和价值头,提供行动决策和局势评估。实验中,将改进后的GNN算法与传统CNN算法在多轮自对弈中进行对比,结果显示,GNN在局势预测、策略控制及训练效率方面均优于CNN,随着训练次数的增加,该方法在效率提升方面表现出更显著的优势。GNN的应用提升了爱恩斯坦棋博弈模型的效率与策略能力,为进一步探索GNN在完美信息博弈中的潜在价值提供了理论支持和实践基础。
This paper investigates the application of Graph Neural Networks(GNN)in Einstein Chess by enhancing deep learning algorithms and replacing Convolutional Neural Networks(CNN)with GNN to improve model performance.The board and movement rules of Einstein Chess are represented as a graph structure and a GNN is built to capture both local and global features.Additionally,Monte Carlo Tree Search(MCTS)is combined to provide decision-making and position evaluation through the policy and value heads of the neural network.In experiments,the improved GNN algorithm is compared with the CNN algorithm in self-play.Results show GNN performs better in position prediction,strategy control,and training efficiency,especially with increased training rounds.It is concluded that GNN enhances the efficiency and strategic capability of the Einstein Chess model,providing theoretical support and a practical guide for exploring GNN’s potential value in perfect information games.
作者
王志明
胡洋成
蔡彪
陈宣儒
李欣蕊
WANG Zhiming;HU Yangcheng;CAI Biao;CHEN Xuanru;LI Xingrui(College of Computer Networks and Security,Chengdu University of Technology,Chengdu 610059,China;College of Industrial Technology,Chengdu University of Technology,Yibin 644000,China;College of Metallurgy and Energy,North China University of Science and Technology,Tangshan 063210,China)
出处
《重庆理工大学学报(自然科学)》
北大核心
2025年第8期111-117,共7页
Journal of Chongqing University of Technology:Natural Science
基金
国家自然科学基金项目(2019JDR0117)。
