基于机器学习的RTS游戏实时胜率预测  

RTS game real-time winning rate prediction based on machine learning

作　　者：温叶廷 黄海于[1] Wen Yeting;Huang Haiyu(School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学信息科学与技术学院

出　　处：《信息技术与网络安全》2020年第3期35-39,共5页Information Technology and Network Security

摘　　要：游戏胜负预测可用于自适应游戏AI的设计,策略层面强化学习的反馈参数等。使用SC2LE公开的数据集,首先通过游戏时间、MMR和AMP指标进行数据预处理,得到质量较高的数据集;然后使用pysc2解析,提取游戏数据;最后进行特征分析,得到基础特征和统计特征,完成游戏特征数据集的构建。最终采用机器学习方法XGB分类模型,利用10次十折交叉验证法进行模型评估与优选。结果表明使用基础特性与统计特性的组合,可以使得实时胜率预测准确率在不同匹配对局情况下均超过80%。Game outcome prediction can be used in the design of adaptive game AI and feedback parameters for reinforcement learning at the strategic level.In this paper,the data set disclosed by SC2LE(StarCraft II Learning Environment)is used to preprocess the data through game time,MMR(matchmaking rating),and AMP(actions per minute)indicators to obtain a high-quality data set;then the pysc2 is used to analyze and extract the game data,and finally the feature analysis is carried out to obtain basic features and statistics features and complete the construction of game feature datasets.Finally,the machine learning method XGB classification model is used,and 10 times 10-fold cross-validation method is adopted for model evaluation and optimization.The results show that using the combination of basic characteristics and statistical characteristics,the real-time win rate prediction accuracy rate can exceed 80%under different matching games.

关 键 词：AI 游戏 实时 机器学习 XGB 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]