中圖分類號(hào)：TP391
文獻(xiàn)標(biāo)識(shí)碼：A
DOI： 10.19358/j.issn.2096-5133.2020.03.007
引用格式：溫葉廷，黃海于.基于機(jī)器學(xué)習(xí)的RTS游戲?qū)崟r(shí)勝率預(yù)測(cè)［J］.信息技術(shù)與網(wǎng)絡(luò)安全，2020,39（3）：35-39.

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

Abstract： 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.

Key words : AI;game;real time;machine learning;XGB