A Study Of Player Winners Prediction In RTS Games Based On Machine Learning And Replay Data

Real-time strategy(RTS)game is a kind of fast-paced war simulation game,in which players make real-time decisions by managing resource and control units.Due to the uncertainty of the opponent unit position and the dynamic nature of the game environment,the characteristics of games are very complex,which provides an ideal platform for the study of game-based artificial intelligence(AI).One of the main research areas is RTS game players winner prediction.This paper firstly describes the process of generating replay data sets and sampling replay data for RTS game AI robots with different search strategies based on μRTS,then introduces the data structure of sampling point data.Secondly we designed a method of encoding the state information and action information in sampling point data.Thirdly,the coding process and the process of RTS game AI robot winner prediction in the competition with CNN,SVM and KNN machine learning algorithms are described in detail.Finally,this paper introduces the processing methods of replay data sets provided by StarCraft Ⅱ learning environment SC2 LE using PySC2,including data acquisition,scaling,filtering,parsing,sampling,feature extraction and coding.We describes the process of StarCraft Ⅱ game players winner prediction by three optimized machine learning algorithms on the coded data sets.At the same time,it was analyzed and comparedIn the experiment of this paper,we calculate the accuracy of different machine learning algorithms in RTS game players winner prediction,draw the ROC curve and calculate the corresponding AUC value,so as to evaluate the performance of different machine learning algorithms in RTS game players winner prediction.