Winner Prediction Of RTS Game Players Based On Deep Learning And Replay Data

Game provides a useful researching and testing platform for Artificial Intelligence(AI)algorithms.Predicting the winner of players in the game by using deep learning algorithms is an important part of AI planning.Real-Time Strategy(RTS)games are popular real-time battle simulation games.Due to the huge state space,limited decision time and dynamic confrontation environment,RTS games provide ideal environment for predicting the winner of players in the game by using deep learning algorithms.Games(especially RTS games)and winner prediction of game players have practical applications in the design of military combat simulation systems.This paper designs encoding methods for two replay datasets:RTS game AI robots which using different search strategy techniques in p RTS(a typical agent-based RTS game simulator)and anonymous game players which provided by SC2LE(StarCraft ?reinforcement learning environment),to generate encoded datasets.This paper introduces the methods of selecting three groups AI robots in ?RTS,generating replay datasets,sampling datasets and encoding states information and actions information at the sampling point using one-hot encoding.This paper contains details of preprocessing replay datasets in StarCraft ?,then grouping them into six combat combinations in three races.For each preprocessed replay data,this paper describes the procedure of analyzing observations information and actions information at all time points by PySC2,filtering effective time point sets,then extracting and encoding spatial features at each effective time point.Two encoded datasets are generated respectively through the encoding methods.This paper uses Convolutional Neural Network(CNN),Multi-Size Convolutional Neural Network(MSCNN),Conditional Neural Process(CNP),Long Short-Term Memory(LSTM)and Bayesian Neural Network(BNN)to predict the winner of players in the game,and calculates the accuracy of prediction,plots ROC curves and calculates the corresponding AUC values to verify the prediction performance of designed methods and analyze the prediction results.The experimental results show that the accuracy of winner prediction using five deep learning algorithms is around 0.8 in RTS game AI robots encoded datasets in ? RTS.In anonymous game players encoded datasets in StarCraft ?,the highest accuracy of winner prediction using five deep learning algorithms is 0.8.The results indicate the feasibility and good performance in winner prediction of RTS game players by using designed encoding methods and deep learning algorithms.The results of research show that RTS games provide a specific applicable platform for studying AI simulation of human decision-making,exploring search algorithms,data analysis,state evaluation and macro management.