Research And Implementation Of Starcraft ? Agent Based On Artificial Intelligence

In recent years,with the development of artificial intelligence technology,artificial intelligence technology has achieved achievements in many game fields beyond the level of the top human players.However,it did not achieve satisfactory results in the StarCraft real-time strategy game.Because StarCraft is a complex system problem consisting of incomplete information game,huge action space,real-time demand,long-term planning and multi-agent collaboration.At present,many of the outstanding StarCraft II agents rely on hard-coded strategies,and can hardly make appropriate adjustments depending on the strategies adopted by the opponents.This topic uses the deep learning algorithm to learn the player's production strategy from the game play of the top players,and applies the learned production strategy to the existing complex agents to improve their competitiveness.This paper first constructs a data set from more than 1.2 million game playbacks provided by SC2LE(StarCraft II Learning Environment)through processes such as preprocessing,parsing,and extracting game features.Then,using the LSTM network to train the network model for predicting the next generation of human activities against the Terran,Zerg,and Protoss,the accuracy rates of 80.19%,80.05%,and 79.58% were obtained.Then,the trained network model of the predicted human race against the next generation of various family actions is integrated into the StarCraft II open source agent,CommandCenter,to improve its original Production Manager module and implement its unfinished BuildOrder Prediction module,to form a new agent,newbot,that uses the human race.Finally,we submitted newbot to participate in the StarCraft 2 AI Ladder,a ladder race for the StarCraft II AI,to test its effectiveness.By the end of the seventh season of StarCraft 2 AI Ladder,newbot had the eighth place in 40 participating agents.This paper first introduces the research significance,background,domestic and international research status of the StarCraft II agent and the main content of the paper and the chapter arrangement of the paper.Then it introduces the game feature data set based on SC2 LE in detail;then introduces an algorithm that predicts the next step of Terran against various ethnic groups.The next step is to produce an action algorithm.Finally,the method of integrating the trained model into the agent and the test results against other agents are introduced,and the work of the thesis is summarized,the future research work is prospected.