Research On Command Decision Method From RTS Perspective On Deep Learning

Since its birth,artificial intelligence technology has continued to grow and develop,and has achieved good results in practical applications in many fields.Among them,deep reinforcement learning algorithms with both deep learning perception capabilities and reinforcement learning decision-making capabilities have gradually become a key research direction,and video games have also been increasingly recognized because they can provide a research environment for verifying various artificial intelligence algorithms.Favored by many researchers.As one of many video games,Real-Time Strategy Game(RTS)covers a variety of artificial intelligence research areas,such as resource management,incomplete information games,and long-term planning.At the same time,the complex and changeable environment in real-time strategy games can simulate real-world problems to a certain extent.This property makes it an excellent test platform for studying general artificial intelligence.Therefore,this article uses StarCraft games as the research environment for intelligent command and decision methods."StarCraft ?"(SC?)is a simulated war strategy game,which contains many strategies.These strategies affect the game's trend.In turn,the strategy will continue to change over time and the environment.This requires a certain degree of flexibility and effectiveness in decision-making.The traditional script-based game AI technology obviously cannot meet this need.Therefore,how to make the command and decision system make optimal decisions under complicated and changeable conditions is the main research content of this paper.This article aims to solve the problem of how to make optimal decisions in real-time strategy games,and designs related algorithms.The main work and research results are:1.According to the experimental environment,determine the network model used to obtain global state information.Considering the number and type of each unit in the game,and in order to ensure the integrity and accuracy of the information,by comparing the experimental data of the two networks of Inception V1 and Resnet 152,the network that is most suitable for the global state information acquisition of the experiment in this paper is selected model.2.For 7 mini map tasks,the DQN algorithm is improved.After preliminary iterations,Agents with certain operational capabilities in local tasks are obtained.3.In order to solve the "dimensional disaster" problem caused by the huge strategy space and state space in the experiment,the hierarchical reinforcement learning idea is used to deal with the state space and strategy space in a hierarchical manner.It is verified by experiments that agents using hierarchical task networks have achieved good results in solving high-dimensional decision-making problems.