【作者】 范腾； 邵坤； 唐振韬； 赵冬斌； 庞中华；

【Author】 Teng Fan;Kun Shao;Zhentao Tang;Dongbin Zhao;Zhonghua Pang;The College of Electrical and Control Engineering,North China University of Technology;The State Key Laboratory of Management and Control for Complex Systems,Institute of Automation,Chinese Academy of Sciences;University of Chinese Academy of Sciences;

【机构】 北方工业大学； 中国科学院自动化研究所复杂系统管理与控制国家重点实验室； 中国科学院大学；

【摘要】 本文针对不完全信息下的决策问题,提出一种基于无监督学习的决策方法。我们选择著名的即时战略游戏星际争霸2作为研究环境,介绍一种分析游戏录像的方法,利用无监督学习分析即时战略游戏中的宏观决策。首先,我们定义一个能够准确描述游戏状态的特征向量,为不同的宏观决策采取不同的提取方式构建数据集;然后,利用K均值聚类算法训练样本并得到分析结果;最后,基于分析结果提取的人类专家宏观决策经验构造智能体,并与游戏内置AI(Artificial Intelligence)对抗。结果证明,智能体宏观决策接近了人类玩家排名系统中前40%的水平。更多还原

【Abstract】 Aiming at the problem of decision-making under incomplete information, this paper proposes a decision method based on unsupervised learning. We focus on the famous real-time strategy game StarCraft II as the learning environment, and introduce a method of analyzing game replays with unsupervised learning to explore the decision-making in macro-management.Firstly, we define a feature vector to accurately describe the state of the game, and use different extraction methods for different macro-management to build a dataset. Then we train the samples to obtain experimental results. Finally, we construct an agent based on the macro-management experiences from human experts, to play against the build-in game AI. Experimental results show that the macro-management ability of the agent reaches the level of top 40% in human players.更多还原