系统仿真学报 ›› 2022, Vol. 34 ›› Issue (10): 2181-2193.doi: 10.16182/j.issn1004731x.joss.21-0429

• 仿真建模理论与方法 • 上一篇    下一篇

基于卷积神经网络的陆战兵棋战术机动策略学习

徐佳乐1,2(), 张海东2,3, 赵东海4, 倪晚成2,3()   

  1. 1.中国科学院大学 人工智能学院, 北京 100049
    2.中国科学院 自动化研究所, 北京 100190
    3.中国科学院 人工智能创新研究院, 北京 100190
    4.国防大学 联合作战学院, 河北 石家庄 050000
  • 收稿日期:2021-05-13 修回日期:2021-07-20 出版日期:2022-10-30 发布日期:2022-10-18
  • 通讯作者: 倪晚成 E-mail:xujiale2020@ia.ac.cn;wancheng.ni@ia.ac.cn
  • 作者简介:徐佳乐(1999-),女,硕士生,研究方向为人工智能理论与方法。E-mail:xujiale2020@ia.ac.cn
  • 基金资助:
    国家自然科学基金(61906197);中国科学院战略性先导科技专项资助(XDA27000000)

Tactical Maneuver Strategy Learning from Land Wargame Replay Based on Convolutional Neural Network

Jiale Xu1,2(), Haidong Zhang2,3, Donghai Zhao4, Wancheng Ni2,3()   

  1. 1.School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China
    2.Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
    3.Innovation Academy for Artificial Intelligence, Chinese Acdemy of Sciences, Beijing 100190, China
    4.Joint Operations College, National Defense University, Shijiazhuang 050000, China
  • Received:2021-05-13 Revised:2021-07-20 Online:2022-10-30 Published:2022-10-18
  • Contact: Wancheng Ni E-mail:xujiale2020@ia.ac.cn;wancheng.ni@ia.ac.cn

摘要:

针对从“人在回路”兵棋推演的复盘数据中提取推演者战术经验高价值知识的问题,提出一种基于深度神经网络从复盘数据中学习战术机动策略模型的方法。将战术机动策略建模为在当前态势特征影响下对目标候选位置进行优选的分类问题:梳理总结影响推演者决策的关键认知因素,定义了由机动范围和观察范围等7个属性构成的基础态势特征建立了带有正负样本标注的态势特征数据集;设计了基于卷积神经网络的分类器,以分类概率实现了单个棋子战术机动终点位置的预测。实验结果表明:该模型的预测准确率可达到78.96%,相比其他模型提高至少4.59%。

关键词: 兵棋推演, 复盘数据, 战术机动策略, 态势特征, 卷积神经网络

Abstract:

Aiming at collecting the high valuable knowledge of action decisions in "man-in-the-loop" wargame's replay data, a method of using convolutional neural network to learn the tactical maneuver strategy model from the replay data of wargame is proposed. In this method, the tactical maneuver strategy is modeled as a classification problem of making a good choice from the target candidate locations under the influence of current situation. The key factors affecting commander's decision-making are summarized, and the basic situation features are defined, which are composed of seven attributes such as "maneuverability range and observation range". The feature dataset with positive and negative labels is established. The classifier based on convolutional neural network is designed, which can predict the maneuver terminal position of a single piece by the classification probability. Experimental results show that the prediction accuracy of the tactical maneuver strategy model based on the convolutional neural network is up to 78.96%, which is improved by at least 4.59% compared with other models.

Key words: wargame, replay data, tactical maneuver strategy, situation feature, convolutional neural network

中图分类号: