基于异构图神经网络的可解释兵棋态势预测方法

陈露; 尚家兴; 刘大江; 张玉芳; 倪晚成

doi:10.16383/j.aas.c240468

基于异构图神经网络的可解释兵棋态势预测方法

doi: 10.16383/j.aas.c240468 cstr: 32138.14.j.aas.c240468

陈露^{1, 2,},
尚家兴^{1, 2,},
刘大江^{1, 2,},
张玉芳^{1, 2,},
倪晚成^{3, 4,}

1.
重庆大学计算机学院重庆 400044
2.
信息物理社会可信服务计算教育部重点实验室重庆 400044
3.
中国科学院自动化研究所北京 100190
4.
中国科学院大学北京 100049

基金项目: 中国科学院自动化研究所复杂系统认知与决策实验室开放基金(CASIA-KFKT-10)资助

详细信息

作者简介:
陈露：重庆大学计算机学院硕士研究生. 主要研究方向为数据挖掘. E-mail: percy_cl@foxmail.com

尚家兴：重庆大学计算机学院教授. 主要研究方向为数据挖掘, 行业大数据分析和可解释人工智能. 本文通信作者. E-mail: shangjx@cqu.edu.cn

刘大江：重庆大学计算机学院副教授. 主要研究方向为可重构计算, 编译优化. E-mail: liudj@cqu.edu.cn

张玉芳：重庆大学计算机学院教授. 主要研究方向为数据挖掘, 集群系统和计算机网络. E-mail: zhangyf@cqu.edu.cn

倪晚成：中国科学院自动化研究所研究员. 主要研究方向为数据挖掘与知识发现, 复杂系统建模, 群体智能博弈决策平台与评估. E-mail: wancheng.ni@ia.ac.cn

计量
- 文章访问数: 16
- HTML全文浏览量: 12
- 被引次数: 0
出版历程
- 收稿日期: 2024-07-01
- 录用日期: 2025-02-14
- 网络出版日期: 2025-04-24

An Interpretable Wargame Situation Prediction Method Based on Heterogeneous Graph Neural Networks

CHEN Lu^{1, 2
,},
SHANG Jia-Xing^{1, 2
,},
LIU Da-Jiang^{1, 2
,},
ZHANG Yu-Fang^{1, 2
,},
NI Wan-Cheng^{3, 4
,}

1.
College of Computer Science, Chongqing University, Chongqing University, Chongqing 400044
2.
Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing 400044
3.
Institute of Automation, Chinese Academy of Sciences, Beijing 100190
4.
University of Chinese Academy of Sciences, Chinese Academy of Sciences, Beijing 100049

Funds: Supported by Open Fund of the Laboratory of Cognition and Decision Making for Complex Systems, Institute of Automation, Chinese Academy of Sciences. (CASIA-KFKT-10)

More Information

Author Bio:
CHEN Lu　Master student at the College of Computer Science, Chongqing University. Her main research interest is data mining

SHANG Jia-Xing　Professor at the College of Computer Science, Chongqing University. His research interest covers data mining, industrial big data analytics, and explainable artificial intelligence. Corresponding author of this paper

LIU Da-Jiang　Associate professor at the College of Computer Science, Chongqing University. His research interest covers reconfigurable computing and compiler optimization

ZHANG Yu-Fang　Professor at the College of Computer Science, Chongqing University. Her research interest covers data mining, cluster systems, and computer networks

NI Wan-Cheng　Researcher at the Institute of Automation, Chinese Academy of Sciences. Her research interest covers data mining and knowledge discovery, complex system modeling, platform and evaluation of swarm intelligence gaming and decision making

摘要

摘要: 复杂多变的现代兵棋模拟中, 精准的战局预测与战场态势解读是提高决策质量的关键. 针对兵棋推演中复杂态势表达困难和模型可解释性不足的挑战, 提出基于异构图神经网络的可解释兵棋预测模型WarGraph, 模型由多关系图建模、时序分析、预测解释三个模块构成. 首先综合复盘数据与先验知识, 将环境与算子之间的多元复杂关系建模为多关系异构图, 从而捕捉作战单元之间以及与环境的复杂交互关系, 实现复杂推演态势的表征; 然后利用Transformer时序分析方法, 动态捕捉整体态势演变, 并通过注意力机制抽取关键决策时刻. 该模型不仅能在复盘推演中精准预测战局胜负, 注意力机制的引入能更好地解释决策中的关键因素. 以“庙算·智胜”实时兵棋对抗平台2021年的108场陆战对局复盘数据作为实验数据集, 结果显示本文提出的模型预测准确率可达90.91%, 相比其他模型提高大约9.09%, 通过对注意力系数的可视化分析, 模型在决策过程中捕捉到关键时刻, 进一步验证模型的可解释性.
- 兵棋推演 /
- 态势预测 /
- 图神经网络 /
- 可解释分析 /
- 深度学习
Abstract: In complex and changeable modern wargame simulations, accurate battlefield situation prediction and interpretation are crucial for high-quality decision-making. To address the challenges of difficult expression of complex situations and insufficient model interpretability in wargame deduction, this paper proposes an interpretable wargame prediction model WarGraph based on heterogeneous graph neural networks. The model consists of three modules: Multi-relational graph modeling, temporal analysis, and interpretable prediction. We first combine replay data with prior knowledge to construct a multi-relational heterogeneous graph, effectively modeling the intricate relationships between the environment and the operators. This enables capturing the complex interactions between combat units and the environment, realizing the representation of complex deduction situations. Then by leveraging Transformer-based temporal analysis, we dynamically track the overall situation evolution and use attention mechanisms to identify key decision-making moments. This model can not only accurately predict the outcome of battles in wargame replays, but also the introduction of the attention mechanism enables a better explanation of the key factors in decision-making. Using replay data of 108 matches from the “MiaoSuan·ZhiSheng” wargame platform in 2021, the results show that the proposed model achieves a prediction accuracy of up to 90.91%, about 9.09% higher than the baseline models. Visualization of the attention coefficients demonstrates that the model captures critical moments in the decision-making process, which further validates its interpretability.
- Wargame deduction /
- situation prediction /
- graph neural network /
- interpretable analysis /
- deep learning

HTML全文

图 1 某一时刻兵棋战场

Fig. 1 Wargame battlefield at a specific moment

下载: 全尺寸图片幻灯片

图 2 不同阶段特征相关性变化

Fig. 2 Changes in Feature Correlations at different stages

下载: 全尺寸图片幻灯片

图 3 四种关系图描述

Fig. 3 Graph description of four relations

下载: 全尺寸图片幻灯片

图 4 通视同构图

Fig. 4 Intervisibility homogenous graph

下载: 全尺寸图片幻灯片

图 5 基于GCN的通视关系表征学习模块

Fig. 5 GCN-based intervisibility relation representation learning module

下载: 全尺寸图片幻灯片

图 6 侦察图表征学习模块

Fig. 6 Scout graph representation learning module

下载: 全尺寸图片幻灯片

图 7 打击图表征编解码器模块

Fig. 7 Encoder/decoder module of attack graph representation

下载: 全尺寸图片幻灯片

图 8 动态演化模式学习框架

Fig. 8 Dynamic evolution pattern learning framework

下载: 全尺寸图片幻灯片

图 9 部分比赛关键时刻权重分布

Fig. 9 Weight distribution at critical moments of partial wargames

下载: 全尺寸图片幻灯片

图 10 四场比赛对应的注意力权重

Fig. 10 Attention weights corresponding to four matches

下载: 全尺寸图片幻灯片

图 11 关键时刻趋势图

Fig. 11 Trend chart at critical moments

下载: 全尺寸图片幻灯片

图 12 图嵌入维度及迭代次数敏感性

Fig. 12 Graph representation dimensionality and epoch sensitivity

下载: 全尺寸图片幻灯片

图 13 激活函数—批大小—学习率敏感性

Fig. 13 Activate function-batch-LR sensitivity

下载: 全尺寸图片幻灯片

图 14 不同Epoch下各模型变体性能

Fig. 14 Model variants performance under different epochs

下载: 全尺寸图片幻灯片

表 1 六角格节点属性

Table 1 Features of hexagonal lattice node

参数名	数据类型	说明
pos	整型	4位整数坐标
elev	整型	高程
node_id	整型	ID
cond	整型	地形
can_hide	整型	是否可掩蔽
has_road	整型	有无道路
has_river	整型	有无河流

下载: 导出CSV

表 2 算子节点属性

Table 2 Features of operator node

参数名	数据类型	说明
op_id	整型	算子ID
color	整型	算子阵营
type	整型	算子类型
sub_type	整型	算子细分类型
basic_speed	整型	基础速度
armor	整型	装甲类型
speed	整型	当前机动速度

下载: 导出CSV

表 3 不同LR和epoch下的对比实验结果

Table 3 Comparative experimental results under different LR and epoch settings

LR	模型	Accuracy (epoch = 20)	Accuracy (epoch = 40)	Accuracy (epoch = 60)
0.01	WarGraph	81.82%	86.36%	90.91%
	Trans-CNN	54.55%	77.27%	72.73%
	Trans-MLP	45.45%	81.82%	81.82%
	LSA-Trans	72.73%	68.18%	77.27%
	LSA-CNN	81.82%	77.27%	63.64%
	LSA-MLP	36.36%	81.82%	81.82%
	ESA-Trans	68.18%	54.55%	50.00%
	ESA-CNN	50.00%	63.64%	63.64%
	ESA-MLP	50.00%	50.50%	55.00%
0.005	WarGraph	81.82%	90.91%	86.36%
	Trans-CNN	77.28%	81.82%	72.73%
	Trans-MLP	68.18%	81.82%	81.82%
	LSA-Trans	68.18%	77.27%	81.82%
	LSA-CNN	63.64%	81.82%	77.27%
	LSA-MLP	54.54%	81.82%	81.82%
	ESA-Trans	54.54%	40.90%	45.45%
	ESA-CNN	63.64%	54.55%	63.64%
	ESA-MLP	36.36%	50.00%	54.55%

下载: 导出CSV

表 4 超参数选择

Table 4 Selection of hyperparameters

超参数	选取范围
图融合表征维度	16, 32, 64, 128, 256
激活函数	ReLU, sigmoid, tanh
批大小	8, 16, 32
学习率	0.001, 0.005, 0.01, 0.05
迭代次数	10, 20, 30, 40, 50, 60

下载: 导出CSV

参考文献(35)

[1]	Mittalf V, Davidson A. Combining wargaming with modeling and simulation. to project future military technology requirements. IEEE Transactions on Engineering Management, 2020, 68(4): 1195−1207
[2]	Ouriques L, Barbosa C E, Xexéo G. Understanding military collaboration in wargames. In: Proceedings of 2023 26th International Conference on Computer Supported Cooperative Work in Design (CSCWD), Piscataway, USA: IEEE Press, 2023. 1920−1925
[3]	Li L, Yang R, Li H, Lv M, Yue L, Wu A. Unsupervised contrastive learning for automatic grouping of aerial swarms. IEEE Transactions on Vehicular Technology, doi: 10.1109/TVT.2023.3342186
[4]	Rosen A M, Kerr L. Wargaming for learning: how educational gaming supports student learning and perspectives. Journal of Political Science Education, doi: 10.1080/15512169.2024.2304769
[5]	姚昌华, 毕珊宁, 马茹飞, 余晓晗, 李家强, 陈金立. 兵棋智能体兵力协同动态联盟形成方法. 系统仿真学报, doi: 10.16182/j.issn1004731x.joss.24-0058 Yao Chang-Hua, Bi-Shan-Ning, Ma Ru-Fei, Yu Xiao-Han, Li Jia-Qiang, Chen Jin-Li. Method for dynamic coalition formation of wargame agent for force cooperation. Journal of System Simulation, 2024, doi: 10.16182/j.issn1004731x.joss.24-0058
[6]	胡晓峰, 齐大伟. 智能化兵棋系统: 下一代需要改变的是什么. 系统仿真学报, 2021, 33(9): 1997−2009 Hu Xiao-Feng, Qi Da-Wei. Intelligent wargaming system:change needed by next generation need to be changed. Journal of System Simulation, 2021, 33(9): 1997−2009
[7]	胡晓峰, 贺筱媛, 陶九阳. AlphaGo的突破与兵棋推演的挑战. 科技导报, 2017, 35(21): 49−60 Hu Xiao-Feng, He Xiao-Yuan, Tao Jiu-Yang. AlphaGo's breakthrough and challenges of wargaming. Science & Technology Review, 2017, 35(21): 49−60
[8]	曹占广, 陶帅, 胡晓峰, 等. 国外兵棋推演及系统研究进展. 系统仿真学报, 2021, 33(9): 2059−2065 Cao Zhan-Guang, Tao Shuai, Hu Xiao-Feng, et al. Research progress of wargaming simulation methods and system abroad. Journal of System Simulation, 2021, 33(9): 2059−2065
[9]	吴琳, 胡晓峰, 陶九阳, 等. 面向智能成长的兵棋推演生态系统. 系统仿真学报, 2021, 33(9): 2048−2058 Wu Lin, Hu Xiao-Feng, Tao Jiu-Yang, et al. Wargaming eco-system for intelligence growing. Journal of System Simulation, 2021, 33(9): 2048−2058
[10]	刘海洋, 唐宇波, 胡晓峰, 乔广鹏. 基于兵棋推演的联合作战方案评估框架研究. 系统仿真学报, 2018, 30(11): 4115−4123 Liu Hai-Yang, Tang Yu-Bo, Hu Xiao-Feng, Qiao Guang-Peng. Research on evaluation framework of COA based on wargaming. Journal of System Simulation, 2018, 30(11): 4115−4123
[11]	刘海洋, 唐宇波, 胡晓峰, 刘戎翔, 崔文华. 面向联合作战评估的兵棋推演实验研究. 指挥与控制学报, 2018, 4(4): 272−280 doi: 10.3969/j.issn.2096-0204.2018.04.0272 Liu Hai-Yang, Tang Yu-Bo, Hu Xiao-Feng, Liu Rong-Xiang, Cui Wen-Hua. Wargaming experiment oriented evaluation of joint operations. Journal of Command and Control, 2018, 4(4): 272−280 doi: 10.3969/j.issn.2096-0204.2018.04.0272
[12]	邢思远, 倪晚成, 张海东, 等. 基于兵棋复盘数据的武器效用挖掘. 指挥与控制学报, 2020, 6(2): 132−140 doi: 10.3969/j.issn.2096-0204.2020.02.0132 Xing Si-Yuan, Ni Wan-Cheng, Zhang Hai-Dong, et al. Mining of weapon utility based on the replay data of war-game. Journal of Command and Control, 2020, 6(2): 132−140 doi: 10.3969/j.issn.2096-0204.2020.02.0132
[13]	尹奇跃, 赵美静, 倪晚成, 张俊格, 黄凯奇. 兵棋推演的智能决策技术与挑战. 自动化学报, 2023, 49(05): 913−928 Yin Qi-Yue, Zhao Mei-Jing, Ni Wan-Cheng Zhang Jun-Ge, Huang Kai-Qi. Intelligent decision making technology and challenge of wargame. Acta Automatica Sinica, 2023, 49(05): 913−928
[14]	Chen L, Zhang Y, Feng Y, Zhang L, Liu Z. A human-machine agent based on active reinforcement learning for target classification in wargame. IEEE Transactions on Neural Networks and Learning Systems, doi: 10.1109/TNNLS.2023.3236944
[15]	周雷, 尹奇跃, 黄凯奇. 人机对抗中的博弈学习方法. 计算机学报, 2022, 45(09): 1859−1876 doi: 10.11897/SP.J.1016.2022.01859 Zhou Lei, Yin Qi-Yue, Huang Kai-Qi. Game-theoretic learning in human-computer gaming. Chinese Journal of Computers, 2022, 45(09): 1859−1876 doi: 10.11897/SP.J.1016.2022.01859
[16]	Sun Y, Yuan B, Xiang Q, Zhou J, Yu J, Dai D, Zhou X. Intelligent decision-making and human language communication based on deep reinforcement learning in a wargame environment. IEEE Transactions on Human-Machine Systems, 2022, 53(1): 201−214
[17]	刘满, 张宏军, 程恺, 郝文宁, 王之腾. 知识与数据互补的战术级兵棋行为决策框架设计与实现. 指挥与控制学报, 2023, 9(02): 182−191 doi: 10.3969/j.issn.2096-0204.2023.02.0182 Liu Man, Zhang Hong-jun, Cheng Kai, Hao Wen-Ning, Wang Zhi-Teng. Framework design and application for tactical-level wargame behavior decision-making based on complementary knowledge and data. Journal of Command and Control, 2023, 9(02): 182−191 doi: 10.3969/j.issn.2096-0204.2023.02.0182
[18]	Xue Y, et al. Multi-attribute decision-making in wargames leveraging the entropy–weight method in conjunction with deep reinforcement learning. IEEE Transactions on Games, doi: 10.1109/tg.2023.3236065
[19]	许霄, 李东, 郭圣明, 吴琳, 胡晓峰. 面向联合作战兵棋推演的智能决策建模框架. 指挥与控制学报, 2023, 9(04): 449−456 doi: 10.3969/j.issn.2096-0204.2023.04.0449 Xu Xiao, Li Dong, Guo Sheng-Ming, Wu Lin, Hu Xiao-Feng. An intelligent decision-making modeling architecture for joint operation-oriented wargaming. Journal of Command and Control, 2023, 9(04): 449−456 doi: 10.3969/j.issn.2096-0204.2023.04.0449
[20]	Chen L, et al. Online intention recognition with incomplete information based on a weighted contrastive predictive coding model in wargame. IEEE Transactions on Neural Networks and Learning Systems, doi: 10.1109/TNNLS.2022.3144171
[21]	Sun D, et al. Relation extraction on the wargame based on pre-trained models. In: Proceedings of the 9th International Conference on Dependable Systems and Their Applications (DSA), IEEE, 2022
[22]	Xuan L, et al. Tactical intention recognition in wargame. In: Proceedings of the 6th International Conference on Computer and Communication Systems (ICCCS), IEEE, 2021
[23]	孙宇祥, 赵俊杰, 解宇轩, 喻车澄, 周献中. 自生成兵棋AI: 基于大语言模型的双层Agent任务规划. 控制与决策, 2024, 39(12): 3927−3936 Sun Yu-Xiang, Zhao Jun-Jie, Xie Yu-Xuan, Yu Che-Cheng, Zhou Xian-Zhong. Self generated wargame AI:double layer agent task planning based on large language mode. Control and Decision, 2024, 39(12): 3927−3936
[24]	Dong L, Li N, Yuan H, Gong G. Accelerating wargaming reinforcement learning by dynamic multi-demonstrator ensemble. Information Sciences, 2023
[25]	Sun Y, Yu J, Li Y, Zhou X. Predicting wargame outcomes and evaluating player performance from an integrated strategic and operational perspective. IEEE Transactions on Games, 2024
[26]	Ma R, Liu S, Xu Z, Zhang Y, Ni Y. Research on fuzzy dynamic route choice model and algorithm of wargame. International Journal of Machine Learning and Cybernetics, 20241−18
[27]	Wang T, et al. MFSCD* Lite: a situation-based path planning algorithm for agent in wargame. In: Proceedings of the 6th International Symposium on Computer Science and Intelligent Control (ISCSIC), IEEE, 2022
[28]	Jensen B, Valeriano B, Whitt S. How cyber operations can reduce escalation pressures: evidence from an experimental wargame study. Journal of Peace Research, 2024
[29]	黄洪原. 浅议兵棋系统的发展及军事运用. 中国军转民, 202253−55 doi: 10.3969/j.issn.1008-5874.2022.04.026 Huang Hong-Yuan. A brief discussion on the development and military applications of wargaming systems. Defence Industry Conversion in China, 202253−55 doi: 10.3969/j.issn.1008-5874.2022.04.026
[30]	张大永, 杨镜宇, 马骏, 宋晨烨. 面向兵棋推演复盘分析的机器学习数据集构建. 系统仿真学报, 2024, 36(03): 608−624 Zhang Da-Yong, Yang Jing-Yu, Ma Jun, Song Chen-Ye. Construction of machine learning data set for the analysis of the replay of the wargaming. Journal of System Simulation, 2024, 36(03): 608−624
[31]	余晓晗, 王启迪, 于坤. 基于图神经网络GraphVAE的兵棋态势预测方法. 指挥控制与仿真, 2023, 45(5): 129−136 doi: 10.3969/j.issn.1673-3819.2023.05.018 Yu Xiao-Han, Wang Qi-Di, Yu Kun. A wargame situation prediction method based on a graph neural network GraphVAE. Command Control & Simulation, 2023, 45(5): 129−136 doi: 10.3969/j.issn.1673-3819.2023.05.018
[32]	陈锦阳, 郝文宁, 靳大尉, 陈刚, 余晓晗. 基于注意力机制的兵棋对抗态势预测方法. 火力与指挥控制, 2023, 48(3): 18−24 doi: 10.3969/j.issn.1002-0640.2023.03.003 Chen Jin-Yang, Hao Wen-Yu, Jin Da-Wei, Chen Gang, Yu Xiao-Han. Research on the prediction method of wargame confrontation situation based on attention mechanism. Fire Control & Command Control, 2023, 48(3): 18−24 doi: 10.3969/j.issn.1002-0640.2023.03.003
[33]	Kipf T, Welling M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv: 1609.02907, 2016.
[34]	Hamilton W, YING Z; Leskovec J. Inductive representation learning on large graphs. In: Proceedings of Neural Information Processing Systems, 2017.
[35]	Vaswani A, et al. Attention is all you need. In: Proceedings of Neural Information Processing Systems, 2017.