摘要
为了充分捕获算子之间复杂的动态时空依赖性,理解和量化态势预测的不确定性,帮助决策人员做出正确的决策,该文提出了一种新的模型,称为不确定性感知的动态多图态势序列预测方法(Uncertainty-aware Prediction Method of Wargame Situation Based on Dynamic Multi-Graph Convolution,UaDMG),它联合多变量时空序列预测及其不确定性估计。首先构建了融合不同类型算子动态特征、地图环境信息和算子实时空间临近性的动态多图生成器;然后利用动态多图卷积捕捉算子之间的动态空间相关性,利用空洞因果卷积捕获游戏发展的时间特征;最后分别通过贝叶斯分类方法和Monte Carlo Dropout近似贝叶斯网络的方法来分别量化偶然不确定性(Aleatoric Uncertainty,AU)和认知不确定性(Epistemic Uncertainty,EU),产生了UaDMG。在多个真实不同场景的兵棋复盘数据中的实验结果表明,该模型的预测精度不仅优于现有的模型方法,而且能够通过量化不确定性来提高预测的可靠性。
To fully capture the complex dynamic spatiotemporal dependencies between operators,understand and quantify the uncertainty of situation prediction,and assist decision-makers in making correct decisions,we propose a new model called the Uncertainty-aware Prediction Method of Wargame Situation Based on Dynamic Multi-Graph Convolution(UaDMG),which integrates multivariate STS prediction and its uncertainty estimation.First,a dynamic multi-graph generator is constructed to fuse operators dynamics,map environmental information,and real-time spatial proximity.Then,dynamic multi-graph convolution is used to capture the dynamic spatial correlation,and dilated causal convolution(DCCN)is used to capture the temporal characteristics.Finally,Bayesian classification methods and Monte Carlo Dropout approximations of Bayesian networks are used to quantify aleatoric uncertainty(AU)and epistemic uncertainty(EU),respectively,resulting in UaDMG.Experimental results on multiple real-world and diverse scenario wargame replay datasets show that the proposed model is not only superior to existing model methods in terms of predictive accuracy but also enhances the reliability of predictions through the quantification of uncertainty.
作者
孙菲艳
郝文宁
靳大尉
邹傲
张燕
曲爱妍
SUN Fei-yan;HAO Wen-ning;JIN Da-wei;ZOU Ao;ZHANG Yan;QU Ai-yan(School of Command and Control Engineering,Army Engineering University of PLA,Nanjing 210000,China;School of Software Engineering,Jinling Institute of Technology,Nanjing 211199,China)
出处
《计算机技术与发展》
2025年第8期119-127,共9页
Computer Technology and Development
基金
国家自然科学基金资助项目(61806221)
国防基础科研计划资助项目(JCKY2020601b018)。
关键词
兵棋
态势预测
动态多图
不确定性估计
贝叶斯
wargame
situation prediction
dynamic multi-graph
uncertainty estimation
Bayesian
作者简介
孙菲艳(1991-),女,博士研究生,讲师,CCF专业会员(R3817M),研究方向为时空序列预测、数据工程;通讯作者:郝文宁(1971-),男,教授,博导,研究方向为军用数据、知识工程、文本处理。
