现有基于贝叶斯网络的威胁评估采用专家经验确定的朴素结构,其推理评估结果精度欠佳.为此,提出一种融合专家经验与数据观测的基于Stacking策略的集成贝叶斯网络(ensemble Bayesian network,EBN).首先使用不同搜索空间内的评分优化算法获得数据观测模型集并进行模型平均;然后使用专家经验朴素模型对平均网络进行修剪,形成威胁约束集合;最后以动态规划为基础,通过该集合限制节点序图扩展,以求取全局最优威胁评估网络.在作战想定中,EBN模型单目标威胁概率推理精度比朴素贝叶斯模型高出10%,在多目标威胁排序任务中,其Spearman系数分布亦优于朴素模型.
The existed threat assessment based on Bayesian networks adopts a naive structure determined by expert experience,and its inference evaluation results have poor accuracy.Thus,a Stacking strategy based ensemble Bayesian network(EBN)is proposed that integrates expert experience and data observation.Firstly,scoring optimization algorithms in different search spaces are used to obtain the data observation model set and perform model averaging.Then,expert empirical naive models are used to prune the average network to form a set of threat constraints.Finally,based on dynamic programming,the set is used to limit the expansion of the node order graph,with the aim of obtaining the global optimal threat assessment network.In combat scenarios,the EBN model has a 10% higher inference accuracy for single target threat probability than the naive Bayesian model,and the Spearman coefficient distribution is also better than the naive model in multi target threat ranking tasks.