基于深度贝叶斯网络学习的不确定性建模方法

战场态势评估涉及很多不确定因素,对不确定性进行仿真建模能够提高态势评估的能力。针对参战对象多元、不确定性增多导致的无法全面准确表达不确定性问题,提出了基于记忆模块和变分自编码器的深度贝叶斯网络模型。采用生成模型设计了基于深度贝叶斯网络学习的态势评估模型;阐述了融合记忆模块的深度生成模型原理和模型的学习与推理过程;以某空袭行动为例构建贝叶斯网络,对所提方法进行了验证。结果表明：深度神经网络能够逼近隐变量的非线性变换,设计的记忆模块能存储深度神经网络提取的大量局部特征,通过学习自动得到了贝叶斯网络条件概率,增强了不确定性建模能力。

Uncertainty Simulation Method Based on Deep Bayesian Networks Learning

There are lots of uncertain elements in battlefields situation assessment and the uncertainty simulation would enhance the ability of situation assessment. A deep variational autoencoder bayesian networks (BN) model with memory module is proposed aiming at the problem of being unable to represent the uncertainties exactly caused by the various combat objects and more uncertain elements. Based on the deep BN learning, the situation assessment model is designed from the deep generative model. The principle of deep generative model mixing with the memory module is discussed and the leaning and reasoning process of the model is explained. The proposed model is verified by an air strike BN construction example. The results show that the deep neural networks can approximate to the nonlinear transform of the latent variables and the designed outside memory module can store lots of local features extracted by the neural networks. The BN conditional probabilities are attained by the automatic learning and enhance the uncertainty simulation ability of BN.