基于稀缺数据集下BN参数学习的目标识别

针对贝叶斯网络（BN）在目标识别参数建模中常常面临特征数据样本相对稀缺的问题，研究了将稀缺数据集与定性专家经验相融合来估算BN模型参数的方法——CSDE，并据此提出了一种目标识别算法。该算法在BN结构已知的情况下，将定性专家经验转化为BN条件概率之间的约束集合；随后引入凸优化求解方法完成BN目标识别模型参数的估算。在实验研究中，先通过对经典的BN模型的参数学习问题验证了CSDE算法的有效性；随后，针对实际稀缺样本数据集目标识别问题，进行了建模及识别实验。实验结果表明：所提出的算法能够较好地解决样本数据集相对稀缺条件下的目标识别参数建模问题。

Target recognition based on BN parameter learning under scarce data set

Aiming at the problem of Bayesian networks modeling for target recognition with the relative scarce characteristic data, the method of the parameters estimation on the BN model is studied by fusing the scarce data set with the qualitative expert knowledge. Based on this, a target recognition algorithm is proposed. With known BN structure, the algorithm transforms the qualitative expert acknowledge into the BN conditional probability constraint set, and the convex optimization algorithm completes the estimation of BN target recognition model parameters. In the experimental study, the effectiveness of CSDE algorithm is verified by classical BN parameter modeling problem. Furthermore, the modeling and recognition experiments are carried out for the target recognition in real scarce sample dataset. The experimental results show that the proposed algorithm can solve the problem of target recognition parameters modeling better even when the modeling data set is scarce.