小样本目标识别问题下的BN参数学习  

BN Parameter Learning in Target Recognition of Small Sample

作　　者：田浩[1] 高晓光[1] 

机构地区：[1]西北工业大学电子信息学院,西安710129

出　　处：《电光与控制》2014年第8期47-49,共3页Electronics Optics & Control

基　　金：全国高校博士点基金(20116102110026)

摘　　要：目标识别问题中存在大量不确定信息,利用BN可以对不确定信息及其相互关系进行学习与推理。但是,目标识别问题的样本量较小,在参数学习过程中,常因观测数据不足产生误差,需要引入单调性信息等专家经验,针对这一问题,提出最小元算法。首先,利用最小元表达单调性信息,将其转化为参数学习可以直接利用的先验信息;然后,基于保序回归思想,对参数学习结果进行优化,消除误差,得到相对准确的网络参数。以空中目标识别为仿真背景,与最小子集算法比较,验证了该算法在准确度与复杂度等方面的优势。There is a lot of uncertain information in target recognition which can be learned and reasoned by the use of BN.However,the target recognition problem is of small sample size,and there are often some errors due to lack of observational data during parameter learning .Therefore,it is needed to introduce monotonic expertise .Focusing on the above problem,the minimum unit algorithm was proposed .By use of the minimum unit,the monotonic information was translated into priori information,which could be directly used in parameter learning .Then,based on the thinking of isotonic regression,the parameter learning outcomes were optimized and the errors were eliminated .The relatively accurate network parameters were obtained.On the background of aerial target recognition,the advantages of minimum unit algorithm compared with the minimum lower sets algorithm in accuracy and complexity are illustrated .

关 键 词：目标识别 参数学习 参数优化 最小子集算法 最小元算法 

分 类 号：V271.4[航空宇航科学与技术—飞行器设计] TP181[自动化与计算机技术—控制理论与控制工程]