摘要
提出基于数据场高斯势约简概率神经网络结构,基本思路:引入数据场估计训练集各类概率密度,选择局部极大密度估计样本构造网络;对初始网络迭代训练,依次扩展各类具有最大密度估计值的误分样本至模式层并调整权重参数,直至满足指定精度.采用增量密度计算,保证快速迭代和高概率收敛.基于重采样技术进一步提升泛化精度.实验表明,提出的算法解释精练、拟合优度适中、计算高效.
This paper proposed to decrease the structure of probabilistic neural network based on Gaussian potential of data field.Core idea is following:Introduce data field to estimate probabilistic density of training set of each class and select their maximum to construct the network;iteratively train the initial network by appending the maximum density sample unrecognized of each class to pattern layer and modify the weight of samples until satisfying desired accuracy.Incremental computing density ensures faster iteration and higher possible convergence.And introduce resampling technique to boost the generalization accuracy.Experiments show that the proposed algorithms have concise explanation,moderate fitness and effective calculation.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2011年第8期1739-1745,共7页
Acta Electronica Sinica
基金
国家863高技术研究发展计划(No.2005AA113040)
北京市教育委员会共建项目建设计划(No.JD100060630)
河北省高等学校自然科学研究项目(No.Z2010279)