摘要
为了构建径向基函数 ( RBF)神经网络模型 ,研究一批取代芳香族化合物结构与其大型蚤生物活性参数 log( 1 /EC50 )的关系 ,利用正交最小平方算法逐一选择非线性高斯函数的中心 ,并将归一化参数 σ和容差 ρ作为网络的系统参数 ,通过广程扫描确定其最佳值 .采用种子聚类分析与简单随机抽样相结合的方法将化合物合理地划分为训练集和预测集 .对构建的 RBF网络模型的预测质量进行了不确定性分析 .与回归模型相比 ,RBF网络具有较好的预测性能 ,模型预测的不确定性降低了 30 % .
A Neural Network Based on Radial Basis Functions(RBF)was applied to study the relationship between structure and bioactivity (log(1/EC50)) to Daphnia magna for some substituted aromatic compounds.The centers of Gauss functions were selected one by one with the orthogonal least squares algorithm.The optimum values of the two parameters of the system,σ and ρ,were determined by largescale scanning.The compounds were divided into learning set and predicting set by seeded clustering and random sampling.The predicted quality of the RBF network model was evaluated by an uncertainty analysis method,which shows a 30% decrease of the prediction uncertainty compared with a regression model.This result indicates that the RBF model has a better predicting performance.
出处
《环境科学》
EI
CAS
CSCD
北大核心
2000年第4期11-15,共5页
Environmental Science
基金
国家自然科学基金
关键词
径向基函数网络
生物活性预测
有机化合物
Radial Basis Function Network
Gauss function
prediction of bioactivity
uncertainly analysis
