摘要
针对在建立定量构效关系(QSAR)模型中,单个人工神经网络模型难以确定参数,容易产生“过拟合”;一般神经网络集成模型虽然建立过程简单,但由于个体差异度小而导致泛化能力相对单个神经网络没有明显改善等问题,提出了一种基于随机梯度法的选择性神经网络二次集成方法。在建立除草剂(苯乙酰胺类化合物)的QSAR模型的实验研究中表明,该方法设计过程简单,能够以较小的运算代价明显地提高了模型的泛化能力,是建立QSAR模型的一个有效方法。
Though neural networks have been widely used in the modeling of quantity structure-activity relationship ( QSAR), there are still many problems puzzling engineers such as the complex design procedure of single neural network, the 'over-fitting' problem and etc. Contrast to single neural network, neural network ensembles are easy to build, but sometimes they could not improve the generalization ability due to small differences among individuals. To solve these problems, a new model is proposed named with two-level selective neural network ensembles based on stochastic gradient select method. Based on this new model, a QSAR for herbicides ( N-phenylacetamides) is established. Simulation results show that the new model is easy to build and promotes the generalization ability of neural network system.
出处
《计算机与应用化学》
CAS
CSCD
北大核心
2005年第2期153-156,共4页
Computers and Applied Chemistry
基金
兵器科技预研项目(42001060402)
关键词
构效关系
选择性神经网络二次集成
随机梯度
quantity structure-activity relationship (QSAR)
two-level selective neural network ensembles
stochastic gradient method
