摘要： 针对使用贪心法、遗传算法等方法实现选择性神经网络集成时出现的“局部最小点”和“过拟合”问题,提出了一类基于随机梯度法的选择性神经网络二次集成方法。理论分析和实验表明,与上述选择性神经网络集成方法相比,该方法易于实现且效果明显。

关键词: 神经网络集成, 二次集成, 贪心法, 随机梯度法

Abstract: In the application of greedy method or genetic algorithm for the selection of the components of neural network ensembles, local minima and over fitting problems occur frequently. To solve such problems, a kind of method of two-level stochastic gradient-based selective neural network ensembles is proposed in this paper. Theoretical analyses and experimental results show that the method is easy to be constructed and performs well.

中图分类号: 

• TP183