摘要
基于模糊聚类思想,提出了一种神经网络集成方法.由训练数据的模糊聚类结果,把训练数据划分成相交子集,基于各子集生成集成的个体神经网络.由于各子集所包含的数据和数据的类别各不相同,因而个体神经网络性能和结构存在差异.子集个数确定集成中个体神经网络个数.另外,基于隶属度函数计算公式,提出了个体神经网络输出结论结合方法.理论分析和实验结果表明,此方法对模式分类能取得较好的效果.
Neural network ensembles can significantly improve the generalization ability of learning systems. To do so, an ensemble must consist of many individual neural networks, which make different errors for pattern recognition. A popularly used scheme in generating individual neural network is to perturb the tratining data such as Bagging and Boosting. However, such schemes don't explore the characteristic of pattern distribution. Based on fuzzy C-means clustering, a new ensemble algorithm named Ensemble-FP is proposed. Firstly, training data set is partitioned into a number of intersection subsets, an approach to produce individual neural network is proposed. The characteristic of data and the number of data classes included in each subset are different, so the ability and construction of individual neural networks trained on these subsets are diversity. The number of subsets determines the number of individual neural networks in an ensemble. Second, a method of combining the output of individual neural networks is given based on membership function. Theoretical analyses and experimental results show that this neural networks ensemble algorithm Ensemble-FP is efficient for pattern classification.
出处
《南京大学学报(自然科学版)》
CAS
CSCD
北大核心
2006年第1期63-68,共6页
Journal of Nanjing University(Natural Science)
基金
国家自然科学基金(60273033)
江苏省自然科学基金(BK2004079)
扬州大学自然科学基金(KK0413160)
关键词
神经网络
神经网络集成
模糊聚类
分类
neural network, neural network ensemble, fuzzy c-means clustering, classification