一种改进的结构自适应自组织神经网络算法

An Improved Structure-Adapting Self-Organizing Neural Network Algorithm

针对结构自适应自组织神经网络神经元分布的特点和节点生长-删除算法的局限性,提出了一种改进的结构自适应自组织神经网络。根据神经元的学习速率、领域、兴奋度及相似阈值等状态信息,该算法可适应地调整神经网络的结构,克服了现有结构自适应神经网络的局限性,具有很好的聚类和泛化能力。实验结果表明该算法对样本的聚类有很好的效果并且具有很好的泛化能力。

According to the characteristics of neural unit distribution and the drawbacks of generating deleting neural unit node in structure adapting self organizing neural network （SASONN）, an improved SASONN is proposed. The proposed algorithm could automatically adapt SASPNN network structures on the base of study rate, field, excite degree and resemble threshold for current neural units. Thus the shortcomings of current SASONN could be avoided with good generalization ability. Numeric results show that the improved SASONN works well in clustering and classification and has a good generalization capability.