基于PCA的概率神经网络结构优化

PCA-based probability neural network structure optimization

为了改善概率神经网络(PNN)在训练样本数量较大冗余度较高时存在的结构复杂的问题,提出一种基于主成分分析(PCA)的结构优化方法。以概率乘法公式为理论依据,根据训练样本的PCA结果对PNN进行结构优化,并引入学习算法减小PNN的参数不确定性。实验结果表明:在训练样本数量较大冗余度较高的情况下,优化后的PNN能够使用比传统PNN更简单的网络结构达到相近的结果。

The structures of probability neural networks （PNN） are quite complicated when trained with large, highly redundant training samples. A principal component analysis （PCA）-based structure was developed to optimize the PNN structure. A probability multiplication formula was used as the theoretical foundation. The PNN structure was optimized based on statistical results from the PCA for the training samples. A learning algorithm was introduced into the PNN to reduce uncertainties parameter. Test results show that with large, highly redundant training samples, the optimized PNN has a simpler structure than the traditional PNN to get a similar result.