改进型遗传神经网络在模式分类中的应用

Evolving neural network based on improved genetic algorithm for pattern classification

为研究图像和语音的模式分类,提出一种采用可变长度串遗传算法(VGA)的进化神经网络.该算法可以全局搜索优化神经网络的结构,找到神经网络接近最优的连接权,再通过反向传播算法(BP),在该优化结构中找到最优连接权.对语音数据和SPOT图像数据的验证结果表明,在模式分类中,采用该算法的分类器(VGA-BP)的分类性能较贝叶斯(Bayes)分类器、最近邻规则(k-NN)分类器具有更高的分类精度.

An evolving neural network classifier using variable string genetic algorithm （VGA） was developed to study pattern classification for image and speech. The classifier could automat ically evolve the appropriate architecture of neural network and find a nearoptimal set of connection weights globally. Then the conformable connection weights for pattern classification could be found with backpropagation （BP） algorithm. Simulations on vowel data and SPOT multispectral image data show that the VGABP classifier has higher classification precision comparing with Bayes classifier and kNN classifier in pattern classification.