基于改进遗传算法的RBF神经网络结构优化研究

针对RBF神经网络隐含层节点数过多导致网络结构复杂的问题,提出了一种基于改进遗传算法(IGA)的RBF神经网络优化算法。利用IGA优化基于正交最小二乘法的RBF神经网络结构,通过对隐含层输出矩阵的列向量进行全局寻优,从而设计出结构更优的基于IGA的RBF神经网络(IGA-RBF)。将IGA-RBF神经网络的学习算法应用于电子元器件贮存环境温湿度预测模型,与基于正交最小二乘法的RBF神经网络进行比较的结果表明:IGA-RBF神经网络设计出来的网络训练步数减少了44步,隐含层节点数减少了34个,且预测模型得到的温湿度误差较小,拟合精度大于0.95,具有更高的预测精度。

RBF neural network structure optimizationbased on improved genetic algorithm

In order to deal with the complex structure of radial basis function (RBF) neural networks due to the excessive number of hidden layer nodes, we propose a RBF neural network structure optimization algorithm based on an improved genetic algorithm (IGA). The IGA is used to optimize the RBF neural network structure based on orthogonal least squares. We globally optimize the column vectors of the output matrixes of the hidden layers to design a RBF network with better structure based on IGA optimization (IGA RBF). The algorithm is applied to a temperature and humidity prediction model for the electronic components storage environment. Results show that compared with the RBF neural network structure based on orthogonal least squares, the number of the hidden layer nodes of IGA RBF network is reduced by 34, and the number of training steps is reduced by 44. The errors of temperature and humidity of the prediction model for the electronic components storage environment are smaller, and the fitting accuracy is greater than 0.95, thus having better prediction accuracy.