摘要
应用Matlab 7.0中神经网络工具箱建立BP神经网络的最优化求解方法,采用共轭梯度法对网络的权值和阈值进行优化计算,实现网络权值和阈值的快速计算,为分析神经网络的合理结构提供了必要条件。对BP神经网络的传统梯度下降法与共轭梯度算法进行了仿真。这里通过对算法的训练速度,容错泛化能力等方面加以讨论,多方面印证共轭梯度算法的优越性,仿真结果凸显了训练速度的大幅提高,尤其对训练后网络受损情况下的泛化能力,采用线性回归的方法进行了仿真验证,同样得到满意结果,从新的角度支持了共轭梯度BP算法。
The optimized solution method of BP neural network is estabished by neural network toolbox in Matlab 7.0, using the conjugate gradient method carry on the optimized computation to the network weight and the threshold value, the network weight and threshold value＇s rapid calculation are realized, the essential condition is provided for analysing neural networkrs reasonable structure. BP neural networkts tradition gradient descent law and the conjugate gradient algorithm are simulated. Through the algorithm training speed, aspects and fault- tolerant pan ability, superiority of conjugate gradient algorithm is verified, the simulation result highlights the improvement of training speed, especially after training the network suffers injury in situation panability,the linear regression method is used to carry on the simulation confirmation, satisfaction result is obtained,the conjugate gradient BP algorithm is supported from the new angle.
出处
《现代电子技术》
2009年第18期125-127,共3页
Modern Electronics Technique
基金
山西省青年科技研究基金资助项目(20031008)
