摘要
提出了一种新的神经网络初始权值的优化方法,该方法首先对样本做K-L变换,将所得到的变换矩阵作为BP网络输入层到隐层的初始权值,然后开始BP算法对多层感知器的训练学习,以缩短样本学习时间.最后分别选取线性可分的样本和非线性可分的样本在MATLAB中进行了仿真,仿真结果证明,该权值优化方法是合理的.
This paper presents a novel method of initial weights optimization method in neural network. Firstly, the samples are transformed by K-L transformation, and then K-L converting matrix is used to initialize the weights between input and hidden layer. Secondly the multiplayer perceptron network is trained with BP algorithm to decrease the training time. Lastly, linear divisible and linear indivisible samples are chosen to he simulated in MATLAB. The result shows that the method is reasonable.
出处
《重庆工学院学报》
2007年第3期103-105,共3页
Journal of Chongqing Institute of Technology
关键词
K—L变换
BP网络
权值优化
K-L transformation
BP algorithm
optimization of weights
