基于Chebyshev神经网络的衍生算法

A deriving arithmetic based on Chebyshev neural network

根据辨识对象 ,神经网络的输入层和输出层神经元的个数易于确定 ,而隐层数和隐层神经元个数难以确定 .借助于生物的生长发育知识 ,在正交基函数神经网络的基础上提出衍生算法 ,该算法的基本思想是 :先选取较少个数的隐层神经元作为初始发育细胞 ,训练K次后 ,如果目标函数J不再变化且大于给定的ε ,则网络自动衍生 ,以上过程反复进行 ,直至J小于等于ε ,则停止衍生和训练 .仿真实验表明 ,该算法在训练过程中改善了收敛速度 ,并自动调整网络的拓扑结构 ,解决了隐含神经元个数难以确定的问题且具有优良的逼进任意非线性特性的能力 .图 2 ,表 2 ,参 7.

In the research of neural networks, the numbers of the neurons in the input-layer and output-layer are easily fixed with the identification objects. But it is difficult to determine the numbers of hidden-layer and hidden neurons. By employing the knowledge of biologic growth, this paper presents a deriving algorithm based on the neural networks. Firstly, less hidden neurons are chosen to be the original cells; secondly, the network is continually deriving if the target function J dose not change and J > Ε, after the network is trained K-times. This process does not stop until J &le Ε. The simulation results show that the deriving algorithm have improves the convergence rates in the training process, corrects the top structure of network automatically, and solves the problem on determining the number of the hidden neurons. The giver algorithm is proved of excellent performance of approximating nonlinear properties.