Hermite插值神经网络权值和结构确定理论探讨

On weights and structure determination of Hermite-interpolation neural network

为了克服BP神经网络固有的缺陷,基于Hermite插值理论,构造了一种新型的前向神经网络模型(即Hermite插值神经网络模型)。针对该网络模型,提出了一种基于矩阵伪逆的权值直接确定方法,并在此基础上探讨了隐神经元数目自动确定的方法(即网络结构自确定方法)。计算机仿真结果表明,相比于传统的BP神经网络,使用权值与结构双确定方法的Hermite插值神经网络具有更好的收敛速度和校验能力。同时,也验证了该神经网络良好的降噪和预测能力。

In order to overcome the inherent drawbacks of BP neural network,based on the Hermite-interpolation theory,this paper constructed a novel type of feed-forward neural-network model,which could be termed as Hermite-interpolation neuralnetwork model.For this model,it presented a pseudo-inverse based weights determination method（or termed,weights-directdetermination method） ;and further investigated the determination of the hidden-layer neuron number（i.e.,structure-automatic-determination method）.Computer-simulation results demonstrate that the presented Hermite-interpolation neural network with the above two methods can converge faster,and has a better testing performance（as compared to BP neural network）,as well as the great de-noising and forecasting capabilities.