多输入Sigmoid激励函数神经网络权值与结构确定法

摘要

In order to overcome the inherent drawbacks of the BP neural network as well as its learning algorithms, this paper proposed a two-stage automatic search and determination method for optimal structure, which was based on the pseudoinverse-type weights direct determination (WDD) method. Taking function approximation for example, numerical experiment results demonstrate the proposed two-stage algorithm is effective and timesaving. They further show that the neural network has relatively excellent performance of approximation (i. e. , training and testing) on multivariate functions.%结合伪逆直接计算得到神经元之间最优权值的方法,提出了一种双阶段自动搜索与确定最优网络结构的算法,克服了原有BP神经网络模型及其学习算法的固有缺陷.以函数逼近为例,计算机数值实验结果显示了算法有效且耗时短,证实了由该算法得到的网络对于多输入函数具有较优良的逼近(学习与校验)性能.