论文部分内容阅读
提出一种训练椭球基函数神经网络(EBFNN)的混合学习算法.此算法首先使用期望最大化算法初始化EBFNN中椭球基函数节点的参数,而网络的连接权重和偏差项则用线性最小二乘方法进行初始化.然后用梯度下降法对EBFNN中所有参数同时进行优化.与其他3个相关的模型相比,用混合学习方法训练的梯度下降椭球基函数神经网络(GDEBFNN)能够取得更优的分类性能.此外,与支持向量机对比表明,GDEBFNN取得与之接近的泛化能力.与基于Adaboost的决策树模型比较表明,GDEBFNN可以取得更优的泛化性能.
This paper proposes a hybrid learning algorithm for training elliptic basis function neural networks (EBFNN). The algorithm first uses the expectation maximization algorithm to initialize the parameters of ellipsoid basis function nodes in EBFNN, while the connection weights and deviations of the network are given by the least linear two Then all the parameters of EBFNN are optimized by gradient descent method.Compared with the other three related models, gradient descent ellipsoidal basis function neural network (GDEBFNN) trained by hybrid learning method can achieve better performance In addition, compared with SVM, GDEBFNN obtains the generalization capability close to it.Compared with Adaboost-based decision tree model, GDEBFNN can achieve better generalization performance.