一种新的对角回归神经网络快速学习算法

New method for diagonal recurrent neural networks with local error-backpropagation

提出一种新的动态对角回归神经网络学习算法——局部动态误差反传算法 (L DBP) ,该算法定义了一种新的局部均方差函数 ,并为回归单元建立一种新的学习结构。如果估计出各层的期望输出值 ,多层回归网络便可分解成一组自适应单元 (Adaline) ,而每个单元可通过二次优化方法进行训练。采用可在有限步内找出全局最优解的共轭梯度法 (CG)进行寻优。由于学习过程采用超线性搜索 。

A fast new local dynamic error backpropagation algorithm(LDBP) is presented for the training of diagonal recurrent neural networks(DRNN). This algorithm is based on the definition of a new local mean squared error function. The approximation to a recurrent node has the similar construction as the Adaline (Adaptive linear element). When the local desired outputs of the elements have been estimated, the DRNN can be decomposed into a set of node elements that can be trained by quadratic optimization methods. The conjugate gradient (CG) method is used. The simulation result shows the advantages of the new algorithm.