论文部分内容阅读
本文提出了一种新的多层前馈神经网络快速训练方法。该算法是基于指数加权局部最小二乘（ＥＷＬＬＳ）目标函数及殴几里得方向集（ＥＤＳ）方法的，在训练过程中，通过估计局部期望输出，多层神经网络可以被分解成若干个自适应线性神经元（Ａｄａｌｉｎｅ），而Ａｄａｌｉｎｅ是通过ＥＤＳ方法进行训练的。该算法的性能是通过将其应用于系统辩识中加以说明的。
In this paper, a new fast training method of multi-layer feedforward neural network is proposed. The algorithm is based on the exponentially weighted local least squares (EWLLS) objective function and the strike direction set (EDS) method. During the training, by estimating the local expected output, the multi-layer neural network can be decomposed into several Adapt to linear neurons (Adaline), while Adaline is trained by the EDS method. The performance of the algorithm is illustrated by applying it to system identification.