多层前馈神经网络的快速学习算法及其仿真研究

本文主要研究多层前馈神经网络的快速学习问题。文中首先概述了多层前馈神经网络的B—P学习算法,并分析了这种算法的主要缺陷,在此基础上提出了若干克服和缓解这些缺陷的技术方法,由此构造了一种新的多层前馈神经网络的快速B—P学习算法即FB—P算法。通过对FB—P算法学习过程进行较详细的分析,本文还建立了一种改进的FB—P学习算法即MFB—P算法。最后本文以三层前馈神经网络识别五类地面目标图像为例,对文中提出的FB—P和MFB—P学习算法的性能(即学习速度与推广特性)进行了计算机仿真实验,同时与B—P学习算法的性能作比较,理论分析与仿真实验表明:MFB—P与FB—P学习算法比B—P学习算法具有更快的收敛速度,且MFB—P算法收敛最快;MFB—P算法比FB—P和B—P学习算法具有更好的推广特性,而后两者的推广特性则大致相同。

Fast Learning Algorithms and Their Simulations for Multi-Layered Feedforward Neural Network

In this paper, the problem of the fast learning algorithm for multi-layered feedforward neural network (MLFNN) is considered. The classical error back-propagation (B-P) learning algorithm is first reviewed, and its several drawbacks such as slow convergence and no better generalization are discussed. Some techniques are suggested to compensate these drawbacks and thereby a new fast B-P learning algorithm (FB-P) is proposod. Based on the detailed analysis of the learning process of FB-P algorithm a modified FB-P learning algorithm (MFB-P) is established in the paper. The network trained by MFB-P has better generalization than that trained by FB-P or B-P algorithm. Some simulations are ran and the corresponding results indicate that: i) MFB-P or FB-P converges much more quickly than B-P and ii) the MLFNN trained by MFB-P has better generalization than that trained by FB-P of B-P and the last two (FB-P and B-P) have similar generalization.