基于模型逼近度和接受概率的一个变步长快速BP学习算法

A FAST VARIABLE STEPSIZE BP LEARNING ALGORITHM BASED ON MODEL APPROXIMATION DEGREE AND ACCEPTANCE PROBABILITY

本文给出模型逼近度的概念．在此基础上并利用模拟退火法的思想，提出了一个改进的变步长快速ＢＰ学习算法．数值结果表明该算法不仅明显地提高了ＢＰ网络的学习收敛速度，而且在一定程度上还能避免陷入局部极小．

This paper presents an improved variable stepsize BP learning algorithm by making use of the concept of model approximation degree and by taking advantages of the merits of simulated annealing method. Numerical experiments show that the proposed algorithm not only enhances a lot the convergence rate of learning process of BP network, but also avoids the stagnation problem to some extent.