摘要
前馈神经网络BP算法的改进方案中,对网络训练(学习)过程中学习率和惯性系数进行模糊自适应调节,以提高收敛速度,是一项很有效的措施。文中具体分析了如何根据设计者的先验知识确定模糊规则和隶属函数,并以三比特异或函数(或称奇偶分类)的实现为例,验证了这种算法的改进、加速了BP网络的学习过程。
In the improved approaches of BP algorithm for feedforward neural networks, it is very effective for accelerating training that the learning rate and the momentum coefficient are updated by the fuzzy logic controller during the learning process. This article analyzes how to choose the fuzzy control rule table and the membership function by using a priori knowledge about the process. We also apply the improved algorithm to the experiment of 3-bit XOR problem, the results show that the convergence rate is much higher than the classical BP algorithm.
基金
国家自然科学基金课题!( 6 96 75 0 0 5 )资助
关键词
前馈神经网络
BP算法
模糊自适应算法
feedforward neural network
BP algorithm
learning rate
momentum coefficient
fuzzy control rule table
membership function
