平滑l_1模神经网络学习算法

Smoothed l_1-norm learning algorithm for neural networks

神经网络误差函数的不同构造将导致不同的学习敛速。本文基于递推辨识技术,对于绝对误差指标函数提出了一种新的平滑方法作为改进的神经元权值修正算法,该平滑思想也适用于其它绝对误差指标下的最优化问题。通过网络对Feigenbaum映射和XOR逻辑的学习算例,说明了算法的快速收敛性能。

The fast learning algorithm for feed-forward neural networks can be approached by focusing on constituting optimization criteria which are used for their training. It was recently found that the rate of convergence of the error back propagation algorithm can be accelerated with the least absolute value (l_1-norm) criterion. Based on the recursive identification technique, this paper addresses a new smoothed absolute error objective function,which leads to some fast training algorithm which update the weights of the each neuron in the network in an effective parallel way. The smoothed idea can be used for solving other l_1-norm optimization problems. Its performance is illustrated using examples from Feigenbaum function and XOR logical operation.