基于连续时间系统仿真的神经网络学习算法

Learning Algorithm for Neural Networks Based upon Continuous-time System Simulation

从连续时间动力学的角度,研究了多层前馈神经网络的学习问题。基于李雅普诺夫稳定性分析方法,建立了一种神经网络权重参数连续调整的学习算法,并基于连续时间系统仿真的思想,给出了一种算法实现的自适应策略。算法实现中,通过估计截断误差估计自动调整步长,几乎不需要人工确定任何参数,而且可以保证算法的稳定性及计算精度。最后,给出了两个典型的应用算例。

The learning issue in feedforward neural networks was investigated and analyzed from the viewpoint of continuous-time system dynamics. A continuous learning algorithm for the multilayer feedforwad networks was set up based on Lyapunov stability theory, and an adaptive updating law for the network weights was presented upon the continuous-time system simulation. In the algorithm implementation, the estimated tnmcation error was applied to adjust the step-size. Almost no preset parameters are required to operate the discrete-time adaptive learning laws, and numerical stability and satisfactory accuracy are assured. Finally, two illustrative examples were given.