递归神经网络的模糊随机学习算法

A Fuzzy Alopex Learning Algorithm of Recurrent Neural Networks

提出了一种新的基于模糊逻辑的Alopex学习算法 (FLA)。FLA算法利用模糊逻辑推理实时获得适应于学习过程的适当的算法修正值 ,克服了Alopex算法中修正值固定不变的缺点 ,使得随机学习过程在速度、精度和稳定性之间获得平衡。将该算法应用于神经网络的训练 ,可以无需神经网络的梯度信息和结构信息 ,因此可以用于具有各种结构特性的递归神经网络的学习。在本文给出的仿真实验中 ,采用FLA算法实现带输出反馈的递归神经网络对动态系统的学习过程。

A novel leaning algorithm, fuzzy logic alopex learning (FLA), is proposed. In this algorithm, fuzzy logic inference is used to modify the weight bias on-line, so the defect of bias unchangeable in Alopex is corrected, and the weight change relevant to the learning process and errors are properly made. This algorithm can be used to recurrent neural networks learning without the need of information about the network structure and grads. It is used to train a recurrent neural network with output feedback for a dynamic system learning. The simulation results show the validation of this structure and learning algorithm.