一种基于极点配置稳定的新型局部递归神经网络

摘要

针对局部全局前馈递归动态神经网络的稳定性问题提出了一种新的采用极点配置稳定方法的局部递归全局前馈(Locally recurrent global forward,LRGF)神经网络.由于动态神经元的极点有存在于实轴上和一对共轭复数极点两种情况,为了避免神经元无限脉冲响应滤波器(Infinite impulse response filter,IIR)的系数投影到稳定区域的复杂性,构造的神经网络将动态神经元分成实数极点IIR和共轭复数极点IIR两部分,通过函数权值的方法将这两部分加权输出.同时针对这种新的神经网络采用了梯度下降的学习算法.通过仿真对本文提出的神经网络的可靠性和有效性进行验证,并分析这种新的神经网络在稳定投影计算上的复杂度.%This paper derives a new stable locally recurrent global forward (LRGF) neural network with pole assignment projection approach. The pole in the hidden neurons of the LRGF neural network can be classified into two situations. One case is that the pole is on the real axis, and the other case is that the pole is a conjugate complex. We divide the dynamic hidden neuron into two parts according to the kind of the pole, so that it can avoid the complexity of the projective computation. A weight function is used to fuse the two parts. The learning method is based on the gradient decent approach, which has been modified to be fit for the proposed neural network. At last, the simulation is given to demonstrate the reliability and effectiveness of the new neural network, and the complexity of the projection computation is also be analyzed.