改进的神经网络及其自适应学习速率的研究

从提高神经网络泛化能力的角度提出一种改进方法.利用Taylor级数展开的思想,用线性和非线性组合构成函数映射关系,即改进的神经网络是用原神经网络的非线性映射和关于输入信号的线性映射的和来逼近期望值.文中还给出了该神经网络学习速率的自适应调节方法.对线性对象和非线性对象分别进行建模仿真,结果表明,改进的神经网络比基于正则化方法的神经网络具有更好的泛化能力.

Study on an Improved Neural Networks and its Adaptive Learning Rate

To enhance the generalization ability of neural networks, an improved method is studied.Based on Taylor series expansion, the mapping relation is set up by combining a linear part with a nonlinear part. That is, the new neural network utilizes the nonlinear mapping of the original neural network and the linear mapping that corresponds to the input signal to approach the target value. And an adaptive method to regulate the learning rate of this improved neural network is given.Simulations are carried out by modeling a linear plant and a nonlinear plant.Results show that the new neural network has better generalization ability than the ordinary neural network with regularization method.