具有最优学习率的RBF神经网络及其应用

A RBF neural network with optimum learning rates and its application

传统固定学习率的RBF神经网络在金融时间序列预测方面已经有比较成功的应用,但网络学习率的选择问题却给传统RBF神经网络的使用带来了不便.利用梯度下降法及优化方法推导出了RBF神经网络的动态最优学习率并将其应用于网络学习算法,具有最优学习率的RBF神经网络能够在保证网络稳定学习的同时兼顾网络的收敛速度.为了检验具有动态最优学习率的RBF神经网络的预测效果,对沪深300指数波动率进行了预测实验.实验结果表明,具有动态最优学习率的RBF神经网络比传统的固定学习率的RBF神经网络有着更快的收敛速度,同时也避免了人为选定学习率的不便.

In this study,a radial basis function（RBF） neural network learning algorithm with optimum learning rate is proposed.In this learning algorithm,the dynamic optimum learning rates which are determined by gradient descent and classical optimization technique are used to adjust the weight changes of RBF neural networks in an adaptive way.Using the dynamic optimum learning rates,the RBF neural networks can learn faster and more stable than the RBF neural networks with fixed learning rates.In order to verify the effectiveness of the proposed algorithm,volatility forecasting experiments on HuShen 300 index in Chinese stock market are conducted.The experimental results show the proposed RBF neural network learning algorithm with dynamic optimum learning rates could learn faster and avoid subjective selection of learning rates.