广义回归神经网络的改进及在交通预测中的应用

A modified general regression neural network with its application in traffic prediction

本研究基于k近邻的方法通过网络性能评价指标来对平滑因子进行选择确定。通过k近邻法找出使得网络性能评价最好的平滑因子,不再仅依赖于一个均方误差数值,而根据均方误差组的排序来选择最优的平滑因子。该算法能够在保持较好的预测效果的前提下解决因数据波动性大而最终得不到最优平滑因子的难题。通过预测交通数据的实验验证了算法的有效性。结果表明通过k近邻方法得到的最优平滑因子会使网络预测误差降至最小。

Based on the method of k nearest neighbors algorithm, the optimum smoothing parameter was found by means of network performance evaluation. The approach depended not only on the value of mean square error, but also could sort the mean square error without affecting the forcasting performance. The optimum smoothing parameter was difficult to be found because of the volatility of the data solved by the modified algorithm. Finally, a traffic forcasting experiment was provided to analyze the effectiveness of the proposed algorithm. The results revealed that the optimum smoothing parameter found by means of k nearest neighbors could obtain the minimum prediction error.