基于Bayesian正则化算法的非线性函数拟合

Approximation of non-linear functions based on bayesian-regularization algorithm

为克服常规BP算法在解决非线性函数拟合时泛化能力不强的问题,本文研究了用贝叶斯正则化算法来提高网络泛化能力的问题,结果表明在相同网络规模或误差条件下,Bayesian正则化算法泛化能力明显优于基本BP算法及其它改进的BP算法,且收敛速度较快,拟合效果好。

To overcome the problem that the general BP algorithms have weak generalization capacity on doing the non-linear function approximation, we study the Bayesian Regularization algorithm to enhance the neural network's generalization capacity. Based on the same network size and error probality, the results show that Bayesian Regularization algorithm has better generalization capacity than the basic BP algorithms and other improved BP algorithms, and furthermore, it has higher convergence speed and better effects of approximation.