针对近似建模的前馈神经网络训练算法

Algorithm to train feed-forward NN for approximately modeling

为克服现有神经网络训练算法在建模精度方面的不足,提出了一种专门面向近似建模的前馈网络训练算法——GA-BP贝叶斯算法.该算法以提高网络的泛化性能为主旨,以获取对应于后验分布最大值的权值向量为训练目标,并采用遗传算法和L-M(Levenberg-Marquardt)BP算法相结合的权值搜索策略.其中,L-M BP算法是当前最流行的前馈网络训练算法.结合一个典型算例,对GA-BP贝叶斯算法和L-M BP算法进行了对比研究.结果表明:与L-M BP算法相比,GA-BP贝叶斯算法所建立的神经网络近似模型具有更高、更稳定的精度.

To overcome the shortcomings in model accuracy of the existent NN training algorithms, a new algorithm to train feed-forward networks, i.e. GA-BP Bayesian algorithm, was proposed, and it was special for establishing approximate model. This algorithm was developed to improve the generalization of neural networks, and its objective is to obtain the weights corresponding to the maximum posterior probability. And it adopts both genetic algorithm and L-M （Levenberg-Marquardt） back-propagation （BP） to search optimal weights. L-M BP is the most popular algorithm to train feed-forward networks. Combined with typical computing instance, GA-BP Bayesian algorithm was compared with L-M BP and the result indicated that the approximate model established based on it has higher and more stable accuracy.