一种基于因素贡献率的自适应前馈网络算法

前馈网络具有很强的信息处理能力，但是实际应用中的样本量有限，训练出来的网络效果不太理想，本文提出一种自适应前馈网络算法，通过调节参数α向量，使网络结构按精确度和推广能力来综合考虑，若训练出来的网络满足先决条件，就认为网络规模是合适的，此算法不仅能删除隐层中的节点也能删掉输入层中若干总贡献率小的节点，从而避免了维数灾难，并给出一种调整剩余权重的算法，避免重新训练，文中讨论了网络结构的重要性，并用该方法进行仿真实验，结果证明了自适应前馈网络算法具有较强的建模能力。

Adaptive feed-forward network algorithm based on factor contribution ratio

Feed-forward neural network has powerful information processing capability, but in the practical applications, the trained network is not satisfactory owing to the small sample size. In this paper we propose a adaptive feed-forward network algorithm, through adjusting the parameter a (a vector), making the network architecture considered by precise and generalization capability. If the trained network meets the precondition, consider the network scale proper. This algorithm not only can cancel the nodes of the hidden layers but also can cancel the nodes of the inputs layer whose gross contribution ratio are lower, so avoiding the dimensionality disaster. And thearticle proposes an algorithm to adjust the remanent weights to avoid training the network again. In this paper, we discuss the importance of the network architecture, and use the present method to do simulation experiment, the result show that this method has a satisfactory modeling capability.