基于隐层生长策略的前馈神经网络隐层性能评测

Hidden Layer Performance Evaluation for Feedforward Neural Networks Based on Hidden Layer Growing Strategy

针对最佳平方逼近三层前馈神经网络模型,讨论了以逐一增加隐单元方式构建隐层时隐层性能的评测方法。分析了影响前馈神经网络性能的相关空间,引入了表示空间、误差空间、目标空间和耗损空间的概念,研究了每个隐单元的误差补偿性能,提出了网络隐层性能的评测参数,并通过对传统BP算法和正交化算法的考查验证了其合理性与有效性。

A hidden layer performance evaluation method is dis cussed according to the model of least-squares approximation feedforward neural networks based on hidden layer gro wing strategy. Firstly, some spaces which affect the performance of feedforward neural networks are analyzed and four concepts of subspace, i.e. representation space, error space, target space and expend, are introduced. The error compensation performance of the hidden unit is analyzed. Finally, evaluation parameter of hidden layer performance is proposed, and the rationality and validity of proposed method are validated by reviewing classical BP algorithm and orthogona l algorithm.