基于超立方体覆盖的构造性网络学习算法

Constructive training algorithm based on hypercube covering

该文提出了一种基于超立方体覆盖的构造性神经网络学习算法,以解决二值型输入变量的K分类问题。该算法分两步来动态地构造一个三层前馈网络。首先,对于每一类的所有训练样本,用尽可能少的超立方体来覆盖它们,并为每一个超立方体构造一个隐层单元;其次,用"或"操作把这些隐单元连接到相应的输出单元上。文章给出了相应的理论分析和一个具体的实现。实验结果表明,该算法优于常用的一些归纳学习算法。

A constructive training algorithm based on hypercube covering was developed for classification problems involving binaryvalued input attributes and multiple output classes. The algorithm dynamically constructs a threelayer feedforward neural network in two stages. First, for all the training examples in each class, the algorithm finds a set of hypercubes to cover the examples with each hypercube corresponding to a hidden neuron. Then, these neurons are connected to the corresponding output unit using the OR operation. The algorithm is analyzed theoretically and illustrated by an example. The results show that the method is superior to other commonly used inductive algorithms.