基于空间分割算法的自适应多层感知器网络

THE MLP NETWORKS BASED ON SPACE-PARTITIONING ALGORITHM

本文提出了一种基于空间＾１）优化分割算法的自适应多层感知器网络，以求解一些复杂的非线性分类问题。这些问题用一般前馈网络难以得到好的结果。我们提出的网络具有特殊的网络结构。第一隐层单元可以自适应地增加与删减，而每个隐单元都表征一个空间分割超平面。文中引入一种优化算法来提高网络容错能力。本文对５圈“双螺线”的分类样本进行了学习与测试，并与其他方法进行了比较。还给出了手写体识别得到的结果。

In this paper an adaptive multi-layer perceptron network based on space-partitioning algorithm (SP-MLP) is proposed for complex nonlinear classification problem, which normal feed-forward neural networks can not get satisfying results. The proposed network has its special structure. The number of the first hidden layer units which represent the space-partitioning hyperplanes. can be updated adaptively. An optimal algorithm is used to increase the tolerance of the network. We apply our method to the ' two spirals ' classification problem. From the result and comparison, we can see that the SP-MLP network shows better performance than other ones introduced in the references. At last we implement it in the handwriting recognition problem and get good results.