Abstract: Neural Network(NN)’s incomprehensible quality is always an inherent defect that limits its self-development. This paper, from the functional point of view, describes a method that employs a ICS algorithm to extract rules from NN. The method uses tactics of immune clone in rules extraction from NN and clusters outputs of hidden neurons of NN, to reduce the searching scale and extract succinct sign rules of good understanding. The method can be used in all kinds of NN based classifier not depending on concrete net structure and training algorithm. The result of experiment shows its practicability and feasibility.

Key words: Neural Network(NN), rules extraction, ICS algorithm, clustering

摘要： 神经网络的不可解释性一直是限制其发展的固有缺陷，该文从神经网络的功能性观点出发，提出基于免疫克隆选择算法的神经网络规则抽取方法。将免疫克隆策略用于神经网络的规则抽取中，对已经训练好的神经网络隐层神经元输出值进行聚类，缩小搜索空间，抽取出理解性好、简洁的符号规则。该方法不依赖于具体的网络结构和训练算法，可以方便地应用于各种分类器型神经网络。实验结果表明该方法的实用性和可行性。

关键词: 神经网络, 规则抽取, 免疫克隆算法, 聚类

CLC Number: 

• TP183