论文部分内容阅读
针对训练好的神经元网络进行解释这一难以解决的问题，提出了一种从神经元网络中抽取规则的新的抽取方法——二阶段法，从隐含层到输出层，利用学习方法从整个隐含激励空间中抽取出有效区域，形成规则；从输入层到隐含层，利用搜索方法，通过分析其间的权值关系抽取出规则，使得所有被这些规则覆盖的实例所产生的隐含激励向量均位于上述有效区域内。实验证明了此方法产生的规则比Ｃ４．５产生的规则的抗干扰力强，同时其可信度比传统的基于搜索的抽取方法——ＫＴ算法要高。
Aiming at the hard-to-solve problem of well-trained neural network, this paper proposes a new method of extracting rules from neural networks-the two-stage method, from the hidden layer to the output layer, using the learning method The entire implicit incentive space is extracted from the effective area, the formation of rules; from the input layer to the hidden layer, the use of search methods, by analyzing the relationship between the weights extracted rules, so that all the rules covered by these rules generated by the implied The excitation vectors are all located in the above effective area. Experiments show that this method produces stronger anti-jamming rules than those generated by C4.5, while its credibility is higher than that of the traditional search-based extraction method - KT algorithm.