基于规则的神经网络在模式分类中的应用

Application of rule-based neural network in pattern classification

针对模式分类任务,提出一种基于粗糙集规则的神经网络构造方法.首先,利用粗糙集理论和遗传算法约简输入特征,在尽量保持分类能力不变的情况下降低条件属性维数,并推导出简练的分类规则集合.然后,以规则集为基础构造BP神经网络结构、确定网络层数、输入输出节点数等,并计算规则的条件属性重要度和依赖度2个参数对连接权值进行初始化.最后,通过一个实例验证了方法的有效性,结果表明该方法能有效解决传统神经网络构造难、解释难、过拟合等问题,提高了分类精度,降低了训练时间.此外,初步探讨了网络训练时对知识提炼的影响.

To solve the problems of pattern classification,a method to construct the rough knowledge-based neural network is presented.First,the rough set theory and the genetic algorithm are used for feature reduction of the conditional attributes in order to reduce the number of the inputs without losing too much classification capacity,and the concise rules are generated.Based on the rules,BP（back propagation） network parameters,including number of layers,number of the input/output,connection weights etc.can be determined and initialized together with the two parameters,i.e.the conditional attribute significance and the dependence degree of rules.The method is proved to be valid by an experiment.The result shows that it not only overcomes the disadvantages of traditional neural network such as difficulty in determining the structure,incomprehensibility and over-fitting,but also improves the classification accuracy and reduces the training time.Additionally,the effect of network training on rule refining is discussed.