基于神经网络结构学习的知识求精方法

知识求精是知识获取中必不可少的步骤．已有的用于知识求精的ＫＢＡＮＮ（ｋｎｏｗ ｌｅｄｇｅ ｂａｓｅｄ ａｒｔｉｆｉｃｉａｌｎｅｕｒａｌｎｅｔｗ ｏｒｋ）方法，主要局限性是训练时不能改变网络的拓扑结构．文中提出了一种基于神经网络结构学习的知识求精方法，首先将一组规则集转化为初始神经网络，然后用训练样本和结构学习算法训练初始神经网络，并提取求精的规则知识．网络拓扑结构的改变是通过训练时采用基于动态增加隐含节点和网络删除的结构学习算法实现的．大量实例表明该方法是有效的

A KNOWLEDGE BASE REFINEMENT METHOD BASED ON STRUCTURAL LEARNING OF NEURAL NETWORKS

Knowledge base refinement is necessary for knowledge acquisition in building expert systems. The KBANN (knowledge based artificial neural network) for knowledge base refinement has been proposed in literatures. The key limitation of KBANN is that there is no mechanism for changing the topology of the network. In this paper, a novel approach to knowledge base refinement based upon structural learning of neural networks is presented. In this approach, a set of rules are mapped into a neural network (i.e., the initial neural network), and then this reformulated knowledge is refined using structural learning (i.e., the initial neural network is trained by structural learning algorithm and a set of training examples). Finally, the refined rules are extracted from the trained neural network. To accomplish the topology change of the initial neural network, structural learning algorithms based on dynamic node creation and network pruning are used in this approach. Many simulation experiments have proved the effect of this approach.