基于人工神经网络的知识求精方法

A KnowledgeBase Refinement Method Based on Artificial Neural Networks

将神经网络用于知识求精的主要局限性，是训练时不能改变网络的拓扑结构。本文根据结构学习算法提出了一种基于神经网络的知识求精方法，训练时采用动态增加隐含节点和网络删除改变拓扑结构，然后提取求精后的规则知识。大量实例表明，该方法是有效的。

Knowledgebase refinement is necessary for building and maintaining efficient expert systems. KBANN(Knowledge based artificial neural networks), proposed by G.G. Towell et al for doing this refinement, suffers from the limitation that there is no mechanism for changing the topology of the network. In our approach, we use dynamic node creation and network pruning, which we discuss quite thoroughly, to accomplish topology changing, thus making refinement better and expert system more efficient. Essentially we do three things: (1) we map the knowledge base into neural networks; (2) we refine the reformulated knowledge by using structural learning which involves dynamic node creation and network pruning; (3) we extract the improved refinement rules from neural networks.Figs.2 and 3 are results of simulation experiments. They show that structural learning can indeed change the topology of the network, i.e., we can indeed obtain improved refinement rules.