【作者】 李军；

【导师】 阮晓钢；

【作者基本信息】 北京工业大学 ， 模式识别与智能系统， 2003， 硕士

【摘要】 目前，发展较为成熟的两种人工智能系统是基于符号处理机制的专家系统和基于非符号处理机制的神经网络，它们的应用，使得计算机具有智能成为现实，解决了一大批工程实践中的问题。本文将这两种方法相结合，提出了一种基于神经网络的专家系统(BPNES)的设计及实现方案。 论文研究的核心是作为知识库的人工神经网络ANN(Artificial Neural Network)和基于ANN的推理机制。论文的主要工作有： 1．在对现有的知识表示方法进行深入分析的基础上，研究了一种知识表达的二元产生式规则方法。知识经过二元标准化后，利于神经网络上的知识存储和维护。 2．知识的存储选用神经网络中的BP(BackPropagation)模型。采用了基于数值优化的Leverberg-Marquardt(简称LM)算法，对网络进行训练，完成知识库的组建。针对神经网络的记忆容量，本文提出建立一个包含多个子网的系统，每个子网记忆一个对象，多个子网可以并行工作。 3．深入分析了基于神经网络的推理算法的运行机理，讨论了正向推理和逆向推理算法的特点。基于ANN的知识库，论文提出了两种不同的推理算法。正向推理算法从已有的事实出发，与相邻及后继事实匹配，进行推理循环，推出的新事实自动添加到事实库中；以知识的目的为驱动机制，逆向推理算法则从根节点顺着推理的二叉树往下搜索，直到找到用户提供的事实节点为止。 北京工业大学工学硕士学位论文 4．提出系统的总体设计方案，在总体设计的框架下实现了一个基于神经网络的专家系统的原型系统，以动物识别为背景，对系统进行合理的方案设计和软件开发，并对所提出的理论方法进行了仿真和实验研究。验证了总体设计的正确性和可行性。更多还原

【Abstract】 At present, two systems in artificial intelligence that is developed well are expert system and artificial neural network. The former is based on symbol processing mechanism and the latter based on connection mechanism. Their applications not only make the fact that computer has intelligence become reality, but solve most problems in engineering. This paper combines the two methods and proposes a design of Expert System Based On Artificial Neural Network.The main contents of this paper as follows:1 .After deeply analyzing the current knowledge representation methods, the paper studies a knowledge representation method called binary production rule. The production rules are standardized into binary rules, each rule containing only two antecedents and one conclusion. After standardizing, knowledge can be easily stored in the ANN.2. Backpropagation model of the ANN is used in knowledge storage and the Leverberg-Marquardt algorithm based on numerical optimization is used in training the neural networks. As to the memory capability of the neural network, the paper propose to set up a system containing a couple of subnets, while each net storing an object and subnets working parallel.3. This paper analyzes the inference mechanism based on neural network and discusses two inference methods which are forward and backward inference. The forward inference algorithm starts from existing facts while the backwardinference starts from result fact.4. On the basis of the research mention above, the master plan and software development of the prototype system has been accomplished, the whole system mainly includes two parts: scheme design and software development. Relevant experiments have been accomplished to the system prototype, including knowledge storage and inference based on neural network. The experiment results show that the scheme and algorithms of system are usable and effective.更多还原