Study Of Knowledge Refinement Combining Symbolic System With Neural Networks

Knowledge Refinement is very important but difficult promblem in machine learning and the development of intelligent system feilds. As one part of "Management Decision-Oriented Intelligent Systems Study of Isomeric Knowledge Representation and Knowledge Management" which is sponsored by National Natural Science Foundation of China (No.70171002), this paper aims at the study of knowledge refinement technology in order to promote the research development in this field.In the development of intelligent systems, such as Expert System, there is one problem encountered that knowledge is always conflict or inconsistent with each other in the knowledgebase, which defects the reasoning efficiency and accuracy of intelligent systems. Hense, some researchers put forword the concept of knowledge refinement in order to solve this problem.Knowledge refinement could not only simplify the initial knowledge base, but also correct error and inconsistent knowledge, hence, promoting reasoning efficiency and accuracy of knowledge base. As a result, knowledge refinement is one of hot topics for oversea researchers, and several refinement methods have been proposed, such as KBANN and KBCNN. Some of the exist methods are too complex and difficult to realize, and rules refined sometimes are too complex to understand and use. The paper proposes a new knowledge refinement method which combining symbolic system with artificial neural network (ANN), as called KRSNN, which remedies the limitations of former methods.KRSNN first translates the initial rule base to neural network, which employscross-entropy as error function, and then trains the network with given sample data;the training process stops if the network achieves the predefined accuracy, network is pruned to remove reluctant connections in order to simplify the network. After that, this paper proposes a revised algorithm of Subset, called SubsetII, which extracts rules from pruned network.In order to evaluate the quality of the refined rules, and compare it with other refinement methods, this paper proposes two criterions, which are defined as reasoning accuracy and reasoning efficiency of the refined rulebase.Two experiments are conducted to demonstrate how effectively KRSNN can learn and revise initial rules. First, this paper use the domain of molecular genetics to recognize promoters in DNA nucleotide strings, the data is widely used by knowledge refinement researchers and it is easy and convictive to compare results with other algorithms using the same initial rules base and data which are obtained from the public domain concerning machine learning. The second experiment is applied in personal bank loan evaluation. All the experiments show that KRSNN works quite well but much simpler, and the refined rules are simpler but easy to understand, and KRSNN surpasses other method in several specified aspects.This paper do a lot work both on theoretical and practical field, the research accords with trends of intelligent system development, which has great significance on actual practice.