Fuzzy-arithmetic-based approach for extracting positive andnegative linguistic rules from trained neural networks

摘要

Our method extracts linguistic rules from trained neural networksfor high-dimensional pattern classification problems with continuousattributes. Characteristic features of our rule extraction method are asfollows: (I)It can extract fuzzy if-then rules with linguisticinterpretation. Extracted fuzzy if-then rules are always linguisticallyinterpretable. (2) It can handle existing feedforward neural networksthat have already been trained. Neither specific learning algorithms nortailored network architectures are assumed. It does not change weightvalues of the trained neural networks during the rule extractionprocess. (3) It is based on fuzzy arithmetic. Linguistic values such as“small” and “large” are presented to neuralnetworks, and corresponding fuzzy outputs are calculated by fuzzyarithmetic for extracting linguistic rules. (4) Negative linguisticrules can be extracted from trained neural networks as well as positiverules. After briefly describing our method, we discuss the accuracy offuzzy arithmetic and show subdivision methods for decreasing the excessfuzziness in fuzzy outputs from neural networks. We also discuss thehandling of negative linguistic rules such as “If x₁ issmall and x₂ is not large then Class 3” and “If x₁ is large then not Class 2”