يادگيري سيستم هاي رده بندي مبتني بر قانون فازي سلسله مراتبي توسط الگوريتم تكاملي

چكيده انگليسي :

Learning Hierarchical Fuzzy Rule Based Classification Systems by Evolutionary Algorithm Azam Amouzadi a amouzadi @ec iut ac ir Date of Submission 2011 04 18 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Abdolreza Mirzaei mirzaie@cc iut ac irAdvisor Maryam Zekri mzekri@cc iut ac irAbstractThe purpose of this research is introducing fuzzy rule based classifiers by means of evolutionary algorithms One of the most common methods to express classification knowledge in data mining applications is the useof if then rules These kinds of rules are capable of converting the hidden information in the dataset to theform of human understandable knowledge Systems based on if then rules help human expert to mergeobtained rules of systems into his own information to make better decisions Fuzzy logic is able to describenonlinear input output relationship Furthermore it is capable of reasoning uncertain and ambiguousknowledge As a consequence it appears that combining fuzzy logic with if then rules causes to extractedrules similar to the human behavior In general there are at least two different kinds of fuzzy rule basedsystems in terms of the inference process including Mamdani and Takagi Sugeno Kang which are differentin the composition of the rule consequences Using each of Mamdani or Takagi Sugeno Kang systemsdepends on whether the interpretability is more important or the accuracy is If the accuracy is moreimportant using Takagi Sugeno type is profitable but Mamdani type is used when interpretability is moreessential So these two criteria play a great role in evaluating rule based classification systems Consideringthis fact that generating highly interpretable rules is prefered the goal of this research is to increase theaccuracy of Mamdani type systems by proposing a novel hierarchical structure In all the existing methodsgenerating rules from input data consists of two main phases pattern space partitioning into fuzzy subspacesand determining associated fuzzy if then rules The choice of a fuzzy partitioning plays an important role inthe performance of a fuzzy rule based classification system and an inappropriate partitioning leads toimproper rules It seems that if too coarse fuzzy partitioning is used because of generating overgeneral fuzzyrules the performance of classification system may become low In the case of fuzzy partition being too fine lack of the training patterns in the corresponding fuzzy subspaces causes that many of the if then rules wouldnot be generated So in this research a novel hierarchical structure is proposed to divide problem space in away that fine and coarse subspaces are covered simultaneously Also two novel methods based on geneticalgorithms are proposed in order to increase the accuracy of obtained rules from Mamdani type system Thefirst proposed method uses genetic algorithm based on Michigan approach Adaboost method and theproposed hierarchical method Using them the proposed method introduces a new hierarchical fuzzy rulebased classifier which is called Michigan halving hierarchical fuzzy rule based classifier MHH FRBC The second proposed method which is named Pittsburgh halving hierarchical fuzzy rule based classifier PHH FRBC is able to create fuzzy rules by the aid of genetic algorithm based on Pittsburgh approach andemploying the proposed hierarchical structure The performance of both proposed hierarchical fuzzy rule based classifiers is evaluated by comparing them with other sequential fuzzy and non fuzzy classificationmethods and the latest version of hierarchical one on a set of benchmark classification tasks Experimentalresults show that the proposed algorithms accomplishe high quality results with small number of rules incomparison to the several classification algorithms Keywords Hierarchical fuzzy rule classification boosting algorithm evolutionaryalgorithm Michigan approach Pittsburgh approach

عموزادي، اعظم

يادگيري سيستم هاي رده بندي مبتني بر قانون فازي سلسله مراتبي توسط الگوريتم تكاملي

الگوريتم تقويت , الگوريتم ژنتيك