Research And Application Of Gene Expression Programming Based Classification Algorithms

Data Mining(DM) is one of the hottest research topics in current computer application technology and theory research area. With the development of more than 20 years, it has established systematic theory and mature technology. Classification is one of the most important branches of DM and it has gotten wide attention by experts from various fields. Evolutionary computation (EC) which is represented by Genetic Algorithms (GAs) and Genetic Programming (GP) has become a special branch for its unique combination of intelligence, parallelism. Gene Expression Programming (GEP) is a kind of new genetic algorithms combining the advantages of both GAs and GP, which has achieved good performance in mathematical modeling and has been applied in many engineering fields successfully.In this thesis, our main attention focused on GEP and classification, we did research on GEP neural network (GEPNN) and GEP decision tree (GEPDT) algorithms as well as its application in classification. The main work and achievements are as follows:1. The encoding characteristic and the hypostasis of advantages were analyzed based on the introduction of the main idea of GEP. Some influential GEP based classification methods were summarized, especially discussed and compared the basic GEP classification method and an accurate and compact GEP classification method.2. The defect that traditional GEPNN can't be used in quadratic or more complex modeling was pointed out, and we proposed a kind of hybrid GEP neural network method (HGEPNN), further more, an improved method for evolving GEP neural network was proposed based on HGEPNN. Experiments showed that good results can be achieved both in symbolic regression and classification problem.3. According to the trouble of splitting data sets in multi-classification problem for GEP neural network methods, GEP decision tree algorithm was introduced. We proposed a constants uniform distribution based GEP decision tree algorithm (UDC-GEPDT) based on the deficiencies of constants array strategy for standard GEP decision tree algorithm. Experiments showed that UDC-GEPDT algorithm was better than traditional C4.5 algorithm and standard GEPDT algorithm.4. The HGEPNN algorithm was implemented in open source data mining platform WEKA, as well as standard GEPDT algorithm and UDC-GEPDT algorithm in evolutionary computation research platform ECJ, which supported the validation and comparison of different algorithms.