The Research On Genetic Programming Techniques For Multi-classification

Classification, particular Multi-classification problem has been a hot topic in datamining. Practical problems such as object recognition and text classification need to dealwith huge and mutil-category data. It is one of the important challenges how to solve largescale and mutil-class problems in recent years. This paper describes the use of the newgenetic study algorithm——Genetic Programming to solve multi-classification problems,by making a bit tentative improvement in the algorithm.Genetic Programming(GP) is a new technology for optimization, which simulatesinherit and evolution in the nature and gets optimal solutions through reproduction,crossover and mutation operations. This paper describes the basic principle of geneticprogramming and the knowledge of evolutionary computation, introduces the relatedclassification technology, analyses the characteristics of GP and the implementationmethods of classification problems, indicates the limitation of GP in multi-classificationproblems, improves GP based its disadvantages.This thesis focuses on three sides to improve the multi-classification technology in GP.The main research contribution of the thesis can be summarized as follows: First, the newdynamic strategies, including Centered Dynamic Range Selection and Slotted DynamicRange Selection, were found to improve the performance of the system, over the standardSRS. Second, Gradient-decent search on individual programs is introduced to GP in thepaper. GP is still used as a global beam search, but local gradient-descent search is madeon programs. Third, Simplification of programs during evolution is introduced in thisarticle. An algorithm is described that removes redundancy from a program, and is used onall programs periodically during evolution. Finally, the experiments through five imagedatasets (shape, coin) are set. The experimental results show that: SRS was seen to performwell on the two-classification and easy problems, but Centered Dynamic Range Selectionand Slotted Dynamic Range Selection were seen to outperformed the standard SRS on theharder multi-classification problems; the use of gradient-decent search produced excellentresults, training times were shortened for easy problems, and test set accuracies wereimproved for harder problems; Simplification improved the classification accuracy,particularly for the harder problems.The paper in the there sides has improved the classification performance in GP. But that is only a preliminary exploration, needing further study.