Study On Classifiers Ensemble Based On Evolutionary Computation And Fuzzy Clustering

Classifiers combination can also be called classifiers ensemble. In pattern recognition, classifiers ensemble technology attracts great attention of researchers, because it can remarkably improve the performance of pattern recognition. The researchers have proposed many classifiers ensemble methods, which have achieved great success in many fields.The purpose of classifiers ensemble is to make full use of the advantage of each component classifier in their respective classification performance, and obtain higher recognition rate more than any single component classifier. Classifiers ensemble technique makes use of the complementarities amongst different classifiers to improve the performance of classifiers after combination. Generally, there are two methods to improve the performance of classifiers ensemble: improving the classification ability of each component classifier and increasing the diversity among the component classifiers. But traditional classifiers ensemble methods still have some disadvantages: firstly, they do not fully mine the valuable information contained in the sample distribution when generating a classifier; secondly, there is not a good selection criterion to select the component classifiers. Researches show that the performance of individual classifier does not connect with the performance of classifiers ensemble. In addition, we must take into account both the accuracy and diversity of component classifiers in the same time, but usually they contradict with each other; thirdly, the role of each component classifier in classifiers ensemble is fixed, that is, all datasets use the same ensemble method.This method can not work well in all kinds of datasets. In order to achieve the best performance of classifiers ensemble, we should select the proper component classifiers for combination according to the target dataset, and use different ensemble methods according to the different types of sample.This dissertation studies how to achieve the diversity of component classifiers in the condition of assuring the performance of the component classifiers. In the aspect of achieving the diversity, we fully consider how to utilize the training sample distribution to improve the performance of classifiers ensemble and increase the diversity of the component classifiers.The main contributions of this dissertation are summarized as follows:Firstly, an adaptive policy gradient algorithm (APG) is proposed and implemented. The algorithm deflects the object-function through applying a deflection technology after having searched one NE, in order to learn all the existing NE for a finite strategic game. APG is evaluated on the bench mark games problems included in GAMBIT. Compared with covariance matrix adaptation (CMA) evolution strategies and particle swarm optimization (PSO) algorithm, the experimental results show that APG can search all Nash Equilibrium of games more efficiently.Secondly, a classifiers ensemble algorithm based on Fuzzy Clustering (FuzzyBoost) is proposed. We get the distribution characteristics of the training samples through applying Fuzzy C-Means(FCM) into classifiers ensemble algorithm. The different training datasets are sampled according to the proposed concept of information entropy to achieve diverse component classifiers, resulting in the performance improvement of the ensemble classifiers. We implement this algorithm on the Weka platform, and compare the results with that of AdaBoost and Bagging. The experimental results on 20 data sets show that FuzzyBoost has higher accuracy and better generalization ability.Thirdly, an adaptive FuzzyBoost algorithm (AFB) is proposed. Inspired by APG algorithm, AFB gets multiple distribution characteristics of training samples using the adaptive deflection technology, and then apply them into FuzzyBoost algorithm. This method remarkable improves the diversity of the component classifiers. We implement AFB on the Weka platform, and test on 20 data sets. The experimental results show this method has higher classification accuracy and better generalization ability than FuzzyBoost.