【作者】 赵明渊；

【导师】 周明天；

【作者基本信息】 电子科技大学 ， 计算机应用技术， 2011， 博士

【摘要】 智能优化算法(Intelligent optimization algorithms)通过模仿生物的智能行为来实现优化功能。例如,遗传算法(Genetic Algorithm, GA)模仿生物种群中自然选择的机制来求解优化问题;粒子群优化算法模仿鸟群觅食迁移中,个体与群体协调一致的机理来指导优化搜索等。智能优化方法具有简单通用、鲁棒性强和适于并行等优点,在模式识别、智能控制、并行搜索、联想记忆等方面得到了广泛的应用。支持向量机(Support Vector Machines, SVM)是一种应用广泛的数据分类技术,当高斯核支持向量机应用于各种各样的数据分类领域时,首先面临的至关紧要的问题是:怎样选择惩罚参数C和核参数γ(即怎样进行支持向量机模型选择),怎样优化输入特征子集以提高分类准确率和减小特征子集。通过用户使用枚举法进行参数选择,往往降低支持向量机分类性能,得出的分类准确率很低,不能满足分类要求。采用网格搜索算法进行支持向量机参数优化,由于一定程度地提高了分类准确率,初步满足分类要求。随着数据分类领域对分类准确率的要求不断提高,采用智能优化算法和支持向量机的结合同时优化输入特征子集和支持向量机参数,进一步提高了分类准确率。本论文在上述研究工作的基础上,将支持向量机的渐进性能融入智能优化算法,提出了基于特征染色体的遗传算法、基于特征粒子的粒子群优化算法、基于特征抗体的克隆选择算法,分别构建提出的算法和支持向量机的混合系统以同时优化输入特征子集和支持向量机参数,取得了更高的准确率、更小的特征子集和更少的处理时间。通过将支持向量机的渐进性能融入智能优化算法,构建智能优化算法和支持向量机的混合系统解决上述问题是本论文的基本研究方法。本论文以分类问题的智能优化算法及其应用研究、智能优化算法和支持向量机混合研究为课题,本论文的贡献和创新点概括如下:(1)基于特征染色体的遗传算法基于遗传算法的原理和搜索机制,将支持向量机的渐进性能融入遗传算法,通过生成特征染色体操作将遗传算法的搜索导向到超参数空间中的最佳泛化误差直线,提出一种基于特征染色体的遗传算法,构造基于特征染色体的遗传算法和支持向量机混合系统以同时进行特征选择和参数设置的优化。对提出的算法的收敛性进行了分析。与类似方法相比,提出的算法不仅具有更高的分类准确率和更小的特征子集,而且具有更少的处理时间。(2)基于特征粒子的粒子群优化算法基于粒子群优化算法的原理和搜索机制,通过生成特征粒子操作将支持向量机的渐进性能融入粒子群优化算法,提出一种基于特征粒子的粒子群优化算法,构造基于特征粒子的粒子群优化算法和支持向量机混合系统以同时进行特征选择和参数设置的优化。分析了提出算法的复杂度。实验指出,提出的算法具有更高的分类准确率、更小的特征子集和更少的处理时间。(3)基于特征抗体的克隆选择算法基于克隆选择算法的原理和搜索机制,通过生成特征抗体操作将支持向量机渐进性能引入克隆选择算法,提出一种基于特征抗体的克隆选择算法,构造基于特征抗体的克隆选择算法和支持向量机混合系统以同时进行特征选择和参数优化。通过算法性能对比实验,得出提出的算法具有更高的分类准确率和更小的特征子集。(4)混合智能优化算法研究提出一种基于特征染色体的遗传算法和量子遗传算法的混合算法,构造基于特征染色体的遗传算法、量子遗传算法和支持向量机混合系统,给出了详细的实验结果和算法性能对比,验证了该算法是一种有效的方法。除此之外,还提出一种基于特征抗体的克隆选择算法和差分进化算法的混合算法,构造基于特征抗体的克隆选择算法、差分进化算法和支持向量机混合系统,给出了详细的实验结果和算法性能对比,实验指出提出的算法是一种有用的方法。(5)智能优化算法应用研究根据微阵列基因表达数据的特点,提出一种基于特征染色体的遗传算法包装法、信噪比过滤法和支持向量机混合系统的信息基因搜索方法,以同时搜索到基因数量少而分类准确率高的信息基因子集,实验结果表明,与其它优秀的肿瘤分类方法相比,提出的方法在信息基因数量及分类准确率方面具有明显的优越性。另外,提出一种基于特征染色体的遗传算法、共空间模式和支持向量机混合系统以进行脑-机接口分类参数优化,取得了提高分类准确率的明显效果。更多还原

【Abstract】 Intelligent optimization algorithms implement optimal functions by imitating biologic intelligent behavior. For example, genetic algorithm (GA) imitates the mechanism of natural selection in biologic population to solve optimal problem, and particle swarm optimization (PSO) imitates the coordinated mechanism of individual and swarm in migration of birds foraging to guiding the optimization search. Intelligent optimization algorithms have the characteristics of simple operations, generalization, robustness and parallelism, which have been extensively applied in various fields of pattern recognize, intelligent control, parallel search and associative memory.Support vector machines (SVM) are a kind of popular data classification technique. When SVM with Gaussian kernel are used in various fields of classification problems, first, some crucial problems to be confronted are: how to select the error penalty parameter C and the Gaussian kernel parameterγ(namely how to conduct SVM model selection), how to optimize input feature subset so as to improve classification accuracy for SVM and reduce the feature subset. Parameter selection by user Enumeration method tends to reduce the classification performance of SVM and obtains very low classification accuracy, so that it can not meet the classification requirements. Grid search algorithm using SVM parameter optimization improves the accuracy of classification to a certain extent, initially to meet the classification requirements. As the requirements of data classification field to the classification accuracy are ever increasing, the use of the combination of intelligent optimization algorithm and SVM optimize the input feature subset and SVM parameters at the same time so as to further improve the classification accuracy. In the basis of above work, the asymptotic behaviors of SVM are fused with intelligent optimization algorithm. The genetic algorithm based on feature chromosomes, the particle swarm optimization algorithm based on feature particles and the clonal selection algorithm based on feature antibodies are proposed by this dissertation to construct the proposed algorithm and SVM hybrid system respectively to simultaneously optimize the feature subset and the SVM parameters, so as to obtain higher classification accuracy, smaller feature subset and shorter processing time. The basic methods of this dissertation are to construct hybrid systems of intelligent optimization algorithms and SVM to solve the above problems by fusing asymptotic behaviors of SVM into intelligent optimization algorithms.The topic of this dissertation is the study on intelligent optimization algorithms and its application for classification problems and the hybrid study of intelligent optimization algorithms and SVM. The contributions and innovations of this dissertation are summarized as follows.(1) Genetic algorithm based on feature chromosomesBased on the basic principles and search mechanism of genetic algorithm, the asymptotic behaviors of support vector machines are fused with genetic algorithm, which thereby directs the search of genetic algorithm to the straight line of optimal generalization error in the superparameter space by generating feature chromosomes operation. A genetic algorithm based on feature chromosomes, termed GAFC, is proposed to construct GAFC-SVM hybrid system so as to simultaneously optimize the feature subset and the parameters for SVM. The convergence of the proposed algorithm is analyzed. Compared with the similar approaches, the proposed approach not only has higher classification accuracy and smaller feature subsets, but also has fewer processing time.(2) Particle swarm optimization algorithm based on feature particlesBased on the basic principles and search mechanism of particle swarm optimization algorithm, the asymptotic behaviors of SVM are fused with particle swarm optimization algorithm by generating feature particles operation. A particle swarm optimization algorithm based on feature particles, called PSOFP, is proposed to construct PSOFP-SVM hybrid system so as to simultaneously optimize the feature subset and the parameters for SVM. The complexity of the proposed algorithm is analyzed. The experimental results indicate that the proposed algorithm has higher classification accuracy rates, smaller feature subsets and fewer processing time.(3) Clonal selection algorithm based on feature antibodiesBased on the basic principles and search mechanism of clonal selection algorithm, the asymptotic behaviors of SVM are fused with clonal selection algorithm by the generating feature antibodies operation. A clonal selection algorithm based on feature antibodies, termed CSAFA, is proposed to construct CSAFA-SVM hybrid system so as to simultaneously optimize the feature subset and the parameters for SVM. The experimental results indicate that the proposed algorithm has higher classification accuracy rate, smaller feature subset and better performance compared with the existing clonal selection algorithm and other classification ones.(4) Study on hybrid intelligent optimization algorithmsA hybrid algorithm of genetic algorithm based on feature chromosomes and quantum-inspired genetic algorithm, called GAFC-QGA hybrid algorithm, is proposed to construct GAFC-QGA-SVM hybrid system. The detailed experimental results are given to validate that the proposed algorithm is an effect method. In addition, a hybrid algorithm of clonal selection algorithm based on feature antibodies and differential evolution, termed CSAFA-DE hybrid algorithm, is proposed to construct CSAFA-DE-SVM hybrid system. The detailed experimental results are given to indicate that the proposed algorithm is a useful one.(5) Study on intelligent optimization algorithms applicationAccording the characteristics of microarray gene expression data, a hybrid system of wrapper method based on GAFC, filter method based on SNR and SVM, called GAFC-SNR-SVM hybrid system, is proposed so as to search the informative gene subset of small number of genes and high classification accuracy simultaneously. Experimental results indicate that the proposed algorithm has obvious advantages on the number of informative genes and classification accuracy rate. Moreover, a hybrid system of genetic algorithm based feature chromosomes, common spatial patterns and SVM, termed CSAFA-DE-SVM hybrid system, is proposed to perform classification parameter optimization of brain-computer interface, which obtains obvious effect in improving classification accuracy.更多还原