摘要
commie当今人类社会已经进入了大数据时代，数据大多呈现出维数高、规模大、结构复杂等特性。在大数据的研究当中，许多数据如媒体数据、遥感数据、生物医学数据、社交网络数据、金融数据等都是高维数据，尤其是在人类生产生活中，含高维数据的无解析模型或一次候选解的评价计算成本十分巨大的昂贵多目标问题，对其仿真求解势必面临维数灾难。因此，寻找合适的降维方法处理高维数据已是迫切需求。
神经网络是模拟人脑的结构和功能而建立起来的分布式信息处理系统，面对高维多目标优化等非线性问题，与其他降维方法相比，神经网络具有巨大的优势，这得益于神经网络具有高度非线性、结构复杂、自学习、自适应等特点。RBF神经网络是一种新颖有效的前馈式神经网络，它具有很强的非线性映射能力，能以任意精度全局逼近一个非线性函数，而且学习速度快。利用RBF神经网络实现对高维数据的降维预处理，不仅有充分的理论依据，而且更具优越性。本文在对RBF神经网络算法进行优化研究的基础上，研究了基于数据驱动的特征选择RBF 神经网络降维方法，并将其应用在高维多目标优化决策空间降维预处理及Pareto 优劣性预测中。
为了提高RBF神经网络的学习效率，本文首先对RBF神经网络进行改进研究。通过自适应调节RBF神经网络的学习率和动量因子，加快了RBF神经网络的收敛速度；同时，利用遗传算法对RBF神经网络的三we remain
个参数初始值进行优化设计，提出了一种遗传自适应RBF神经网络算法。将改进算法分别应用于故障诊断和UCI数据集的分类实验上，验证了改进RBF神经网络算法的有效性和优越性。锫
针对无解析模型的高维多目标优化问题，提出了一种最大信息系数与最大相关最小冗余相结合的特征选择方法，利用遗传自适应RBF神经网络算法在高维特征空间中选取出了一个低维的特征子集，从而实现对高维特征空间的降维。通过在UCI数据集上的分类实验，证明了该降维算法在保证较好分类精度的前提下，大大减少了计算成本。
为了降低高维多目标优化的维数灾难，将本文提出的基于最大冗余最小相关的遗传自适应RBF神经网络特征选择算法用于多目标优化中的决策空间降维预处理，进行Pareto优劣性预测并将其嵌入MOEAs算法。通过与NSGA-II的实验效果对比，结果证明了本文提出的遗传自适应RBF神经网络特征选择算法在保证得到一个可接受的Pareto最优解的前提下，大大减小了计算成本，避免了维闲逸的意思
数灾难。
关键词：RBF神经网络；高维数据降维；最大相关最小冗余；特征选择；Pareto 优劣性预测
Abstract
普京收入曝光
Nowadays,human society has entered the big data era,and most of the data are characterized by hig
h dimensionality,large scale and complex structure.In the study of big data,many data such as media data,remote nsing data,biomedical data, social network data and financial data are high dimensional data,especially in the human production and living,an containing high dimensional data expensive multi-objective problem with no analytic model or high cost of an candidate solutions evaluation,and its simulation must cau dimensional disasters,and its simulation must cau dimensional disasters.Therefore,it is urgent to find a suitable method to deal with high dimensional data.
Neural network is a distributed information processing system bad on simulating the structure and function of brain,In the face of high-dimensional nonlinear multi-objective optimization problems,neural networks have great advantages over other dimensionality reduction methods,which are due to their highly nonlinear,complex structure,lf-learning and lf-adaptive characteristics. Radial basis function(RBF)neural network is a novel and effective feedforward neural network,which has strong nonlinear mapping ability,can approximate a nonlinear function globally with arbitrary accuracy,and has a fast learning speed.Using RBF neural network to reduce the dimensionality of high dimensional data not only has sufficient theoretical basis,but also has more advantages.This paper focus on the dimensionality reduction method bad on data-driven featur
e lection RBF neural network,and applies it to classification and Pareto dominance prediction.
In order to improve the learning efficiency of RBF neural network,this paper firstly studies the improvement of RBF neural network Algorithm.By adjusting the learning rate and momentum factor of RBF neural network adaptively,the convergence rate of RBF neural network is accelerated.At the same time,the initial values of three parameters of RBF neural network are optimized by genetic algorithm, and a genetic adaptive RBF neural network algorithm is propod.The improved algorithm is applied to fault diagnosis and the classification experiments of UCI data ts respectively,and the effectiveness and superiority of the improved RBF neural network algorithm is verified.
Aiming at the no analytic model high dimensional multi-objective problem,this paper propos a feature method combining the maximum information coefficient
with the maximum correlation minimum redundancy,and then using genetic adaptive RBF neural network algorithm lect a low dimensional feature subt in high dimensional feature space,so as to realize dimension reduction of high dimensional feature space.Through the classification experiment on the UCI data t,it is proved that the dimensionality reduction algorithm can greatly reduce the calculation cost on the premi of ensuring better classification accuracy.
In order to reduce the dimension disaster of high-dimensional multi-objective optimization problems,the feature lection algorithm of genetic adaptive RBF neural network bad on maximum redundancy and minimum correlation was applied to the dimension-reduction preprocessing of decision space,then predict the Pareto dominance and embed the prediction algorithm to MOEAS.By comparing with the experimental results of NSGA-II,the results show that the feature lection algorithm of the genetic adaptive RBF neural network propod in this paper greatly reduces the calculation cost and avoids the dimension disaster on the premi of obtaining an acceptable Pareto optimal solution.
badminton怎么读
Key Words:RBF neural network;Dimensionality reduction of high-dimensional data; Maximum correlation minimum redundancy;Feature lection;Pareto dominance prediction
目录
摘要........................................................................................................VI 第1章绪论 (1)
1.1研究背景与意义 (1)
1.2研究现状 (2)
1.2.1高维数据降维研究现状 (2)
1.2.2RBF神经网络研究现状 (9)
1.3本文研究内容 (10)
第2章RBF神经网络结构与算法分析 (11)
2.1RBF神经网络结构 (11)
2.2RBF神经网络学习算法 (12)
2.2.1参数计算 (12)
2.2.2学习步骤 (14)
2.3RBF神经网络逼近理论 (15)
2.4RBF神经网络存在的缺陷及其原因 (16)
2.5小结 (16)
第3章一种RBF神经网络改进算法 (17)
shampoo3.1RBF神经网络学习率和动量因子的优化方法 (17)
3.1.1学习率和动量因子的自适应调整 (17)
3.1.2仿真实验结果及分析 (17)
3.2遗传自适应RBF神经网络学习算法 (21)
3.2.1算法框架 (21)
3.2.2实验结果及分析 (22)
3.3小结 (24)
第4章基于最大相关最小冗余的RBF神经网络降维方法 (25)
4.1最大信息系数和最大相关最小冗余 (25)
4.1.1基于决策分量与目标分量二维投影的网格划分 (27)
4.1.2最大信息系数与最大相关最小冗余特征选择 (28)
六一儿童节英文4.2最大相关最小冗余RBF神经网络降维算法 (28)六级口语考试内容
zombi4.2.1最大相关最小冗余RBF神经网络降维算法构造 (28)
4.2.2仿真实验结果及分析 (29)
