基于支持向量机优化RBF神经网络的算法及应用研究

【作者】 雷剑；

【导师】 任金霞；

【作者基本信息】 江西理工大学 ， 控制理论与控制工程， 2009， 硕士

【摘要】 学习是人类的基本智能活动,学习能力是人类智能的根本特征。机器学习是指机器在人工智能系统中模拟并实现各种学习行为的过程。传统的机器学习方法主要有神经网络、小波网络、模糊系统及建立在统计学习理论基础的新的通用机器学习方法——支持向量机等。径向基函数神经网络是一种新颖有效的前馈式神经网络,它具有其他前向网络所不具有的最佳逼近的性能和全局最优的特性,并且结构简单,训练速度快。在RBF神经网络中,隐层中心的数量和位置是整个网络性能优劣的关键,直接影响着网络的性能。中心的数量即隐层节点数量选得太多,容易导致过拟合,使得推广能力下降;中心数选得太少,所学习的网络对样本中包含的信息学习得不充分,也会使得推广能力下降。在实际应用中,RBF网络的优势在于用线性学习算法来完成以往非线性学习算法所做的工作,同时又能保持非线性算法所具有的准确率高等特点的神经网络。但是,在解决高维数据问题时,用传统方式确定的RBF网络在推广能力上有着很明显的缺点。基于统计学习理论的支持向量机算法具有坚实的数学理论基础和严格的理论分析,具有理论完备、全局优化、适应性强、推广能力好等优点,它在很大程度上解决了以往的机器学习模型的选择与过学习、非线性、维数灾难、局部极小点等问题,由于支持向量机在模式识别、回归估计、函数逼近、风险预算、金融序列分析、密度估计、新奇性检验等各个领域获得了巨大成功,立刻成为了机器学习、神经网络、人工智能等方向的专家与学者研究的对象。它使用结构风险最小化原则,综合了统计学习、机器学习和神经网络等方面技术,在最小化经验风险的同时,有效地提高了算法泛化能力。它与传统的机器学习方法相比,具有良好的潜在应用价值和发展前景。本文以径向基函数神经网络和支持向量机为主要研究对象,在介绍了机器学习方法的基础理论以及RBF神经网络和支持向量机的机理后,分析研究了这两种学习方法的内在联系。本文在研究这种内在联系并阐述遗传算法的流程和基本原理的基础上,提出了基于支持向量机和遗传算法RBF神经网络优化算法,即使用遗传算法为支持向量机进行模型参数选择,再利用所建立的支持向量机来构造RBF神经网络。此算法避免了传统算法易陷入局部极小点的缺点,又不需要通过大量实验或凭经验预先指定网络结构。最后,将用本算法优化的RBF神经网络用于非线性系统辨识,通过仿真实验表明,该RBF网络具有较好的辨识精度和泛化能力。更多还原

【Abstract】 Learning is mortal foundational intellect activity.Learning ability is fundamental feature of mortal intelligence. Machine learning means the process that machine(computer or intellect machine)simulates and implements various learning behavior in artificial intelligence system.Traditional machine learning method include neural network、wavelet network、fuzzy system、bayes categorizer and fuzzy division and the general novel machine learning method,support vector machine,which is based on statistics learning theory.RBF neural network is a new and effective neural network.It has the best and universal approximation property,simple structure and fast training speed.The key point in design of radial basis function networks is to specify the number and the locations of the centers.If the number of centers(or the hidden layer units)is chosen too much,over-fitting results and the generalization are getting worse.On the contrary,if the centers chosen is too little,the network is not enough to study the training samples that the performance of networks,for example,generalization will become bad.The advantage of radial basis function neural networks lies in achieving high accurary by taking place of nonlinear algorithm with linear algorithm.So,radial basis function neural network is a kind of neural networks with the performance of high convergence and accuracy.But,when the radial basis function neural network is used to solve the problem of high-dimension data,the generalization of neural network determined by past center-chosen algorithms is very poor.Support vector machine based on the Statistical Learning Theory is a new approach and research field in machine learning because of its advantage such as firm mathematic theory foundation,strict theory analysis,complete theory,global optimization as well as good adaptability and generalization.SVM prodigiously solves many problems encountered by machine learning methods,such as model selection,overfitting,nonlinear and dimension curse in high dimension.Because of the successful application in the fields of pattern recognition,regression estimation,function approaching,risk budget,finance series analysis,density estimation and so on,SVM became the research hotshot in many study fields.SVM improves the algorithm generalization effectively and minimizes the empirical risk simultaneously by using Structural Risk Minimization and synthesizing the techniques including the statistical learning,machine learning and neural networks,etc.It also has good latent applincation values and development prospects compared with the conventional machine learning methods.The thesis emphasis on these two sorts of primary machine learning methods:RBF neural network,support vector machine.After introducing fundamental theory of machine learning and mechanism of neural network,support vector machine,the thesis studies and analyzes inward link between them.In this paper, thoroughly researching this inward link and expatiating the flow and basic principle of genetic algorithm,a new optimization algorithm based on support vector machine and genetic algorithm for RBF neural network is presented,in which GA is used to choose the SVM model parameter and SVM is used to help constructing the RBF.It avoids the disadvantage of traditional algorithms which are often trapped to local minima.Another advantage of the method is that it is not required to designate the network structure in advance by experience or plenty of trials.Finally,the network based on this algorithm is applied on nonlinear system identification. According to simulation,the method has higher precision,good generalization ability and classification ability.更多还原