SVM Predictive Optimization Algorithm And Application Based On Ensemble Learning

Support vector machine (Support Vector Machine, SVM), which optimization objective isminimize the range of values of confidence, is a new pattern recognition method base on theframework of statistical learning theory. It uses learning methods based on the criterion ofstructural risk minimization, has been proved the good promotion ability. SVM is good at solvingproblem which is with small sample size, low-dimensional linear space and other issues that cannot be classified, and can be applied to other machine learning method like function fitting and soon. However, due to the SVM is aid quadratic programming to solve support vector, and solvingquadratic programming will involve the calculation of m-order matrix, it is difficult to implementwhen applies to large-scale training samples.The traditional support vector machine in dealing with large volumes of data samples, thereare shortcomings in terms of predicting time, aims that this problem, this paper make optimizationalgorithm improvements. The goal is to make sure that ensure the accuracy while learning time isgreatly reduced. The main research work is listed as follows:First, this paper discusses the basic concepts of machine learning and support vectormachine basic theory, including statistical learning theory, VC dimension theory, structural riskminimization principle, linearly separable and non-linear separable case, the kernel function andso on. Focused on support vector machine forecasting established forecasting regression model,and do simulation experiments using Matlab in LIBSVM toolkit. Use cross-validation method todetermine the parameters of the kernel function and punishment function. When gettingalgorithm’s performance indicators such as executed mean square error and running time. In termsof prediction fluctuate range, the paper uses method of information granularity, and splittime-series into a number of sub sequence, and then blur into three particles to doing prediction.Finally get result which is correspond to the maximum, minimum and average prediction theresults.Secondly, for the shortcomings of training time in aspects of large-scale sample, paper raisedimproved optimization algorithm based on dynamic clustering algorithms and integrated learning ideas. Dynamic fuzzy clustering method can obtain approximate initial clustering results byselecting size and calculated the distance between each two samples, membership function is thejudging criteria and determines belonging cluster of each data, to ensure that inter-object distanceis minimum in the same kind, while distance is maximum between samples in different kinds.Measure criteria of different distance levels of granularity function get different number of clusters,each time clustering results constitute a sub-sample study. Carry out predict training separately foreach child, then integrity output of each sub-learning machine in accordance with the size of theerror integration, and obtain the final output. Then after comparing with training all the samplesbefore optimization and conducting randomized sampling to training, paper analyzes algorithmperformance advantages and disadvantages.Next, set up software system platform based on LabVIEW and Matlab. Concerning theadvantages of LabVIEW in interface design and reading data interface, combined with theadvantages of Matlab to simulation experiments. Platform mainly consists of five functionalmodules: the login module, the data read module, algorithm execution module, data storagemodule, the drawing results module.Finally, this paper summarizes and Prospects.