Research On Improvement Of Online Support Vector Machine And Its Application

SVM (Support Vector Machines, SVM) have been proposed as a new idea, a new way tosolve problems in machine learning. Because of its solid theoretical foundation, not onlyeffectively overcome the traditional machine learning methods in many shortcomings, butalso has a high generalization convergence properties, in addition to its solving globaloptimization. The least squares support vector machine (Least Squares Support VectorMachines, LS-SVM) compared to the support vector machine, the choice of a different lossfunctions and constraints. Not only SVM has many advantages, but also one of the SVMalgorithm improvements. LS-SVM to solving the problem is to solve a system of linearequations problem, change the inequality constrained problem support vector machine,reducing the amount of calculation and improve the solution speed. Aiming at some problemsexisted in LS-SVM, also need to conduct in-depth research on its theory. In this paper, theLS-SVM noise immunity and its lack of sparsity issues in-depth study. The main work is asfollows:(1) In-depth analysis of the basic LS-SVR theory based on support vector machine datadescription (Support Vector Domain Description SVDD) method for determining the fuzzymembership for each sample a different definition of fuzzy membership. Simulation resultsshow that this method improves the LS-SVR noise immunity.(2) Clipping algorithm proposed training sample vector density based on the nearestneighbor. Kind of training samples of each sample cluster, delete the noise data to improve theaccuracy of SVM training. Similarity by calculating the average density of each sample andthe mean of the class, the class similarity threshold to obtain a sample, and according to thesimilarity threshold value, the sample will be less than the threshold value for the similarity ofthe class combined to reduce the total number of training samples. The UCI classificationalgorithm is applied to the reference data set, the experimental results show that: the proposedalgorithm to ensure the accuracy of the training, the experimental results show that: thealgorithm in the case that the training precision, reduces the number of support vectors,effectively achieves a sparse solution for LS-SVM.(3) For general least squares support vector machine trained to handle large data setappears slow, computing capacity, the disadvantage is not easy online training. First, based onthe density of the nearest neighbor training sample vector clipping algorithm for datapreprocessing get a new set of samples, and then use the online training methods incrementalsliding window on the LS-SVR of training.(4) In-depth analysis of glutamic acid fermentation process, the main factors and variablesthat influence from glutamic acid fermentation process, combined fermentation processcontrol technology, select the appropriate input and modeling steps. The improved algorithmis applied to online LS-SVR glutamic acid fermentation process modeling, cell concentrationand glutamate concentration on modeling, and from the training time, the model accuracy,model comparison of learning ability test results, the results show improved online LS-SVRmodel in ensuring the accuracy of the situation and improve training speed, and has a strong learning ability, suitable for the fermentation process modeling.