Application Of SVM To Sensory Evaluation

Support Vector Machine (SVM) as a machine learning method is based on the solid theory foundation of Statistical Learning Theory, and focuses on the small samples. Its good generalization have aroused many researchers' concentration and received good applications. The thesis begins from the theory analysis of SVM, expatiates its basic principles and characteristics, and then makes a detailed exploration in the model designs and realistic performance about the application of SVM regression to sensory evaluation. At the same time, this thesis makes comparison between SVM and BP neural network.From the presentation of learning problems and statistical learning theory, the thesis describes and explores the theory foundation of SVM, and makes a summarization for present studies and a comparison between main improved algorithms. On this basis, the thesis analyzes the main problems in SVM, and especially in the selection and tuning of SVM parameters in specific methods.For the application of SVM to sensory evaluation, the thesis which based on the characteristics of sensory sample data points out that the kernel about computer aided sensory evaluation is constructing a machine learning model, which can adapt the small sample. Moreover, this model should be pay more attention on the generalization ability. To solving the multi-class recognitioh, the thesis proposes a method which is using SVM regression to construct the learning model for sensory evaluation, and takes a selection and design for SVM algorithm and its parameters.For testing the learning and generalization performance of SVM in actual industrial environments of application, the thesis puts SVM into this data condition: small sample, high dimension, big noise and nonlinear relationship. Experimentally the performances of SVM have been validated in sensory evaluation. According to these experiments, SVM regression has been compared with BP neural networks, and gained some valuable conclusions.The whole thesis consists of six chapters. In the first chapter, the background and significance of topic selection are shown. Meanwhile it introduces the present studies about SVM and sensory evaluation. In the second chapter, some important concepts such as learning problem, statistical learning theory and SVM are explained. It describes the nonlinear SVM regression method, and puts forward the learning problem of sensory evaluation. The third part of this thesis explores and compares the present main proved algorithm of SVM, analyzes the key problems which existin SVM, and especially expatiates the selection of parameters.The fourth chapter firstly points out the shortcoming of traditional sensory evaluation and validates the multi-class SVM. The results of experiment show that multi-class recognition does not satisfy the actual application. Then it analyzes the data characteristics about sensory evaluation samples, brings forward the kernel idea of modeling in sensory learning problem, and proposes the method which is using SVM regression to solving the multi-class classification in sensory evaluation. Meanwhile it designs the SVM and its parameters, and validates its performance using the tobacco data. The results of experiments shows this method could satisfy the actual demand of sensory evaluation. It is a practical and better technique for sensory evaluation.The comparison between SVM regression and BP neural networks is laid stress in chapter five. At first, the relationship to SVM and neural networks has been expatiated, and the performances about above methods have been compared from the theory level. Then it compares from the accuracy and generalization in sensory evaluation. The comparison shows that SVM regression and BP neural networks have the close performance in sensory evaluation. The rules which learned from sample data by above techniques are consistent and accord with actual law. At generalization, SVM regression is superior to BP neural networks. The results of this chapter are helpful to research for the SVM and multiple layer feed-forwa...