Dissertation > Excellent graduate degree dissertation topics show

Research on the Graphical Models in Intelligence Data Processing

Author: WangShuangCheng
Tutor: YuanSenZuo
School: Jilin University
Course: Communication and Information System
Keywords: Bayesian networks Markov networks Chain Figure Moral Figure Markov blanket Markov boundary Neighborhood Discrete variables Continuous variables Complete data Missing data Structure Learning Parameter learning Gradual learning Classifier Classification accuracy Hidden variables Clustering Gibbs sampling MDL standard Bayesian standard Cross-entropy Mutual Information Conditional entropy Feature subset selection
CLC: TP391.41
Type: PhD thesis
Year: 2004
Downloads: 340
Quote: 5
Read: Download Dissertation


Graphics mode is a combination of probability theory and graph theory, graphical representation of dependencies between random variables. Node in the graph represents a random variable, the existence of the edge (directed or undirected) conditional independence between random variables. It has a visual image of the knowledge representation form of reasoning, as well as closer to human thinking features, has been widely used in expert systems, decision analysis, pattern recognition, machine learning and data mining, the recent years and abroad intelligent data processing research focus. Graphics mode consists of two parts, the part of the structure (graphics), the other part is a parameter (conditional or marginal probability distribution), respectively, for the qualitative and quantitative description of the dependencies between random variables. More graphics mode research, the core of Bayesian network (directed acyclic graph), Markov network (undirected graph) and chain diagram (directed and undirected mixed graph). This paper studies the Bayesian networks and Markov networks, and chain diagram for a brief introduction. Focus on research graphics mode intelligent data processing applications, how to transform data to knowledge (graphical mode of learning) and knowledge into intelligent reasoning (based graphics mode). The specific contents are as follows: 1. Graphics mode with complete data and discrete variables learning overview and analysis of representative methods and algorithms. Rely on analytical thinking and causal the semantic orientation Bayesian network structure learning methods and Bayesian networks based on the basic dependent relationship between variables, basic structure and dependency analysis of ideological and Markov network structure learning methods were established. These two methods can avoid the the existing scoring - search method of exponential complexity and local optimal structural problems, and to rely on the analysis method of a large number of high-end conditional probability calculation and edge orientation limitations. Two Bayesian network learning algorithm accuracy evaluation method. 2. Graphics mode of learning is widespread because of the incomplete data (or missing data) with incomplete data and discrete variables, and learning but also due to the presence of missing data can not be directly graphics mode, the graphics mode of learning with missing data has been a attention an important and difficult research topic. Mainly combined with the EM algorithm (or based gradients lt; WP = 153 gt; optimization method) and scoring - search method with missing data graphics mode learning, low efficiency, and easy to fall into local optimum structure. This paper presents a new graphics mode of learning with missing data. The method combines the graphics mode and Gibbs sampling, the the iterative correction with optimal adjustment of the graphics mode of the data lost by random initialization graphics mode with missing data iterative learning. Avoid using the EM algorithm (or gradient-based optimization method) local optimal and deception convergence of the Gibbs sampling process smooth convergence to the global distribution, In every iteration, based on the graphic mode decomposition joint probability can significantly improve the sampling efficiency through optimal adjustment of the graphics mode, the graphics mode in the iterative process gradually close to the stationary distribution of the graphics mode, until the end of the iteration termination condition is met. In this paper, the three cases: (1) random data loss situation with incomplete data. Each column contains some random lost data with variable dimension (range), and some examples of information; (2) hidden variables (or clustering variables) lost data. Hidden variables (or clustering variables) column of data is completely lost, does not have the information and examples of the hidden variables (or clustering variables) dimension; (3) small sample set of missing data. A large number of rows of data is completely lost (not observed), with all variable dimension information and some examples of information. Were established based on the analysis of existing methods and algorithms of these three cases, for some problems, new methods and algorithms, and the the necessary theoretical arguments and Test Analysis. Graphics mode of learning with continuous variables can also be converted to incomplete data problems, learning is an iterative process. Iterative process, we use a hybrid data clustering method discretization of continuous variables, adjustments to optimize the graphics mode on the basis of new discrete variable, until convergence. 3. Graphics mode the progressive learning assimilation and responsive to the two basic mechanisms of human learning new knowledge, human learning process can be seen as a continuous assimilation of new knowledge and responsive process. Based on the structure and parameters of the basic mechanisms of human learning new knowledge and graphics mode changes are not synchronized, progressive way to learn a new graphics mode. The method first adaptive testing to determine whether the structural adjustment of the original structure of the data set in graphics mode. If required, the adaptation of the structure, and adjust the parameters on the basis of a new structure, or only in the parameter adjustments based on the original structure, to obtain a new graphics mode. The learning process in line with the basic mechanisms of human learning new knowledge, and be able to effectively describe the graphics mode structure and parameters of the dynamic changes, and do not need to smooth the existing methods and two Markov assumptions. 4. Basic theories of graphics mode and graphics-based mode of reasoning from conditional independence between random variables, graphical mode probability model between nodes lt; WP = 154 gt; nature of the d-separation (or s-separation), as well as an overview of the basic theory of the graphics mode, the link between the three aspects. , Given the definition of non-negative form (the original definition given to the negative form, it is difficult to understand), and introduced the basic theory of Bayesian network core concept of d-separation standard helps to understand the d-separation standard two Bayesian network model (information pipeline model and small ball model). Inferred from the probability of evidence passed and causal analysis in terms of the graphical model-based reasoning system described and analyzed, and combined with examples give the necessary instructions. 5. Graphic pattern classification in graphics mode learning method based on, respectively, the establishment of a learning and optimization method based on class the constraint graphics mode classifier learning method and graphical pattern classifier gives a graphical mode, and in the 0-1 loss The optimal classifier prove. Classifiable accuracy estimation methods and the accuracy of different classifiers comparison. 6. Sub-sets based on the characteristics of the graphics mode select feature subset selection is a as much as possible to exclude irrelevant and redundant features to optimize the performance of the classifier, machine learning, mold

Related Dissertations

  1. Some Strong Deviation Theorems for Information Sources and a Class of Limit Theorems on an Extended Tree,O236
  2. New Algorithms of 3D Medical Image Rigid Registration,R310
  3. Nonlinear Dynamical Time Series Analysis Methods and Its Application,O19
  4. Entanglement Dynamics in the Double Jaynes-Cummings Models,O431.2
  5. Nonlinear spatial and temporal characteristics of the data based on the observations of the complexity,P467
  6. Arbitrary random variables Multivariate Function Sequences deviation theorems,O211.5
  7. Method Study of Human Resource Selection Based on Support Vector Machine,F224
  8. Protein Secondary Structure Prediction Modeling Research Based on Neural Network,Q51
  9. The Research of Chaos Synchronization and Its Applications,O415.5
  10. Rough Sets and Rough Entropy of a relational database,TP311.132.3
  11. Research on Term Automatic Translation Technique for Japanese-Chinese Machine Translation System,TP391.2
  12. Fuzzy Joint Entropy and Fuzzy Conditional Entropy Based on Credibility Distributions,O236
  13. Hidden Markov model - based intrusion detection system research,TP393.08
  14. Research on Incremental Reduction Algorithm Based on Rough Sets,TP18
  15. Studies on Classfication Algorithms of Motor Imagery-based Brain-computer Interfaces,R319
  16. Research on Network Intrusion Detection Based on Data Mining,TP393.08
  17. The Study on KDD Technologies Based on Rough Set Theory,TP18
  18. Research of Network Security Audit Based on Data Mining,TP311.13
  19. Differential Evolution Algorithm and Its Application,O224
  20. A Research on Different Feature Expression in OFFSS,TP181

CLC: > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device
© 2012 www.DissertationTopic.Net  Mobile