Improvement On Relevance Vector Machine And Its Applications

The problem of classification and recognition is the most important branch in the fieldof data mining. There are many common methods in classification and recognition, such asthe Support Vector Machine (SVM) and Relevance Vector Machine (RVM) are the twomethods with good development prospects at present.RVM is a sparse probabilistic model based on Bayesian framework. Compared withSVM and another classification and recognition methods, RVM has less computation kernelfunction, sparse model and higher classification accuracy. However, there are someproblems in RVM need to be solved, such as the kernel parameters need to be optimized,and so on. So the improved RVM methods are proposed and applied in gas and oil welllogging. Related work and innovations are as follows:(1) Study and improvement on RVM based on Quantum Culture Particle SwarmOptimization (QCPSO-RVM). In order to solve the kernel parameters optimization problemin RVM and local minima problem for RVM training with PSO algorithm, the RVM basedon Quantum Culture Particle Swarm Optimization (QCPSO-RVM) is proposed. Based onPSO algorithm, quantum computing and cultural evolutionary thought are brought inQCPSO algorithm which can achieve the global convergence. The simulation results of atypical dataset show that the classification accuracy of QCPSO-RVM is higher thanPSO-RVM and traditional RVM, and QCPSO is faster than PSO at convergence rate.(2) Propose and study on RVM based on Second-Order Cone Programming(SOCP-RVM). The RVM model with multi-kernel functions is constructed based on thecombination idea of multi-kernel functions, in order to solve the problem of combinationparameters selection on multi-kernel RVM, the SOCP method is adopted to solve theparameters selectiom, and a novel RVM based on Second-Order Cone Programming(SOCP-RVM) is proposed. The simulation results of a typical dataset show that theclassification accuracy of SOCP-RVM is higher than that of QCPSO-RVM, PSO-RVM andtraditional RVM.(3) Application on recognition of oil and gas layer in actual logging. In order to testthe practical application effect of SOCP-RVM, three typical oil and gas wells in severaloilfields are selected. Through sample selection and data pre-processing, the SOCP-RVM is used in the recognition modeling with the sample data after the attribute reduction, and thetrained SOCP-RVM model is applied in the classification and recognition of all thewell-section. Compared with the conclusions of the oil and gas trial, the recognition effectof SOCP-RVM is very good, the recognition accuracy is higher than that of QCPSO-RVM,PSO-RVM and traditional RVM, and the proposed SOCP-RVM fully meets the actuallogging logging industry standards and actual logging requirements.