Optimization Methodology Research Of Relevance Vector Machine

Relevance vector machine(RVM) was a machine learning algorithm based on sparse Bayesian framework and it absorbed the advantage of wonderful generalization and precision from support vector machine as a result of RVM approximately had the same architecture and functionality as SVM. Meanwhile, because RVM constructed learning machine based on sparse Bayesian framework, unlike the SVM based on Structural Risk Minimization, RVM possessed the advantages of better generalization, probabilistic predictions and non-’Mercer’kernels. Research on theory and application of RVM will be significant, so this paper focused on the fundamental theory of RVM and explored the method to optimize it. Coming in research work from following aspects:1. Analyzed the theory of RVM and SVM and comparing RVM to SVM, then took the SVM as reference and explored the optimization by the analysis of disadvantages on RVM.2. By utilizing the feature of low portion, proposed a optimized method based on Rapid Estimation. That was preprocessing the training set and excluding the non-Relevance Vector with iterative estimation integrated by threshold factor and maximum reduction upper limit, then decreased the sample scale and reduced the training time. The results of experiments on UCI datasets illustrated that the novel algorithm possess faster training rate while maintaining training precision.3. As the precision and generalization of the traditional RVM with single kernel function was not ideal, proposed a optimized methodology based on Kernel Alignment. Constructed homogeneous and heterogeneous multiple kernel RVM by kernels combination and testified kernels and parameters by Kernel Alignment, then acquired optimized multiple kernel RVM. The result of experiments showed that the multiple kernel RVM possessed better precision and generalization.