TIME SERIES PREDICTION USING RELEVANCE VECTOR MACHINE

Abstract

The Relevance Vector Machine (RVM) introduced by Tipping is a probabilistic model similar to the state-of-the-art Support Vector Machine (SVM), but where the training takes place in a Bayesian framework and predictive distributions of the outputs instead of point estimates are obtained. This thesis focuses on the use of RVM for regression and classification in general and time series prediction in particular. The standard training method for RVM has been modified to automatically adapt the width of the basis function to the optimum. To speed up things multithread programming has been used. The hardware used is a dual core processor and the software used is the parallel processing toolbox in MATLAB v201 Ob. The model developed has been tested for predicting Mackey-Glass time series and Tree-ring data. The performance of the model is found to be superior to both standard RVM and SVM.