Incremental learning of Gaussian process models

Dejan Petelin (2009) Incremental learning of Gaussian process models. EngD thesis.

PDF Download (2696Kb)

Abstract

Gaussian processes modeling is a relatively new modeling method which is due to its good features more and more applied. Unfortunately the computation time grows with third power as for size of training data set. Therefore this method is not convenient for online learning in principle. Variety of approximation methods and chose the convenient one for online learning were reviewed. The chosen method is described and demonstrated on a simulated and a real life problem. The method is based on the combination of a Bayesian online algorithm together with the sequential construction of a relevant data subsample which specifies the model. It was found on the basis of experiments (figures and errors measured with several measures) that the model obtained with this method depends on sequence of incorporated data. That is a result of the inclusion criterion which depends on the score of contribution regarding to the current model. Surprisingly, we also found the method is more effective on one-input problems than on multi-input problems.

Actions (login required)