On-Line Learning Methods for Gaussian Processes

Shigeyuki OBA  Masa-aki SATO  Shin ISHII  

IEICE TRANSACTIONS on Information and Systems   Vol.E86-D   No.3   pp.650-654
Publication Date: 2003/03/01
Online ISSN: 
Print ISSN: 0916-8532
Type of Manuscript: LETTER
Category: Pattern Recognition
Gaussian process,  on-line learning,  Bayesian estimation,  

Full Text: PDF(351KB)>>
Buy this Article

We propose two modifications of Gaussian processes, which aim to deal with dynamic environments. One is a weight decay method that gradually forgets old data, and the other is a time stamp method that regards the time course of data as a Gaussian process. We show experimental results when these modifications are applied to regression problems in dynamic environments. The weight decay method is found to follow the environmental change by automatically ignoring the past data, and the time stamp method is found to predict linear alteration.